Map4Scrutiny – a linked open data solution for politicians interest registers

Universidade do Minho Escola de Engenharia Map4Scrutiny - A Linked Open Data Solution For Politicians Interest Registers Map4Scrutiny - A Linked Open Data Solution For Politicians Interest Registers Inês Catarina Barreira Lopes Map4Scrutiny – A Linked Open Data Solution For Politicians Interest Registers Masters Dissertation Master’s in Information Systems Work developed under the supervision of: Professor Doctor Ana Alice Baptista Professor Doctor Óscar João Atanázio Afonso December 2021 ii This is an academic work that can be used by third parties if the internationally accepted rules and good practices on the scope of the creators’ and other rights are followed. Therefore, the work is available for re-use under the terms defined by the license present below. For any other uses not contemplated by the license, the user should contact the author via University of Minho’s RepositoriUM. This work is licensed under a Creative Commons Attribution 4.0 International License. Attribution CC BY https://creativecommons.org/licenses/by/4.0/ iii STATEMENT OF INTEGRITY I hereby declare having conducted this academic work with integrity. I confirm that I have not used plagiarism or any form of undue use of information or falsification of results along the process leading to its elaboration. I further declare that I have fully acknowledged the Code of Ethical Conduct of the University of Minho. iv RESUMO O trabalho desenvolvido no âmbito desta dissertação descreve o processo de recolha, uniformização e transformação de dados abertos em formato de texto e tabelas (CSV) para dados abertos ligados ( Linked Open Data ). Especificamente, dados sobre os registos de interesses dos deputados à assembleia da república portuguesa e contratação pública, ligados pelas organizações que são mencionadas em ambos. O estado da arte inclui uma análise de fundo aos conceitos de corrupção, transparência, dados abertos, e dados abertos ligados, tal como a projetos de dados abertos e dados abertos ligados relevantes. A seleção dos dados a utilizar, com respeito aos tópicos de conjuntos de dados relevantes e ao interesse público, o desenho da solução proposta e a seleção de ferramentas, métodos e processos, seguiu a proposta de três ciclos de Hevner para uma abordagem ao desenho de investigação na ciência. O processo de implementação é iniciado com a recolha de dados das fontes utilizando bibliotecas Python para web Scraping e a transformação dos mesmos em tabelas (CSV). Estes dados são depois limpos e uniformizados com auxílio do OpenRefine. Esta ferramenta é também usada para mapear os dados da tabela para triples que são exportados em ficheiros Turtle . Este mapeamento foi previamente desenhado num perfil de aplicação que serviu de base para a criação das formas dos dados (ShExC) usadas para conduzir o processo de validação nos ficheiros Turtle . Esta validação assegura que os ficheiros gerados pelo OpenRefine são conformes com o perfil de aplicação. Para descrever adequadamente os dados foram usados vocabulários já existentes complementados, quando necessário, com a criação de novas classes, propriedades e valores. Este processo está também descrito e os vocabulários estão disponíveis para consulta e reutilização. Por fim, foram feitas consultas modelo em SPARQL para ilustrar a diferença entre os dados originais e o conjunto de dados transformado. O objetivo deste trabalho é contribuir para as áreas de dados abertos ligados e dados abertos para a transparência e escrutínio público. Os contributos principais para o primeiro são um novo esquema de dados e a descrição de todos os passos do processo de transformação. Para o segundo o contributo que se destaca é mais uma implementação que demonstra o potencial do escrutínio de dados no aumento da transparência através da comparação entra as consultas possíveis aos conjuntos de dados originais e ao resultante da solução proposta. O processo de implementação está documentado abaixo e os ficheiros resultantes disponibilizados para consulta. Palavras-chave: Dados Governamentais Abertos, Design Science Research , Escrutínio , Linked Open Data , Transparência v ABSTRACT The work developed in the scope of this dissertation describes the process of sourcing, uniformizing, and transforming text and tabular (CSV) open data to linked open data. More exactly, data on Portuguese parliamentarians’ interest registers and public procurement, linked by the organisations mentioned in both. The state of the art presented includes a background analysis on the concepts of corruption, transparency, open data, and linked open data and an analysis of relevant open data and linked open data projects. The research was conducted using Hevner’s three-cycle design science research approach which led to the definition of the data scope concerning relevant dataset topics and the public’s interest, the design of the proposed solution, and the selected tools, methods, and processes. The implementation process starts with Scraping the data from the sources with the aid of python libraries and generating tabular (CSV) outputs. These are cleaned and uniformized in OpenRefine. OpenRefine is also the tool used to map the data on the tables into triples and generate outputs in Turtle. The map was designed in an application profile that also served as a base for writing the shapes (in ShExC) and conducting validation on the exported Turtle files. This validation ensures that the data is conformant with the application profile. To successfully describe the data in triples, on top of the external vocabularies used, new classes, properties and values had to be created. This process is also thoroughly described, and the outputs are open to access and reuse. Finally, sample SPARQL queries were made to showcase the difference between the sourced data and the resulting dataset. The goal is to contribute to the field of linked open data and open data for transparency and public scrutiny. The main contributions to the first are a new data scheme and the description of every step in the transformation process, while to the latter the contribution is a further implementation showcasing the scrutiny potential of data in improving transparency by comparing the querying possibilities of the final dataset with the originals. Every step taken is documented below and the resulting outputs of the different stages are available for consultation. Keywords: Design Science Research, Linked Open Data, Open Government Data, Scrutiny, Transparency vi TABLE OF CONTENTS Resumo.............................................................................................................................................. iv Abstract............................................................................................................................................... v 1 Introduction ................................................................................................................................ 1 2 State of the art ............................................................................................................................ 4 2.1 Background Analysis - Corruption ........................................................................................ 6 2.1.1 European and Portuguese Contexts ................................................................................... 7 2.1.2 Syndromes of Corruption .................................................................................................. 9 2.1.3 Good Practices .................................................................................................................. 9 2.1.4 Final considerations ........................................................................................................ 11 2.2 Background Analysis – Open Data ..................................................................................... 12 2.2.1 Availability, Status and Evaluation ................................................................................... 12 2.2.2 Portugal’s Open Data and Transparency Scenario ........................................................... 15 2.2.3 Open Data Re-use ........................................................................................................... 17 2.2.4 Open Data Standards ...................................................................................................... 19 2.2.5 Data Literacy ................................................................................................................... 22 2.2.6 Open Data Re-Use Projects ............................................................................................. 23 2.3 Linked Open Data ............................................................................................................. 26 2.3.1 Semantic Web ................................................................................................................. 26 2.3.2 Linked Data .................................................................................................................... 27 2.3.3 Advantages of LOD .......................................................................................................... 30 2.3.4 Linked Open Data Implementations – With a Focus on Government Data ....................... 31 2.3.5 LOD Challenges .............................................................................................................. 33 2.3.6 Implementation ............................................................................................................... 34 3 Data And Methods .................................................................................................................... 37 3.1 Method - Design Science Research .................................................................................... 38 3.2 Proposed Solution - Description ......................................................................................... 42 3.2.1 Data Sources .................................................................................................................. 42 3.2.2 Base.gov.pt and Parlamento.pt ....................................................................................... 46 vii 3.2.3 Interests Declaration ....................................................................................................... 47 3.2.4 Selecting Data ................................................................................................................. 47 4 Implementation ......................................................................................................................... 49 4.1 Data Sourcing, Cleaning, and Uniformization ..................................................................... 51 4.1.1 Sourcing ......................................................................................................................... 52 4.1.2 Cleaning and Uniformization ........................................................................................... 60 4.2 Designing the Linked Data App Profile ............................................................................... 64 4.2.1 Reification ....................................................................................................................... 68 4.2.2 Dublin Core Application Profile (DCTAP ) ......................................................................... 70 4.2.3 Vocabulary Creation ........................................................................................................ 74 4.2.4 Shape Expressions .......................................................................................................... 77 4.3 Transformation, Validation and Publishing.......................................................................... 79 4.3.1 Transformation ................................................................................................................ 80 4.3.2 Validation ........................................................................................................................ 82 4.3.3 Publication ...................................................................................................................... 84 5 Conclusions .............................................................................................................................. 87 5.1 Context ............................................................................................................................. 87 5.2 Contributes ....................................................................................................................... 88 5.2.1 Knowledge Base .............................................................................................................. 88 5.2.2 Application domain ......................................................................................................... 89 5.3 Future Work ...................................................................................................................... 91 5.4 Closing Note ..................................................................................................................... 93 6 Appendix .................................................................................................................................. 94 7 Attachments ........................................................................................................................... 106 8 References ............................................................................................................................. 107 2 That constitutes the research problem at hand: The data opened by the government on Parliamentarians Interest Registers and Public Procurement is isolated, difficult to link, and neither machine-readable nor interoperable by semantic web standards. Thus, it is not ready to be used and reused by third parties. Therefore, the main objective of this research is to propose a solution for linking open data on Portuguese parliamentarians interest registers and public procurement, transforming said data from their current level (one or two) to level five, LOD [2]. To achieve this objective, the datasets may be transformed to a machine-readable format, the two datasets may be linked to each other and external LOD vocabularies, and the final combined dataset may be uploaded into a Triplestore with an online SPARQL endpoint so it can be queried by anyone. Also, the process of transformation should be transparent so that any third party could replicate the proposed solution. The research method used is Hevner’s Three Cycle Design Science Approach [6]. The developed solution attempts to respond to the stated problem by basing the design decisions on known literature and previous projects. This work is developed in the domain of information systems, more specifically semantic web and LOD with the area of the application being transparency. Identifying the area of application as transparency is more accurate than corruption because further work needs to be developed to understand the impacts of LOD implementations as corruption-fighting tools. The prototype and results presented here are promising but not enough. The identified stakeholders are on first level journalists, corruption-fighting organizations, and researchers, and on a second level the general public. The expected contributions include state-of-the-art research comprising a background analysis in corruption, open data and LOD illustrated with sample projects. A definition of the methodologies, a description of the implementation from the sourcing of the data to the process of mapping it to RDFTurtle and publication in a Triplestore. The files resulting from each phase are also shared and include: LOD Application Profile, code written to source the data, logs from the cleaning data process, transformed RDF-Turtle data, ontology, and value vocabularies, the shapes of the data in ShExC, the shape maps, the validation results, and sample SPARQL queries. Each of these outputs is also a contribution to the knowledge base in LOD. The sample queries will illustrate the potential of an LOD profile to apply scrutiny on the relationships between persons and organizations of interest to the public. The ability to easily apply scrutiny to entities is useful when there is suspicion of corruption. 3 The remainder of this document is structured as follows: The State of the art, chapter two, is divided into Corruption, Open Data, and Linked Open Data. The first defines the area of application and explores transparency as a corruption-fighting tool, and open data or even LOD as transparency enablers. Open Data is discussed in the second sub-chapter with further analysis on the concepts of what makes good open data and what datasets are important. Throughout, open data availability, quality, and abiding laws are largely discussed with a special focus on Portugal. This information is then put against already ongoing projects focused on open data re-use to select national data sources. The third part goes into the domain of this work by exploring LOD in the scope defined by the previous two sub-chapters and the overall good practices and implementation methods. Chapter three, Data And Methods, established the methods used to conduct the research and feeds on the previous chapter to set the data scope of the proposed solution. Then in Implementation , every step taken in sourcing, cleaning and uniformizing, designing the LOD application profile, writing the shapes, selecting, and creating vocabularies, transformation, validation and publishing the data is described and generated outputs presented. At last, chapter 5, Conclusions consolidates the work developed, compares the objectives with the results, and presents recommendations and future work. 4 2 STATE OF THE ART The goal of the state of the art is to understand where the research stands in terms of perception of corruption, ways to fight it, the role of open data in this scenario and more specifically the role of LOD. Being this a dissertation, the first searches are conducted in academic paper databases looking for research work. When this first approach is not fruitful enough the search is complemented with reports and other online sources created by credible organizations such as the European Union, Transparency International, and The World Bank. The search for similar projects in the area begins with an academic perspective by refining a query revolving around the keywords in the base idea: ontology, corruption, and LOD Table 1. Transparency is used as an alternative to corruption because it is a way to prevent it. According to TI, it is the duty to act visibly, predictably, and understandably to promote participation and accountability and allow third parties to easily perceive what actions are being performed. It is related to corruption in the sense that “institutions should make a commitment to report annually on the measures they are adopting to strengthen risk management, especially in relation to bribery and corruption. “ [7] For every document with an adequate title and abstract the list of references and citations is also viewed and adequate documents are selected. Most of these documents are not as helpful as one would expect. There is vast and relevant work on both ontologies and LOD, and then in corruption and public scrutiny. The scarcity begins to be noted when these concepts are mixed Table 1. 5 Table 1: The number of publications found in Scopus for the filtered queries. Search Keywords Domains Number of Results Source Ontology AND Corruption Computer Science, Engineering 24 Scopus Computer Science (all available) 1 Web Of Science ("linked data" OR "linked open data" OR "open data") AND (corruption OR transparency) Computer Science, Engineering 578 Scopus Computer Science (all available), Information Science Library Systems, Public Administration 368 Web Of Science ("linked data" OR "linked open data") AND (corruption OR transparency) Computer Science, Engineering 156 Scopus Computer Science (All), Information Science Library Systems, Public Administration 70 Web Of Science ("linked data" OR "linked open data") AND (corruption) Computer Science, Engineering 12 Scopus Computer Science (All), Information Science Library Systems, Public Administration 7 Web Of Science Since the results are limited the next step is to broaden the search. This led to a search on OD as a corruption-fighting tool. The main sources for understanding this subject are the EU, the Organisation for Economic Co-operation, and Development (OECD), the World Wide Web Foundation (WWWF), Transparency International (TI), and The World Bank (WB). The following section funnels down from a broad context of corruption to the Portuguese context followed by symptoms, and good practices. 6 2.1 Background Analysis - Corruption Table 2 Definitions of Corruption as in [8] Corruption Description Primary Source “Corruption is the abuse of entrusted power for private gain. It hurts everyone who depends on the integrity of people in a position of authority” [1], [9] Transparency International “The extent to which public power is exercised for private gain, including both petty and grand forms of corruption, as well as “capture” of the state by elites and private interests” [8], [10] The World Bank Description “Abuse of power for private gain.” [8], [11] European Commission “Active corruption or ‘active bribery’ is defined as paying or promising to pay a bribe (OECD, 2008). The ‘active or passive misuse of the powers of Public officials (appointed or elected) for private financial or other benefits” [8] OECD “Requesting, offering, giving or accepting, directly or indirectly, a bribe or any other undue advantage or prospect thereof, which distorts the proper performance of any duty or behaviour required of the recipient of the bribe, the undue advantage or the prospect thereof.” [8] Council of Europe All in all, every description involves a form of using power for private gain, with a focus on TI’s addition of “entrusted power”, this is important when considering that corruption is a practice here applied to those who hold a public office and were in some way elected or entrusted with that position. [12] The definitions are alike and seem to point in the same direction, however, there are some caveats. The concept of harming the public is ambiguous and dependant on context since there is no guideline to know what is in the interest of the public. [8] Also, corrupt practices operate behind the curtains, except for when cases are exposed [12]. This is an obstacle to any measure that tries to access corruption on its extent, impact, or perception as is seen further on. Nevertheless, corruption was still conceptualized, and, according to [11] there are several known types defined depending on the actor. Grand corruption is applied when the actors hold high offices, while petty corruption refers to small bribes given, for instance to doctors for better treatment or examiners for a driving license. Corruption is also classified as local and central separating the range of the actions, such as in the municipality (local) or on a national level (central). As seen above bribery is often mentioned meaning that the private gain is of the financial sort. Corrupt practices are estimated to annually cost the European economy 120 billion euros, out of these, 5.33 billion euros are associated with public procurement alone [12]. The latter is one of the most observed types of corruption. However, the financial cost alone is not enough of a measure, the practices mess with the social tissue and one of the most prominent and extreme connections with it is society’s lack of trust in public institutions [12], [11]. Growing untrust becomes a problem in itself with consequences such as tax avoidance by discouraged taxpayers and even possibly a threat to democracy. [12] 7 2.1.1 European and Portuguese Contexts Corruption measurements are either based on subjective data, such as the perception of corruption, or objective data such as judicial prosecutions. [8] All the indicators mentioned below fit the first category. The WB and Corruption Perception Index (CPI) are the two most widely used corruption indicators [8]. In common they have the fact that both present their results ordered by country ranks, creating a direct comparison between countries. In 2014 a study developed by Malito comparing measures of corruption criticized both for different reasons: CPI questionably displays the perception of corruption as being connected to the extent of corruption, while the WB displays a biased perspective from large business elites. There was expressed concern over the fact that both aggregate many sources, risking the loss of clarity. It’s important to keep these limitations in mind alongside the fact that more questions were raised than answered by this report. [8] In Table 3 they are both described alongside two Eurobarometer’s developed by the European Commission. Table 3 Measures of Corruption Surveys and Indexes Institution Goal Control of Corruption The World Bank (2019) It is one of the six worldwide governance indicators used by WB. In 2019 it accessed 209 countries using on average 9 sources per country ranging from 16 to only 1. It reflects the use of public power for private gain [13] and aims to aid in instrumentalizing government assistance. [8] Corruption Perception Index Transparency International (2019) Focused on how corruption is seen and perceived in 180 countries around the world. It feeds from 13 sources from 12 institutions and only considers countries appearing on at least three of them. [5] The aim is to raise awareness and demand accountability from political leaders. [8] Eurobarometer on businesses’ attitudes to corruption European Commission (2015) A flash survey to EU countries focusing on their perception of corruption for companies when doing business. Focuses on bribery and public procurement management and overall corrupt procedures. [12] Eurobarometer on corruption European Commission (2014) A survey was carried in 2013 for the commission’s report on corruption in 2014. It was meant to be the first of a bi-annual practice, but it ended up being the only one.[11] They surveyed 27 786 persons on their perception and experience with the matter. The definition from TI introduces the concept of accountability which is very often mentioned when discussing matters related to corruption. Accountability in the context of a business, or entity is the willingness to be judged on actions taken. To be accountable for something is to own upon what is in cause and accept consequences. [14] Across the EU only 23% of citizens consider their governments’ anti-corruption efforts to be effective, while 26% consider being personally affected by corruption in their daily lives, and 67% advocate for more 8 transparency and oversight over the financing of political parties [11]. Bribery and the use of connections are often considered the easiest way to get certain public services by 73 % of the respondents. [11] Companies also struggle against corruption, 56% say corruption in public procurement is widespread on a national level and 60% on a regional/local authority level. In Portugal, this number goes up to 68%. [11] Transparency in election processes and campaign donations are associated with low levels of corruption. Countries, where campaign finance regulations are comprehensive and systematically enforced, have an average CPI score of 70. Moreover, enforcing better practices in this matter was part of the policies in 60% of the countries that raised their CPI between 2012 and 2019. [5] Although the data used by CPI is subjective, corruption scandals and prosecutions (absolute data) have been shown to directly influence the respective countries scores. Table 4 Corruption Indexes Scores Geographical Area CPI (2019) WB - Control of Corruption (2019) Rank (1-180) Score (100-0) Rank (0–100) Score (-2.5–2.5) EU - Average 34 64 77.2 0.95 Portugal 30 62 77.4 0.76 Highest EU Score Denmark - 1 87 Finland 2.15 99.04 Lowest EU Score Bulgaria - 74 43 Bulgaria -0.16 50.43 The results presented in Table 4 place Portugal shortly below the EU average score for both CPI and WBCC (World Bank Control of Corruption). WBCC is only one of six indicators used by the WB, from these, the one where Portugal performed worst was Regulatory Quality which “Reflects perceptions of the ability of the government to formulate and implement sound policies and regulations that permit and promote private sector development.” [13] These scores, alongside information about businesses' high perception of corruption, raise the question of whether there is a lack of trust in the government? And if there is, what is causing it? There is not enough data to answer these questions now, but further research will keep them on the table. 9 2.1.2 Syndromes of Corruption Some wrongdoings are more common than others, or at least appear so based on the corruption indicators results and reviews. Businesses often struggle with public procurement mainly in four fields: 1) Tailor-made specifications for certain bidders; 2) Splitting of public tenders into smaller bids to seem less attractive to competitors; 3) Prominence of conflicts of interest; 4) Excessive use of emergency claims to avoid publication. [12] Deep-rooted corruption is an obstacle to economic development. However, the “Mobility of labour between the public and private sectors are essential for the functioning of modern society and can bring major benefits to both the public and the private sector.” [11] For a healthy relationship between the private and public sectors and a balance between both statements above, the risk of public officials’ giving preferential treatment to former or current private-sector connections must be mitigated.[11] 2.1.3 Good Practices Before listing good practices, one should be aware that corruption-fighting measures are highly dependent on social, political, and legislative contexts. [11] Policies efficient in one country might not be on its neighbour if there are social nuances, and this is especially true when it comes to law and third-party agencies, as is explained further down. The impact of legislation and administrative measures is dependent on the cultural mentality of both public bodies and society. To sum it up, recommendations fit contexts, not countries. [12] One of the suggested ways to improve cultural mentality is by further involving the public in political decisions while diminishing the influence of money. Political consultation needs to be fair, impartial, and non-partisan. [5] A key pillar of the fight against corruption is prevention. [12] If, as seen in the previous chapter, public officials economic interests are a risk factor, then there are two options: either to prohibit certain activities deemed susceptible to corruptive practices, establishing “cooling-off” periods for an efficient and verified transfer, and applying sanctions; Or to increase accountability and transparency through an effective asset and interests declaration, making possible transgressions easier to spot and possibly corruptive practices accessible to public scrutiny. This also improves public trust. [12], [11] Both [12], [11] advert for thorough verification of asset declarations, preferably done by third party agencies provided with enough power, tools, and access to conduct investigation efficiently and spot the lack of or incorrect information. When this verification is involved in red tape, or it is not sufficiently 10 endorsed it can be a problem. Verification really is the key here. When assets declarations exist but lack verification, they can create the illusion of transparency and accountability while not fully doing their part. Many member states lack these tools, but a good example is Romania where a third-party agency is capable of scrutinizing and validating more than half a million declarations a year. This work has led to finding both omitted and wrong information and caused public officials to lose their offices or seats. Preventive policies have had great success when implemented effectively, and not so much when it is done in fragments. 2.1.3.1 Open and Transparent “Transparency is a particularly strong tool in the fight against corruption. Freedom of access to information improves good governance and helps to make government more accountable.” [12] Above, trust in the government is linked to corruption levels, preventive measures are deemed mandatory, and transparency seems to be linked to both. The remaining good practices were therefore narrowed down to those focusing on transparency and openness in a preventive way. The more transparency there is in the public sector, the fewer opportunities for corruption there are, making markets more competitive in a healthy way. For these practices to be transparent the processes, results, and documents must be open. Preventive measures are showcasing public contracting and expenditures in at least Croatia, Estonia, Portugal and Slovenia [12]. Portugal’s platform is Base.gov.pt where all public contracts are uploaded by the hiring company. “The publication of contracts in both BASE and the Official Gazette is now mandatory for direct adjustments, increases of 15 % in the price of already concluded contracts and potential penalties.” [11] To this day, the fact that data is not time-sensitive, and contracts being published with months delay is still a limitation. High standards of transparency in public procurement are a recommendation of the EU as of 2014. It is included in the preventive policies alongside easy access to public interest information and the assertion that it helps mitigate corruption in public administration and political party financing. Achieving transparency using OD became a trend in such a way that 21 EU members are part of the Open Government Partnership whose goals are to promote transparency, empower citizens, fight corruption, and harness new technologies to strengthen governance. There is no simple guideline that works in every situation. However, reductions in corruption have been seen in countries lacking regulation and strategic programmes but with a preventive and tactical high transparency standards approach. The same cannot be said for the implementation of complex legal and institutional programs in countries with serious corruption problems. [12], [11] 11 2.1.4 Final considerations 2.1.4.2 Agencies and government programs These complex programs and the act of anti-corruption agencies are not a lost cause, they both work but are highly dependent on context. Agencies run the risk of creating the appearance that corruption is being dealt with but not being given the tools to do this job. Simultaneously, agencies achieved amazing results when their powers and legitimacy were not under constant pressure, and they have the cooperation of public entities.[11] Strategies and programs are more complicated since they have, in some cases, been essential and in others had little effect. The objective reasons why they succeed, or fail are not explained. 2.1.4.3 Legislation To conclude this chapter, and thus turn solely to data, it is essential to look at some legislation. “An effective policy response needs to be based on evidence about its prevalence and forms in a given country, the conditions that enable it and the institutional and other incentives that can be used against it.” [12] Studies have found a positive link between the quality of the legal system and the level of institutional trust. [15] Also, good data is needed to hold governments accountable for their actions [16]. When the rules are not effectively enforced systemic problems remain unattended, and there are no concrete results. “Track records of prosecutions and final court decisions are weak, and few cases of public procurement corruption are finalised with dissuasive sanctions.” [11] Formal cooperation is not enough if people remain under pressure to be silent. Corruption networks are often dismantled through whistle-blowers, so there needs to be a culture aided and supported by the law to protect these people. The United Kingdom (UK) and Ireland are great examples of legislation about whistleblowing. There, whistle-blowers are protected and encouraged to make any wrongdoing public no matter the intentions behind the revelations or the employment status. There is also financial aid if whistleblowing is the cause of dismissal. Another example is the Netherlands, one of the few member states where integrity policies are successfully an active part of public administration both locally and nationally. [12] For law enforcement to work, it must be both independent and equipped to fight corruption. “Likewise, striking the right balance between privileges and immunities of the public officials and ensuring that these are not used as obstacles to effective investigation and prosecution of corruption allegations is still an issue in some Member States.” A workshop organised by the Commission in 2015 showed that law is more likely to be efficient when it comes out of public discussions. [12] Public interest and awareness is, once more, key for the success of law enforcement. 18 For each sector of corruption at least one project is given as an example. These served as a starting point to find keywords and references that lead to other similar projects. This chapter explores the connection between the good OD practices laid out in 2.2 Background Analysis – Open Data and the projects that materialize some or all these practices. First and foremost, let us settle on common ground. Although OD can be used as a corruption prevention tool it is not always aligned with this interest. [18] The databases made available to the public are not always the ones the public is more interested in, or the ones needed to create value. [32] The motive for making data available is also relevant. Often, OD serves as a way of showing transparency over what has already been done. This is where the difference between transparency and OD lays. Transparency can be achieved with OD, but it is only one of its many possible uses. For instance, for OD to be proactive and enable decision making it should be released and available for scrutiny before decisions are made, not after to show transparency over what is already done. This is one of the reasons why it is important to create OGD that sustains and aids in decision making, having a direct impact on governance instead of being a final display of carried out actions. [33] An example of data that has the potential to aid in decision making is lobby registers, however, this is a rare dataset, for instance, a report looking into five G20 countries found that only one published lobby registers. [17] This is different from the EU lobby register mentioned in 2.2.1 Availability, Status and Evaluation which is not mandatory on a national level. The impact of these projects and the amount created are hard to assess, in part because documentation on impact is lacking. [32] Most data re-use projects are government independent initiatives Table 6. An attempt to evaluate public interest in re-using data would also be faulty for two factors: 1) As seen above and again on these reports, training stakeholders to produce and use data is crucial. [32] As long as this gap exists, the access to data of public interest cannot be assessed; 2) As long as governments are creating only human-readable and digested data displays with attractive visualizations, instead of providing access to raw data, remains a barrier to re-use, and might not even pass the criteria for accessibility [21]. Creating final reports with charts is a very common practice useful to give the public an overall idea and a sense of transparency but it is not useful for exploring the data. Moreover, by the standards of the ODC, data visualizations are not OD, since they do not have the technical features needed to be re-used by anyone. The remainder of the chapter dives more into the relation between OD re-use and the prevention of corruption, how to incentive and apply it, what different reports agree on, and the main challenges. 19 2.2.4 Open Data Standards Good quality OD is emerging as an anti-corruption tool. On top of the requirements for OD established above, uniformization, standardization, machine-readable formats, and metadata are important to compare different datasets and look for patterns, trends, and anomalies that can expose corruptive practices. [18] A comparison between the CPI and the ODB for 2016 (most recent ODB) shows a connection between the values in Figure 1 in the form of a strong correlation (0,71) between the two indexes 2 . A report by TI points out that OD “should contribute to anti-corruption reform” [17] and the ODC states that “Open data can play an effective role in dismantling corruption networks” [34]. Figure 1 Scatter plot of CPI and ODB scores [17] 2 For this comparison only the countries scored in both indexes are evaluated, that makes a total of 113 countries. 20 OD is a vehicle for achieving transparency and citizen trust. These are values that impact the perception of corruption, however, it is not common for the impact of OD and the Perception of Corruption to be officially connected [33] as is showcased in Figure 1. To further understand the relationship between OD and the perception of corruption, six different reports on the matter, ranging across America, Asia, and Europe 3 , were investigated. The remainder of this chapter is based on these reports. There is no recipe for corruption fighting, but there are good practices, and they are highly dependent on context. The role of OD is tricky here. Theory points to transparency and OD being enablers of accountability, which is a way to fight and prevent corruption. But OD has not been widely used in corruption investigations. [33] Despite it not being common, it is still a recommendation. The G20 established a set of principles directly connecting the anti-corruption measures and OD [18], and the ODC makes it clear that OD has an important part to play in the anti-corruption reform. [17] The impact of OD in anti-corruption is related to the third pillar from the OECD, data reuse. This is the stage that allows for further investigation from third parties which are then able to mine data for incorrect practices. [33] These are even more powerful when the publishing format allows and incentivize triangulation, linkage, and comparison. [17] Besides the technical features of the data, the content is also important for re-use. In these scenarios, more data is not the same as better data. Certain datasets are more relevant for anti-corruption than others. The “Open Up Guide” divides these priority datasets into three types: registers, public disclosure, and transactions. Amongst the total of 30 datasets, the top 5 are all registers, namely: interest declarations, lobbying register, company register, charity register, and politically exposed people's list. [34] The G20 also listed 10 key datasets [18] that are of the public interest [34]. Below in Table 6 are the top ten datasets from the Open Up guide and the ten datasets from the G20. 3 [32] Malaysia [1] EU [20] Latvia, Sweden and Finland [18] Brazil [19] Indonesia [35] India 21 Table 6 Priority datasets [34] and key datasets [18] Open Data Charter [34] G20 [18] Interest declarations Lobbying Registers Lobbying register Company Registers Company registers Beneficial Ownership Registers Charity register Directories of Public Officials Politically exposed people's list Government Budget Records Public officials register Government Spending Records List of government contractors Public Procurement Records Corruption-sensitive postings Political Financing Records Council / advisory board members Legislative Voting Records Contracts register Land Registers The four shaded titles on each column represent the datasets that are both in the list from the G20 and the top ten from the ODC. From the six key datasets in the right column of Table 6 that are not shaded, only the Beneficial Ownership Registers is not mentioned at all by the ODC. These recommendations are extremely valuable. Opening government data is a process, and, since it cannot be done all at once, institutions must prioritize release. With this guide, prioritization can be done strategically and governments should be choosing the datasets known for creating greater impact upon release. [1] One thing is knowing what, in theory, should be available to apply scrutiny on the public sector. Another thing is knowing if it is working. The reports analysed all showed similar approaches and challenges. Amongst the main challenges identified the one that pops out the most is lack of training. This applies both to public officials and the general population. [20] When evaluating datasets, the challenges were also quite similar. Overall, the published datasets are not always the ones needed and the formats are not easy to re-use resulting in a lot of time spent in data treatment and transformation. [32] The conclusions are also similar, in Indonesia, a report states that there is an effort, but it is not enough. The country needs a dynamic platform to create value. [19] An analysis in Brazil adds that functioning legislative frameworks on Data Protection and the Right to Information are key for policies to evolve and transparency to become cultural. [20] The main challenges related to re-use are the technical and content quality of the data and the data literacy in civil society. The first challenge reflects itself in that low quality and non-centralized data makes it harder and more time consuming to follow connections, names and stories across different sectors [33]. As for data literacy, it is connected to the technological gap affecting public officials, investigative 22 journalists and civil society that disables them from using data to prevent and detect corruption. [17], [18] 2.2.5 Data Literacy The lack of data literacy has been repeatedly mentioned in the last sub-chapters as a challenge. Although there is also no quick fix for this technological gap, the paragraphs below display some examples of both training sessions and incentives to data re-use that have been successful in different countries. Finland, Latvia and Sweden formed a Coalition to foster the use of ICT technologies that resulted in both training projects and programs to raise awareness in civil society. [20] In Brazil, several e-learning courses, online resources for government staff, and guidelines on how to use OD are only some examples of the tools available to boost the impact of OD.[18] In Malaysia, the “Sinar Project”, on a limited budget, managed to create consumer applications and research notes based on their combined database. The database connects politicians to the companies they own, and the constructing deals these companies landed. The project includes data on 4000 people of interest, 200 organizations/committees, 900 government posts and 222 current parliament members. This project was expanded to allow exploration of local politicians, private and public sectors. [32] In India, “Project Background” displays arranged data visualizations for the whole country and, one instance on a state level (Maharashtra). This project used data from different national and international databases to create a diagnostic tool. The description was completed with the outcomes of a Mumbai workshop to include public opinion. The remarkable thing about this project is that it includes public opinion as the main part of it and allows for replication on any state inside or outside India as long as they hold a workshop similar to the one in Mumbai. [35] The model proved to be scalable, as promised, through successful implementation in Mexico and Mexico City. In conclusion, the main challenges related to re-use are the technical and content quality of the data and the data literacy in civil society. The first challenge reflects itself in that low quality and non-centralized data makes it harder and more time consuming to follow connections, names and stories across different sectors [33]. As for data literacy, it is connected to the technological gap affecting public officials, investigative journalists and civil society that disables them from using data to prevent and detect corruption. [17], [18] 23 2.2.6 Open Data Re-Use Projects All the OD re-use projects mentioned in the introduction to 2.2.3 Open Data Re-use are analysed below in Table 7. The analysis is part of this project and aims to better understand how the scenario for data reuse laid out above materializes in online projects. Description of Table 7: Title – This is the name of each project. Corruption area – One or more areas are attributed to each project. The areas were picked according to [1] which based them on the ODB and describes them as “forms of corruption and affected sectors as dictated by leading anti-corruption indices and surveys. (…) it rather focuses on sectors and corruption-related issues that are most pressing and evident.”[1] Then, there are six Boolean variables where (X) means it exists and (blank) is not found. Due to language barriers in certain projects, the information had to be translated online. This means that there is a possibility that some (blanks) exist but were simply not found. Each of these six fields is now further explained: Is Active – Whether or not the website is up, and the information updated. “Diferentonas” is the only case on which the website is unavailable, but the information is found in the report. The others have the website up, but the information is old or discontinued. Private Initiative – Boolean on whether the project is a private initiative or public. All public initiatives are governmental on either a local or national level. Public initiatives are still data re-use projects when they feed on other portals. For instance, “Portal Municipal” feeds on Dados.gov.pt. Data source – Whether the data sources for the projects are specified so it is possible to trace the data re-use back to its source it is also considered true if it is made evident that the data is generated by the project. OD Re-use may be partial. Source Code – Whether the source code is made available. Almost every true case has a repository on GitHub 4 . Shows Method – Whether or not the website explains how the data is treated and analysed. Some positives are more specific than others almost allowing replication. Has documentation – Whether articles (academic or not) and reports were found about this project. It is very helpful since reports tend to be very comprehensive and justify the means and the gap the project came to fill. Data Display Mode – Most of these are projects that handle a lot of data and considering the principles of OD mentioned above it is important to evaluate how the user data is displayed. This column shows 4 https://github.com/ 24 how the data can be accessed and if it enables further re-use. For instance, visualization projects may draw an interesting conclusion but create a barrier for further work with the same data. Table 7 Analysing projects Title Corruption Area Active Private Initiative Data Source Source Code Shows Metho d Has Docs Data display mode Influence Explorer Political parties Business Lobbying X X X API & Bulk Abgeordneten Watch Parliament / Legislature X X X Visualization & Q&A STIRNA Media Corruption X X X Visualization, JSON, XLXS & CSV Open public contracts in Slovakia Business Public Procurement X X X X API (paid) & Bulk Open Courts Judiciary X X X X X API Rekvizitai Public Officials Civil Servants Public Procurement X X X XLXS and CRM integration (if requested) Tender Public Procurement X X X X Visualization Project Background Public Procurement X X X X X Visualization As diferentonas Public Procurement X X - RECORD: Reducing Corruption Risks with Data project Public Procurement X X X Visualization (links source) Portal Municipal Public Procurement X X X CSV OpenSpending Public Procurement X X X X X X API, CSV, bulk, Visualization (personalized) Where does my money go? Parliament/Legislature X X X X X Bulk Vouliwatch Parliament/Legislature X X X CSV, Visualization & Q&A Lisboa Aberta Public Procurement X X CSV Votaciones – La Nacion Parliament/Legislature X X X CSV & Visualization Dollars for docs – How industry dollars reach your doctors Public Procurement X X X X Visualization & CSV (paid) NarcoData – Animal Politico Public Procurement X X X X Visualization & Tabular AtuServicio.uy Public Procurement X X X X X Visualization (compare) Open Contracting Public Procurement X X X X X X Tabular, JSON, CSV Quien es Quien Business, Lobbying X X X X X API & Visualization (filter) The sample of projects selected turned out to support the conclusions depicted above. The most prominent corruption is depicted in Public Procurement, which is in focus on 12 of the projects, it is followed by Parliament/Legislature appearing in 4 projects. When it comes to the Boolean variables, being a private initiative (18) and displaying their sources (18) are the two most common characteristics. Amongst the first, there are mainly newspapers, third party corruption fighting agencies, and independent citizens. The source code of the project is, unfortunately, 25 only available for 6 projects. Having some kind of methodology description is a more common practice taken by 12 of the projects. Nine projects also disclose documentation including both reports and news articles. From the 21 projects most showed only visualizations and only four had an API. These projects are, at least partially, re-using OD, which means that is available on its source. Nevertheless, it is a good practice to make your data available, so others can replicate similar projects or triangulate the project's data with new datasets. All in all, this short analysis allowed for two things: understanding the practical side of the OD challenges and recommendations discussed above; designing the first sketch of the final project based on what exists online and what is working with people. A very interesting project is “Abgeordneten Watch” which was founded in Germany and lead to the creation of the Parliament Watch Network 5 , from which “Vouliwach” above is also a part. This project consists of a website with information on members of the German Parliament (data re-use) and a question and answers feature where the public can directly ask questions to the politicians, and they can respond. The implementation was very successful, and the project is active. Further down in the “Methodologies” chapter, the impact of this project on defining the structure for the project developed in this thesis is further explained. OpenSpending is a free, open, and global platform to search, visualise and analyse fiscal data in the public sphere. It comprises over 3,527 data packages from 85 countries with over 187,226,528 fiscal records. NarcoData is a tool that recompiles trusted information on the last forty years of drug cartel history in Mexico. It is managed by Animal Politico, a native digital Mexican newspaper. It is especially interesting because it touches on a topic that runs in the shadows and back streets and about which it is not easy to find organized and structured data. “Quien es Quien” is about contracts between public entities, businesses, and politicians. This tool makes it easier for investigators, journalists, and the common citizen to access OD from different sources hence enhancing transparency and accountability. This is a pure re-use project that has the simple aim of making research easier. The four projects singled out in the previous paragraphs are not necessarily the best of more complete projects, they are simply examples of variated the application of OD can be. 5 https://parliament.watch/ 26 2.3 Linked Open Data “Info that is not linkable is not used, information that is not used is not valuable” [36] This is a bold statement and one that is fit to start the discussion about LOD. 2.3.1 Semantic Web First things first, the concept of LOD is preceded by that of the semantic web. The latter was introduced by the creator of the world wide web, Tim Berners-Lee, and refers to the existence of a web that is readable both to humans and machines. What this means is that the web as we know it today is limited, it provides access to enormous amounts of information, but this information has already been processed and isolated. To compare it to more online available information a traceback must be done and do all that work is manual. This is the online equivalent of roaming a library for different books mentioning the same theme. The break of barriers and the unleashed potential of the internet are in being able to have all this information linked and available for comparison. This is what is discussed in this chapter. Having semantic in the web context essentially means that information is represented in a graph-based data model that facilitates extension, integration, and inference. [4] Giving semantics to data is the process of linking it to more data to be as understandable for a computer as a written text would be for a human. The preferred data format to achieve this goal is RDF, which allows the links connecting data to be described. Ultimately, the connection between documents is implicit, something that is often lost in regular data formats found on the Web of Documents. [37] The web of documents is the web of today, with separate websites and pages showcasing their own information. The Web of Data is an extension to the web of documents applying semantic principles. So, they naturally share some properties: There are no restrictions on to what data types can be contemplated; Everyone can publish data; Publishers can choose their vocabularies; Entities are connected by RDF links. [37] From a development perspective, in “Linked Data – The Story so Far”, the web of data is further described: Data is strictly separated from visualization aspects; Data is self-describing; HTTP is a data access mechanism and RDF a data model; Applications are not stuck to a fixed set of data sources, new data sources can be discovered at run-time by following RDF links. [37] Understanding the Web of Data is understanding the ultimate, long-term goal of LOD implementations. LOD is a set of best practices, and those are what is discussed further on. 27 2.3.2 Linked Data Linked Data refers to “a set of best practices for publishing and connecting structured data on the Web.” [37] In practice, it means that all data is linked to instances or connections from different sources. It can be applied on an organizational level connecting databases that are not interoperable, yet. Or, on a transorganizational level linking data from different organizations in a unique shared database, as is the case of data sharing in Alzheimer's disease research. [38] The Alzheimer data share case started from the following premise: Professionals agreed that sharing data on Alzheimer research was valuable but did not know how to do it because every organization’s internal database had a different structure. They found in LOD a method for tagging data from different organizations around the globe with the same tags and lists of terms, therefore, having access to more data and supporting better research. This short description shows that, by nature, the requirements for LOD present a solution for some of the barriers identified above by open-data re-use projects. Namely, the challenges related to scattered noncentralized data. Back in 2006, Tim Berners-Lee defined a set of four base principles for publishing LOD on the web: Use URIs as names for things; Use HTTP URIs so that people can look them up; Provide useful information on URI’s using the standards (RDF, SPARQL); Include links to other URIs. [37] Overall, they lead to one of the main cornerstones in LOD, interoperability. In Table 8 the main technologies needed for understanding LOD are explained. These technologies are “semantic”, so the information is “represented on a graph-based data model that facilitates extension, integration, inference and uniform querying.” [4] Table 8 Main Linked Open Data Concepts Explained based on [37] and [39] URIs Uniform Resource Identifiers use the http:// scheme to identify any entity in the world. These entities can be looked up by dereferencing the URI over the HTTP protocol. [37] RDF Resource Description Framework “A graph-based data model that facilitates extension, integration, inference and uniform querying.” [4] RDF is based on triples: subject (S), object (O) and predicate (P). S and P can be strings but should be both URI’s and O is what connects them. For instance, S is the mother of P. “is the mother” is the object that gives context to the link between the two. [37] SPARQL SPARQL is the language used to query an RDF dataset and the semantic Web. A SPARQL endpoint is a service that processes SPARQL queries and returns results. [4] Its logic is similar to SQL. It enables any user to query the web as if it is their local database. [40] Vocabularies “Vocabularies provide the semantic glue enabling Data to become meaningful Data.”[41] They are essential to LOD. Whenever possible well-established vocabularies such as FOAF, SIOC, SKOS, DOAP, vCard, Dublin Core, OAI-ORE or GoodRelations should be used to make the 34 The second challenge is the LOD knowledge curve challenge. There are plenty of tools out there but most of them are not user-friendly for querying LOD datasets without knowing a query language such as SPARQL. [50] SPARQL knowledge is considered a limitation since it is not a widely dominated skill, its potential for next to unlimited queries ends up being limited by the need to learn the language. [43] Moreover, the lack of general knowledge about LOD itself ends up also being a barrier for the users to take advantage of the full potential of this data model. [50] However, usability can be prioritized for non-specialized users through the implementation of natural text search and other possibilities. Another technique, used by [43], was creating redundant properties to enable less complex and shorter queries and avoid reasoning engines. This practice is ill-advised for large datasets. The third challenge is asserting trust in data. Therefore, it is so important to connect to known vocabularies and publish metadata. Linking helps to assure maintenance and reliability through being transparent about data provenance. [37] To conclude, four main challenges need special attention: Data quality and availability; Lack of a standardized methodology for implementation; LOD knowledge curb; Data Trust. 2.3.6 Implementation Although there is no set methodology, there are guidelines commonly followed. “How to Use Linked Data” [4] proposes a checklist for creating LD. This checklist, represented in Table 9 is a decomposition of the three basic steps for implementing LD suggested by Bizer, Heath, and Berners-Lee. [37] the three basic steps are in the first row separating the columns. Table 9 Linked Data Checklist from "How to use Linked Data". Creating LD Interlinking LD Publishing LD Verify that all relevant entities/concepts were effectively extracted from the raw data? Link dataset to other RDF datasets Provide dataset’s metadata. Provide information about licensing. URIs are dereferenceable Created terms are linked to other vocabularies Create alternative access methods: SPARQL endpoint; Data dump. Terms used are from widely accepted vocabularies. Only non-existing terms were created. Register the dataset in LD catalogues via a Data Hub or CKAN. 35 This checklist helps guide the implementation process. Despite not being mentioned above, one last step that can improve this checklist is the addition of other Visualization techniques. [50] When doing a re-use project, the step that comes after selecting the data sources the first step is extracting the data from said sources. The complexity of the next step depends on the format of the data file (XML, CSV, JSON, SQL, PDF). For sources with data dumps, a common procedure is simply standardizing all the sources to the same data format file, for instance, JSON. For other web sources with no data dumps a common practice is crawling the websites for the information needed. This can be aided by tools such as GATE, Zemanta and DBpedia Spotlight. The result of this procedure is then transformed into the same type as the remainder data. [49], [3], [4] Depending on the formats available the next step may be cleaning and normalizing the data or jumping forward to transforming it to an RDF format. The latter is a crucial part of any LD or LOD implementation. This is when vocabularies are chosen to enrich the data, relationships between triples are defined and properties are chosen. [48] Increasing interoperability relies largely on the choice of well-known and used vocabularies like FOAF and Dublin Core. [43] The relations defined should also have inheritance between them, this can be partially achieved with the use of properties such as rdfs:subPropertyOf and rdfs:subClassOf to relate properties or classes from different vocabularies. On the schema level, RDFS, OWL and the SKOS are useful for the same goal. [40], [4] Another important thing is, when possible, linking to external knowledge sources already in the LOD cloud. [43] One of the main tools to do so is DBpedia. [50] It is important to remember: LOD can never be isolated. This process of transformation is tricky but there are tools that aid in the transformation of both tabular data (OpenRefine 7 ) and ER (R2Rml 8 ) to linked data. [4] Moreover, whenever possible the URIs should be human-readable, using specific references to entities. When working with LOD the application profile should not be restricted by local usage samples. The data should be published “as is” with as few logical constraints as possible to allow flexibility. For instance, when defining an Application Profile it would appear acceptable stating that “the triplet: "author is Academic Member" requires that any author is personal of the institution.” [40] This would be acceptable in local implementations, and restraints such as this are common in ER models but the reality is that, in many cases that restraint does not apply. [40] In LOD restraints that have a local or project, the specific scope should be avoided to increase re-use of the ontologies. 7 https://openrefine.org/ 8 https://www.w3.org/TR/r2rml/ 36 Next is the implementation phase. Here we must remember the importance of metadata to describe the dataset, element definitions, value lists, provenance information, construction and transformation methodology, and potential use cases, and limitations. [3] After the data source is translated into RDF, it needs to be made available for further consumption. When everything is ready to be published, a specialized LOD server such as Virtuoso is recommended. [50] This step enables querying the data with SPARQL. It is also recommended to implement enhanced services based on semantic methods, visualization techniques and other tools to increase human usability and improve the final user’s experience. [3], [48] The process is complex and there are numerous guides and recommendations for implementation. For that reason, this chapter did not go further into exploring tools. That is done later in the methodologies with everything being applied to this project’s case scenario. 37 3 DATA AND METHODS Inside Data And Methods lays the description of every decision made designing the proposed solution and how it was based on the state of the art. From the data sources to the tools and methodologies. In Figure 2 Concepts Map there is a recap on the concepts presented in the State of the Art, how they are connected and the navigation throughout them that led the researchers to the tools and materials. Figure 2 Concepts Map The state of the art starts with the definition of corruption, how it is measured, the syndromes, and the good practices to fight it. For the scope of this project, the focus is on good practices, in specific prevention. Literature on prevention showed that OD is a corruption-fighting tool through scrutiny enabling transparency, hence the title of this document. The State of the art also showed that one of the measures for the impact of OD is the level of Re-Use and that the highest level of OD Is LOD. A considerable number of projects defining themselves as an “Open Data Project to Prevent Corruption” were analysed. As for LOD projects making the same statement, not many were found. Nevertheless, in spite of not having the prevention of corruption as a stated goal, the topics of the governmental LOD 38 projects analysed are in line with the areas of interest for Open Data in the prevention of corruption defined by the ODC. Before moving to the characterisation of the proposed solution it is important to point out that this is not a one-way path. Meaning that prevention is not the only way to tackle corruption, transparency is not the only way to act preventively, and OD is not the only way to achieve transparency. It is, however, a path well documented and sustained by literature and the one that best fits the scope of this work. For context and research credibility the state-of-the-art chapters put these concepts against their neighbours and alternative paths. This information is summarized in Figure 2 in a concept map resulting from the research done above. The proposed solution places itself in a gap found in the literature: There are few Open Data Re-Use projects with machine-readable and linked data. The data selected is aligned with the datasets of interest found in the state of the art for corruption and transparency, which means it is a scrutiny enabling project. The technical field is the semantics web, more specifically LOD. 3.1 Method - Design Science Research This project follows Hevner’s three-cycle approach to design science research. [6] The author states that “practical utility alone does not define good design science research. It is the synergy between relevance and rigour and the contributions along both the relevance cycle and the rigour cycle that define good design science research.” [6] From a broad perspective, the idea behind the original model Attachment 2 A Three Cycle View of Design Science Research [6] is that there are three pillars: Internal Environment; Build Design Artifacts and Processes; Foundations. They all iterate with the pillar immediately before and after through the Relevance and Rigor Cycles. The Build and Design pillar also iterates with itself through the Design Cycle. In 2016 the optional External Environment pillar and Change/Impact Cycle were added to the model [53] as an expansion to the original three cycles of 2007. [6] This expansion is not used because interaction with the external environment is out of scope for this project. The guidelines behind this approach date back to 2004, [54] and each is thoroughly explained therefore aiding in applying the model to any project. Table 10 Design Science Guidelines according to Hevner shows every guideline applied to the project at hand. 39 Table 10 Design Science Guidelines according to Hevner Guidelines Map4Scrutiny Domain Guideline 1: Design as an Artifact The main artifact is the final linked open data dataset. Guideline 2: Problem Relevance Stakeholders’ concerns about corruption aligned with readiness and potential of Open Data. Guideline 3: Design Evaluation Linked Data Principles, Open Data Quality, Shape Validation, and Querying capability. Guideline 4: Research Contributions Improve the impact and scrutiny potential of already Open Data by transforming it to LOD. Guideline 5: Research Rigor Methods used in other Linked Data applications with adjacent themes, best practices, and guides. Guideline 6: Design as a Search Process Multiple-step and prior-knowledge based implementation. Guideline 7: Communication of Research The proposed solution is open to the public and articles will be published. Each guideline is an important step of academic work. Design as an artifact created the final product, whether it is a theory, a framework, or a model. In this case, it represents the final LOD implementation. Every other guideline either justifies, adds value, or reflects the impact of the artifact. The initial idea of using LOD as a tool to fight corruption was broad and had little supporting research. It took a high level of abstraction, going back to the isolated concepts of LOD and Corruption to understand if there is a connection, if this connection has causality and how it can be approached. Below, Figure 3 shows Hevner’s model applied to this project with the information from the guidelines above in Table 10. The Background and Gap Analysis on both Corruption and OD established the need for projects of this nature along with an existing connection with LOD, the extent of which is further explored in the Internal Environment. As showcased in Figure 2 the research goes from an overall search including three different concepts to a specific prototype. A key factor in fighting corruption is preventing it, which can be achieved through transparency that can be achieved with OD, which’s the highest quality is LOD. Topic-wise, Portuguese citizens, one of the 40 stakeholders, are interested in data about the activity of politicians and public officeholders. Thus, the proposed solution will transform OD that fits these criteria into LOD. This implementation is also aligned with the notion of F.A.I.R Data. Data that is Findable, Accessible, Interoperable, and Reusable. Figure 3: Drechsler, Hevner (Ed.) 2016 - A three-cycle view of design science research [6] Going from left to right in Figure 3 we have the three cycles applied to this project. The people, societal systems and technical systems set the idea for the project. In this case, we are talking about the first half of the background research that shows how corruption is a concern if it is of interest to the public, and that OD is a possible solution, but most implementations lack quality or incentive to re-use. This takes us to the problems and opportunities where the scope is defined. Ideally, this project aids in sustaining the arguments for publishing OD with better quality and in machine-readable formats. Thus, showcasing the potential relevance of the proposed solution. These are the requirements for the chosen datasets, the main problems, and the opportunities. [6] Understanding this environment is crucial to designing a proper solution in the next pillar. Here the topics of interest to the public and the barriers to the existence of similar projects are the focus points. 41 The building pillar is where the design and build of the proposed solution happen. This pillar alone iterating itself multiple times is the application development. It is its connection with the other two that justify the relevance and create knowledge. The work being developed must remain relevant to the internal environment and improvements must be explained and measured. This is the difference between design science research and simple application development. [53] All the work happens within this cycle, but every decision is based on knowledge brought from the rigour cycle. The relationship works both ways. Existing projects contribute with methods and knowledge and this project contributions back to the knowledge base for future projects. The foundations include inspiration sources, for tackling opportunities/problems, existing artifacts, analogies/metaphors, and theories. “Research contributions to the knowledge base are key to selling the research to the academic audience just as useful contributions to the environment are the key selling points to the practitioner audience.” [6] By following this approach, one can create and develop while keeping a good grasp on research, inspiration, and relevance. The pillars iterate with each other multiple times in each step of the development. A sample of the result of this iteration is the use of web scraping and reification techniques. These are mentioned but not further explored in the state of the art because the need to use them was caused by the characteristics of the selected data. The use of these tools was not planned before research, it was a consequence of the first iterations between the second and the third pillars. The need was felt by an obstacle in the design cycle, the solution for the obstacle was found in the foundations and therefore included. The evaluation on good practices and technical correctness of the approach is done by basing the decision making in literature: Is the dataset considered “Of Interest” by the ODC?; Is the Linked Data checklist complete? These two questions are answered based on Table 6 and Table 9. The conformance of the data transformed with the application profile is validated via ShEx by parsing the data against a defined shape schema. At the bottom of the last column, the tools gathered are part of the knowledge base to build the solution. These contribute back to the knowledge base in the form of output files that help future researchers in following the same approach. These output files also materialize the objectives set in the Introduction. 42 3.2 Proposed Solution - Description The Overall design and implementation of the solution fit in the second column of Hevner’s four-cycle view of design science research portrayed in Figure 3. 3.2.1 Data Sources Before the LOD implementation, the first step was defining a data scope for the prototype. This is subdivided into two more specific steps: Narrow the choice down to a few databases; Select what data is being used from each selected dataset. Considering the importance of data provenance, trust, the public interest, and the concept of F.A.I.R data that is explained above, the following list of requirements for the datasets emerged: • Open Data is free to access, extract and Re-use. • About entities or activities susceptible to corruption. • Of interest to the public. • About Portuguese entities. • Aligned with the key datasets presented in Table 6 Table 11 Seven Up to Date Open Databases in Portugal Name Focus Search Options Export Options Base.gov.pt Public Procurement Keywords Search bar and filters Bulk with multiple data and PDFs for contracts dados.gov.pt National Open Data Portal (multiple themes) Keywords Search Bar and Navigation through categories Bulk and API Pordata.pt Municipalities, Portugal, and Europe (multiple themes) Navigation Through categories and Advanced Filtering Bulk (data can be filtered before exporting) transparencia.sns. gov Health Keywords Search Bar and Navigation through categories Bulk datasets estatisticas.justica. gov Judicial and Law Navigation through categories Mostly PDF’s dadosabertos.turis modeportugal Tourism Navigation Through categories and filters Bulk and API INESCtec Research (multiple themes) Keywords Search bar and filters Bulk INE Social Statistics (Multiple Themes) Keywords, Navigation Through categories, and Filters Bulk and API 43 Since no document was found with a list of the main or most relevant databases in Portugal, the databases portrayed in Table 11 are the ones often mentioned in European documents (dados.gov.pt, and Base.gov), laws, articles, and the ones that are easy to find by simply querying a standard search engine for OD Portugal. From the bottom up, INE is the national institute for statistics and offers a diverse collection of data on the most varied social and demographic topics. Most datasets are available for bulk download and the source is credible. Data from this portal is essential for any project that needs data on the demographic properties of Portuguese society. INESCtec is not of interest since it is not governmental or directly related to entities or activities susceptible to corruption. The Tourism portal feeds on Dados.gov.pt, therefore, it is irrelevant for this project. Justice.gov has data on judicial decisions, court cases, and more. The fact that most data is available only in pdf is a disadvantage since it would mean taking data all the way from level 1 to 5. Transparencia.sns.gov is very organized, has a large variety of datasets about health, and makes most of them available for bulk download. Pordata.pt is a private initiative lead by the Francisco Manuel dos Santos Foundation. It has very powerful embedded visualization tools and allows the crossreference of different tables from their dataset. They also offer metainformation with every dataset and allow for bulk or a filtered download. The other two: dados.gov.pt and Base.gov, are the most mentioned Portuguese data portals, they have been mentioned in the Open Data Chapter as good examples, which does not rid them of limitations. Dados.gov.pt has better usability and an API, on the other hand, Base.gov.pt has data on public procurement and fits all requirements set at the beginning of this chapter. Base.gov.pt is chosen as one of the main databases for this project, which is now lacking a second database and the connection between them. Some municipalities’ open data portals were also found by the same keywords search, OD Portugal. They are not included in Table 11 Seven Up to Date Open Databases in Portugal because they are not centralized in the sense that each municipality publishes their data for public scrutiny separately. With the key datasets from Table 6 at hand, the next step was to search for these datasets in the order of priority to understand if they are made available in the country. The idea is to link to Base.gov.pt above a dataset as high as possible in the ODC hierarchy. The first searched term was “Interests Registers” which returned Parlamento.pt. This is a website focused on the activity of the Portuguese parliament with a page dedicated to OD with XML dumps. As for the Interest Declarations, those are only made available in HTML or PDF depending on whether they belong to members of parliament or ministers. 50 Overall, the phases of an LOD implementation are Creating, Interlinking and Publishing. To better explain what work is done on each phase this chapter makes an overview of the three main processes that are then described in further detail. The three main processes are in Figure 4: Data Sourcing, Cleaning, and Uniformization; Designing the Linked Data App Profile; Transformation, Validation and Publishing. Phase one, creating, takes in the first and second processes, the phase two, interlinking, is designed on the second process, and implemented on the third, lastly, the publishing phase occurs entirely on the last step of the third process. Figure 4 Data Journey - From Sourcing to LOD Figure 4 represents a Macro perspective meant to illustrate the data journey from the source to being published as LOD. The important idea to take away from this view is that the steps in Source, Clean, and Uniformize and the steps in Design Linked Data Profile happened simultaneously but separately. Meaning that no outputs from any of the processes were needed to continue the other one. They did not interact with each other but both processes needed to be finished to start the Transform, Validate, and Publish process. All decisions and steps are now described in the following sub-chapters: 4.1 Data Sourcing, Cleaning, and Uniformization, 4.2 Designing the Linked Data App Profile, 4.3 Transformation, Validation and Publishing. To aid in understanding the processes there are a set of Activity Diagrams below following the norms presented in chapter 15 “Activities” of the Unified Modelling Language Specifications [57]. 51 4.1 Data Sourcing, Cleaning, and Uniformization Figure 5 Source, Clean, and Uniformize Ideally, a transformation to LOD would start from Tabular files or an ER database. However, since the data selected from Parlamento.pt is in HTML, the first step is to scrape/retrieve it and turn it into a CSV. Figure 5 shows the journey from the Data Sources on top “Parlamento.pt” and “Base.Gov” to the cleaned tabular data (CSV format) ready to transform into RDF. On the left side, the swim lanes divide the Scrapper, and OpenRefine + RDF Extension. The first refers to every Python script to retrieve data from the web and the second, OpenRefine, is a recommended powerful tool for transforming data, reconciling, and exporting in RDF formats with the extension of the same name. [4]. The course of the activities flows throughout the sources and the tools to better highlight which data came from where and how it was handled. The steps here are both chronological and causal, one cannot be done before its ancestor. Below, 4.1.1 Sourcing explains the processes behind the activities in the scrapper’s horizontal swim lane, and 4.1.2 Cleaning describes the processes behind the Open Refine + RDF extension swim lane. 52 4.1.1 Sourcing The process of data sourcing describes everything in Figure 5 Source, Clean, and Uniformize that crosses the swim lane “Scrapper” with “Parlamento.pt” and “Base.gov”. All the code written for this purpose can be found in the support files. 10 It is written in python with aid of the libraries listed in Table 15. Automating the data extraction not only avoids human error but also helps others re-use the process and update the data in the future due to new elections, updated documents, changes of the statute, or switches in parliament seats. Table 15 Python Libraries Library Purpose Beautiful Soup Parses the data from XML/HTML and navigates the parse tree. Requests Sends HTTP requests. Selenium Interacts with browsers like a puppet. Needs a web driver to run and is very useful with ASP.Net pages. Pandas Aids in reading/writing tabular Data frames. Timer Creates sleeping times to avoid overloading the servers with requests. Web Scraping was born from the need to analyse data from the internet, and the incapability to do so manually due to the large volume of available information. It essentially is the practice of automating and systemizing the extraction and organization of relevant data from the web to analyse or combine said data. [58], [59] Scraping data from HTML structures is also not new for LOD implementations. In [49] they used crawlers with built-in text extraction for sources without dumps, and OpenRefine itself offers a simple built-in scraper that parses the HTML of any given URL [60]. The practice is as old as the internet itself. Currently, setting up a scrapper is straightforward and requires minimal programming effort. [59] The most commonly used languages are R and Python largely due to their multiple libraries destined for this purpose. [58] For most scrappers the macro process is defined in three phases: website analysis/navigation and access, website crawling/parsing relevant data, and organization/output [58], [59]. The first phase is analysing the website (or websites) to understand how the scrapper is going to be built, from here, a software agent mimics browsing interactions and retrieves the data needed. This is done by 10 https://doi.org/10.34622/datarepositorium/KWFSU2 53 parsing each page. The last phase is retrieving and keeping only what is necessary and organizing the data, usually giving it an output. [58], [59] Naturally, the process of automating the extraction and reusing data published by others raises discussions over the legality and ethics of web scraping. Throughout the years, practitioners have come up with a set of good practices and indicators to look out for that are always presented next to articles and tutorials on the matter. In [58] Krotov and Silva identify a preliminary set of ethical and legal questions and considerations based on both legal and information systems literature that aid in keeping the scrapper harmless to a website. Even in technical books, such as “Web Scraping with Python” there is a chapter dedicated to the good practices for ethical Scraping. All the formal and informal articles, books and tutorials mostly agree on the main aspects that keep web scraping legal and ethical. Respecting Robots.txt – These files cannot be enforced, they have no blocking power, scrapers can read and obey them or go around them. [61] That being said, the latter is likely not a good idea. This is a simple syntax text file that states what directories of the website the developers want or not to be accessed by robots. This file should always be on the first level of the website tree, this is also how it can be found, by appending “/robots.txt” to the URL. These files are very useful for scrapers since by telling you “where” not to go the developer is implying he is aware of the existence of web scraping and accepts the practice anywhere else on the website, except for the denied directories. [61] In this project - Parlamento.pt has no Robots.txt file and Base.gov.pt has one as of the launch of the new version. This file states a 10-second delay for all crawlers and a few out of access URL’S. Both conditions are respected by the code developed. Avoid Overloading with Requests – Just because code can go through all the pages of the website and get the needed data at a rate of a page every 2 seconds, it does not mean it should do it. It is always best to delay the scrapper and let it run longer, so it mimics human interaction and avoids overloading the server with requests. The consequences for not doing this could be damaging a website and that can be problematic for the author. [61] Despite the possible problems, overloading a server with requests is never morally correct regardless of the consequences. If the website is primarily accessed by people in the same time zone (case of both sources) it is also better to run the crawler late in the night when it is less likely to be used by humans. On top of these two basic practices, Krotov and Silva [58] came up with a set of questions to which negative answers mean there is a “decreased likelihood of legal problems and ethical controversies in 54 their work”. If there are negative answers it does not mean Scraping is completely out of the table, there might be an alternative solution such as contacting the owner of the website and getting permission. The questions and their answers in the scope of this project are presented in Table 15. Table 16 Questions on legal and moral pointers for data Scraping [58] Questions Base.gov Parlamento.pt Justification Legal Is Web Crawling or Web Scraping explicitly prohibited by the website’s “terms of use” policy? x x Neither have explicit Terms of Use that need to be accepted to navigate the website. Is the website’s data explicitly copyrighted? x x There is no obvious copyright information. Does the project involve illegal or fraudulent use of the data? x x Anyone can access the data without registration or payment. Can crawling and scraping potentially cause material damage to the website or the Web server hosting the website? x x Scrappers are slowed down and run at night to avoid any damage. Ethics Can the data have obtained from the website compromise individual privacy? x x Does not apply because the data used is open and the public has the right to this information. Can the data obtained from the website reveal confidential information about the operations of the organizations providing data or the company owning the website? x x It can reveal connections but not confidential information or trade secrets. Can the project requiring the Web data potentially diminish the value of the service provided by the website? x x Financial losses and concurrence do not apply since there are no ads, and profit is not the goal of making this information linked. All questions on legality and ethics in Table 16 Questions on legal and moral pointers for data Scraping [58]have a negative answer in both sources. None of the questions in legality applies to either of the sources. As for ethics, the data used is meant to be public, the project is only adding context by linking existing data making the datasets interoperable. As for the last question concerning the value of the site, the website is always the source and the project states that clearly by redirecting to the source. For this reason, the prototype is not considered to affect the value of the websites. 55 Some other important practices to apply when developing a scrapper are: • When an API exists, it should always be prioritized over a scrapper. [59] For instance, Twitter’s robots.txt prohibits access to directories containing data that can be obtained via their API. [61] • Scraping information that is of open access to any human is unlikely to be a legal problem. The same does not apply to information protected by passwords or other means of access. [61] • When it comes to the risk of copyright infringement this usually only applies to creative work, therefore leaving out company names, prices, names, and any other facts. [61] The data used in this prototype is already Open by Default, so the protection of the information does not apply. With all of this in mind, two scrappers were written for this project. One that navigates through Parlamento.pt and retrieves data from the parliamentarians’ interests registers, and another one that looks up companies on Base.gov.pt and returns the contracts they are involved with when there are any, then looks up the NIFs of all entities involved in the returned contracts and returns a link to their profile in Base.gov. The full code is available in the support files. 11 The scrapper for Parlamento.pt was built on two phases: First, based on selenium, the code simulates human navigation opening every interest register page. Selenium and a chrome web driver were essential because the Parlamento.pt website uses asp.net, meaning the content is generated at the moment and there are no permanent links to follow, so the solution is to simulate a click. When testing showed the scrapper was able to open every single interest register page and move on to the next, the second phase started. Figure 6 shows the part of the code that identifies the number of registers per page and follows the link for each one of them in a loop until it reaches the last link. Then it repeats the process on the next page until it reaches the last link on the last page. Row 68 identifies the link by the text it contains, which is always the same, and row 69 is the click simulation. 11 https://doi.org/10.34622/datarepositorium/KWFSU2 56 Figure 6 Parlamento.pt scrapper extract: Looping through Interest Registers The second part of the code was written based on three downloaded pages with the interests register so it could be tested without sending multiple requests to the website. This essentially selects the needed data on each page, checks for blanks and identifies them, and appends the data to an array. This phase also includes the writing of scraped data in a CSV file with Pandas. The fact that every page follows the same structure made this process simpler and a matter of jumping between “if-else” conditional structures. The resulting product is a CSV with one row per parliamentarian and one column per attribute. For attributes with multiple values, such as “Job Positions”, where one person can have multiple job positions, the values are all in the same cell separated by two consecutive colons “::”. Separating these values again is then handled in the data cleaning phase. The data about interest registers in parlamento.pt is displayed as a form result. There is a title for current positions, then the information about the positions, then a title about old positions, and information about those. This, at times, leads to repeated information. For instance, if a parliamentarian works in a company since 2017, instead of the position showing only once as a current position that started in 2017, sometimes it appears, once as a past position “last three years”, and once as a current one “accumulation”. A translation by the author example is displayed in Table 17 Example of the information displayed on Parlamento.pt (city name withheld). The name of the city was removed to avoid singling out individuals. 57 Table 17 Example of the information displayed on Parlamento.pt (city name withheld) III – Data About jobs/positions/activities Last Three Years Job/Position/Activity Member of Local Parliament Entity Local Parliament of City 1 Start 2017 -10--22 End Accumulation with political/public chair position Job/Position/Activity Member of Local Parliament Entity Local Parliament of City 1 Start 2017-10-22 End Although duplicate this information was kept, ensuring that important data was not accidentally excluded. When transforming the data sampled in Table 17 into a tabular format where there are four columns identifies whether: the job position, the entity, the start date, and the end date which have as values whichever value is on the website. Then, to avoid losing information, two columns were added, one to specify if the job is a social position, a working position, or a society, in Table 17 it is a working position because it is under “III – Data About jobs/positions/activities”. The second column identifies whether the position happens in the accumulation with political work or not, in Table 17, that information is on lines two and seven. Figure 7 shows how these two columns are added. Line 172 makes sure the data is only being retrieved in the section of the form with info on “social Positions”, therefore, in line 189 the property “Social Position” is added to make sure this information is not lost. Then for the accumulation, in lines 190 to 196, the scrapper checks whether the information is under the accumulation separator. If yes, then the value is True, if not then the value for isAccumulation is false. 58 Figure 7 Parlamento.pt scrapper extract: Identifying Data on Social Positions The last six lines in Figure 7 show the sum that is made to make sure the data is aligned and that every multivalued cell in one line has the same number of values. At first sight, these additions seem redundant because one should be able to tell whether it is an accumulation by the start and end date, however a quick analysis of the registers showed that these dates are not always given so extracting as much context information from the page as possible was considered the better option. Uniformize then proceeds with several data cleaning steps that result in a list of entities. This list then feeds a scrapper that looks for the entities in Base.Gov. The scraper for Base.gov.pt was simpler to build. There is an array with company names where the spaces are substituted by “+” and one at a time they are substituted in the URL. Selenium is only used here for clicking the download button and checking the sum of contracts for control purposes. For extracting the contracts from Base.gov.pt there were two options: Search the entity names as simple text, this would return contracts with the entity name in any of the contract’s fields; Search for the entity name as a contractor and then as a supplier separately, this returns only contracts where the entity appears as contractor or supplier of the contract. The first option would return results faster, but the results were worst because if the name of the entity appeared in a field of the contract that was not the Supplier or the Contractor the contract would be downloaded anyway. 59 For instance, let’s consider a company named “Creativity”. A search for the text “Creativity” returns contracts where the supplier or contractor have the name “Creativity” and a contract from a public entity that contracted a professional school to give “A workshop to increase the creativity of the team”. The last result is not of interest because the entity researched is not involved. Considering the above, on phase one all entities are run in the “contractor” field three times, then in the “contracted” field three times and the ones that are not found in either are checked one last time on the “text” field. All 725 entities were searched as suppliers and contractors three times. The time the driver waited for an answer before moving to the next result increased in each attempt. After the third try, all entities that were not found were searched only once as free text. All downloaded files were then verified manually to make sure the entities mentioned correspond to the entities on the Interests Registers. Whenever it was not possible to verify if it was the same entity due to lack of information the contracts from said entity were discarded. This process is described in Figure 8. Figure 8 Retrieve Contracts 66 In this matrix, the parliamentarian is linked to the entity through the subject membership. This linking option did was discarded for the following reasons: • Parliamentarians are not directly connected to the entity. They are only connected to the entity on a second level by the place of the membership. • Obeying the domains and ranges creates too many resources to describe and there is not enough information for this to be reasonable since it creates several subjects with less than three predicates. • Having several levels of data creates more namespaces and more complex queries. This impacts query response time in the future and generates a heavier dataset. The idea of merging all these vocabularies would have been ideal if there was control over the information and the scope of the data sources was larger. What ended up happening was that by having three classes of resources on the dataset and having only first level information about each of them, representing the data on multiple levels adds more complexity and heavier datasets but does not add more information. From this experience the process re-started with three goals: give priority to well-established property schemas, having a property with a parliamentarian as subject and an entity as an object or an entity as object and parliamentarian as subject, and using fewer vocabulary namespaces. This second attempt uses only 50 properties (6 had to be created), 9 namespaces, 5 described subjects and 2 blank nodes. No data was discarded, the reduction of properties is due to the reduction of subjects caused by the choice of less restrictive properties. Selecting the properties to re-use, defining the shapes of the data, and trying to avoid excessive blank nodes constituted one of the biggest challenges of the implementation phase. There are many ontologies in the LOD universe, but when one tries to take from different ontologies to create a multidisciplinary database, obeying the ranges and domains becomes a considerable challenge. With the Public Procurement and Public Contracts Ontology, this was the case because the properties are very restrictive. The description and the names of the properties were adequate to every situation but then the data available was not enough to justify diving into sublevels of subjects for the price, legal documents, or framework agreements. To make querying more straightforward, and to keep the number of blank nodes as low as possible in the prototype, the final model opts for having as many properties on the first level as possible. To aid in visualizing the properties mentioned, and how the subjects connect to one another, an adaptation of an Entity Relational model was elaborated and can be seen below in Figure 12 Adapted Linked Data Entity Relation Model. This Image does not follow the specifications for an ER model diagram, 67 the correspondences are that each table is a subject. The class of the subject is the title of the table, inside each table the left side has the properties and on the right side, the value types or controlled vocabulary stems. The arrows connecting the tables point to the value: A pc:Contract has as pc:supplier a schema:Organization. For the values with vocabularies, skos:Concept represents the class of the value, followed by the suffix of said vocabulary and the “~” is a placeholder meaning that any value that starts with that root applies. Figure 12 Adapted Linked Data Entity Relation Model The main vocabulary of properties used for describing what is seen in Figure 12 is Schema.org. This is a well-maintained project with community participation and an up-to-date GitHub for support and discussions. Moreover, it was created as a joint effort from Google, Microsoft, and Yahoo! as a markup language to be read by search engines that would use it to provide richer results. [4] Schema.org is used on 23 properties and replaced every instance of the organization ontology that was used on the first matrix. 68 4.2.1 Reification In Figure 12 the table further to the left has a note on top saying “Reified Statement”. This chapter explains why it is needed and how its format was selected. Besides the creation of new properties and vocabularies, one challenge was the representation of the connection between person and entity with reification. The first is discussed in the next chapter, the latter is explained below. Let us consider the following information as an example: The parliamentarian “John Doe” had a Role of consultant in “Society A” until 2019. If it were not for the year and the role consultant, the representation of this as a triple would be simple: John Doe – Subject > Member of – Predicate > Society A – Object However, both the data about the year and the position are important. This is a case of handling knowledge about knowledge which is known as metaknowledge [62] and can be represented in several ways in RDF. Metaknowledge usually refers to information about data provenance, reliability, and timestamp. However, understanding its usability is complex [62] and the same techniques used to describe it are also used for describing different data. One of the biggest and most common examples are the qualifiers in Wikidata. Wikidata has qualifiers on most resources, they have a large collection of resources, and they include politicians, even some politicians from the Portuguese parliament. The way they organized the information about parliamentarians served as a base to organize the application profile, (even though their properties were avoided for lack of human readability), especially the use of qualifiers to add context, and extra information about statements. Wikidata offers an RDF dump which shows that the syntax they use in RDF for describing their qualifier is a form of reification that does not follow the standard RDF reification. The RDF standard for reification is sometimes controversial causing the appearance of uses like this that are inspired in the norm, use the type “rdf:Statement” but don’t follow the guidelines. [63] Wikidata introduced the RDF dump in April 2015, before that, both [64] and [63] discussed the representation of Wikidata’s syntax in RDF and went through different reification options. The first compared four ways of achieving the result: standard RDF reification, n-ary relations, singleton properties and named graphs. They tested all four versions by querying them with SPARQL and uploading them to different Triplestore and concluded that there was no clear winner in terms of query performance. However, some Triplestores did struggle with processing singleton properties. [64] The second ended up narrowing it down to standard RDF reification (without obeying the formal guidelines) or named graphs as the best options for representing the qualifiers. [63] 69 Below in Table 20 is an example of what the triple in the example above looks like as a Wikidata qualifier, as a named graph, and as standard RDF reification. Table 20 Reification samples Wikipedia Named Graphs and Standard Reification Wikidata Qualifiers Named Graphs Standard RDF Reification :JohnDoe :hasRole :Consultant :JohnDoe :hasRole :S1 . :S1 :hasRole :Consultant; :S1 :ended “2019”; :S1 :entity :SocietyA . :JohnDoe :hasRole :Consultant :Graph1 . :Graph1 :entity :SocietyA; :Graph1 :ended “2019” . :JohnDoe :hasRole :Consultant . S1 rdf:type rdf:Statement S1 rdf:object :John doe S1 rdf:predicate :hasRole S1 rdf:Object :Consultant S1 :ended “2019” S1 :entity :SocietyA More recently there is also the work of [65] that mapped a relation between Wikidata and DBpedia. This included relating the properties with “owl:equivalentProperty” as well as the information. When facing the issue of reification, they make an important note of the fact that DBpedia has a more stable ontology than Wikidata that was not native RDF. This is one of the reasons why they select simple RDF reification as the go-to technique for mapping Wikidata with DBpedia. [65] Both for the sake of scientific work and the reputation of DBpedia in the world of linked data this prototype is going to use standard RDF reification. Literature shows it works, it has solid guidelines and a standard way to be described 18 , and the most widely discussed downside, which is the number of triples it generates does not have as big an impact as one would expect when it comes to query time response. [64] Named graphs also seemed like an adequate solution and a largely supported one, however, OpenRefine is selected as the tool to transform the data into RDF and that also presents a constraint since a way to create named graphs in OpenRefine was not found. Figure 13 shows what a standard reification mapping looks like in OpenRefine and what the outputs looks like in Turtle. Querying reified data will be discussed further down in Publication. 18 https://www.w3.org/TR/rdf-primer/#reification 70 Figure 13 OpenRefine: Reification Map and Turtle Sample (identifiers blurred) 4.2.2 Dublin Core Application Profile (DCTAP ) The DCTAP , presented above in 2.3.2 Linked Data, being an ongoing project naturally means it is not finished, however, their tabular approach is already very comprehensive and was, therefore, used to describe the application profile for this project as can be seen in Table 21. The template for the table as well as the guidelines followed are all available on their GitHub. 19 The representation in Table 21 is only missing the label for the shapes since they are self-explanatory and the label for the properties that can be found in Table 14. This format is preferred to the Matrix and the ER adaptation mentioned above because it is a format with guidelines for implementation available to the public, meaning that it can be interpreted by anyone. The other two are working documents that aid in planning and visualizing but are not supported by research as rich as this version. 19 https://github.com/dcmi/dctap/blob/main/TAPprimer.md 71 Table 21 Application Profile - DCTAP Template Prefix Namespaces dbo: http://dbpedia.org/ontology/ pc: http://purl.org/procurement/public-contracts# pproc: http://contsem.unizar.es/def/sector-publico/pproc# rdf: http://www.w3.org/1999/02/22-rdf-syntax-ns# schema: http://schema.org/ m4s: <www.example.org/map4scrutiny/core#> m4sv: <www.example.org/map4scrutiny/value#> shapeID propertyID mandatory repeatable Value Node Type Value Data Type valueConstraint valueShape note Parliamentarian rdf:type TRUE TRUE IRI m4s:Parliamentarian schema:url TRUE FALSE IRI IRIStem www.Parlamento.pt/De putadoGP/Paginas/Bio grafia.aspx?BID=~ schema:sameAs FALSE TRUE IRI schema:description TRUE FALSE LITERAL schema:alternate Name TRUE FALSE LITERAL schema:name TRUE FALSE LITERAL schema:gender TRUE FALSE IRI IRIStem schema:GenderType schema:birthDate TRUE FALSE LITERAL schema: Date ISO 8601 date format. schema:spouse FALSE FALSE LITERAL m4s:property Regimen FALSE FALSE IRI IRIStem m4sv:PropertyRegimen schema:has Credential FALSE TRUE IRI IRIStem http://data.europa.eu/ snb/isced-f/ schema:has Occupation TRUE FALSE IRI IRIStem http://data.europa.eu/ esco/isco/ m4s:parliamentary Comissions FALSE TRUE IRI IRIStem m4sv:ComissoesParlam entares dbo:parliamentary group FALSE TRUE LITERAL schema:memberOf TRUE TRUE IRI Organization m4s:Parliamentary Term TRUE TRUE LITERAL roman numbers m4s:constituency TRUE TRUE IRI IRIStem http://vocab.getty.edu/ tgn/ schema:comment FALSE FALSE LITERAL 72 shapeID propertyID mandatory repeatable Value Node Type Value Data Type valueConstraint valueShape note _:Roles rdf:type TRUE TRUE IRI rdf:Statement rdf:type TRUE TRUE IRI schema:Role rdf:subject TRUE FALSE IRI Parliamentarian rdf:predicate TRUE FALSE IRI schema:memberOf rdf:object TRUE FALSE IRI Organization schema:roleName TRUE FALSE LITERAL IRIStem http://data.europa.eu/ esco/isco/ schema:endDate FALSE FALSE LITERAL Schema: Date schema:startDate FALSE FALSE LITERAL Schema: Date schema:description FALSE FALSE LITERAL m4s:company Shares FALSE FALSE LITERAL m4s:is Accumulation FALSE FALSE LITERAL Schema: Boolean shapeID propertyID mandatory repeatable Value Node Type Value Data Type valueConstraint valueShape note Organization rdf:type TRUE TRUE IRI IRIStem schema:Organization schema:legalName TRUE FALSE LITERAL schema:alternate Name FALSE TRUE schema:hasOffer Catalog FALSE FALSE LITERAL schema:location FALSE TRUE IRI IRIStem http://vocab.getty.edu/ tgn/ schema:taxID FALSE FALSE LITERAL schema:url FALSE FALSE IRI 73 shapeID propertyID mandatory repeatable Value Node Type Value Data Type valueConstraint valueShape note Contract rdf:type TRUE TRUE IRI pproc:Contract schema:description TRUE FALSE LITERAL pc:procedureType FALSE TRUE IRI IRIStem m4sv:TipoProcedimento pc:kind FALSE TRUE IRI IRIStem m4sv:TipoContrato pc:mainObject TRUE FALSE IRI IRIStem http://purl.org/cpv/20 08/ pc:contracting Authority TRUE FALSE IRI Organization pc:supplier TRUE FALSE IRI Organization pc:awardDate FALSE FALSE LITERAL Schema: Date pc:actualEndDate FALSE FALSE LITERAL Schema: Date pproc:duration FALSE FALSE LITERAL schema:datePublis hed FALSE FALSE LITERAL Schema: Date pc:location FALSE TRUE IRI IRIStem http://vocab.getty.edu/ tgn/ pproc:legal Document Reference FALSE FALSE LITERAL pc:framework Agreement FALSE FALSE LITERAL number and description pc:agreedPrice FALSE FALSE LITERAL m4s:Euros in euros pc:actualPrice FALSE FALSE LITERAL m4s:Euros in euros schema:url FALSE FALSE IRI pproc:contract Modification FALSE FALSE bNode _:Cmodification pproc:contractOr Procedure Extinction FALSE FALSE bNode _:Cmodification shapeID propertyID mandatory repeatable Value Node Type Value Data Type valueConstraint valueShape note _:Cmodification rdf:type FALSE FALSE IRIStem pproc:ContractModificat ion rdf:type FALSE FALSE IRIStem pproc:ContractOr ProcedureExtinction pproc:modification Reason FALSE FALSE LITERAL pproc:duration Change FALSE FALSE LITERAL pproc:extinctionCau se FALSE FALSE LITERAL 74 4.2.3 Vocabulary Creation Everything in Table 21 with the prefix (m4s:) represents a vocabulary that is going to be created. This chapter explains what led to this decision and, when appliable, to what other properties the new ones are being linked to, and why these properties were not used in the first place. Table 22 Created Properties and Classes Object Properties Linked to Motive propertyRegimen schema:spouse The property regimen is important for the data and connected to the marriage, but no property was found to represent this. For this reason, a new one is being created as a sub-property of “schema:spouse” since it is a further specification of the marriage. constituency wiki:P768(electoral district)a admingeo:constituency The only options found for this property are not ideal and for this reason, the decision was to create a new one that has a “same as” connection to the ones mentioned. parliamentaryComissions ocd:rif_ufficioParlamentare b The only property found has incompatible range and domain being used as a same as instead of simply reused. parliamentaryTerm ocd:rif_leg b wiki:P2937 (parliamentary term) a The only options found for this property are not ideal and for this reason, the decision was to create a new one that has a “same as” connection to the ones mentioned. Datatype Properties Linked to Motive companyShares schema:owns No property was found to represent this so schema:owns was selected as a super property since having shares is a type of ownership. isAccumulation - This property is a Boolean describing whether the role it is used on happened in accumulation with public office. It is needed because start and end dates are not always specified. Classes Linked to Motive Parliamentarian ocd:deputato b schema:Person It is a subclass of schema:Person and the same as the OCD class. The properties in the columns “linked to” were excluded for one of the following reasons: a – Using codes for coding properties jeopardizes human readability. b – These resources are only described in Italian and do not follow the norm of capitalizing classes and using lowercase on properties. Admingeo:constituency was discarded because the description specifies a country. Every re-used property is annotated in the ontology with a translation to Portuguese since that is the language of the original data, a schema:definedBy property with the link to the original description as a 75 value, and, when necessary, an addition to the domain and range. Extensions of the domain and range were only applied when the property made perfect sense, but the original range was for instance an IRI and most objects in the dataset are IRI’s, but some are strings, or when a datatype was more suitable. The properties used for this purpose were schema:rangeIncludes and schema:domainIncludes. These properties were also used to define the range and domain of the new properties and classes. This way, the intended use is clear, and third parties have the basis to interpret whether the assets are appropriate. This solution is preferred because the description of an asset can suit a different, but very similar, range or domain than the one mentioned that makes more sense in a different implementation. For the same reason, despite being identified as “Object” and “Datatype” properties in Table 22 in the ontology, the new properties have as rdf:type only rdf:Property to avoid excessive restraints. The ontology also includes a new datatype named “Euros” to be used in the price spaces to identify the currency. Figure 14 Extract from the Map4Scrutiny Ontology with three new properties Figure 14 shows the implementation of Table 22 with triples for three of the new properties all linked to external vocabularies in different ways. ‘’Parliamentarian commissions” has an equivalent property in an external vocabulary, the “Property Regimen” does not have a super property nor an equivalent one, however, the concept is semantically related to that of a spouse in the sense that an application profile needing this property, probably also needs schema:spouse, thus the rdfs:seeAlso. The case for “Company shares” is different, holding actions from a society is a type of ownership, and for that reason schema:owns is an appropriate super-property. The full ontology with new and re-used assets is available in the support files 20 . 20 https://doi.org/10.34622/datarepositorium/KWFSU2 82 Figure 19 Map4Scrutiny Data Print From Turtle File - One Parliamentarian 4.3.2 Validation Validation is done essentially by parsing the data generated in OpenRefine and comparing it to the previously defined shapes in ShExC. For this process three files are needed, a Turtle file with the description of all data, a ShEx file with the schema of the data, and a Shape Map. A Shape Map is a simple text file matching each node to its shape, thus informing the validator that Subject A should pass as a parliamentarian while Subject B should pass as an entity. [47] The syntax is very simple, and an example is displayed in Figure 20 Sample of Shape Map for ShEx validation. Figure 20 Sample of Shape Map for ShEx validation The actual validation was run on local implementation of ShEx-java 22 made available on GitHub and it generated an output text file indicating whether the shapes are conformant or not. This file is available for consultation in the support files 23 . The Reified Roles presented a challenge for validation. When parsing the Turtle file with the data, the implementation of ShEx-Java mentioned above gives a new ID to the blank nodes. Let’s consider the sample data in Figure 21. The validator starts by parsing the Turtle and ShEx files. When it finds a blank node in the TTL file it gives it a random ID, so <_:node123> could after parsing be <_:nodeA>. The 22 https://github.com/iovka/shex-java 23 https://doi.org/10.34622/datarepositorium/KWFSU2 <Subject_A_URI>@<Shape_Parliamentarian_URI>, <Subject_B_URI>@<Shape_Organization_URI>, <Subject_C_URI>@<Shape_Contract_URI>, … 83 challenge of this characteristic is that when the validator reads the shape map below, _:node123 does not exist anymore and it is not validated. Figure 21 Sample Validation of Blank Nodes Since the reified statements are blank nodes that do not exist as a value for any property, the solution to validate them against the defined shape was using an online validator 24 that does not give blank nodes new IDs when parsing them and accepts blank node IDs in its shape map. The online validator accepts a shape map such as the one in Figure 21 and outputs a list of conformant and non-conformant shapes. Above, Figure 16 shows the last version of the shape for the reified roles. The reason why the shape identifier for the Parliamentarians and Organizations had to be removed, was because this data was not present in the online validator. The dataset was too large to be validated entirely there. Nevertheless, all organizations and politicians were validated in the local implementation of ShEx-Java. The resulting file from the role’s validation returned 10 non-conformant roles that were not uploaded because the original data has neither an entity where the role is played nor a role name. Not having an entity means there is no Object which is a part of the reification vocabulary and therefore is mandatory. Not having a Role Name renders the reified statement useless since its purpose is to describe the role. As may be seen in Figure 17 after the validation two different things can happen to non-conformant data. Mapping and transformation errors refer to mistakes in the previous processes that caused a problem. Two specific samples are the following: A typo when uploading the schema ontology. In the data, it was “HTTP” and in the shapes “HTTPS” this caused several nonconformant nodes that were easily fixed; Some whitespaces were left in the Contracts URI’s, this was another error that was easily fixed by removing the spaces and exporting the data again. After the mapping errors were corrected the data regarding the politicians and the entities all passed the validation process. The data regarding the contracts has 135 nonconformant values. 11 of these are 24 https://shex.io/webapps/shex.js/doc/shex-simple.html TTL _:node123 rdf:type rdf:Statement. ShEx :roles BNODE { rdf:type rdf:Statement; } Shape Map _:node123@:roles 84 nonconformant because there is no link for the supplier. The supplier is simply a string and has no unique fiscal identifier (NIF), for this reason, it was not possible to include it in the automatic search for entity’s links and it does not have one. The remainder 124 are contracts that do not have a common procurement vocabulary (CPV) number. Both situations could have been made valid by making the original shapes even more open, however, having a link to the contractor and supplier is the minimum one could ask for when describing a contract as LOD. Since these are public, having a CPV should also always be required. So, these nodes are not valid against the predicted shape of the data and were not uploaded into the Triplestore. 4.3.3 Publication The second to last step mentioned above is uploading the data to a Triple Store. Following the recommendations in the literature, this project was uploaded to a local version of OpenLink Virtuoso [50], [4], [37] to test SPARQL queries and verify if the subjects were properly linked to each other. Table 25 The Final Triples Database in Numbers Types Subjects/Entities Triples Parliamentarians 224 5 850 Organizations 25 519 101 866 Contracts 131 531 1 898 482 Roles 1650 14 942 Contract Modification and Extinction 52285 143 958 Total 211 209 2 165 098 After discarding the non-conformant data the uploaded triples describe the entities, parliamentarians, and contracts with a total of 2 165 106 triples. Table 25 considers only the data about these but that is not all that was uploaded to the database. To make querying the data in a local implementation a friendlier experience all the controlled vocabularies for values and vocabularies for properties used to describe the data are also uploaded. The summary of their dimension is represented in Table 26. The TGN Location Taxonomy is not fully available. Due to the large volume of the Taxonomy, only the used locations are uploaded in the local implementation. A folder with all the data needed to upload the same data in any local machine is available in the support files. 25 25 https://doi.org/10.34622/datarepositorium/K1DQIT 85 Table 26 Auxiliary Vocabularies Vocabulary Name Subjects Triples Map4Scrutiny New Properties 58 255 Map4Scrutiny New Values 84 466 Getty Taxonomy Geographical Locations 30341 357697 ESCO – European Occupations Taxonomy 2950 14076 Common Procurement Vocabulary 10420 298583 The SPARQL endpoint offered by OpenLink Virtuoso allowed running a set of queries both to verify if the data is linked properly and explore the potential of the added vocabularies. With all the vocabularies linked it is possible to query for information that was not available before. For instance, it is possible to have a query retrieving the location of an Organization, the ISO code of that location, what broader location it belongs to and any other property that is described by the Geographical Places Taxonomy. Another example is the common Procurement Vocabulary where it is possible to navigate to narrower or broader connections of a given code. Figure 22 shows a long query and its return with only six selected values. This query was formulated to showcase several possibilities of the data, not to answer a specific question. The query can de be deconstructed as follows: Any parliamentarian that fits the criteria is represented by the wildcard ?s. The parliamentarian should have a name and a memberOf property. The Organization he is a member of, represented by ?o needs to have a legal name, a NIF and a location. The location, ?heads, cannot be a string because it needs to have a label. Then to get further information on the member of relation, the ?b represents de reified statement that has the same ?s as a subject, the ?o as an object, has also a role name, and the description is “Acionista”. So far, the query is pulling every parliamentarian, that is a shareholder in an organization and that has all the properties listed above. From the reified statement the query also pulls the value for the held shares. Then, the subject changes to ?c, which stands for contracts. The query is now pulling subjects that have ?o as a supplier. These also need to have an actual price, a CPV, a description, and an IRI location that has a label. Both the location and the CPV vocabularies are external and available in several languages. CPV was left as an IRI in these samples, but labels for the location of both the organization’s headquarters and the contract execution were pulled and are in the “SPARQL | HTML 5” table. The labels for the locations have a language filter because otherwise random languages would be presented. 86 For this sample, only six values are selected to be shown, but any instance of an “?” could be selected and printed. The names, NIFs, and roles were left out of the selection on purpose to avoid singling out personal data. Figure 22 Map4Scrutiny Print OpenLink Virtuoso - Sample SPARQL Query Other samples of queries are available in Appendix D Sample SPARQL Queries. These samples are meant to showcase the added exploration possibilities that the transformation brought to the original data, they are “link” oriented instead of results-oriented. They can be run by anyone that uploads the bulk data in a Triplestore. Being able to query the data across different vocabularies is the final step of the implementation process. 87 5 CONCLUSIONS This chapter is an overview and direct comparison between what was initially planned and what was achieved. There is also a focus on the development experience, the challenges encountered, and the value it added to the design science research’s knowledge base and the application domain. 5.1 Context In the Introduction one main objective and two sub-sequent goals were set: transform open data on Portuguese parliamentarians interest registers and public procurement from their current level (one or two) to level five [2], or LOD; Publishing the resulting data online via a SPARQL endpoint; Make the process transparent so that any third party could replicate this solution. The first objective is technical. The datasets used are part of the Table 6 Priority datasets [34] and key datasets [18] defined by both the Open Data Charter and the G20. The interest registers were available in HTML and the contracts in CSV, so they were considered one star, and three stars, respectively [2]. In the final dataset, the data from both sources is connected to each other and to external data that it had no contact with before, such as the CPV, locations, and occupations vocabularies. Moreover, it provides further context on the original with new vocabularies created within the scope of the solution. It is, therefore, by definition, five stars OD [2]. The final solution is currently missing one characteristic of LOD which is that the dataset is not yet published on the web. This step was not implemented because the dataset, ontology and vocabularies are not published online RDF files. They are only uploaded in a local instance of the Virtuoso Triplestore and available in bulk for anyone else to upload them to their local Triplestore or to publish them online. 26 The data was meant to be published in a university’s server with a working SPARQL endpoint. This publication was not possible due to technical constraints not related to the work developed. Nevertheless, the data is available in bulk for download and ready to be uploaded to a local or online instance of a Triplestore by any party. For the same reason, the implementation process employed all the features in Table 9 Linked Data Checklist from "How to use Linked Data"., except for not yet having the data registered in Linked Data Catalogues. 26 https://doi.org/10.34622/datarepositorium/K1DQIT 88 As for the third objective, the whole process has been described and all the data and code are available for both consultation and re-use. By fulfilling the objectives, the proposed solution is also a suitable solution for the stated research problem: The data provided by the government on Parliamentarians Interest Registers and Public Procurement is neither machine-readable nor interoperable by semantic web standards. The transformed dataset is both interoperable and connected to external data, and machine-readable. In summary, the first and third objectives were fully accomplished, the second partially accomplished in the sense that the steps that could be taken within the scope of this work to fulfil that objective are completed. The remainder of chapter five explores how the outputs that contributed to designing and building the proposed solution contribute back to the domain application and the knowledge base, followed by a set of recommendations for the data sources. At last, based on the questions raised by the research and implementation stages, future work is proposed to further expand bot the knowledge base and application domain. 5.2 Contributes The research and outputs that fulfilled the objectives above also constitute contributions both to the knowledge base and the application domain. 5.2.1 Knowledge Base This chapter makes an overview of the contributions the work developed has for the knowledge base identified in Figure 3: Drechsler, Hevner (Ed.) 2016 - A three-cycle view of design science research [6]. The relevance of this work affects mainly the LOD and Semantic web research fields by providing a detailed description of an implementation that enables replication independent of the data sources and themes. Every step of the implementation is described, the transformed dataset is open to the public and so are the intermediate code and files needed to achieve the final dataset. For the grounding of the work, every software, step, and decision was based on literature and prior art. The goal was to create a sustained solution that follows LOD and Semantic Web guidelines and to document a full LOD transformation process. The work developed contributions back to the foundations on semantic web with new Linked Data vocabulary of properties, classes and a datatype properly connected to existing vocabularies, controlled 89 vocabularies of values for the suitable topics, annotations with translation in re-used properties and classes. Still in the semantic web field fits the final Dataset in Turtle, the prior-art research, the overview on reification methods, an LD rendition of the European vocabulary for occupations in Portuguese, the tabular application profile for the data, the shapes of the data expressed in ShExC, a description of the validation process together with the files used and needed to replicate the said process. In the field of transparency and OD fit the python scrappers used to retrieve data both from Parlamento.pt and Base.Gov, the added value to data when the quality is increased, the application profile, and the research on prior-art. 5.2.2 Application domain The interest of the public in the originally selected datasets was reinforced by the results of the Global Corruption Barometer 2021 and the reactions to a recent blackout of the Base.gov.pt portal. The barometer shows that 90% of the Portuguese population believes there is corruption in the government and 27% stated that most parliamentarians are corrupt. This places parliamentarians as the perceptibly more corrupt government seats. The drop is significant when it comes to members of government, these are only believed to be corrupt by 16% of the population. [66] As for the Base.gov.pt portal blackout, between the first and the fourth of October 2021, the portal Base.gov.pt was unavailable due to GDPR infringements related to the disclosure of personal information in the procedure pieces. On the fourth, the portal was put back online without access to any procedure features. [67] This episode reminds the public of the frailty of the access to OD that is currently available. The data in the contracts is often incomplete, and without the procedure pieces, that are still not available, access to public procurement is limited. [67] As a result of this episode the property “schema:URL” of the contract's description that linked to the procedure pieces was also removed from the bulk data made available as an output of this project. The research conducted in this dissertation and the proposed solution both show that the current data made available on parliamentarians could be much richer and valuable without having to disclose more information, only by disclosing it with context, quality, and in a machine-readable format. The same could be said for the contracts, these could take advantage of ranking as five stars, having better quality control and not remaining so dependent on procedure pieces and other unstructured data documents. In the areas of OD and Transparency, the solution shows how information that is already available to the public could have added value if on top of availability there is also work on quality and linking. Choosing data that, according to surveys, is likely to be of interest to the public, despite the low quality it is made 90 available in, presents challenges and ads work but also creates a higher step between what was scraped or retrieved and what is uploaded to the Triplestore in the end. 5.2.2.5 Recommendations One of the largest challenges when developing the proposed solution was the quality of the data, but despite the difficulties faced, the cleaning process generated a set of observations on the data. These were then compiled with good practices on OD and resulted in a set of recommendations for both Parlamento.pt and Base.gov.pt presented below, preceded by a review on the data quality issues that propelled the recommendations. In Parlamento.pt that data appears to not be reviewed at all and there is no coherence between the registers. The values for certain properties were not appropriate and many values resulted in entities or positions that could not be disambiguated. An example of information that may be misleading is: There are values with the comment “no salary” under “roles” in social and other working positions. This comment was added in front of the role name. Not receiving a salary is not part of a job name, and since this information is not asked for by the form, it is wrong to assume that every other role that does not specify being salaried is therefore salaried. As for the contracts, the most noteworthy gaps in information related to the existence of contracts without a CPV or without a supplier being identified with at least a NIF. The database for Base.gov.pt is a lot larger than the interest registers from Parlamento.pt and, for this reason, the challenges in maintaining data quality are in no way comparable. Nevertheless, some basic requirements should be met. Some examples found here that caused ambiguity problems are: Several profiles for the same entity, the same entity existing with different NIFS, and several variations of “Nothing changed” strings written as values for price and procedure changes. Combining the experience in cleaning the data with the good practices retrieved from the literature resulted in a further contribution to the knowledge base in the form of a set of recommendations both Base.gov.pt and Parlamento.pt could implement that would be an incentive to data re-use and analysis by third parties. In Parlamento.pt both the registers from the parliamentarians and the members of the government should be made available in a bulk format as they do with biographical information. As for the quality of the data, there should be a set of rules, explanations, and standards to avoid misplaced data, the registers should all be checked for quality before being published and some controlled vocabularies could be added such 91 as the EU taxonomies for qualifications and occupations. It would also be ideal that all mentioned organizations are identified uniquely, this could be done with a NIF. In Base.gov.pt the most important recommendation would be that all contracts should have at least the name and NIF of the involved parties and a CPV to be published. For the latter, if there are any specific cases where a CPV is not appliable, this should have a justification comment. As for the remainder of the inconsistencies found, they relate closer to overall data quality. Nevertheless, they affect data analysis, for instance, the price and date modification feature with variations of the string “nothing changed” create noise and create an obstacle to simple statistics such as querying “how many contracts had a modification” by counting not empty strings. Experience suggests that to have good quality LOD on public procurement and interest registers, the most crucial factor, is the quality of the data available. It is necessary for OD portals to improve their standards for quality and to display data with context that may be analysed by any individual without the risk of making assumptions. 5.3 Future Work Future work is also divided into two sections. Refinements that were not implemented because the time it would take was found to outweigh the advantages, or due to time constraints. Despite this, the actions needed to implement these additions are clear and described below. The second section relates to more challenging work and dissemination. This work is now in a laboratory stage, and a further process would also be to test its impact with the identified stakeholders and evaluate the relevance of future improvements. Following the review of the objectives, the first task on the future work would be to upload the data to a server that maintains an online SPARQL endpoint for the public to explore. Following that, and being this a proposed solution, not all challenges were solved in a manner that can be simply mirrored into a broader implementation. Below are the features identified during the research and implementation processes that are suitable to constitute future work. The first suggestions are related to maintaining the dataset. A lot of manual work went into creating the proposed solution and it is not viable to keep the solution up to date with the same process. Automating the data sourcing process - The Scraping process could be easily improved and automated into a scrapper that runs regularly and retrieves new data when there is any. 98 Descrição Prestação de serviço para a criação de campanha de Marketing Digital não incluída na exportação CSV e Redundante com A20 Fundamentação Artigo 20.º, n.º 1, alínea d) do Código dos Contratos Públicos Fundamentação para recurso ao ajuste direto (se aplicável) ausência de recursos próprios não incluída na exportação CSV Entidades adjudicantes Universidade Nova de Lisboa (501559094) Entidades adjudicatárias Mosca Azul Publicidade, Lda. (510855679) Objeto do contrato Prestação de serviço para a criação de campanha de Marketing Digital para promoção e divulgação do novo programa de bolsas dos SASNOVA. Procedimento centralizado - não incluída na exportação CSV CPVs 79342000-3, Serviços de marketing Data do contrato 10/02/2021 Preço contratual 19.860,00 € Prazo de execução 182 dias Local de execução Portugal, Lisboa, Lisboa Entidades concorrentes - não incluída na exportação CSV Anúncio - não incluída na exportação CSV Peças do procedimento - não incluída na exportação CSV Modificações contratuais - não incluída na exportação CSV Documentos Contrato_Mosca_rasurado.pdf é uma cópia do contrato não normalizada Observações - não incluída na exportação CSV Critérios ambientais - não incluída na exportação CSV Justificação para não redução a escrito do contrato - não incluída na exportação CSV Causa da extinção do contrato - Data do fecho do contrato - Preço total efetivo - Causas das alterações ao prazo - Causas das alterações ao preço Contratos CSV Extra Estado N.º registo do Acordo Quadro Não aplicável. Descrição do Acordo Quadro Não aplicável. 99 Ligação para Peças do Procedimento Lista de Fornecedores (cocontratantes) FALSE Entidade NIF 510855679 redundante A18 e A19 Descrição Mosca Azul Publicidade, Lda. redundante A18 e A20 Localização da sede Portugal Redundante info parlamento Nº de contratos como adjudicante 0 count function Total gasto 0,00 € sum function Contratos como adjudicante - count function Nº de contratos como adjudicatária 5 count function Total ganho 194.185,00 € sum function Contratos como adjudicatária Lista dos contratos link para toda a info A9 - A39 100 Appendix B CSV headers with selected data for the Application Grupo de dados Propriedades Tipo de Dados Exemplo preenchimento Perfil Nome Completo string (1) Adão José Fonseca Silva Sobre o Deputado como descritivo Data de nascimento data (1) 01/10/1957 Habilitações literárias string (1,n) Licenciatura em Línguas e Literaturas Modernas - Estudos Portugueses e Franceses Profissão string (1,n) Professor Comissões Parlamentares a que pertence String (0,1,n) Comissão de Defesa Nacional Grupo Parlamentar / Partido String [0:1] PSD Círculo eleitoral String [1:n) Bragança Legislatura string [1,n] V::VI::VIII:IX::X::XI::XII::XIII::XIV Registo de interesses Estado civil String (1) Casado(a) Nome completo do cônjuge String (0,1) Ana Maria Afonso Silva Regime de bens String (0,1) Comunhão geral Perfil Cargos/funções/atividades String (1,n) Deputado à Assembleia da República Conexão entre Deputado e Entidade Acumulação Boolean(True/False) FALSE Início data (1) 17/09/2020 Fim data (0,1) 24/10/2019 Tipo Cargo Cargo Social, Cargo, Função ou Atividade, Sociedade Cargo Social Participação social Percentagem (n) (só aplicável quando tipo = Sociedade) 74% Entidade String (0,1,n) Assembleia da República Sobre entidade Natureza e área de serviço String (0,1,n) Medicina Dentária e Estomatologia; Saúde Humana Local da sede String (0,1,n) Rua D. Afonso Henriques, 125 - 1º andar - sala N - 4435-005 Rio Tinto Apoios ou benefícios será sempre uma string porque é preenchido em forma de comentário Senhas de Presença Comissão de Ética para a Investigação Clínica: 160€ Senhas de Presença Conselho Nacional para a Procriação Medicamente Assistida: 70€ Sobre deputado Serviços prestados será sempre uma string porque é preenchido em forma de comentário Comentadora política TVI Fevereiro 2018 a julho 2018 Comentadora política SIC Novembro 2018 a setembro 2019 Outras situações será sempre uma string porque é preenchido em forma de comentário As atividades acima referidas em ONG, como investigadora e como Presidente das Mulheres Socialistas não são remuneradas. Declaração sobre exclusividade Boolean: Exclusividade/ Não exclusividade Exclusividade 101 Grupo de dados Propriedades Tipo de Dados Exemplo preenchimento (CSV gerado CSV Contratos por entidade Objeto do Contrato string comentário Prestação de serviço para a criação de campanha de Marketing Digital para promoção e divulgação do novo programa de bolsas dos SASNOVA. Tipo de Procedimento string (vocabulário na forma de lista d opções Consulta Prévia Tipo(s de Contrato string (vocabulário na forma de lista d opções Aquisição de serviços CPVs string (vocabulário na forma de lista d opções 79342000-3, Serviços de marketing Entidade(s Adjudicante(s String com NIF ( Universidade Nova de Lisboa (501559094 Entidade(s Adjudicatária(s String com NIF ( Mosca Azul Publicidade, Lda. (510855679 Entidade(s Adjudicante(s string (Transformed from F) Universidade Nova de Lisboa Entidade(s Adjudicatária(s int (Transformed From F) 501559094 NIF(s Adjudicante(s string (Transformed from G) Mosca Azul Publicidade, Lda. NIF(s Adjudicatária(s int(Transformed from G) 510855679 Preço Contratual int (normalized) 19.860,00 € Data de Publicação date 24/02/2021 Data de Celebração do Contrato date 10/02/2021 Prazo de Execução (dias) string 182 dias Local de Execução string Geo (pode ser mais que 1 separado por </br>) Portugal Fundamentação string (vocabulário baseado na lista) Artigo 20.º, n.º 1, alínea d do Código dos Contratos Públicos Causa de Extinção do Contrato string (campo normalizado - lista valores Data de Fecho do Contrato date Preço Total Efetivo int (normalized) Causas das Alterações ao Prazo (string Comentário) Causas das Alterações ao Preço (string Comentário) Estado N.º registo do Acordo Quadro (é um vocabulário) Não aplicável. Descrição do Acordo Quadro (é um vocabulário) Não aplicável. Procedimento Centralizado Boolean (true or false) FALSE Ligação para Peças do Procedimento URL Lista de Fornecedores (cocontratantes Bollean (true or false) FALSE 102 Appendix C Constraint Matrix rdfs:domain m4s:Parliamentarian subclassOf: foaf:Person same as: ocd:Parliamentarian Attribute ER Property Description Original Domain Original Range Allowed values/Datatypes Cardi nality Description schema:description known for...equal for all of them schema:Thing schema:Text "Politico Português" 1 Parliamentaria n Name schema:alternateName Selected two or three names schema:Thing schema:Text LITERAL 1 Name schema:name Full name schema:Thing schema:Text LITERAL 1 Gender schema:gender Bio Gender schema:Person schema:GenderType or Text Schema:Gender 1 Birthdate schema:birthDate Date of birth dd-mm-yyyy schema:Person schema:Date schema:Date 1 Spouse's Name schema:spouse Full name of the partner schema:Person schema:Person LITERAL 0-1 Matrimonial property regimen m4s:propertyRegimen Distribution of assets in the relationship schema:Person x skos:Concept 0-1 Education Level schema:hasCredential Bachelor, Post doc.. schema:Person schema:EducationalO ccupationalCredential skos:Concept 0-n Profissão schema:hasOccupation Main job schema:Person schema:Occupation skos:Concept 1 Parliamentary commissions m4s:parliamentary Comissions Commissions, workgroups and subcommissions from the Portuguese parliament schema:Person x skos:Concept 0-n Parliamentary group dbo:parliamentaryGroup Mainly same as party (except:Coligações) dbo:Person x schema:Text 0-1 Political Party schema:memberOf militant this party schema:Person schema:Organization schema:Organiza tion 1 Parliamentary Term m4s:ParliamentaryTerm governance term x LITERAL 1-n Electoral cycle m4s:constituency electoral district x Skos:Concept 1-n? Entity schema:memberOf entity where a position existed/exists schema:Person schema:Organization schema:Organiza tion 0-n Support and benefits / Services / Commentary schema:comment hand writen note on other types of support or perks, on services and commentaries. schema:CreativeWork, schema:RsvpAction schema:Comment LITERAL 1 metainformation for Parliamentarian> memberOf > Organization type:Role Jobs/Activities schema:roleName the position held in a given entity Role / Event Text or URL LITERAL 0-1 End Date schema:endDate Role Date Schema:Date 0-0 Start Date schema:startDate Role / Event Date Schema:Date 0-1 Shares m4s:companyShares owned shares in % or € schema:Person, schema:Organization LITERAL 1 Accumulation m4s:isAccumulation happened at the same time as the current mandate schema:Boolean schema:Boolean 1 Work Relation schema:description If it is a social position, a job or a society schema:Thing schema:Text LITERAL schema:Organization 103 company name schema:legalName name of company schema:Organization schema:Text LITERAL 1 area of activity schema:hasOfferCatalog typed area of activity for societies only schema:Organization schema:OfferCatalog LITERAL 0-1 headquarters schema:location where it is located schema:Organization Place,Text and other skos:Concept 0-1 NIF schema:taxID 9 digit unique identifier number schema:Organizatio, schema:Person schemaText LITERAL 0-1 Website schema:url Base.gov.pt page schema:Thing schema:Url IRI 0-1 Domain pproc:Contract subclassOf: pc:Contract Contract object name schema:description title of contract schema:Thing schema:Text LITERAL 1 Procedure Type pc:contractProcedure how it was dealt pproc:Contract pc:procedureTypes skos:Concept 0-n type of contract pc:kind what was acquired pproc:Contract pproc:public-contractskinds skos:Concept 0-n Contract object pc:mainObject connection to further info pproc:Contract skos:Concept skos:Concept 1 Buyer/Contract or pc:contractingAuthority buying company pproc:Contract gr:BusinessEntity schema:Organiza tion 1 Seller/Service Provider pc:supplier supplying entity (not found) (not found) schema:Organiza tion 1 Celebration Date pc:awardDate date when contract was closed pproc:Contract xsd:date Schema:Date 0-1 End date pc:actualEndDate when contract was completed or extinct pproc:Contract xsd:date Schema:Date 0-1 Working timeframe (days) pproc:duration days to fulfil contract x xsd:duration LITERAL 0-1 Publication date schema:datePublished publication of contract notice (not contest) schema:CreativeWork schema:Date, schema:DateTime Schema:Date 0-1 Place pc:location where was fulfilled pproc:Contract schema:Place skos:Concept 0-n Grounding pproc:legalDocument Reference Articles sustaining contract pproc:Contract x LITERAL 0-1 Framework Agreement pc:frameworkAgreement is a framework agreement pproc:Contract pc:Framework Agreement LITERAL 0-1 Contract Price pc:agreedPrice contracted price gr:PriceSpecification m4s:Euros 0-1 Actual price pc:actualPrice paid price x gr:PriceSpecification m4s:Euros 0-1 Modifications to contract pproc:contract Modification all modification reasons pproc:Contract pproc:Contract Modification bNode ntractModification 0-1 Extinction of contract pproc:contractOr ProcedureExtinction extinction information pproc:Contract pproc:ContractOr ProcedureExtinction bNode 0-1 domain pproc:ContractModification Why price Changed pproc:modification Reason Reasons for price shift pproc:ContractModific ation xsd:string LITERAL 0-1 Why timeframe changed pproc:durationChange Reasons for time shift pproc:ContractModific ation xsd:string LITERAL 0-1 104 Appendix D Sample SPARQL Queries SELECT ?person ?organization ?nif ?role ?heads WHERE { ?s <https://schema.org/memberOf> ?o . ?o schema:taxID ?nif . ?o schema:location ?heads . ?o schema:legalName ?organization . ?s <https://schema.org/name> ?person . ?b rdf:subject ?s . ?b rdf:object ?o . ?b <https://schema.org/roleName> ?role . } LIMIT 100 SELECT DISTINCT ?person ?organization ?nif ?role ?shares ?heads ?contract ?price WHERE { ?s <https://schema.org/memberOf> ?o . ?o schema:legalName ?organization . ?o schema:taxID ?nif . ?o schema:location ?heads. ?s <https://schema.org/name> ?person . ?b rdf:subject ?s . ?b rdf:object ?o . ?b <https://schema.org/roleName> ?role . ?b <https://schema.org/description> "Acionista" . ?b m4s:companyShares ?shares . ?c pc:supplier ?o . ?c pc:actualPrice ?price . ?c schema:description ?contract . } LIMIT 100 SELECT DISTINCT?s ?CPVlabel ?broadLabel ?sisterLabel WHERE { ?s rdf:type pproc:Contract. ?s pc:mainObject ?CPV. ?CPV rdfs:label ?CPVlabel. ?CPV skos:broader ?broad. ?broad rdfs:label ?broadLabel. ?broad skos:narrower ?sister. ?sister rdfs:label ?sisterLabel FILTER ( lang(?broadLabel) = "pt" ). FILTER ( lang(?CPVlabel) = "pt" ). FILTER ( lang(?sisterLabel) = "pt" ). }LIMIT 100 105 Appendix E Shape Expressions :politician IRI EXTRA a { a [m4s:Parliamentarian] ; schema:description LITERAL ; schema:url IRI ; schema:sameAs IRI *; schema:alternateName LITERAL ; schema:name LITERAL ; schema:gender [<https://schema.org/>~] ; schema:birthDate schema:Date ; schema:spouse LITERAL ? ; m4s:propertyRegimen [<https://example.org/map4scrutiny/value#>~]?; schema:hasCredential LITERAL OR <http://data.europa.eu/snb/eqf/>~]*; schema:hasOccupation LITERAL OR [<http://data.europa.eu/esco/>~] * ; m4s:parliamentaryComissions [<https://example.org/map4scrutiny/value#>~]*; dbo:parliamentaryGroup LITERAL ? ; schema:memberOf IRI @:organization + ; m4s:parliamentaryTerm LITERAL + ; m4s:constituency LITERAL OR [<http://vocab.getty.edu/tgn/>~] ; schema:comment LITERAL ? ; } :organization IRI EXTRA a { a [schema:Organization] ; schema:legalName LITERAL ; schema:alternateName LITERAL * ; schema:hasOfferCatalog LITERAL * ; schema:location [<http://vocab.getty.edu/tgn/>~] OR LITERAL * ; schema:taxID LITERAL ? ; schema:url IRI ? ; } :contract IRI EXTRA a { a [pproc:Contract] ; schema:description LITERAL ; pc:procedureType [<https://example.org/map4scrutiny/value#>~] ; pc:kind [<https://example.org/map4scrutiny/value#>~] OR LITERAL * ; pc:mainObject [<http://purl.org/cpv/2008/>~] + ; pc:contractingAuthority LITERAL OR IRI @:organization ; pc:supplier LITERAL OR IRI @:organization ; pc:awardDate schema:Date ? ; pc:actualEndDate schema:Date ? ; pproc:duration LITERAL ; schema:datePublished schema:Date ; pc:location LITERAL OR [<http://vocab.getty.edu/tgn/>~] * ; pproc:legalDocumentReference LITERAL * ; pc:frameworkAgreement LITERAL ? ; pc:agreedPrice m4s:Euros ; pc:actualPrice m4s:Euros ? ; schema:url IRI ? ; pproc:contractModification BNODE @:modification ? ; pproc:contractOrProcedureExtinction BNODE @:modification ? ; } :modification BNODE EXTRA a { a [pproc:ContractModification] ? ; a [pproc:ContractOrProcedureExtinction] ? ; pproc:modificationReason LITERAL ? ; pproc:durationChange LITERAL ? ; pproc:extinctionCause LITERAL ? ; } :metaRole BNODE EXTRA a { a [rdf:Statement] ; a [schema:Role] ; rdf:subject IRI ; rdf:predicate [schema:memberOf] ; rdf:object IRI ; schema:description LITERAL ? ; schema:startDate schema:Date ? ; schema:endDate schema:Date ? ; schema:roleName LITERAL OR [<http://data.europa.eu/esco/>~] + ; m4s:companyShares LITERAL ? ; m4s:isAccumulation ["true"^^schema:Boolean "false"^^schema:Boolean] ; } 106 7 ATTACHMENTS 1 Supply oriented. The goal is to publish data in quantity. 2 Data needs to be requested or requires action to access. No API keys. 3 Values user engagement and feedback. Metadata emerges. 4 User incentive with hackathons and open communication channels. 5 Public interest, high-value data taxonomies are published 6 Linked Open Data and investment in digital skills development. 7 Data is high quality, timely and used for data-driven decisions. 8 Value creation and multiple sources working together with data communities and much more. An ecosystem is created, and data is no longer a goal but a means for a platform-based government. Attachment 1 8 Steps to Mature Data according to the OECD [12] Attachment 2 A Three Cycle View of Design Science Research [6] 107 8 REFERENCES [1] K. Granickas, “Open Data as a Tool to Fight Corruption,” European Public Sector Information Platform EPSI platform, Topic Report No. 2014 / 04, 2014. Accessed: Feb. 24 2021. [Online]. Available: http://35.158.62.204/sites/default/files/2014_open_data_as_a_tool_to_fight_ corruption.pdf [2] J. Kim and M. Hausenblas, "Homepage: 5-star Open Data,". [Online]. Available: https:// 5stardata.info/en/ (Accessed: Jan. 27 2021). [3] E. Folmer, S. Ronzhin, J. van Hillegersberg, W. Beek, and R. Lemmens, “Business Rationale for Linked Data at Governments: A Case Study at the Netherlands’ Kadaster Data Platform,” IEEE Access , vol. 8, pp. 70822–70835, 2020, doi: 10.1109/ACCESS.2020.2984691. [4] E. Fagnoni, E. Norton, B. Acosta, M. Maleshkova, M. Domingue, J. Mikroyannidis, and A. Mulholland., “How to use Linked Data,” Linkeddata.center wikitolearn.org , Oct. 2020. Accessed: Feb. 24 2021. [Online]. Available: https://en.wikitolearn.org/Course:How_to_use_Linked_Data [5] Transparency International, “Corruption Perception Index 2019,” Transparency International Berlin, Germany , 2020. Accessed: Feb. 24 2021. [Online]. Available: www.transparency.org/cpi [6] A. Hevner, “A Three Cycle View of Design Science Research,” Scandinavian Journal of Information Systems , pp. 1–7, 2007. [Online]. Available: https://www.researchgate.net/publication/ 254804390 [7] Transparency - Corruptionary A-Z - Transparency.org,". [Online]. Available: https:// www.transparency.org/en/corruptionary/transparency (Accessed: Nov. 1 2021). [8] D. V. Malito, “Measuring Corruption Indicators and Indices,” Robert Schuman Centre for Advanced Studies Italy, EUI Working Paper RSCAS 2014/13, Feb. 2014. Accessed: Nov. 1 2021. [Online]. Available: http://hdl.handle.net/1814/29872 [9] Transparency International, "What is corruption?," Transparency International. [Online]. Available: https://www.transparency.org/en/what-is-corruption (Accessed: Jan. 28 2021). [10] The World Bank, "Control of Corruption: Percentile Rank | Data Catalog," World Bank, 2020. [Online]. Available: https://datacatalog.worldbank.org/control-corruption-percentile-rank (Accessed: Jan. 28 2021). [11] European Comission, “Report From The Commission To The Council And The European Parliament: Eu Anti-Corruption Report,” European Comission Brussels, COM(2014) 38 final, Feb. 2014. Accessed: Feb. 24 2021. [Online]. Available: https://ec.europa.eu/home-affairs/sites/