Representing Texts as LOD: a Systematic Literature Review

Representing Texts as LOD: a Systematic Literature Review Michela Bandini & Valeria Quochi CNR-ILC Pisa ➔Significant growth in LOD interest in digital humanities over the last decade. ➔The majority of linguistic resources in the LLOD cloud are dictionaries, lexica, thesauri, terminologies, and controlled vocabularies. ➔The publishing of textual corpora as LOD is still debated and controversial. ➔Some recent projects have started to explore different approaches to converting into and representing textual corpora as LOD. ➔A systematic literature review to assess benefits and identify existing models to represent textual corpora as LOD. ➔Further goal is to find the most suitable model for ancient language inscriptions for our ongoing project purpose 2 Why a systematic literature review? Project ItAnt: Languages and Cultures of Ancient Italy. Historical Linguistics and Digital Models ❖Main goal: creation of an ecosystem of language resources for ancient “fragmentary” languages (Oscan, Neo Faliscan, Venetic, Celtic). ❖Very small number of texts, lack of knowledge, and epigraphic evidence. ❖The publication of text corpora (in TEI/EpiDoc) on the Semantic Web is still at an early stage: still deciding what approach/model to follow. 3 4 Systematic Literature Review Methodology 4 1. Define terms and questions 2. Define criterias 3. Screening 4. Quality assessment 5. Data analysis ➔What are the most relevant works that have already attempted to transform/represent (text) corpora as LOD? ➔What are the models and formats already in use by the digital humanities community to represent corpora as LOD? ➔How can we classify projects according to the model used for representation or the granularity of the representation? 5 Terms and Questions of the Research ❖Sources: regular/advanced search on digital libraries (e.g. ACL Anthology, DBPL, Google Scholar). ❖Authors: authors strictly related to LOD works on corpora (e.g. C. Chiarcos). ❖Keywords/seed terms: 3 seed keywords combined with other terms, multi-term search, extra keywords (e.g. "POWLA", "NIF”). ❖Filtering criteria: date, language, number of citations, kind of publications (e.g. 2000 and up, ita or eng, conference papers, etc.). 6 Criteria of the Research 7 Screening for Inclusion Title reading Abstract, intro & conclusion reading Deep reading Filtering criteria: We excluded papers that… ❖provided not interesting works or not relevant topics ❖provided papers related to theoretical topics as RDF or XML formats. Filtering criteria: We excluded papers that… ❖illustrated general and theoretical topics; ❖provided a LLOD-compliant model or activity clearly not related to texts. Filtering criteria: We excluded papers that… ❖targeted corpus-derived data represented as lexicons, or CSV / TSV data; ❖mentioned textual data, but dealt with platforms, websites, or ontologies. Results: 219 relevant works Saved on a Zotero Library Results: 136 relevant works Results: 77 relevant works EXAMPLE excluded: Zinn, C., Hinrichs, M., & Hinrichs, E. (2022). Adapting GermaNet for the Semantic Web. In Proceedings of the 18th Conference on Natural Language Processing (KONVENS 2022) (pp. 41-47). EXAMPLE excluded: Tittel, S., & Chiarcos, C. (2018). Historical lexicography of Old French and linked open data: transforming the resources of the Dictionnaire étymologique de l'ancien français with OntoLex-Lemon. EXAMPLE excluded: Van Assem, M., et all (2004). A method for converting thesauri to RDF/OWL. In The Semantic Web–ISWC 2004: Third International Semantic Web Conference, Hiroshima, Japan, November 7-11, 2004. Proceedings 3 (pp. 17-31). Springer Berlin Heidelberg. 8 Quality Assessment & Data Analysis Full-article Reading Data Classification Categorization according to 2 criteria: 1. level of granularity in data representation; 2. model and formats used for the representation. 9 Granularity of the Data Representation Textual data represented at a document level Partial representation of textual data Granular representation of textual data ❖data represented as bibliographic entities or cultural objects; ❖no representation of the textual data contained. ❖only some extracted text parts are represented as RDF triples (e.i. named entities). ❖data represented taking into account detailed info (e.i. sentences, tokens, morphological information, etc.).