What are the structural differences in character networks within Portuguese-language literary works across various genres?

Master Degree Program in Data Science and Advanced Analytics What are the structural differences in character networks within Portuguese-language literary works across various genres? Catarina Ribeiro Duarte Master Thesis presented as partial requirement for obtaining a Master’s Degree in Data Science and Advanced Analytics NOVA Information Management School Instituto Superior de Estatística e Gestão de Informação Universidade Nova de Lisboa MDSAA i NOVA Information Management School Instituto Superior de Estatística e Gestão de Informação Universidade Nova de Lisboa What are the structural differences in character networks within Portuguese-language literary works across various genres? by Catarina Ribeiro Duarte Master Thesis presented as partial requirement for obtaining the Master’s degree in Data Science and Advanced Analytics, with a specialization in Data Science Supervised by Flávio Pinheiro, PhD, NOVA Information Management School João L. M. Pereira, PhD, Universidade de Évora July, 2024 i 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 Rules of Conduct and Code of Honor from the NOVA Information Management School. Lisboa, 15th July 2024 ii DEDICATION Para os meus Pais, Ana Isabel e João Paulo, que levo sempre comigo mesmo sem ser Joana. iii ACKNOWLEDGEMENTS I am deeply grateful to my primary supervisor, Professor Flávio Pinheiro, for his continuous support, invaluable guidance, and ever-present good mood. My thanks also extend to Professor João Pereira for joining as a co-advisor and enriching this work with his insights and knowledgeable point of view. Additionally, I would like to express my heartfelt appreciation to NOVA IMS for providing an excellent educational environment throughout my studies. A distinctive thanks to Tiago Canário, an indisputable surprise and a fantastic teammate. Not only for the hard work, motivation, and companionship in hard challenges but also for the music recommendations that became the soundtrack for most of my writing sessions. If this experience were a book, it would be a mix of genres, but mainly it was an adventure, a tale of perseverance, and an ode to the academic journey that brought me here. I’m so grateful for the characters in this narrative, whose network would be vast and dense, and I would like to thank the dear and important characters in this story. Firstly, my whole family, whose trust and support I have always had and who passed on to me a love for books. Mãe, Pai, Mano, and Magali, you have been my home where I could always rest and find comfort in the middle of the journey in so many ways. Your support was vital to this narrative and kept the story moving forward. Marta, who has been present and essential since the Prologue, even though from a different background, celebrated so deeply every conquest and struggle with me, the only one able to make 300 km seem so close. And Carolina, so intrinsically present in every chapter, always offering words of encouragement and teaching me in so many forms the meaning of 'unconditional'. I also extend my gratitude to all my closest friends who added a touch of comedy to this tale; it would be a mystery how it could have been without all and each of you. Lastly, thanks to Avó Celeste, whose belief in me endures. iv ABSTRACT This study investigates the structural differences in character networks within Portugueselanguage literary works across various genres, including Fantasy, Mystery, and Romance. By exploring how character interactions and relationships differ across genres, we aimed to understand the underlying narrative structures that define each genre. Through Social Network Analysis, specific metrics were extracted for each genre, analyzed, and a Principal Component Analysis (PCA) was conducted, with real-life networks added for comparison and reference. Our analysis revealed that, although no extremely prominent metric differences were observed among the genres, there were notable structural tendencies within each genre. Fantasy narratives tend to focus on a single main character that interacts with many other characters, exhibiting lower density and clustering than other genres, and have less focus on relationship building. Mystery narratives feature moderately sized, denser networks with key bridging characters and the highest centrality on a relationship, translating into a balance of relationship and character building. In contrast, Romance narratives present the most variable network size, emphasizing multiple central characters and cohesive subgroups, focusing especially on relationship building. This makes Romance and Fantasy the most different genres, while Mystery acts as a mix of these structure strategies and is simultaneously closer to what is observed in real-life networks. The study acknowledges limitations related to sample size and data processing, highlighting the need for future research with larger and more diverse datasets to validate these findings. KEYWORDS Social Networks Analysis (SNA); NLP; Genre Theory; Portuguese Literature v TABLE OF CONTENTS 1. Introduction .................................................................................................................. 1 2. Related work ................................................................................................................. 3 3. Methodology ................................................................................................................ 6 4. Results and discussion ................................................................................................ 16 5. Conclusions and future works .................................................................................... 24 Bibliographical References .............................................................................................. 26 Appendix A ...................................................................................................................... 31 vi LIST OF FIGURES Figure 1 - Overview of the generic character network extraction methodology ...................... 7 Figure 2 - Character Network for "O Filho de Mil Homens" .................................................... 10 Figure 3 - Variance explained by each principal component and cumulative variance .......... 19 Figure 4 - Novels PC scores represented in 2D for 2 PC ........................................................... 21 Figure 5 - Character Network for "A Costa dos Murmúrios" ................................................... 31 Figure 6 - Character Network for "A Cidade do Medo" .......................................................... 36 Figure 7 - Character Network for "Limbo" ............................................................................... 41 6 3. METHODOLOGY This section outlines the methodology adopted to analyze the structural differences in character networks within Portuguese-language literary works across various genres. This work necessitates a meticulous selection of genres and literary works, which form the foundational framework for translating textual elements into character networks. These networks will then be measured using relevant metrics chosen to serve as comparison points among genres. The processes detailed below aim to capture the unique narrative dynamics that characterize each genre and highlight their differences. We start by outline the genres that were chosen for comparison. The focus was on subgenres of Fictional Narratives to create richer networks and set more coherent expectations for these subgenres. Sources differ in categorizing the same novels; some novels are labeled with multiple genres, while others are not categorized at all. To address this challenge and enable a more robust and distinct comparison, we selected genres where the reader's expectations are clearly defined. For example, readers anticipate a love story in a romance novel or a crime to be solved in a mystery novel. This contrasts with genres like adventure and fantasy, which may have more overlapping themes and less distinct reader expectations (Foster, J. n.d.). Regarding book selection, the sources for genre tagging included “Projeto Adamastor” 3 and the Goodreads 4 website. “Projeto Adamastor” is a digital library of public domain Portuguese-language works and it has been utilized in Silva et al. (2023) research. Each book is tagged with its genre, allowing for extraction and genre classification for use in this study. This platform is continuously updated, and it is noteworthy not only for converting texts into digital format but also for its meticulous review of each available work to minimize errors and adhere to the current Portuguese Spelling Reform. This is crucial for this study as it provides a base text that is the least error-prone and thus will produce the most accurate character networks. According to the “Collaborator Guide for Projeto Adamastor” by Lourenço, R.F. (2014), each collaborator identifies a genre during the selection and book treatment process, which is then validated by a peer review involving at least one collaborator not involved in the initial review. They conduct a comprehensive reading of the file to ensure the quality of the book and the accuracy of genre tagging. “Goodreads” is a book-based social website where members share, review, and rate literary works and connect with other readers (Thelwall & Kousha, 2017). On this platform, genres are attributed in a manner that considers several factors but remains largely undisclosed to users. It is known, however, that user influence plays a role due to genre tags reflecting the selfcreated names by its community. Shelves are a method for users to organize and group their 3 https://projectoadamastor.org 4 https://www.goodreads.com 7 books, often named after genres, such as romance or crime. These names influence genre assignments for the books, as the genre lists on Goodreads book home pages include up to ten of the most popular assigned book genres (Thelwall, 2019). For the book selection within each genre, it was ensured that the book, whether on “Projeto Adamastor” or among the first five genre tags on Goodreads, was identified with that genre. Although no genre assignment strategy can perfectly categorize genres and ensure a book fits only one category, these precautions and methodological approaches were adopted for this study. The chosen genres, all within the realm of fiction, in the Portuguese language, including Brazilian Portuguese, are presented in table 1 with their respective definitions from Goodreads. This alignment provides a preview of reader expectations and the sample collected focused on presenting more contemporary books. Regarding the pre-processing required for the book text before passing it into the pipeline to extract the networks, only a simple cleanup is needed to remove non-narrative text from the digital versions of the books. Therefore, to correct this issue, the initial or ending unnecessary chunks were manually removed from the TXT files of the books. The pipeline used for this work, TAGGUS (Canário & Duarte,2024) was specifically designed to extract character networks from Portuguese novels, addressing and overcoming all the impediments and constraints associated with the language that have been previously mentioned. Detailed information about the entire pipeline process and the decisions made can be found in it’s Github repository 5 . Nevertheless, Labatut & Bost (2020), in their widely cited work, outlined the extraction process, which also served as the foundation for the pipeline's creation and it’s represented in figure 1. Figure 1 - Overview of the generic character network extraction methodology 5 https://github.com/ticadu/taggus. 8 Table 1 - Chosen Genres and selected books for each Book Genre Goodreads definition Author, Book title and year Romance6 "Two basic elements comprise every romance novel: a central love story and an emotionally satisfying and optimistic ending." Both the conflict and the climax of the novel should be directly related to that core theme of developing a romantic relationship, although the novel can also contain subplots that do not specifically relate to the main characters' romantic love” Valter Hugo Mãe, “O Filho de Mil Homens” (2011)7 Lídia Jorge, “A Costa dos Murmúrios” (1988)8 Miguel Sousa Tavares, “Equador” (2003)9 João Ricardo Pedro, “O teu Rosto será o último” (2012) 10 Fantasy11 Fantasy is a genre that uses magic and other supernatural forms as a primary element of plot, theme, and/or setting. Fantasy is generally distinguished from science fiction and horror by the expectation that it steers clear of technological and macabre themes, respectively […]. Thiago d'Evecque, “Limbo” (2015) 12 Eduardo Spohr, “A batalha do Apocalipse” (2007) 13 Sandra Carvalho, “A Última Feiticeira” (2005) 14 Filipe Faria, “Oblívio” (2011)15 Mystery16 The mystery genre is a genre of fiction that follows a crime […] from the moment it is committed to the moment it is solved. Mystery novels often […] turn the reader into a detective trying to figure out the who, what, when, and how of a particular crime. Pedro Garcia Rosado, “A Cidade do Medo” (2015) 17 Nuno Nepomuceno, “A Célula Adormecida” (2007) 18 Raphael Montes, “Dias perfeitos” (2005)19 Francisco Moita Flores, “O Bairro da Estrela Polar” (2012)20 6 https://www.goodreads.com/genres/romance 7 https://www.goodreads.com/book/show/12395665-o-filho-de-mil-homens 8 https://www.goodreads.com/book/show/3371637-a-costa-dos-murm-rios 9 https://www.goodreads.com/book/show/1140942.Equador 10 https://www.goodreads.com/book/show/13597497-o-teu-rosto-ser-o-ltimo 11 https://www.goodreads.com/genres/fantasy 12 https://www.goodreads.com/book/show/25844730-limbo 13 https://goodreads.com/book/show/18299638-a-batalha-do-apocalipse 14 https://www.goodreads.com/book/show/6333188-a-ltima-feiticeira 15 https://www.goodreads.com/book/show/10474666-obl-vio 16 https://www.goodreads.com/genres/mystery 17 https://www.goodreads.com/book/show/12742439-a-cidade-do-medo 18 https://www.goodreads.com/book/show/32574062-a-c-lula-adormecida 19 https://www.goodreads.com/book/show/21405408-dias-perfeitos 20 https://www.goodreads.com/book/show/16155430-o-bairro-da-estrela-polar 9 The process begins with the preliminary dataset, a single literature work, which is fed into the model's pipeline. The first step is character identification, divided into two sequential stages. Initially, all the different character names are identified and stored. However, not every name represents a unique character, as narratives, like real life, often use various names to refer to the same character. For example, in the work Project Hail Mary 21 , Ryland Grace may also be referred to in the same story as Mr. Grace, representing the same character but by another expression. Therefore, a co-reference resolution step is employed to unify character occurrences. The pipeline used in this work takes precautions to ensures accurate unification by matching gender, nicknames, likelihood of the most popular character, and anaphoric resolution. Following character identification, the next step is to detect interactions between characters. In this work, co-occurrence is used to count interactions. This method decomposes the narrative into smaller units. These units may vary from project to project, considering two characters to interact when they appear together in the same unit (Labatut & Bost, 2020). Although this technique is widely used due to its simplicity and ease of implementation, it has the disadvantage of potentially labeling simply two characters mentioned in the same unit as interactions, leading to false positive interaction counts. However, this remains as the most widespread approach in the literature for this issue (Labatut & Bost, 2020). In this pipeline the narrative unit chosen to detect interaction was two characters mention in the same sentence. Finally, after unifying the characters and detecting their interactions, a graph is constructed. The graph that represents a character network is built with nodes and edges, each having its own definition. The nodes represent the characters, and the edges represent the relationships among them. Moreover, these graphs can also incorporate temporal integration, resulting in either a static network or a dynamic network (Labatut & Bost, 2020). The static network, which is the approach most commonly presented in the literature, captures interactions between characters over the entire narrative period. In contrast, the dynamic network integrates the progression of interactions over time or across different sections of the story. While the static network allows for a better visualization of the novel, it does not capture the evolution of characters throughout the narrative (Labatut & Bost, 2020). For this study, given the goal of understanding and analyzing the global story and the lack of necessity for temporal dimension analysis, static networks were used. The simpler element, nodes, can represent individual characters or groups of characters and also present character characteristics, such as character type (e.g., human, dog, etc.), gender, race, abilities, etc. Moreover, weighted nodes can give us an idea of the relevance of the characters throughout the story. On the other hand, the edges represent the interactions and connections among characters and can contain more information. Interaction intensity can be shown through weighted nodes. Additionally, the edges can have directions, creating a 21 https://www.goodreads.com/book/show/54493401-project-hail-mary 10 directed graph that transmits the speaker/addressee in the interaction. Furthermore, assigned graph properties can present the interaction polarity that conveys the sentiment of the interaction/relationship (Labatut & Bost, 2020). In this work, the static networks will feature weighted nodes, where each node represents a distinct character. While the nodes are weighted to indicate the relevance of each character, the edges in our network will be weighted as well, undirected, unsigned, and unattributed. An example can be seen in figure 2 and a representative network of each genre can be found in Appendix A as well as its respective adjacency matrix. Figure 2 - Character Network for "O Filho de Mil Homens" When comparing character networks, we focused on measurable aspects that could be translated to narrative significance and analyzed for genre-specific differences. For each extracted character network, we computed the network size (N) that informs on the number of nodes of the network, its size. In our context it gives us the number of characters of each book (Silva et al., 2023; Suen et al., 2013). Silva et al. (2023) analysis showed that the network size varies little between Portuguese literary works, with an average of nine characters in her sample. In the case of figure 2 it would be 12 nodes so, 12 characters. Furthermore, the number of components counts the subset of the vertices of a network that are connect, this is, that there exists at least one path from each member of that subset to each other member (Newman, 2010). In the context of character networks, analyzing the number of components can reveal how fragmented or cohesive the narrative structure is. For 11 example, a high number of components might indicate multiple isolated subplots or character groups that do not interact, while a low number of components, and especially a single one, suggests a more integrated and interconnected story. Next, maximum degree shows the highest number of connections a single character has, this is the highest degree of any node in the network (Silva et al., 2023; Labatut & Bost, 2020). This metric helps identify the most central or influential character within the narrative and how connected it is, providing insights into the structure and dynamics of the character interactions. Sudhahar & Cristianini (2013) found in their work that the hero of a narrative always had the highest degree in a network, showing their central position in the narrative. In figure 2 it’s clear that these would be the number of connections Matilde has, due to being the most connected character. In our work, we propose a new metric, “Number of Main Characters” (NºMC) which aims to count, besides the character with the highest degree, how many other central characters exist. This is done by counting any other node with a number of interactions ( 𝑘𝑖) close to the highest degree (𝑘𝑚𝑎𝑥), with a threshold of at least 80% and above a static lower bound of 10 interactions. This metric identifies both central and “secondary” main characters, reflecting narratives that feature multiple significant characters rather than focusing solely on one. The lower bound of 10 interactions ensures that characters considered as main have a meaningful level of engagement in the story. While metrics like maximum degree and single character centrality (SCC) highlight the most central character, they may overlook other important characters. The “Number of Main Characters” metric considers characters with substantial interactions, providing a more comprehensive view of the narrative’s structure and recognizing the importance of key secondary characters alongside the primary one. 𝑁º𝑀𝐶 = ∑[𝑘𝑖≥0.8 𝑘𝒎𝒂𝒙 ∧𝑘𝑖≥10 𝒊](𝟏) Still in the topic of connectivity, density measures how closely the characters are connected in the network. This is possible by calculating the proportion between existing edges (E) in the network and all possible network connections (Silva et al., 2023; Labatut & Bost, 2020; Dekker et al., 2019). The possible connections value is the count of pair of nodes that can form an edge. The density value ranges from 0 to 1, where a value closer to 1 indicates a highly interconnected network, and a value closer to 0 suggests a sparsely connected network. For example, Ardanuy & Sporleder (2015) note that in the case of the Harry Potter 22 books, the earlier books are considerably denser than the later ones, as the community represented in them becomes broader and less tightly knit with every new book. This shows how the density 22 https://www.goodreads.com/series/45175-harry-potter 12 metric can reflect difference in the structure and focus of a narrative, in this case in a series. The equation (2) used to calculate it was retrieved from Bhattacharya et al. (2023). 𝑑 = 2𝐸 𝑁(𝑁−1)(𝟐) The average path length measures the average number of steps to achieve the shortest paths for all possible pairs of nodes in the network (Silva et al., 2023; Labatut & Bost, 2020; Dekker et al., 2019; Newman, 2010). This metric provides insight into how connected the nodes, characters, are within the network, reflecting the overall efficiency of information or interaction flow within the narrative. A lower average path length suggests a close community where any character can be reached from any other character through a few interactions. Conversely, a higher average path length indicates a more spread out or fragmented network. In our study, in the case of when we have a fragmented network, we have in consideration the average path length of the largest component since it’s the best and fairest metric for that narrative. The real worlds concept of six degrees of separation, which suggests that any two people are, on average, six or fewer social connections apart (Milgram, 1967), can illustrate average path length usage in social networks. The defined equation (3) was retrieved from Newman (2010) and 𝑑𝑖𝑗 is the shortest distance between characters 𝑖 and 𝑗. 𝐴𝑃𝐿 = 1 𝑁(𝑁−1) ∑𝑑𝑖𝑗 𝑖≠𝑗 (𝟑) Now segueing into the centrality measures, which help identify the most important or influential nodes within the network, the single character centrality (SCC) measure for individual characters was proposed by Suen et al. (2013) to see how much a story is focused on a single character. This is achieved by measuring the disparity between the character with the highest weighted degree (𝑠𝑖) and the second highest, normalized, and understanding the difference. While maximum degree provides a direct measure of a character’s interactions, SCC quantifies the dominance or disparity in centrality within the network, indicating how much the story focuses on one character above all others. The SCC value ranges between 0 and 1, where a value closer to 1 indicates high dominance of a single character, and a value closer to 0 suggests a more evenly distributed character network 13 𝑆𝐶𝐶 = 𝑚𝑎𝑥𝑖(𝑠𝑖)−𝑛𝑒𝑥𝑡_𝑚𝑎𝑥𝑖(𝑠𝑖) ∑𝑠𝑖𝑖 (𝟒) Also proposed by Suen et al. (2013), single relationship centrality (SRC) measures the prominence of the most central relationship relative to others in the network. This metric quantifies how much a single relationship stands out in terms of interactions compared to the second most significant relationship. It helps identify the most dominant pair of characters, this is, the edge with more weight 𝑤𝑖𝑗, and their influence on the narrative. While single SCC focuses on the dominance of a single character, single relationship SRC highlights the importance of a specific interaction between two characters following the same scale logic. 𝑆𝑅𝐶 =𝑚𝑎𝑥𝑖,𝑗(𝑤𝑖,𝑗)−𝑛𝑒𝑥𝑡_𝑚𝑎𝑥𝑖,𝑗(𝑤𝑖,𝑗) ∑𝑤𝑖,𝑗𝑖,𝑗 (𝟓) The betweenness centrality (BC) measures the extent to which a certain node is on the shortest paths between other nodes (Hansen et al., 2020; Labatut & Bost, 2020; Suen et al., 2013; Hettinger et al., 2015). In other words, it helps identify characters who play a “bridge” role in a network. First, calculate the betweenness centrality for each node by counting the number of times it lies on the shortest path between other nodes (𝜎𝑠𝑡(𝑖)). Then, add up all the betweenness centrality values for the nodes in the network. Next, normalize each value by dividing it by the maximum possible value for any node in the network, ensuring all values fall between 0 and 1. Finally, compute the average betweenness centrality by dividing the total sum of the normalized values by the number of nodes in the network. By averaging it for the whole network, average betweenness centrality (ABC), we can capture the overall importance of intermediary nodes within the entire network. A value of 0 for average betweenness centrality means that, on average, nodes do not act as intermediaries; this could be due to a completely disconnected network or one where all the shortest paths are direct. In contrast, a high positive value means that many characters play a critical role in facilitating interactions or the flow of information between nodes. 𝐴𝐵𝐶 = 1 𝑁 ∑2 (𝑁−1)∗(𝑁−2) ∑𝜎𝑠𝑡(𝑖) 𝜎𝑠𝑡 𝑠≠𝑖≠𝑗 𝑁 𝑖=1 (𝟔) The degree assortativity measures how correlated the degrees of connected vertices are, in other words, it measures the tendency of nodes to connect to similar nodes with respect to their degree (Silva et al., 2023; Labatut & Bost, 2020). This metric helps characterize the relationship between primary and minor characters. Being a coefficient, assortativity is calculated as the ratio of covariance to variance. To calculate the covariance, we look at the 14 connections, degree 𝑘, of every node and then determine the joint degree distribution, 𝑒𝑗𝑘. This joint degree distribution indicates how often nodes with a certain degree 𝑘 are connected to nodes with degree 𝑗. In simpler terms, it measures the frequency of edges existing between nodes of different degrees. Then, this covariance is divided by the variance of the degree distribution, 𝜎2=∑𝑘2𝑞𝑘−[∑𝑘2𝑞𝑘] 𝑘2 𝒌 . The result takes a value between 1, assortative, where nodes with similar degrees tend to connect with each other, and -1, disassortative, where nodes with differing degrees tend to connect, or 0 (neutral) in the case of noncorrelation (Bhattacharya et al., 2023). In Silva et al. (2023), the networks were generally disassortative, reflecting the diversity of social interactions in the analysed works much like in real-world networks. The used formula was collected from Massey (2016). 𝑟 =∑(𝑒𝑗𝑘 −𝑞𝑗𝑞𝑘) 𝑗𝑘 𝜎𝑞 2(𝟖) Lastly, the average clustering coefficient (ACC) measures the degree to which nodes in a network tend to cluster together, providing insight into the overall tendency of characters to form cohesive and closed groups (Newman, 2010; Suen et al., 2013; Ardanuy et al., 2015; Dekker et al., 2019; Labatut & Bost, 2020). This metric is calculated as the average of the local clustering coefficients of all nodes in the network. The local clustering coefficient is the probability, when randomly picking two neighbors of a node, 𝑘𝑖, that there is an edge between them 𝑚𝑖 (Labatut & Bost, 2020) then averaging it by the number of nodes N. The ACC is useful for understanding the general clustering tendency across all nodes in the network. A higher ACC indicates a greater likelihood of strong community structures within the narrative. Dekker et al. (2019) provide a comparative analysis of clustering coefficients in social networks from novels and various well-known networks such as youtube and Flickr, demonstrating similar variation patterns across different network types. The used formula was collected from Massey (2016). 𝐶 = 1 𝑁 ∑ 2𝑚𝑖 𝑘𝑖 (𝑘𝑖−1) 𝑁 𝑖=1 (𝟗) The measures were calculated individually for each novel using Python with the NetworkX package 23 and then averaged by genre. Moreover, a Principal Component Analysis (PCA) was conducted with the metrics. PCA is a multivariate technique that extracts the most important information from the data by 23 https://networkx.org 15 reducing its dimensionality thought the creation of new variables that are linear combinations of the original variables. This process simplifies the dataset and allows for an analysis of its structure (Abdi & Williams, 2010). The goal is to understand the key metrics that distinguish genres while also have a visual representations of their grouping. 22 Analyzing the results, it's clear that although each genre generally gravitates together, each genre has no clear and defined group within the PCA results, indicating diversity in narrative structures within every genre under study. This also means that while the genres naturally distinguish themselves by grouping together novels of the same category, they do not clearly separate from those of different genres, indicating differences but not complete separation. In fact, Mystery presents itself as a bridge having its cloud of novels located between Fantasy and Romance, as these last two are more separated. This could imply that the Mystery genre, despite the thematic differences, employs similar narrative strategies in terms of character networks as the other two genres. Fantasy novels mostly present negative values of “Character Interconnectedness and Similarity” and have more prominent positive values than moderate negatives of “Network Size and Segmentation.” This translates into diverse structural features but a tendency to achieve networks that are more interconnected with fewer main characters interacting with a more diverse set of characters. Regarding network size its balanced, while some focus on larger, more segmented networks with many main characters, others lean towards smaller, denser networks, evidenced by the volatility in scores of “Network Size and Segmentation.” This suggests that this genre is more consensually focused on the creation of character work rather than relationships. Mystery novels gravitate towards the center of the axis, with most novels showcasing moderate results on both PCs, bridging both Romance and Fantasy. This indicates a balance in the attention towards character work as well as relationships. The central positioning on the PCA plot suggests that Mystery novels maintain a moderate network size and density, with a balanced number of main characters and intermediary connecting characters. This balance is crucial for the genre, as it often involves intricate plots that require a network structure that supports both the development of individual characters and the relationships between them. The Romance novels present the highest values of “Character Interconnectedness and Similarity”. This means that Romance novels focus on creating networks with multiple central characters, higher assortativity, and disconnected characters with a high average path length, indicating an emphasis on character relationships and interactions. Each Romance novel seems to have its unique value for “Network Size and Segmentation”, ranging the whole scale indicating a wide variability in network size and density. This variability, along with the presence of both highly connected and isolated characters, reflects the complexity and diversity of relationships typically explored in Romance narratives. This structure supports the genre's exploration of various subplots, providing a rich tapestry of character interactions. In conclusion, the PCA analysis reveals that while each genre exhibits distinct structural characteristics, being Romance and Fantasy the most different in terms of their character network structures. Romance novels focus on creating intricate, interconnected networks with multiple central characters, while Fantasy novels are more variable, ranging from highly interconnected networks to larger, more segmented ones. Mystery novels, acting as a bridge 23 and a compromise of the structure of Romance and Fantasy, showing balanced network structures that closely mimic real-life social networks. 24 5. CONCLUSIONS AND FUTURE WORKS Genre provides an important frame of reference that helps readers identify, select, and interpret texts (Chandler, 1997). This work explored the structural differences among genres in Portuguese texts through their character networks. By combining averaged network metrics, Principal Component Analysis (PCA) and the inclusion of real-life networks, we can draw comprehensive conclusions about how these genres differ and relate to one another. The averaged network metrics highlight some structural features within each genre. Fantasy novels are characterized by a single, highly central hero with high maximum degree and moderate levels of density and clustering, indicating a balanced interconnectedness. Mystery novels feature moderately sized, denser networks that highlight significant relationships between characters and the presence of bridging characters, maintaining connectivity. Romance novels display the most variable networks, with moderate density and high maximum degree, but also show tendencies for isolated characters, leading to disconnected networks. Lastly, real-life networks do not present any centrality on a particular character nor in a relationship. The Principal Component Analysis results support these findings by showing no clear clustering of genres, suggesting that structural differences in character networks are not strongly tied to genre. The disperse in the PCA results within each genre, indicates diversity in narrative structures. While the genres naturally distinguish themselves by grouping together novels of the same category, they do not clearly separate from those of different genres, indicating differences but not complete separation. Fantasy novels show diverse structural features but tend to achieve networks that are more interconnected with fewer main characters interacting with a more diverse set of characters. Most Mystery novels showcase a moderate balance between character and relationship building, while Romance emphasizes relationship building within its narrative dynamics. Mystery presents itself as a bridge between Fantasy and Romance, suggesting that despite thematic differences, similar narrative strategies are employed in terms of character networks from both genres. The real-life networks stand near each other and near the literary networks sharing common structural characteristics while also exhibiting unique features that reflect their distinct contexts, being this group most similar to Mystery. Our conclusion is that the similarities within genres and differences between genres are not prominent in graph-specific measures, even though each genre exhibits unique structural characteristics. Fantasy novels are characterized by a focus on a central hero and less interconnected groups, aligning with their narrative emphasis on individual journeys and less relationship building overall. Mystery novels exhibit moderately sized, denser networks with key bridging characters and a balance between relationships and character development, reflecting their intricate plots and are the most similar to real-life networks. Romance novels 25 display the largest and most variable networks, having more than one central character and cohesive subgroups, indicative of their focus on diverse and intricate relationships. These findings provide valuable insights into the narrative styles of Portuguese-language literature, showing how each genre employs distinct network structures to support their unique storytelling approaches. The lack of pronounced differences could be expected since, as Thelwall & Kousha (2017) point out, genre is a concept that readers, writers, and publishers may interpret uniquely. Booksellers may invent new genres to market books, libraries may categorize books in certain ways to attract readers, and genre types and typical structures can vary immensely in topics and internationally. In interpreting the findings of this study, it is important to consider the limitations associated with the data and methodology. Firstly, all the interpretations relate to the pipeline and its existing limitations, such as the inability to process stories with multiple points of view and some degree of imprecision in the co-occurrence detection, which is not fully robust to nonPortuguese names. Furthermore, one significant concern is the use of averages to describe network metrics for each genre. While averages provide a useful summary, they can sometimes obscure important variations and nuances within individual works, this is avoided by completing the information with the standard deviation. Another limitation, this related to the data, is the underdevelopment of Portuguese themes in genres such as fantasy or other as SiFi that could be studied, with ebook versions that would allow for a larger dataset. In fact, all the small sample sizes could lead to results that are less reliable and harder to generalize to the entire genre. Future research should aim to include a larger and more diverse sample of novels to validate these findings. 26 BIBLIOGRAPHICAL REFERENCES Abdi, H., & Williams, L. J. (2010). Principal component analysis. WIREs Computational Statistics, 2(4), 433–459. https://doi.org/10.1002/wics.101 Alberich, R., Miro-Julia, J., & Rossello, F. (2002). Marvel Universe looks almost like a real social network (arXiv:cond-mat/0202174). arXiv. https://doi.org/10.48550/arXiv.condmat/0202174 Amaral, D. O. F., Fonseca, E. B., Lopes, L., & Vieira, R. (n.d.). Comparative Analysis of Portuguese Named Entities Recognition Tools. Retrieved from https://repositorio.pucrs.br/dspace/bitstream/10923/14041/2/Comparative_Analysis_ of_Portuguese_Named_Entities_Recognition_Tools.pdf Ardanuy, M. C., & Sporleder, C. (2015). Clustering of Novels Represented as Social Networks. Linguistic Issues in Language Technology, 12. https://doi.org/10.33011/lilt.v12i.1379 Bento, F., Tagliabue, M., & Sandaker, I. (2020). Complex Systems and Social Behavior: Bridging Social Networks and Behavior Analysis (pp. 67–91). https://doi.org/10.1007/978-3-03045421-0_4 Bhattacharya, S., Sinha, S., Dey, P., Saha, A., Chowdhury, C., & Roy, S. (2023). Chapter 5— Online social-network sensing models. In D. Das, A. K. Kolya, A. Basu, & S. Sarkar (Eds.), Computational Intelligence Applications for Text and Sentiment Data Analysis (pp. 113– 140). Academic Press. https://doi.org/10.1016/B978-0-32-390535-0.00010-0 Borgatti, S. P., Mehra, A., Brass, D. J., & Labianca, G. (2009). Network Analysis in the Social Sciences. Science, 323(5916), 892–895. https://doi.org/10.1126/science.1165821 Borgatti, S. P., Everett, M. G., & Johnson, J. C. (2018). Analyzing Social Networks. SAGE Publications. https://books.google.pt/books?id=XD1ADwAAQBAJ Cambria, E., & White, B. (2014). Jumping NLP Curves: A Review of Natural Language Processing Research [Review Article]. IEEE Computational Intelligence Magazine, 9(2), 48–57. https://doi.org/10.1109/MCI.2014.2307227 Canário,T., & Duarte, C. (2024). TAGGUS. GitHub. Retrieved July 3, 2024, from https://github.com/ticadu/taggus Chandler, D. (1997). An introduction to genre theory. http://visualmemory.co.uk/daniel//Documents/intgenre/chandler_genre_theory.pdf Dekker, N., Kuhn, T., & Erp, M. van. (2019). Evaluating named entity recognition tools for extracting social networks from novels. PeerJ Computer Science, 5, e189. https://doi.org/10.7717/peerj-cs.189 27 dos Reis, C. A. A. (1993). História crítica da literatura portuguesa. Editorial Verbo. https://books.google.pt/books?id=G8YQAQAAMAAJ Earle, T. F., Parkinson, S., & Alonso, C. P. (Eds.). (2013). A companion to Portuguese literature (Vol. 282). Boydell & Brewer Ltd. https://books.google.com/books?hl=ptPT&lr=&id=TNQTAgAAQBAJ&oi=fnd&pg=PP1&dq=portuguese+literature&ots=t812Ai5 dnC&sig=1cJBKY0nl0i89YP2OJn0VI1EI7Q Fan, C., & Li, Y. (2022). Network extraction and analysis of character relationships in Chinese literary works. Computational Intelligence and Neuroscience, 2022. https://doi.org/10.1155/2022/7295834 Fernández-Peña, R., Ovalle-Perandones, M.-A., Marqués-Sánchez, P., Ortego-Maté, C., & Serrano-Fuentes, N. (2022). The use of social network analysis in social support and care: A systematic scoping review protocol. Systematic Reviews, 11(1), 9. https://doi.org/10.1186/s13643-021-01876-2 Fonseca, B. de P. F. e, Sampaio, R. B., Fonseca, M. V. de A., & Zicker, F. (2016). Co-authorship network analysis in health research: Method and potential use. Health Research Policy and Systems, 14(1), 34. https://doi.org/10.1186/s12961-016-0104-5 Foster, J. (n.d.). Examining mystery, science fiction, romance, and horror: A look at popular fiction genres. Jamie Foster Science. Retrieved June 19, 2024, from https://www.jamiefosterscience.com/mystery-science-fiction-romance-and-horrorare-all-considered/ Fowler, A. (1982). Kinds of literature: An introduction to the theory of genres and modes. Cambridge, Mass. : Harvard University Press. http://archive.org/details/kindsofliteratur0000fowl Freeman, L. C. (2004). The Development of Social Network Analysis: A Study in the Sociology of Science. Empirical Press. https://books.google.pt/books?id=VcxqQgAACAAJ Fronzetti Colladon, A., & Naldi, M. (2019). Predicting the performance of TV series through textual and network analysis: The case of Big Bang Theory. PLOS ONE, 14(11), e0225306. https://doi.org/10.1371/journal.pone.0225306 Gessey-Jones, T., Connaughton, C., Dunbar, R., Kenna, R., MacCarron, P., O’Conchobhair, C., & Yose, J. (2020). Narrative structure of A Song of Ice and Fire creates a fictional world with realistic measures of social complexity. Proceedings of the National Academy of Sciences, 117(46), 28582–28588. https://doi.org/10.1073/pnas.2006465117 Hansen, D. L., Shneiderman, B., Smith, M. A., & Himelboim, I. (2020). Chapter 6—Calculating and visualizing network metrics. In D. L. Hansen, B. Shneiderman, M. A. Smith, & I. 28 Himelboim (Eds.), Analyzing Social Media Networks with NodeXL (Second Edition) (pp. 79–94). Morgan Kaufmann. https://doi.org/10.1016/B978-0-12-817756-3.00006-6 Hettinger, L., Becker, M., Reger, I., Jannidis, F., & Hotho, A. (2015). Genre classification on German novels. In 2015 26th International Workshop on Database and Expert Systems Applications (DEXA) (pp. 249-253). IEEE. https://doi.org/10.1109/DEXA.2015.62 Isinkaye, F. O., Folajimi, Y. O., & Ojokoh, B. A. (2015). Recommendation systems: Principles, methods and evaluation. Egyptian Informatics Journal, 16(3), 261–273. https://doi.org/10.1016/j.eij.2015.06.005 Jarynowski, A., & Boland, S. (2016). Social Networks Analysis in Discovering the Narrative Structure of Literary Fiction. https://doi.org/10.48550/arXiv.1608.05982 Kim, K.-R., & Kim, J. (2012). Recommender system design using movie genre similarity and preferred genres in SmartPhone. Multimedia Tools and Applications - MTA, 61, 1–18. https://doi.org/10.1007/s11042-011-0728-y Labatut, V., & Bost, X. (2020). Extraction and Analysis of Fictional Character Networks: A Survey. ACM Computing Surveys, 52(5), 1–40. https://doi.org/10.1145/3344548 Lourenço, R.F. (2014) Guia do Colaborador [v1.01], 16. https://projectoadamastor.org/wpcontent/uploads/2014/09/Guia-do-Colaborador.docx Mamede, N., & Chaleira, P. (2004). Character Identification in Children Stories. In J. L. Vicedo, P. Martínez-Barco, R. Muńoz, & M. Saiz Noeda (Eds.), Advances in Natural Language Processing (Vol. 3230, pp. 82–90). Springer Berlin Heidelberg. https://doi.org/10.1007/978-3-540-30228-5_8 Massey, S. E. (2016). Social network analysis of the biblical Moses. Applied Network Science, 1(1), Article 1. https://doi.org/10.1007/s41109-016-0012-1 Mata, A. S. da. (2020). Complex Networks: A Mini-review. Brazilian Journal of Physics, 50(5), 658–672. https://doi.org/10.1007/s13538-020-00772-9 Maulud, D. H., Zeebaree, S. R. M., Jacksi, K., Sadeeq, M. A. M., & Sharif, K. H. (2021). State of Art for Semantic Analysis of Natural Language Processing. Qubahan Academic Journal, 1(2), Article 2. https://doi.org/10.48161/qaj.v1n2a44 Middha, K., Munzir, M., Goja, S., & Choudhary, S. (2022). Entertainment content recommendation system using machine learning. International Research Journal of Engineering and Technology (IRJET), 10(1), 279-284. https://www.irjet.net/archives/V10/i1/IRJET-V10I148.pdf 29 Mourchid, Y., Renoust, B., Roupin, O., Văn, L., Cherifi, H., & Hassouni, M. E. (2019). Movienet: A movie multilayer network model using visual and textual semantic cues. Applied Network Science, 4(1), Article 1. https://doi.org/10.1007/s41109-019-0226-0 Network Density—An overview. ScienceDirect Topics. Retrieved June 20, 2024, from https://www.sciencedirect.com/topics/computer-science/network-density Newman, M. (2010). Networks: An Introduction. OUP Oxford. https://books.google.pt/books?id=LrFaU4XCsUoC Norman, M., Jones, C., Watson, K., & Previdelli, R. L. (2023). Social Network Analysis as a Tool in the Care and Wellbeing of Zoo Animals: A Case Study of a Family Group of Black Lemurs (Eulemur macaco). Animals, 13(22), Article 22. https://doi.org/10.3390/ani13223501 Otte, E., & Rousseau, R. (2002). Social network analysis: A powerful strategy, also for the information sciences. Journal of Information Science. 28 (6): 441–453. https://doi.org/10.1177/016555150202800601 Ozella, L., Paolotti, D., Lichand, G., Rodríguez, J. P., Haenni, S., Phuka, J., Leal-Neto, O. B., & Cattuto, C. (2021). Using wearable proximity sensors to characterize social contact patterns in a village of rural Malawi. EPJ Data Science, 10(1), Article 1. https://doi.org/10.1140/epjds/s13688-021-00302-w Park, S.-B., Kim, Y.-W., Uddin, M. N., & Jo, G.-S. (2009). Character-Net: Character Network Analysis from Video. 2009 IEEE/WIC/ACM International Joint Conference on Web Intelligence and Intelligent Agent Technology, 1, 305–308. https://doi.org/10.1109/WIIAT.2009.54 Park, G.-M., Kim, S.-H., & Cho, H.-G. (2013). Structural Analysis on Social Network Constructed from Characters in Literature Texts. Journal of Computers, 8. https://doi.org/10.4304/jcp.8.9.2442-2447 Rahul, Ayush, Agarwal, D., & Vijay, D. (2021). Genre Classification using Character Networks. 2021 5th International Conference on Intelligent Computing and Control Systems (ICICCS), 216–222. https://doi.org/10.1109/ICICCS51141.2021.9432303 Reddy, S. R., Nalluri, S., Kunisetti, S., Ashok, S., & Venkatesh, B. (2019). Content-based movie recommendation system using genre correlation. In S. C. Satapathy, V. Bhateja, & H. S. Saini (Eds.), Smart Intelligent Computing and Applications (Vol. 105, pp. 391-397). Springer. https://doi.org/10.1007/978-981-13-1927-3_42 Rocha, C., Jorge, A., Oliveira, M., Brito, P., Gama, J., & Pimenta, C. (2014). From entity extraction to network analysis: A method and an application to a Portuguese textual source. https://www.fep.up.pt/docentes/cpimenta/textos/pdf/ACM15_Rocha2.pdf 30 Rossi, R. A., & Ahmed, N. K. (2015). The Network Data Repository with interactive graph analytics and visualization. In Proceedings of the AAAI Conference on Artificial Intelligence , 29(1). https://doi.org/10.1609/aaai.v29i1.9277 Silva, M. O., Oliveira, G. P., & Moro, M. M. (2023). Analyzing Character Networks in Portuguese-language Literary Works. Anais Do Brazilian Workshop on Social Network Analysis and Mining (BraSNAM), 115–126. https://doi.org/10.5753/brasnam.2023.230585 Sudhahar, S., Cristianini, N. (2013). Automated analysis of narrative content for digital humanities. International Journal of Advanced Computer Science. 3. https://citeseerx.ist.psu.edu/document?repid=rep1&type=pdf&doi=58e9de78dc9c534 8339e83f774c689bbf7d9c45a Sugishita, K., & Masuda, N. (2023). Social network analysis of manga: Similarities to real-world social networks and trends over decades. Applied Network Science, 8(1), Article 1. https://doi.org/10.1007/s41109-023-00604-0 Suen, C., Kuenzel, L., & Gil, S. (2013). Extraction and Analysis of Character Interaction Networks From Plays and Movies. Digital Humanities Conference. https://www.semanticscholar.org/paper/Extraction-and-Analysis-of-CharacterInteraction-Suen-Kuenzel/a57bf7098ea7adccf626caba682f3b4350aa3955 Tamen, M., & Buescu, H. C. (2013). A revisionary history of Portuguese literature. Routledge. https://api.taylorfrancis.com/content/books/mono/download?identifierName=doi&id entifierValue=10.4324/9780203726877&type=googlepdf Tantardini, M., Ieva, F., Tajoli, L., & Piccardi, C. (2019). Comparing methods for comparing networks. Scientific Reports, 9(1), 17557. https://doi.org/10.1038/s41598-019-53708-y Thelwall, M., & Kousha, K. (2017). Goodreads: A social network site for book readers. Journal of the Association for Information Science and Technology, 68(4), 972–983. https://doi.org/10.1002/asi.23733 Wellek, Ren., & Warren, A. (1956). Theory of literature. Harcourt, Brace & World. Zachary, W. W. (1977). An Information Flow Model for Conflict and Fission in Small Groups. Journal of Anthropological Research, 33(4), 452–473. https://doi.org/10.1086/jar.33.4.3629752 31 APPENDIX A Figure 5 - Character Network for "A Costa dos Murmúrios" Genre: Romance 38 João Deus Jaime Paixão Amélia Alice Domingos Sequeira João Pessoa Nicolau Eunice TVN William Alenquer 0 0 0 0 0 1 1 2 Guilherme Vau 0 1 0 0 0 0 0 0 Alberto 0 0 0 0 0 0 0 0 Estrela 1 1 0 0 1 0 3 0 Afonso Soares 0 0 0 0 0 0 1 0 Joel Franco 0 19 20 13 0 0 0 2 Barradas 0 1 0 0 0 0 1 0 Direito 0 1 0 1 0 0 0 0 João Deus 0 0 0 0 0 0 0 0 Jaime Paixão 0 0 5 0 0 0 0 0 Amélia 0 5 0 8 0 0 0 0 Alice 0 0 8 0 0 0 0 0 Domingos Sequeira 0 0 0 0 0 0 0 0 João Pessoa 0 0 0 0 0 0 0 0 Nicolau 0 0 0 0 0 0 0 1 Eunice TVN 0 0 0 0 0 0 1 0 Ana Paula 0 0 0 0 3 0 0 0 Humberto 0 0 0 0 0 0 0 0 Bernardino Reis 0 3 0 0 0 0 0 3 Sebastião Carvalho 0 0 0 0 0 0 0 1 Inês Albergaria 0 0 0 0 0 0 0 0 Northrope 0 0 0 0 0 0 0 0 Londres 0 0 0 0 0 0 0 0 António Monsanto 0 0 0 0 0 0 0 3 Conceição Bonifácio 0 0 0 1 0 0 0 0 Marlene 0 0 0 0 0 0 0 0 Jorge 0 0 0 0 0 0 0 0 Golden Investments 0 0 0 0 0 0 0 0 Aníbal 0 0 0 0 0 0 0 0 Anton Grisak 0 0 0 0 0 0 0 0 39 Ana Paula Humberto Bernardin o Reis Sebastião Carvalho Inês Albergaria Northrope Londres William Alenquer 0 3 9 0 0 4 3 Guilherme Vau 1 0 33 0 0 13 4 Alberto 0 0 0 0 0 0 0 Estrela 0 0 0 0 0 0 0 Afonso Soares 0 0 15 0 0 0 0 Joel Franco 0 0 20 0 10 1 0 Barradas 0 0 1 0 0 0 0 Direito 0 0 1 0 0 0 0 João Deus 0 0 0 0 0 0 0 Jaime Paixão 0 0 3 0 0 0 0 Amélia 0 0 0 0 0 0 0 Alice 0 0 0 0 0 0 0 Domingos Sequeira 3 0 0 0 0 0 0 João Pessoa 0 0 0 0 0 0 0 Nicolau 0 0 0 0 0 0 0 Eunice TVN 0 0 3 1 0 0 0 Ana Paula 0 0 0 0 0 0 0 Humberto 0 0 0 0 0 0 0 Bernardino Reis 0 0 0 0 1 3 0 Sebastião Carvalho 0 0 0 0 0 0 0 Inês Albergaria 0 0 1 0 0 0 0 Northrope 0 0 3 0 0 0 6 Londres 0 0 0 0 0 6 0 António Monsanto 0 0 12 0 1 3 1 Conceição Bonifácio 0 0 0 0 0 0 0 Marlene 0 0 0 0 0 0 0 Jorge 0 0 0 0 0 0 0 Golden Investments 0 0 1 0 0 5 3 Aníbal 0 0 1 0 0 0 0 Anton Grisak 0 0 0 0 0 1 0 40 António Monsanto Conceição Bonifácio Marlene Jorge Golden Investments Aníbal Anton Grisak William Alenquer 10 4 3 0 0 0 0 Guilherme Vau 17 2 0 0 3 0 1 Alberto 0 0 0 0 0 0 0 Estrela 1 0 0 0 0 0 0 Afonso Soares 4 0 0 0 0 0 0 Joel Franco 9 4 0 0 0 1 0 Barradas 1 0 0 1 0 0 0 Direito 1 0 1 0 0 0 0 João Deus 0 0 0 0 0 0 0 Jaime Paixão 0 0 0 0 0 0 0 Amélia 0 0 0 0 0 0 0 Alice 0 1 0 0 0 0 0 Domingos Sequeira 0 0 0 0 0 0 0 João Pessoa 0 0 0 0 0 0 0 Nicolau 0 0 0 0 0 0 0 Eunice TVN 3 0 0 0 0 0 0 Ana Paula 0 0 0 0 0 0 0 Humberto 0 0 0 0 0 0 0 Bernardino Reis 12 0 0 0 1 1 0 Sebastião Carvalho 0 0 0 0 0 0 0 Inês Albergaria 1 0 0 0 0 0 0 Northrope 3 0 0 0 5 0 1 Londres 1 0 0 0 3 0 0 António Monsanto 0 2 0 0 0 0 1 Conceição Bonifácio 2 0 0 0 0 0 0 Marlene 0 0 0 0 0 0 0 Jorge 0 0 0 0 0 0 0 Golden Investments 0 0 0 0 0 0 0 Aníbal 0 0 0 0 0 0 0 Anton Grisak 1 0 0 0 0 0 0 41 Figure 7 - Character Network for "Limbo" Genre: Fantasy 42 Table 7 - Adjacency Matrix for "Limbo" Genre: Fantasy Verdade Lilith Azazel Rafae Carlos Tomoe Marsílio Olivier Carlos Magno Durandal Verdade 0 1 0 0 0 0 0 0 0 0 Lilith 1 0 12 0 23 1 0 1 1 0 Azazel 0 12 0 2 1 0 0 0 0 0 Rafael 0 0 2 0 0 0 0 0 0 0 Carlos 0 23 1 0 0 1 0 0 0 1 Tomoe 0 1 0 0 1 0 0 0 0 0 Marsílio 0 0 0 0 0 0 0 1 0 0 Olivier 0 1 0 0 0 0 1 0 1 0 Carlos Magno 0 1 0 0 0 0 0 1 0 0 Durandal 0 0 0 0 1 0 0 0 0 0 Musa Musas 0 1 0 0 0 0 0 0 0 0 Finn Cumhail 0 0 0 0 3 0 0 0 0 0 ExGeneral 1 1 0 0 0 0 0 0 0 0 Professor 0 1 0 0 1 0 0 0 0 0 Henrique Xiangu 0 1 0 0 2 0 0 0 0 0 Tio 0 1 0 0 0 0 0 0 0 0 Adolesce nte 1 0 0 0 0 0 0 0 0 0 Ancião 0 0 0 0 0 0 0 0 0 0 Odin 0 1 0 0 0 0 0 0 0 0 Thor 0 0 0 0 0 0 0 0 0 0 Miguel 0 0 1 1 0 0 0 0 0 0 Eleos 0 5 0 0 2 0 0 0 0 0 Artur 0 3 0 0 1 0 0 0 0 0 Samael 0 1 0 0 1 0 0 0 0 0 Gabriel 0 4 0 1 0 0 0 0 0 0 43 Musa Musas Finn Cumhail ExGeneral Professo r Henriqu e Xiangu Tio Adolesce nte Ancião Verdade 0 0 1 0 0 0 1 0 Lilith 1 0 1 1 1 1 0 0 Azazel 0 0 0 0 0 0 0 0 Rafael 0 0 0 0 0 0 0 0 Carlos 0 3 0 1 2 0 0 0 Tomoe 0 0 0 0 0 0 0 0 Marsílio 0 0 0 0 0 0 0 0 Olivier 0 0 0 0 0 0 0 0 Carlos Magno 0 0 0 0 0 0 0 0 Durandal 0 0 0 0 0 0 0 0 Musa Musas 0 0 0 0 0 0 0 0 Finn Cumhail 0 0 0 0 0 0 0 0 Ex-General 0 0 0 4 1 0 0 1 Professor 0 0 4 0 1 2 1 1 Henrique Xiangu 0 0 1 1 0 0 0 0 Tio 0 0 0 2 0 0 1 1 Adolescent e 0 0 0 1 0 1 0 1 Ancião 0 0 1 1 0 1 1 0 Odin 0 0 0 0 0 0 0 0 Thor 0 0 0 0 0 0 0 0 Miguel 0 0 0 0 0 0 0 0 Eleos 0 0 0 0 0 0 0 0 Artur 0 1 0 0 0 0 0 0 Samael 0 0 0 0 0 0 0 0 Gabriel 0 0 0 0 0 0 0 0 44 Odin Thor Miguel Eleos Artur Samael Gabriel Verdade 0 0 0 0 0 0 0 Lilith 1 0 0 5 3 1 4 Azazel 0 0 1 0 0 0 0 Rafael 0 0 1 0 0 0 1 Carlos 0 0 0 2 1 1 0 Tomoe 0 0 0 0 0 0 0 Marsílio 0 0 0 0 0 0 0 Olivier 0 0 0 0 0 0 0 Carlos Magno 0 0 0 0 0 0 0 Durandal 0 0 0 0 0 0 0 Musa Musas 0 0 0 0 0 0 0 Finn Cumhail 0 0 0 0 1 0 0 Ex-General 0 0 0 0 0 0 0 Professor 0 0 0 0 0 0 0 Henrique Xiangu 0 0 0 0 0 0 0 Tio 0 0 0 0 0 0 0 Adolescente 0 0 0 0 0 0 0 Ancião 0 0 0 0 0 0 0 Odin 0 1 0 0 0 0 0 Thor 1 0 0 0 0 0 0 Miguel 0 0 0 0 0 0 1 Eleos 0 0 0 0 0 0 0 Artur 0 0 0 0 0 0 0 Samael 0 0 0 0 0 0 0 Gabriel 0 0 1 0 0 0 0 45