Automatic Extraction of Social Networks from Literary Text

The goal of the research is to extract social networks from literary texts and to analyze structural balance of the resulting graphs. In order to accomplish the goal, we formulated the following list of problems: • to analyze related works on extracting graphs of interactions between characters, • to find (or develop) tools for text annotation, • to extract characters from the text, • to find interactions between extracted characters, • to construct relationship graphs, • to evaluate the quality of constructed graphs, • to analyze structural balance of the text. After reviewing related works on constructing social networks of characters, we created a tool (NovelGraphs) for English-language literary fiction, which uses a new approach of extracting characters and connections between them. As a result we obtained undirected graphs, where nodes represent characters found in the text, and edges connect them to other characters with whom they interact. Every edge has its weight, e.g. average sentiment of the context or frequency of characters pair in the text. If edge has no weight, it will not appear in the graph. We compared our automatically constructed graphs with the «gold standard» (graph made by expert). The best result was achieved with TokenDistance (detects characters if distance between them is less than or equal to 15 tokens) and TokenDependencies (detects characters if they are syntactically connected by the verb) extractors. At the moment, combinations of extractors and aggregators detect characters better than interactions between them. It was established, that analysis of structural balance identifies key passages of the text that correspond to the minima and maxima on the balance plot.