Methodology for analyzing unstructured data (networks and texts)

One of the areas of work with data of this type in the laboratory is devoted to the development of a methodology for identifying interpersonal connections in narrative texts in Russian.

In Yegor Perekhozhev’s dissertation research “Network analysis of (narrative) texts: methodology for extracting data on connections between elements in the Russian language,” two tasks were set. The first part of the work is aimed at creating a program for isolating named entities from texts using the stanza, Natasha, etc. packages in the Python programming language (the project requires a comparative study of tools to better solve the problems), as well as building connections between them with subsequent visualization (using the NetworkX package in Python). The second part of the work is devoted to creating a program for extracting named entities from text data and analyzing the relationships between them taking into account the context (positive - neutral - negative). To implement the second part, it is proposed to use the Selenium or BeautifulSoup packages to download data about the plots of released Russian films, cartoons and TV series from Wikipedia. Next, using coders (that is, manually) or using ChatGPT (automatically), mark up these texts for connections between the main characters, locations and objects. After this, the resulting model can be used to analyze narrative texts, for example, sociological interviews, speeches of public speakers, etc.

The methodology for analyzing unstructured data (networks and texts) is used in a number of projects by laboratory staff: for example, to study the structure of the judiciary in modern Russia and the most typical career trajectories of judges, empirical legal research on judicial decisions and crime in Russia.

The project is aimed at researching methods and algorithms for identifying named entities, their relationships and filling knowledge graphs. For a clearer understanding, it can be divided into the following subtopics:

• Linking of entities extracted from texts

When creating a knowledge base based on entities extracted from a text, the task of comparing the entity with a standard from the knowledge base arises. This task is called entity linking. A successful solution to this problem allows you to:

1. detect entities that denote the same real-world object and combine information related to them (for example, “King of France” and “Louis XIV”);
2. distinguish entities with the same names, but relating to different objects of the real world (for example, see https://ru.wikipedia.org/wiki/Ivanov,_Sergey). External data sources can be used for this, but for some applications this is not possible.

The task itself usually belongs to the field of NLP (input data for analysis comes from texts), but graph methods are used for solution.
In this topic, it is required to develop and implement an algorithm for disambiguation of entities extracted from text using a third-party application, without using external data. Additionally, you should pay attention to the time component of the extracted facts – some of them are valid for a limited time. To simplify the task, it is proposed to use a model array of real organizations.

• Bipartite graph matching method

The task of matching two graphs allows you to find a node from graph A in graph B without having a priori knowledge about their connection.
In applied tasks, this allows you to compare groups and users in different networks. Most existing algorithms consider only homogeneous graphs in which users are connected to nodes of the same type (for example, a friendship graph in two social networks). The ability to analyze a bipartite graph (For example, a graph of user membership in various groups) allows you to use metrics for comparing two probability distributions in this problem, which provides additional information compared to basic analysis algorithms and has not been used by anyone before.
Attempts to solve this problem on synthetic data greatly distort the results, since the algorithm turns out to be sensitive to the method of generating a synthetic graph from the original one and does not describe real data.
It is proposed to compare groups and users for two real-life graphs using the example of comparing the IMDb and film search databases. Both bases have significant overlap in terms of films and actors, but there are also differences. At the same time, it is not difficult to compare the film team and the title in two databases using indirect signs.

• Methods for extracting relations from texts
• Methods for extracting quantitative information from texts
• Methods for classifying goods based on graph models

The project is dedicated to the implementation of primarily graph models in applied text analysis problems.

Given the fairly high complexity of modern language models, it turns out that due to the iterative nature of training, they do not use all the data about the subject area. Some researchers solve this problem by artificially forming a training sample with the required sequence of input documents, but this cannot be used in the general case.

It is proposed to perform preliminary customization of the input array of documents, then supplement the existing basic language model with a label or cluster vector to which the document belongs and study the change in the quality of the solution to the problem of predicting a missing token.

The study aims to analyze available topic modeling methods for the quality of their results on short texts. The relevance of the study is determined by the steady growth in the share of texts produced in social media and limited by platform limits on the volume of publication. Isolation of thematic structure in arrays of short texts is carried out using methods based on probabilistic modeling of the co-presence of words, as well as on methods for identifying communities in bimodal networks and clustering pre-trained word embeddings. The resulting topics are assessed for coherence, diversity, and classification quality for the labeled data.

Social Media Listening as a Research Tool. Monitoring the media space to identify trends, analyze brand perceptions and other CX research.