Cosine similarity measure to process the unfinished sentences (using the example of studying the image of a patriot)

In the context of the intensive development of natural language processing methods, the question arises about the integration of innovative technologies into the work processes of sociologists. Social scientists often face the need to process text data obtained both as part of their own research projects and on the Internet. Obviously, using available online sources as a database places increased demands on the techniques and procedures for processing a huge corpus of documents, often exceeding several hundred thousand lines. However, it is not beyond the scope of attention to work with the materials of author’s sociological research of a much smaller volume, which often require significant labor and time resources if they are processed manually. In this case, the consistency of collective coding and the role of the subjective opinion of experts in the generalization or grouping of data raises questions. The purpose of the article is to show the possibilities and limitations of using the cosine similarity measure to process the results of the unfinished sentences method. The empirical basis of the study was the materials obtained during the study of the image of a patriot in one of the Moscow universities in March 2025. A total of 70 students participated in the study. The article processed responses to a stimulus sentence, which the respondents had to complete in their own words: “A patriot always...” The results of calculating the cosine similarity measure have shown that this metric can be a useful tool in the initial search for statements that are similar in content. In case of doubt and the need to verify their conclusions or solve the problem of consistency of collective coding, the use of a measure of semantic proximity can act as a significant additional quantitative indicator to determine the thematic focus of each respondent’s utterance. Thus, using the cosine similarity assessment, it is possible to group the texts that are closest in semantic load, thereby bringing closer to understanding the general structure of the studied image and the thesaurus of the study participants. In conclusion, a conclusion is drawn about the modern requirements for the training of specialists in the social and humanitarian fields, which raises new methodological questions and opens up discussions about the optimal integration of technological advances in natural language processing into the analytical practices of social scientists and researchers.

unfinished sentence method, semantic similarity, cosine similarity, BERT language model, patriot image