BERT-Based Similarity Measures Oriented Approach for Style Change Detection

Style change detection is the task of identifying the writing style changes in a text. In general, the documents contain paragraphs of text and each paragraph contains sentences. The possibility of writing style changes occurs in a document at paragraph level or sentence level. At paragraph level, two consecutive paragraphs are written by two different authors, whereas at sentence level, two consecutive sentences are written by two different authors. In general, authors are following the same style throughout a document while writing the text in a document. According to this, if two sentences or paragraphs are written by two different authors, it means the similarity between these two sentences or paragraphs is very less. In this chapter, we propose an approach by using BERT architecture and similarity measures. The BERT model is used for representing the sentences or paragraphs as vectors. The similarities among these vectors are computed by using similarity measures. Different similarity measures are used in this experiment, and the performance of these measures for style change detection are compared. The experiment was performed on the dataset provided in the PAN competition 2022 task of style change detection. In this task, three different tasks are introduced based on the writing style changes at sentence level or at paragraph level. The accuracy measure is used for presenting the results of three tasks of style change detection. The proposed method obtained the best accuracies for three tasks of style change detection when compared with various approaches of style change detection.