PROSODIC INDEX: ANALYSIS OF WORD PROMINENCE USING MULTI-HEAD ATTENTION AND BERT

The article highlights a problem of meaning formation in a text, where timbre, as a specific supersegmental colouring of speech, plays a crucial meaning-distinguishing role. In speech, it is timbre that helps the speaker to assign prominence to speech elements. Even though the role of prosodic cues in conveying sentiment was emphasized in early machine learning studies, it did not receive the attention it deserves. A recipient grasps hidden and non-literal meanings and understands the purport of the author through the identification of the prominence of speech elements. Using a multi-head attention mechanism and BERT enables us to calculate the prominence of words. The method that the researcher proposes in this paper is based on modeling the prosodic index, which is a quantitative measure of the prominence of a word that reflects its contribution to the prosodic and semantic aspects of the whole text. The results reveal that the method of prosodic index allows to assess the prominence of words in context more accurately than traditional text analysis methods, such as bag-of-words and TF-IDF.