Emotion Recognition and Sentiment Analysis of Extemporaneous Speech Transcriptions in Russian

Speech can be characterized by acoustical properties and semantic meaning, represented as textual speech transcriptions. Apart from the meaning content, textual information carries a substantial amount of paralinguistic information that makes it possible to detect speaker’s emotions and sentiments by means of speech transcription analysis. In this paper, we present experimental framework and results for 3-way sentiment analysis (positive, negative, and neutral) and 4-way emotion classification (happy, angry, sad, and neutral) from textual speech transcriptions in terms of Unweighted Average Recall (UAR), reaching 91.93% and 88.99%, respectively, on the multimodal corpus RAMAS containing recordings of Russian improvisational speech. Orthographic transcriptions of speech recordings from the database are obtained using available pre-trained speech recognition systems. Text vectorization is implemented using Bag-of-Words, Word2Vec, FastText and BERT methods. Investigated machine classifiers include Support Vector Machine, Random Forest, Naive Bayes and Logistic Regression. To the best of our knowledge, this is the first study of sentiment analysis and emotion recognition for both extemporaneous Russian speech and RAMAS data in particular, therefore experimental results presented in this paper can be considered as a baseline for further experiments.