Style-sensitive Sentence Embeddings for Evaluating Similarity in Speech Style of Japanese Sentences by Contrastive Learning

Since dialogue systems are required to keep its speech style consistency, evaluating the similarity of speech styles is an important task. However, the Japanese language has a wide variety of speech styles, and also for each speech style, huge variety of vocabulary and word usage characteristics are observed, making it difficult to evaluate the similarity of speech styles. This study proposes a speech style embedding model that produces a style-sensitive sentence embedding of Japanese sentences. The speech style embedding model is constructed by fine-tuning a pre-trained BERT model. Here, sentence pairs with similar/dissimilar speech styles are automatically collected on a large scale using a sequence of sentences in web novels, with which con-strastive learning is performed for the training of the speech style embedding model. Using the Ward’s hierarchical clustering method, we also analyze the clusters of speech styles and the characteristic vocabulary/word usage of each speech style. Finally, we focus on the variation in speech styles of each person depending on the situation, and further analyze the variation in style-sensitive embeddings of each character in the novel.