Speaker Salience Based Tagging for Dialogues

Tags or keywords provide an efficient way to manage and retrieve large scale data. This paper proposes an unsupervised method to suggest informative tags for multi-party dialogues by integrating dialogue characteristics. Our model first extracts keywords from dialogue texts under a speaker salience based framework. Then we get keyword bigrams through frequent pattern matching. In order to generate more flexible and meaningful tags, we expand keywords and their bigrams by tag association rules mined from a popular bookmarking web del.icio.us. Finally we rank the three types of tag candidates under a uniform metric. Experimental results validate the effectiveness and the versatility of our method when compared with several strong baseline models like TextRank, TFIDF rank and KNN.