Personalized Extractive Summarization with Discourse Structure Constraints Towards Efficient and Coherent Dialog-Based News Delivery

In this paper, we propose a method to generate a personalized summary that may be of interest to each user based on the discourse structure of documents in order to deliver a certain amount of coherent and interesting information within a limited time, primarily via a spoken dialog form. We initially constructed a news article corpus with annotations of the discourse structure, users’ profiles, and interests in sentences and topics. The proposed summarization model solves an integer linear programming problem with the discourse structure of each document and the total utterance time as constraints and extracts sentences that maximize the sum of the estimated degree of user’s interest. The degree of interest in a sentence is estimated based on the user’s profile obtained from a questionnaire and the word embeddings of BERT. Experiments confirm that the personalized summaries generated by the proposed method transmit information more efficiently than generic summaries generated based solely on the importance of sentences.