Abstract

Meeting summarization is challenging due to complex multi-party interactions with spontaneous utterances. This paper tries to mitigate this challenge by leveraging dialogue acts (DAs). We propose Dialogue Acts enhanced extract–abstract framework for meeting Summarization (DASum). We jointly train meeting summarization with DAs classification to enable our model get aware of DAs and better understand dialogue interactions. Moreover, our DASum model uses DAs to measure the salience of an utterance in the extractor and as prefix prompt to guide how to integrate extracted utterances in the abstractor. We conducted experiments on two public datasets, ICSI (for meeting summarization) and QMSum (for query-based meeting summarization), which compress the meeting transcripts (average 9000 tokens) into a general summary (no more than 700 tokens) or a specific summary (no more than 150 tokens) for a query. Experimental results show our DASum achieves promising results on the ICSI (46.41/ 11.52/ 43.82/ 5.86) and QMSum (36.37/ 11.71/ 32.07/ 19.18) datasets in terms of ROUGE-1/ ROUGE-2/ ROUGE-L/ BERTScore.

Full Text
Published version (Free)

Talk to us

Join us for a 30 min session where you can share your feedback and ask us any queries you have

Schedule a call