Element graph-augmented abstractive summarization for legal public opinion news with graph transformer

Abstract

Automatic summarization for legal public opinion news has been an attractive research problem in recent years. Compared with the open-domain, summarization for legal public opinion news has two essential constraints: (1) the key information (e.g., the case elements) of the news should be summarized; (2) the factual errors should be avoided in the generated summary. To address these challenges, the summarizer should learn a structured representation of the news (event plan), making it better to understand the event information implied in the news. This paper proposes a novel element graph-augmented abstractive summarization model, which first constructs the structural graph by extracting elements from the source document and then produces graph representation via graph transformer network. Finally, the structural representation is taken as an essential complementary component of the conventional sequence-to-sequence model to guide the decoding process simultaneously. Furthermore, the pre-trained language model is introduced to enhance the sequential and structural encoder, which further promotes the summarization model’s performance. For evaluation, we build a large-scale legal public opinion news (LPO-news) corpus. Experimental results on LPO-news and another news-oriented CNN/Daily mail dataset show that our model significantly outperforms other baselines in terms of both ROUGE scores and Bert scores. We also perform a human evaluation to demonstrate our model’s effectiveness by evaluating the generated summary using several subjective metrics.