Identifying Implicit Quotes for Unsupervised Extractive Summarization of Conversations

Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 291–302 December 4 - 7, 2020. c⃝2020 Association for Computational Linguistics 291 Identifying Implicit Quotes for Unsupervised Extractive Summarization of Conversations Ryuji Kano†‡ Yasuhide Miura† Tomoki Taniguchi† Tomoko Ohkuma† †Fuji Xerox Co., Ltd. ‡Institute of Innovative Research, Tokyo Institute of Technology {kano.ryuji, yasuhide.miura, taniguchi.tomoki, ohkuma.tomoko} @fujixerox.co.jp Abstract We propose Implicit Quote Extractor, an end- to-end unsupervised extractive neural summa- rization model for conversational texts. When we reply to posts, quotes are used to high- light important part of texts. We aim to ex- tract quoted sentences as summaries. Most replies do not explicitly include quotes, so it is difficult to use quotes as supervision. How- ever, even if it is not explicitly shown, replies always refer to certain parts of texts; we call them implicit quotes. Implicit Quote Extractor aims to extract implicit quotes as summaries. The training task of the model is to predict whether a reply candidate is a true reply to a post. For prediction, the model has to choose a few sentences from the post. To predict ac- curately, the model learns to extract sentences that replies frequently refer to. We evaluate our model on two email datasets and one so- cial media dataset, and confirm that our model is useful for extractive summarization. We fur- ther discuss two topics; one is whether quote extraction is an important factor for summa- rization, and the other is whether our model can capture salient sentences that conventional methods cannot. 1 Introduction As the amount of information exchanged via online conversations is growing rapidly, automated sum- marization of conversations is in demand. Neural- network-based models have achieved great perfor- mance on supervised summarization, but its appli- cation to unsupervised summarization is not suf- ficiently explored. Supervised summarization re- quires tens of thousands of human-annotated sum- maries. Because it is not realistic to prepare such large datasets for every domain, there is a growing requirement for unsupervised methods. Previous research proposed diverse methods of unsupervised summarization. Graph-centrality Figure 1: Example of a post and a reply with a quote and a reply with no quote. Implicit quote is the part of post that reply refers to, but not explicitly shown in the reply. based on the similarity of sentences (Mihalcea and Tarau, 2004; Erkan and Radev, 2004; Zheng and Lapata, 2019) has long been a strong feature for un- supervised summarization, and is also used to sum- marize conversations (Mehdad et al., 2014; Shang et al., 2018). Apart from centrality, centroid of vectors (Gholipour Ghalandari, 2017), Kullback- Leibler divergence (Haghighi and Vanderwende, 2009), reconstruction loss (He et al., 2012; Liu et al., 2015; Ma et al., 2016), and path scores of word graphs (Mehdad et al., 2014; Shang et al., 2018), are leveraged for summarization. The premise of these methods is that important topics appear frequently in a document. Therefore, if important topics appear …

Content truncated to 500 words