Topic-Aware Pointer-Generator Networks for Summarizing Spoken Conversations

Abstract

Due to the lack of publicly available resources, conversation summarization has received far less attention than text summarization. As the purpose of conversations is to exchange information between at least two interlocutors, key information about a certain topic is often scattered and spanned across multiple utterances and turns from different speakers. This phenomenon is more pronounced during spoken conversations, where speech characteristics such as backchanneling and false-starts might interrupt the topical flow. Moreover, topic diffusion and (intra-utterance) topic drift are also more common in human-to-human conversations. Such linguistic characteristics of dialogue topics make sentence-level extractive summarization approaches used in spoken documents illsuited for summarizing conversations. Pointer-generator networks have effectively demonstrated its strength at integrating extractive and abstractive capabilities through neural modeling in text summarization. To the best of our knowledge, to date no one has adopted it for summarizing conversations. In this work, we propose a topic-aware architecture to exploit the inherent hierarchical structure in conversations to further adapt the pointer-generator model. Our approach significantly outperforms competitive baselines, achieves more efficient learning outcomes, and attains more robust performance.

Multi-View Sequence-to-Sequence Models with Conversational Structure for Abstractive Dialogue Summarization

“…Zhao et al (2019); Zhu et al (2020b) introduced turn-based hierarchical models that encoded each turn of utterance first and then used the aggregated representation to generate summaries. A few studies have also paid attention to utilizing conversational analysis for generating dialogue summaries, such as leveraging dialogue acts (Goo and Chen, 2018), key point sequence (Liu et al, 2019a) or topics (Liu et al, 2019b;. However, they either needed a large amount of human annotation for dialogue acts, key points or visual focus (Goo and Chen, 2018;Liu et al, 2019a;, or only utilized topical information in conversations Liu et al, 2019b).…”

“…As a way of using language socially of "doing things with words" together with other persons, the conversation has its own dynamic structures that organize utterances in certain orders to make the conversation meaningful, enjoyable, and understandable (Sacks et al, 1978). Although there are a few exceptions such as utilizing topic segmentation (Liu et al, 2019b;, dialogue acts (Goo and Chen, 2018) or key point sequence (Liu et al, 2019a) (Gliwa et al, 2019) with its topic view and stage view (extracted by our methods), and the human annotated summary. extensive expert annotations of discourse acts (Goo and Chen, 2018;Liu et al, 2019a), or only encode conversations based on their topics (Liu et al, 2019b), which fails to capture rich conversation structures in dialogues.…”

“…Even one single conversation can be viewed from different perspectives, resulting in multiple conversational or discourse patterns. For instance, in Table 1, based on what topics were discussed (topic view) (Galley et al, 2003;Liu et al, 2019b;, it can be segmented into greetings, today's plan, plan for tomorrow, plan for Saturday and pick up time; from a conversation progression perspective (stage view) (Ritter et al, 2010;Paul, 2012;Althoff et al, 2016), the same dialogue can be categorized into openings, intention, discussion, and conclusion. From a coarse perspective (global view), conversations can be treated as a whole, or each utterance can serve as one segment (discrete view).…”

“…From a coarse perspective (global view), conversations can be treated as a whole, or each utterance can serve as one segment (discrete view). Models that only utilized a fixed topic view of the conversation (Joty et al, 2010;Liu et al, 2019b) may fail to capture its comprehensive and nuanced conversational structures, and any amount of information loss introduced by the conversation encoder may lead to larger error cascade in the decoding stage. To fill these gaps, we propose to combine those multiple, diverse views of conversations in order to generate more precise summaries.…”

MedFilter: Improving Extraction of Task-relevant Utterances through Integration of Discourse Structure and Ontological Knowledge

“…Dialogue Summarization: In addition to the challenges noted earlier in the paper, other linguistic phenomena such as backchannels, false starts, and topic diffusion are prominent in human-tohuman conversations. They add noise, which challenges the capabilities of otherwise effective sumarization approaches such as pointer-generator networks (See et al, 2017;Liu et al, 2019b).…”