Towards Abstractive Grounded Summarization of Podcast Transcripts

Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 4407 - 4418 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Towards Abstractive Grounded Summarization of Podcast Transcripts Kaiqiang Song,†‡ Chen Li,‡ Xiaoyang Wang,‡ Dong Yu,‡ Fei Liu† ‡Tencent AI Lab, Seattle, WA †University of Central Florida, Orlando, FL {riversong,ailabchenli,shawnxywang,dyu}@tencent.com feiliu@cs.ucf.edu Abstract Podcasts have shown a recent rise in popularity. Summarization of podcasts is of practical bene- fit to both content providers and consumers. It helps people quickly decide whether they will listen to a podcast and/or reduces the cognitive load of content providers to write summaries. Nevertheless, podcast summarization faces sig- nificant challenges including factual inconsis- tencies of summaries with respect to the inputs. The problem is exacerbated by speech disfluen- cies and recognition errors in transcripts of spo- ken language. In this paper, we explore a novel abstractive summarization method to alleviate these issues. Our approach learns to produce an abstractive summary while grounding sum- mary segments in specific regions of the tran- script to allow for full inspection of summary details. We conduct a series of analyses of the proposed approach on a large podcast dataset and show that the approach can achieve promis- ing results. Grounded summaries bring clear benefits in locating the summary and transcript segments that contain inconsistent information, and hence improve summarization quality in terms of automatic and human evaluation. 1 Introduction Podcasts are one of the most popular forms of new media. As of today, over 155 million people listen to a podcast every week (Christian, 2021). With the growing interest, there is an increased demand for textual summaries that foretell the content of podcasts. Those summaries help people decide, in a few seconds, if they will listen to a podcast or subscribe to the channel. They are helpful for users who want to find podcasts previously listened to. Furthermore, they can be re-purposed for social me- dia posts or email marketing campaigns, enabling content creators to make their podcasts accessible to a larger audience. It is desirable to generate grounded summaries from podcast transcripts, where spans of summary text are closely tethered to the original audio. Fig- ure 1 provides an example of a grounded abstrac- tive summary. When a user clicks on a summary segment, she will be directed to an audio clip that gives further detail of the conversational context. Grounded summaries give us a preview of notable podcast clips (Shalom, 2019) and they may further release summarization service providers from po- tential legal claims by directing users to the original audio. This is because, speech recognizers induce transcription errors and abstractive summarization models may hallucinate facts that are not entailed by the original (Kryscinski et al., 2020), both can cause podcast summaries to contain misleading or inaccurate information. With grounded summaries, users are able to frame, interpret, and place into context any system-generated summaries, thus re- ducing the barriers to deploy podcast summariza- tion technology. One may attempt to align summary …

Content truncated to 500 words