Exploring Abstractive Text Summarisation for Podcasts: A Comparative Study of BART and T5 Models

Proceedings of Recent Advances in Natural Language Processing, pages 1023–1033 Varna, Sep 4–6, 2023 https://doi.org/10.26615/978-954-452-092-2_110 1023 Exploring Abstractive Text Summarisation for Podcasts: A Comparative Study of BART and T5 Models Parth Saxena Mahmoud El-Haj School of Computing and Communications Lancaster University Lancaster, United Kingdom parth.s.1909@gmail.com m.el-haj@lancaster.ac.uk Abstract Podcasts have become increasingly popular in recent years, resulting in a massive amount of audio content being produced every day. Effi- cient summarisation of podcast episodes can enable better content management and discov- ery for users. In this paper, we explore the use of abstractive text summarisation methods to generate high-quality summaries of podcast episodes. We use pre-trained models, BART and T5, to fine-tune on a dataset of Spotify’s 100K podcast. We evaluate our models using automated metrics and human evaluation, and find that the BART model fine-tuned on the pod- cast dataset achieved a higher ROUGE-1 and ROUGE-L score compared to other models, while the T5 model performed better in terms of semantic meaning. The human evaluation in- dicates that both models produced high-quality summaries that were well received by partici- pants. Our study demonstrates the effectiveness of abstractive summarisation methods for pod- cast episodes and offers insights for improving the summarisation of audio content. 1 Introduction Podcasts are a rapidly growing and popular medium for consuming knowledge and enter- tainment through spoken audio files that can be streamed or downloaded. With the podcast indus- try reaching new heights, there is a need for inno- vative and computationally effective methods for processing, analysing, and summarising podcast content. This need is further highlighted by the fact that Spotify1 acquired Anchor and Gimlet Media, two leading podcast companies, for $340 million in 2018 and has since invested approximately $500 million in this industry, demonstrating the growth and importance of podcasting (Sullivan, 2019). The abundance of data generated in various forms worldwide necessitates the need for meth- ods to compress and understand this data, enabling 1www.spotify.com the discovery of content available globally. One of these methods is automatic text summarisation, a technique used to shorten lengthy input texts to small coherent texts that convey the original meaning, reducing reading time and facilitating faster processing of research information. The field of Natural Language Processing (NLP) has wit- nessed significant progress in the development and research of text summarisation, resulting in an ac- tive research area in both Information Retrieval (IR) and NLP. The successful integration of trans- former architecture and attention mechanisms has also contributed significantly to the advancement of text summarisation techniques. The podcast domain presents unique challenges in text summarisation due to its conversational na- ture and informal language use, which could cause difficulties in extracting salient information. Fur- thermore, podcast episodes can be lengthy, making it a daunting task for listeners to find and iden- tify those of interest. Summarising podcast tran- scripts into short, readable text can alleviate these issues, allowing podcast listeners to make more informed decisions on which episodes to listen to and enabling easier retrieval of information (Tulley, 2011). …

Content truncated to 500 words