Keyphrase Prediction from Video Transcripts: New Dataset and Directions

Proceedings of the 29th International Conference on Computational Linguistics, pages 7147–7155 October 12–17, 2022. 7147 Keyphrase Prediction from Video Transcripts: New Dataset and Directions Amir Pouran Ben Veyseh1, Quan Hung Tran2, Seunghyun Yoon2, Varun Manjunatha2, Hanieh Deilamsalehy2, Rajiv Jain2, Trung Bui2, Walter W. Chang2, Franck Dernoncourt2, and Thien Huu Nguyen1 1Department of Computer Science, University of Oregon, OR, USA 2Adobe Research, USA {apouranb, thien}@cs.uoregon.edu {franck.dernoncourt}@adobe.com Abstract Keyphrase Prediction (KP) is an established NLP task, aiming to yield representative phrases to summarize the main content of a given document. Despite major progress in re- cent years, existing works on KP have mainly focused on formal texts such as scientific pa- pers or weblogs. The challenges of KP in informal-text domains are not yet fully studied. To this end, this work studies new challenges of KP in transcripts of videos, an understudied domain for KP that involves informal texts and non-cohesive presentation styles. A bottleneck for KP research in this domain involves the lack of high-quality and large-scale annotated data that hinders the development of advanced KP models. To address this issue, we introduce a large-scale manually-annotated KP dataset in the domain of live-stream video transcripts ob- tained by automatic speech recognition tools. Concretely, transcripts of 500+ hours of videos streamed on the behance.net platform are manually labeled with important keyphrases. Our analysis of the dataset reveals the challeng- ing nature of KP in transcripts. Moreover, for the first time in KP, we demonstrate the idea of improving KP for long documents (i.e., tran- scripts) by feeding models with paragraph-level keyphrases, i.e., hierarchical extraction. To fos- ter future research, we will publicly release the dataset and code. 1 Introduction Keyphrases are one or multiple consecutive words that could represent the main ideas in a document. Keyphrases are commonly categorized as Present or Absent. A present keyphrase explicitly appears in the document, while an absent keyphrase does not exist in the document. Keyphrases can serve as concise summary for a document, hence benefit- ing various NLP applications Information Retrieval (Hersh, 2021) and Text Summarization (Adhikari et al., 2020). Due to their usefulness, in the more than two decades, KP has been studied in many re- search works (Turney, 2000; Wu et al., 2005; Jiang et al., 2009; Hasan and Ng, 2014; Mahata et al., 2018; Chen et al., 2020; Ye et al., 2021). Whereas traditionally feature engineering has been used for KP (Turney, 2000; Sheeba and Vivekanandan, 2014), recently deep learning is proved to be more efficient for this task (Ye et al., 2021; Ahmad et al., 2021). However, one limitation in the current works is that they are mainly limited to the formal text such as scientific papers (Meng et al., 2017) and web-logs (Xiong et al., 2019). As such, the challenges in other domains are still un- resolved. Among others, video transcript is one of the less-explored domains that could significantly benefit from KP. For instance, it could be used for video summarization and retrieval or benefit peo- ple who are …

Content truncated to 500 words