Unsupervised Open-domain Keyphrase Generation

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 10614–10627 July 9-14, 2023 ©2023 Association for Computational Linguistics Unsupervised Open-domain Keyphrase Generation Lam Thanh Do♠∗ Pritom Saha Akash♡ Kevin Chen-Chuan Chang♡∗ ♠Hanoi University of Science and Technology, Viet Nam ♡University of Illinois at Urbana-Champaign, USA lam.dt183573@sis.hust.edu.vn {pakash2, kcchang}@illinois.edu Abstract In this work, we study the problem of unsu- pervised open-domain keyphrase generation, where the objective is a keyphrase genera- tion model that can be built without using human-labeled data and can perform consis- tently across domains. To solve this problem, we propose a seq2seq model that consists of two modules, namely phraseness and informa- tiveness module, both of which can be built in an unsupervised and open-domain fashion. The phraseness module generates phrases, while the informativeness module guides the generation towards those that represent the core concepts of the text. We thoroughly evaluate our pro- posed method using eight benchmark datasets from different domains. Results on in-domain datasets show that our approach achieves state- of-the-art results compared with existing unsu- pervised models, and overall narrows the gap between supervised and unsupervised meth- ods down to about 16%. Furthermore, we demonstrate that our model performs consis- tently across domains, as it overall surpasses the baselines on out-of-domain datasets.1. 1 Introduction Keyphrases are short word sequences that describe the core concepts of the text. The prediction of keyphrases for a text is a task that has received much attention recently. It is a crucial prob- lem as its outputs can be useful for a variety of downstream tasks such as building digital libraries (Gutwin et al., 1999; Witten et al., 2009), document summarization (Litvak and Last, 2008), document visualization (Chuang et al., 2012) and so on. There are mainly two approaches to keyphrase prediction, namely keyphrase extraction (Mihal- cea and Tarau, 2004; Florescu and Caragea, 2017a; Bennani-Smires et al., 2018) and keyphrase gener- ation (Meng et al., 2017; Chen et al., 2019; Yuan ∗Work done while visiting Cazoodle Inc. 1Code and data will be available at https://github. com/ForwardDataLab/UOKG. et al., 2020; Shen et al., 2022). Keyphrase ex- traction highlights keyphrases that appear within the text. On the other hand, keyphrase generation generates keyphrases based on the understanding of the given text and therefore allows predicting absent keyphrases alongside present ones (Meng et al., 2017). This ability has made keyphrase gen- eration receive more attention than keyphrase ex- traction in recent years, as human also tend to use keyphrases that are absent from the text. Most of the existing keyphrase generation mod- els use manually labeled data for training (Meng et al., 2017; Chen et al., 2018, 2019; Yuan et al., 2020; Ahmad et al., 2021). However, obtaining labeled data is often the most expensive component of any machine learning model, and this is the same for keyphrase generation. Compared to labeled data, access to unlabeled data is easier and mostly available. For example, the arXiv dataset (Clement et al., 2019) containing metadata (e.g., title, …

Content truncated to 500 words