Automatic Keyphrase Generation by Incorporating Dual Copy Mechanisms in Sequence-to-Sequence Learning

Proceedings of the 29th International Conference on Computational Linguistics, pages 2328–2338 October 12–17, 2022. 2328 Automatic Keyphrase Generation by Incorporating Dual Copy Mechanisms in Sequence-to-Sequence Learning Siyu Wang1, Jianhui Jiang1, Yao Huang1 and Yin Wang1,2 1Gusu Laboratory of Materials, Suzhou, China 2Hongzhiwei Technology (Shanghai) Co., Ltd, Shanghai, China {wangsiyu2022, jiangjianhui2021, huangyao2021}@gusulab.ac.cn wangyin@hzwtech.com Abstract The keyphrase generation task is a challenging work that aims to generate a set of keyphrases for a piece of text. Many previous studies based on the sequence-to-sequence model were used to generate keyphrases, and they introduce a copy mechanism to achieve good results. How- ever, we observed that most of the keyphrases are composed of some important words (seed words) in the source text, and if these words can be identified accurately and copied to create more keyphrases, the performance of the model might be improved. To address this challenge, we propose a DualCopyNet model, which in- troduces an additional sequence labeling layer for identifying seed words, and further copies the words for generating new keyphrases by dual copy mechanisms. Experimental results demonstrate that our model outperforms the baseline models and achieves an obvious per- formance improvement. 1 Introduction A keyphrase is a short piece of text that summa- rizes and abstracts the main semantics of a long text (named “document” or “source text” in this study). High-quality keyphrase promotes readers to efficiently understand, summarize and access documents’ content (Meng et al., 2017). Not only that, extracting high-quality keyphrase had been widely applied to many downstream tasks in Natu- ral Language Processing and Data Mining, such as Information Retrieval (Jones and Staveley, 1999), Text Summarization (Zhang et al., 2004), Text Cate- gorization (Hulth and Megyesi, 2006) and Opinion Expression Mining (Berend, 2011). Thus, how to automatically extract high-quality keyphrases has become popular research topics in recent decades (Augenstein et al., 2017; Kim et al., 2010). Due to accessibility of text data, many datasets from scien- tific articles are used as benchmarks for keyphrase extraction algorithms. Thus, our study also focuses Figure 1: An example of keyphrases. The colored words (seed words) in the source text appear repeatedly in keyphrases. on extracting keyphrases from the scientific arti- cles. Generally speaking, keyphrases can be divided into two categories: present keyphrase and ab- sent keyphrase. Present keyphrases are the ex- plicit words that appear directly in source text, and vice versa for absent keyphrase. Many previous studies have focused on how to extract present keyphrases from documents. These studies (Cam- pos et al., 2020; Hulth, 2003; Bougouin et al., 2014; Boudin, 2018; Bennani-Smires et al., 2018) con- sider the keyphrases extraction as a ranking task, which extracts a set of candidate phrases from the source text, and then selects keyphrases from the sorted candidates with the higher importance score. In recent years, some studies have also at- tempted to use deep learning methods for present keyphrase extraction. For example, Alzaidy et al. (2019) considered the present keyphrases extrac- tion as a sequence labeling task by using LSTM- CRF model …

Content truncated to 500 words