FASTMATCH: Accelerating the Inference of BERT-based Text Matching

Proceedings of the 28th International Conference on Computational Linguistics, pages 649–656 Barcelona, Spain (Online), December 8-13, 2020 649 A Joint Learning Approach based on Self-Distillation for Keyphrase Extraction from Scientific Documents Tuan Manh Lai 1, Trung Bui 2, Doo Soon Kim 2, Quan Hung Tran 2 1 University of Illinois at Urbana-Champaign 2 Adobe Research, San Jose, CA Abstract Keyphrase extraction is the task of extracting a small set of phrases that best describe a docu- ment. Most existing benchmark datasets for the task typically have limited numbers of annotated documents, making it challenging to train increasingly complex neural networks. In contrast, digital libraries store millions of scientific articles online, covering a wide range of topics. While a significant portion of these articles contain keyphrases provided by their authors, most other articles lack such kind of annotations. Therefore, to effectively utilize these large amounts of unlabeled articles, we propose a simple and efficient joint learning approach based on the idea of self-distillation. Experimental results show that our approach consistently improves the per- formance of baseline models for keyphrase extraction. Furthermore, our best models outperform previous methods for the task, achieving new state-of-the-art results on two public benchmarks: Inspec and SemEval-2017. 1 Introduction Keyphrase extraction is the task of automatically extracting a set of representative phrases from a doc- ument that concisely describe its content. As keyphrases provide a brief yet precise description of a document, they can be utilized for various downstream applications (D’Avanzo and Magnini, 2005; Litvak and Last, 2008; Kim et al., 2010; Boudin and Morin, 2013). Over the past years, researchers have proposed many methods for the task, which can be divided into two major categories: supervised (Sterckx et al., 2016; Zhang et al., 2017; Alzaidy et al., 2019) and unsupervised techniques (Florescu and Caragea, 2017b; Boudin, 2018; Mahata et al., 2018). In the presence of sufficient domain-specific labeled data, supervised keyphrase extraction methods are often reported to outperform unsupervised methods (Kim et al., 2013; Caragea et al., 2014; Sahrawat et al., 2020). Recently, many deep learning based methods have achieved promising performance in a wide range of NLP tasks (Lai et al., 2018b; Tran et al., 2018; Peters et al., 2018; Devlin et al., 2019; Lewis et al., 2019; Lai et al., 2020). However, most existing benchmark datasets for keyphrase extraction typically have limited numbers of annotated documents, making it challenging to train an effective deep learning model for the task. In contrast, digital libraries store millions of scientific articles online, covering a wide range of topics. While a significant portion of these articles have author-provided keyphrases, most other articles lack such kind of annotations. For example, major NLP conferences (i.e., ACL, EMNLP, COLING) normally do not require authors to provide keywords of their publications. In this paper, to effectively utilize these large amounts of unlabeled articles available online, we propose a novel joint learning approach based on the idea of self-distillation (Furlanello et al., 2018). To evaluate the effectiveness of our method, we use the …

Content truncated to 500 words