CofeNet: Context and Former-Label Enhanced Net for Complicated Quotation Extraction

Proceedings of the 29th International Conference on Computational Linguistics, pages 2438–2449 October 12–17, 2022. 2438 CofeNet: Context and Former-Label Enhanced Net for Complicated Quotation Extraction Yequan Wang1∗, Xiang Li2∗, Aixin Sun3, Xuying Meng4, Huaming Liao4, Jiafeng Guo4 1Beijing Academy of Artificial Intelligence, Beijing, China 2Alibaba Group, China 3School of Computer Science and Engineering, Nanyang Technological University, Singapore 4Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China tshwangyequan@gmail.com, yuanye.lx@alibaba-inc.com, axsun@ntu.edu.sg, {mengxuying, lhm, guojiafeng}@ict.ac.cn Abstract Quotation extraction aims to extract quotations from written text. There are three components in a quotation: source refers to the holder of the quotation, cue is the trigger word(s), and content is the main body. Existing solutions for quotation extraction mainly utilize rule- based approaches and sequence labeling mod- els. While rule-based approaches often lead to low recalls, sequence labeling models cannot well handle quotations with complicated struc- tures. In this paper, we propose the Context and Former-Label Enhanced Net (CofeNet) for quotation extraction. CofeNet is able to extract complicated quotations with components of variable lengths and complicated structures. On two public datasets (i.e., PolNeAR and Riqua) and one proprietary dataset (i.e., PoliticsZH), we show that our CofeNet achieves state-of- the-art performance on complicated quotation extraction. 1 Introduction Quotation extraction aims to extract quotations from written text (Pouliquen et al., 2007). For example, given one instance shown in Fig- ure 1, we extract the quotation with source: some democrats, cue: privately express, and con- tent: reservations about .... As a point of view, quotations provide opinions of the speaker, which is important for analyzing the speaker’s stand. In general, quotation extraction is the first step before any further analysis, e.g., speaker stand detection. In this paper, we focus on the extraction of the three quotation components. As illustrated in the above example, the extrac- tion of content component in a quotation is com- plicated and difficult due to three reasons: variable length, unclear boundary, and indistinguishable *Indicates equal contribution Yet for all the symbolism and feel-good value of such an appointment, some democrats privately express reservations about entrusting a seat that could decide the balance of power in the closely divided senate to a candidate who has never won statewide, is considered less than dynamic and has been an anemic fundraiser. Figure 1: An example of quotations. Text spans with orange, green and gray denote source, cue and content respectively. components. Specifically, the length of content can be over 10, or even more than 50 tokens. More- over, content does not come with a regular pattern, which not only leads to a more unclear boundary of itself, but also affects the estimation of source and cue. For example, content in a quotation can be a complete instance with subject, predicate, and object. It is therefore hard to distinguish a noun (subject or object) representing the source or a part of content. Difficulty also exists in recognition of cue when tackling with a predicate, e.g., verb. Thus, as content may contain another quotation, such a nesting structure …

Content truncated to 500 words