Multi-Level Matching and Aggregation Network for Few-Shot Relation Classification

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 2872–2881 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 2872 Multi-Level Matching and Aggregation Network for Few-Shot Relation Classification Zhi-Xiu Ye, Zhen-Hua Ling∗ National Engineering Laboratory for Speech and Language Information Processing, University of Science and Technology of China zxye@mail.ustc.edu.cn, zhling@ustc.edu.cn Abstract This paper presents a multi-level matching and aggregation network (MLMAN) for few-shot relation classification. Previous studies on this topic adopt prototypical networks, which cal- culate the embedding vector of a query in- stance and the prototype vector of each sup- port set independently. In contrast, our pro- posed MLMAN model encodes the query in- stance and each support set in an interactive way by considering their matching informa- tion at both local and instance levels. The fi- nal class prototype for each support set is ob- tained by attentive aggregation over the rep- resentations of its support instances, where the weights are calculated using the query in- stance. Experimental results demonstrate the effectiveness of our proposed methods, which achieve a new state-of-the-art performance on the FewRel dataset1. 1 Introduction Relation classification (RC) is a fundamental task in natural language processing (NLP), which aims to identify the semantic relation between two enti- ties in text. For example, the instance “[London]e1 is the capital of [the UK]e2” expresses the relation capital of between the two entities London and the UK. Some conventional relation classification meth- ods (Bethard and Martin, 2007; Zelenko et al., 2002) adopted supervised training and suffered from the lack of large-scale manually labeled data. To address this issue, the distant supervision method (Mintz et al., 2009) was proposed which annotated training data by heuristically aligning knowledge bases (KBs) and texts. However, the long-tail problem in KBs (Xiong et al., 2018; Han ∗Corresponding author: Zhen-Hua Ling. 1The code is available at https://github.com/ ZhixiuYe/MLMAN. Support Set class A: mother instance #1 The Queen Consort [Jetsun Pema]e2 gave birth to a son on 5 February 2016 , [Jigme Namgyel Wangchuck]e1. instance #2 He married the American actress [Cindy Robbins]e2 and was stepfather to her daughter , [Kim- berly Beck]e1. instance #3 Edgar married actress [Moyna Macgill]e2 and became the father of [Angela Lansbury]e1. instance #4 In 1845 , [Cemile Sultan]e1 ’s mother , Em- press [Dzdidil Kadn]e2, died. instance #5 Bo ’s wife [Gu Kailai]e2 traveled with their son [Bo Guagua]e1 to Britain. class B: member of ... class C: father ... class D: sport ... class E: voice type ... Query Instance He was married to [Eva Funck]e2 and they have a son [Gustav]e1 . Table 1: A data example of 5-way-5-shot relation clas- sification in FewRel development set. The correct re- lation class for the query instance is class A: mother. The instances for other relation classes are omitted for saving space. et al., 2018) still exists and makes it hard to clas- sify the relations with very few training samples. This paper focuses on the few-shot relation clas- sification …

Content truncated to 500 words