Element Intervention for Open Relation Extraction

Element Intervention for Open Relation Extraction Fangchao Liu1,3, Lingyong Yan1,3, Hongyu Lin1,∗, Xianpei Han1,2,∗, Le Sun1,2 1Chinese Information Processing Laboratory 2State Key Laboratory of Computer Science Institute of Software, Chinese Academy of Sciences, Beijing, China 3University of Chinese Academy of Sciences, Beijing, China {fangchao2017,lingyong2014,hongyu,xianpei,sunle}@iscas.ac.cn Abstract Open relation extraction aims to cluster rela- tion instances referring to the same underlying relation, which is a critical step for general re- lation extraction. Current OpenRE models are commonly trained on the datasets generated from distant supervision, which often results in instability and makes the model easily col- lapsed. In this paper, we revisit the procedure of OpenRE from a causal view. By formu- lating OpenRE using a structural causal mod- el, we identify that the above-mentioned prob- lems stem from the spurious correlations from entities and context to the relation type. To address this issue, we conduct Element Inter- vention, which intervenes on the context and entities respectively to obtain the underlying causal effects of them. We also provide two specific implementations of the interventions based on entity ranking and context contrast- ing. Experimental results on unsupervised re- lation extraction datasets show that our meth- ods outperform previous state-of-the-art meth- ods and are robust across different datasets1. 1 Introduction Relation extraction (RE) is the task to extract re- lation between entity pair in plain text. For ex- ample, when given the entity pair (Obama, the United States) in the sentence “Obama was sworn in as the 44th president of the United States”, an RE model should accurately predict the relation- ship “President of” and extract the corresponding triplet (Obama, President of, the United States) for downstream tasks. Despite the success of many RE models (Zeng et al., 2014; Baldini Soares et al., 2019), most previous RE paradigms rely on the pre- defined relation types, which are always unavail- able in open domain scenario and thereby limits their capability in real applications. ∗Corresponding authors. 1Code available at https://github.com/Lfc1993/EI ORE P E C X P E C X (a) (b) (c) 𝑑𝑜(𝐸) P Relation Prototype E Entity pair C Context X Relation instance Y Y P E C X Y 𝑑𝑜(𝐶) Figure 1: The Structural Causal Model demonstrates the procedure of OpenRE. (a) is the original SCM; (b) Entity intervention that fixes the entity pair and adjusts different contexts; (c) Context intervention that fixes the context and adjusts different entity pairs. Open Relation Extraction (OpenRE), on the oth- er hand, has been proposed to extract relation facts without pre-defined relation types neither annotat- ed data. Given a relation instance consisting of two entities and their context, OpenRE aims to identify other instances which mention the same relation. To achieve this, OpenRE is commonly formulated as a clustering or pair-matching task. Therefore the most critical challenge for OpenRE is how to learn effective representations for relation instances and then cluster them. To this end, Yao et al. (2011) adopts topic model (Blei et al., 2003) to generate latent relation type for unlabelled instances. Lat- er works …

Content truncated to 500 words