Modeling Multi-Granularity Hierarchical Features for Relation Extraction

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5088 - 5098 July 10-15, 2022 ©2022 Association for Computational Linguistics Modeling Multi-Granularity Hierarchical Features for Relation Extraction Xinnian Liang1∗, Shuangzhi Wu2, Mu Li2 and Zhoujun Li1† 1State Key Lab of Software Development Environment, Beihang University, Beijing, China 2Tencent Cloud Xiaowei, Beijing, China {xnliang,lizj}@buaa.edu.cn; frostwu@tencent.com,limugx@qq.com; Abstract Relation extraction is a key task in Natural Language Processing (NLP), which aims to ex- tract relations between entity pairs from given texts. Recently, relation extraction (RE) has achieved remarkable progress with the devel- opment of deep neural networks. Most exist- ing research focuses on constructing explicit structured features using external knowledge such as knowledge graph and dependency tree. In this paper, we propose a novel method to extract multi-granularity features based solely on the original input sentences. We show that effective structured features can be attained even without external knowledge. Three kinds of features based on the input sentences are fully exploited, which are in entity mention level, segment level, and sentence level. All the three are jointly and hierarchically modeled. We evaluate our method on three public bench- marks: SemEval 2010 Task 8, Tacred, and Ta- cred Revisited. To verify the effectiveness, we apply our method to different encoders such as LSTM and BERT. Experimental results show that our method significantly outperforms ex- isting state-of-the-art models that even use ex- ternal knowledge. Extensive analyses demon- strate that the performance of our model is con- tributed by the capture of multi-granularity fea- tures and the model of their hierarchical struc- ture. Code and data are available at https: //github.com/xnliang98/sms. 1 Introduction Relation extraction (RE) is a fundamental task in Natural Language Processing (NLP), which aims to extract relations between entity pairs from given plain texts. RE is the cornerstone of many down- stream NLP tasks, such as knowledge base con- struction (Ji and Grishman, 2011), question answer- ing (Yu et al., 2017), and information extraction (Fader et al., 2011). *Contribution during internship at Tencent Inc. †Corresponding Author Most recent works focus on constructing explicit structured features using external knowledge such as knowledge graph, entity features and depen- dency tree. To infuse prior knowledge from ex- isting knowledge graph, recent works (Peters et al., 2019a; Wang et al., 2020b,a) proposed some pre- train tasks to help model learn and select proper prior knowledge in the pre-training stage. Bal- dini Soares et al. (2019); Yamada et al. (2020); Peng et al. (2020) force model learning entitiy- related information via well-designed pre-train tasks. Zhang et al. (2018); Guo et al. (2019); Xue et al. (2020); Chen et al. (2020) encode dependency tree with graph neural network (Kipf and Welling, 2017) (GNN) to help RE models capture non-local syntactic relation. All of them achieve a remarkable performance via employing external information from different structured features. However, they either need time-consuming pre- training with external knowledge or need an ex- ternal tool to get a dependency tree which …

Content truncated to 500 words