A Unified Architecture for Semantic Role Labeling and Relation Classification

Proceedings of COLING 2016, the 26th International Conference on Computational Linguistics: Technical Papers, pages 1264–1274, Osaka, Japan, December 11-17 2016. A Unified Architecture for Semantic Role Labeling and Relation Classification Jiang Guo�, Wanxiang Che�, Haifeng Wang♩, Ting Liu� and Jun Xu� �Center for Social Computing and Information Retrieval, Harbin Institute of Technology, China ♩Baidu Inc., China {jguo, car, tliu, jxu}@ir.hit.edu.cn wanghaifeng@baidu.com Abstract This paper describes a unified neural architecture for identifying and classifying multi-typed se- mantic relations between words in a sentence. We investigate two typical and well-studied tasks: semantic role labeling (SRL) which identifies the relations between predicates and arguments, and relation classification (RC) which focuses on the relation between two entities or nominals. While mostly studied separately in prior work, we show that the two tasks can be effectively connected and modeled using a general architecture. Experiments on CoNLL-2009 benchmark datasets show that our SRL models significantly outperform state-of-the-art approaches. Our RC models also yield competitive performance with the best published records. Furthermore, we show that the two tasks can be trained jointly with multi-task learning, resulting in additive significant improvements for SRL. 1 Introduction Semantic relation identification and classification are important problems towards the understanding of natural language sentences. Multi-typed semantic relations have been defined between two terms in a sentence in natural language processing (NLP) to promote various applications. For instance, the task of Semantic Role Labeling (SRL) defines shallow semantic dependencies between arguments and predicates, identifying the semantic roles, e.g., who did what to whom, where, when, and how. SRL has been a long-standing and challenging problem in NLP, primarily because it is strongly dependent on rich contextual and syntactical features used by the underlying classifiers (Gildea and Jurafsky, 2002). Another instance is Relation Classification (RC) which assigns sentences with two marked entities (or nominals) to a predefined set of relations (Hendrickx et al., 2010). Compared with SRL, relations defined in RC express much deeper semantics. Figure 1 shows example annotations of SRL and RC respectively. These two problems are typically studied separately in different communities. Hence the connections between them are neglected, both in data resources and approaches. In this paper, we show that SRL and RC have a lot of common ground and can be modeled with a unified model. We start by looking into the key features which have been proven dominant in both SRL and RC. • Contextual features. Words within a proper window size of the target words are important for most statistical models of various NLP tasks, such as Part-of-Speech tagging, Named Entity Recognition and Parsing. They are also important for identifying the semantic relatedness between two terms in a sentence. Consider the RC example in Figure 1(b), the context word “moved” is a strong indicator for classifying the relation of ⟨People, downtown⟩ as Entity-Destination. However, most of the conventional approaches in SRL and RC only considers local context features through feature engineering, which might be incomplete. • Syntactical features. Both state-of-the-art SRL and RC systems employ the syntactic …

Content truncated to 500 words