Neural Maximum Subgraph Parsing for Cross-Domain Semantic Dependency Analysis

Proceedings of the 22nd Conference on Computational Natural Language Learning (CoNLL 2018), pages 562–572 Brussels, Belgium, October 31 - November 1, 2018. c⃝2018 Association for Computational Linguistics 562 Neural Maximum Subgraph Parsing for Cross-Domain Semantic Dependency Analysis Yufei Chen♠, Sheng Huang♠, Fang Wang♠, Weiwei Sun♠♥ and Xiaojun Wan♠ ♠ Institute of Computer Science and Technology, Peking University ♠ The MOE Key Laboratory of Computational Linguistics, Peking University ♥ Center for Chinese Linguistics, Peking University {yufei.chen,huangsheng,foundwang,ws,wanxiaojun}@pku.edu.cn Abstract We present experiments for cross-domain se- mantic dependency analysis with a neural Maximum Subgraph parser. Our parser targets 1-endpoint-crossing, pagenumber-2 graphs which are a good fit to semantic dependency graphs, and utilizes an efficient dynamic pro- gramming algorithm for decoding. For dis- ambiguation, the parser associates words with BiLSTM vectors and utilizes these vectors to assign scores to candidate dependencies. We conduct experiments on the data sets from Se- mEval 2015 as well as Chinese CCGBank. Our parser achieves very competitive results for both English and Chinese. To improve the parsing performance on cross-domain texts, we propose a data-oriented method to ex- plore the linguistic generality encoded in En- glish Resource Grammar, which is a precision- oriented, hand-crafted HPSG grammar, in an implicit way. Experiments demonstrate the ef- fectiveness of our data-oriented method across a wide range of conditions. 1 Introduction Semantic Dependency Parsing (SDP) is defined as the task of recovering sentence-internal bilex- ical semantic dependency structures, which en- code predicate–argument relationships for all con- tent words. Such sentence-level semantic analy- sis of text is concerned with the characterization of events and is therefore important to understand the essential meaning of a natural language sen- tence. With the advent of many supporting re- sources, SDP has become a well-defined task with a substantial body of work and comparative eval- uation. (Almeida and Martins, 2015; Du et al., 2015a; Zhang et al., 2016; Peng et al., 2017; Wang et al., 2018). Two SDP shared tasks have been run as part of the 2014 and 2015 International Work- shops on Semantic Evaluation (SemEval) (Oepen et al., 2014, 2015). There are two key dimensions of the data-driven dependency parsing approach: decoding and dis- ambiguation. Existing decoding approaches to syntactic or semantic analysis into bilexical de- pendencies can be categorized into two domi- nant types: transition-based (Zhang et al., 2016; Wang et al., 2018) and graph-based, i.e., Max- imum Subgraph (Kuhlmann and Jonsson, 2015; Cao et al., 2017a) approaches. For disambigua- tion, while early work on dependency parsing fo- cused on global linear models, e.g., structured per- ceptron (Collins, 2002), recent work shows that deep learning techniques, e.g., LSTM (Hochre- iter and Schmidhuber, 1997), is able to signif- icantly advance the state-of-the-art of the pars- ing accuracy. From the above two perspectives, i.e., the decoding and disambiguation frameworks, we find that what is still underexploited is neu- ral Maximum Subgraph parsing for highly con- strained graph classes, e.g., noncrossing graphs. In this paper, we fill this gap in the literature by developing a neural Maximum Subgraph parser. …

Content truncated to 500 words