Semantic Role Labeling with Associated Memory Network

Proceedings of NAACL-HLT 2019, pages 3361–3371 Minneapolis, Minnesota, June 2 - June 7, 2019. c⃝2019 Association for Computational Linguistics 3361 Semantic Role Labeling with Associated Memory Network Chaoyu Guan†, Yuhao Cheng†, Hai Zhao†∗ †Department of Computer Science and Engineering, Shanghai Jiao Tong University †Key Laboratory of Shanghai Education Commission for Intelligent Interaction and Cognitive Engineering, Shanghai Jiao Tong University, Shanghai, China {351549709,cyh958859352}@sjtu.edu.cn zhaohai@cs.sjtu.edu.cn Abstract Semantic role labeling (SRL) is a task to rec- ognize all the predicate-argument pairs of a sentence, which has been in a performance improvement bottleneck after a series of lat- est works were presented. This paper pro- poses a novel syntax-agnostic SRL model enhanced by the proposed associated mem- ory network (AMN), which makes use of inter-sentence attention of label-known asso- ciated sentences as a kind of memory to fur- ther enhance dependency-based SRL. In de- tail, we use sentences and their labels from train dataset as an associated memory cue to help label the target sentence. Furthermore, we compare several associated sentences se- lecting strategies and label merging methods in AMN to find and utilize the label of associ- ated sentences while attending them. By lever- aging the attentive memory from known train- ing data, Our full model reaches state-of-the- art on CoNLL-2009 benchmark datasets for syntax-agnostic setting, showing a new effec- tive research line of SRL enhancement other than exploiting external resources such as well pre-trained language models. 1 Introduction Semantic role labeling (SRL) is a task to recog- nize all the predicate-argument pairs of a given sentence and its predicates. It is a shallow seman- tic parsing task, which has been widely used in a series of natural language processing (NLP) tasks, such as information extraction (Liu et al., 2016) and question answering (Abujabal et al., 2017). Generally, SRL is decomposed into four classi- fication subtasks in pipeline systems, consisting of ∗Corresponding author. This paper was partially sup- ported by National Key Research and Development Program of China (No. 2017YFB0304100), National Natural Science Foundation of China (No. U1836222 and No. 61733011) and Key Project of National Society Science Foundation of China (No. 15-ZDA041). predicate identification, predicate disambiguation, argument identification, and argument classifica- tion. In recent years, great attention (Zhou and Xu, 2015; Marcheggiani et al., 2017; He et al., 2017, 2018a,b) has been turned to deep learn- ing method, especially Long Short-term Memory (LSTM) network for learning with automatically extracted features. (Zhou and Xu, 2015) pro- posed the first end-to-end recurrent neural network (RNN) to solve the SRL task. (Marcheggiani et al., 2017) studied several predicate-specified embed- ding and decoding methods. (He et al., 2017) de- livered a full study on the influence of RNN train- ing and decoding strategies. Whether to use the syntactic information for SRL is also studied ac- tively (He et al., 2017, 2018b). Since the recent work of (Marcheggiani et al., 2017), which surprisingly shows syntax-agnostic dependency SRL for the first time can be rival of syntax-aware models, SRL has been more and more formulized into standard sequence …

Content truncated to 500 words