Semantic Sentence Matching via Interacting Syntax Graphs

Proceedings of the 29th International Conference on Computational Linguistics, pages 938–949 October 12–17, 2022. 938 Semantic Sentence Matching via Interacting Syntax Graphs Chen Xu1, Jun Xu1,2,∗, Zhenghua Dong3, Ji-Rong Wen1,2 1Gaoling School of Artificial Intelligence, Renmin University of China 2Beijing Key Laboratory of Big Data Management and Analysis Methods 3Huawei Noah’s Ark Lab {xc_chen, junxu, jrwen}@ruc.edu.cn dongzhenhua@huawei.com Abstract Studies have shown that the sentence’s syntac- tic structures are important for semantic sen- tence matching. A typical approach is encoding each sentence’s syntactic structure into an em- bedding vector, which can be combined with other features to predict the final matching scores. Though successes have been observed, embedding the whole syntactic structures as one vector inevitably overlooks the fine-grained syntax matching patterns, e.g. the alignment of specific term dependencies relations in the two inputted sentences. In this paper, we for- malize the task of semantic sentence matching as a problem of graph matching in which each sentence is represented as a directed graph ac- cording to its syntactic structures. The syntax matching patterns (i.e. similar syntactic struc- tures) between two sentences, therefore, can be extracted as the sub-graph structure align- ments. The proposed method, referred to as Interacted Syntax Graphs (ISG), represents two sentences’ syntactic alignments as well as their semantic matching signals into one association graph. After that, the neural quadratic assign- ment programming (QAP) is adapted to extract syntactic matching patterns from the associa- tion graph. In this way, the syntactic struc- tures fully interact in a fine granularity during the matching process. Experimental results on three public datasets demonstrated that ISG can outperform the state-of-the-art baselines effec- tively and efficiently. The empirical analysis also showed that ISG can match sentences in an interpretable way. 1 Introduction Matching two natural language sentences has be- come a fundamental technique in information re- trieval (IR) and natural language processing (NLP). Extensive research efforts have been devoted to the task (Li and Xu, 2014; Xu et al., 2020). Recently, ∗ Corresponding author playing VBG some DT are VBP men NNS a DT sport NN Sentence X Sentence Y similar with IN a DT game NN soccer NN playing VBG multiple JJ males NNS det nsubj aux obj det det compound nsubj obj amod case subgraph similar Figure 1: A semantically similar sentence pair (X = “Some men are playing a sport”, Y =“A soccer game with multiple males playing”) from SNLI. The words and POS tags are shown in the nodes, and syntactic dependencies are shown as edges. researchers found that the sentences’ rich syntac- tic information helps match. Some studies utilize the implicit syntax-encoding methods to learn the sentence embedding based on its syntactic stric- tures (Chen et al., 2016, 2017). Other studies di- rectly utilize the different syntactic tags as syntactic features and fuse them with word features (Mou et al., 2016; Chen et al., 2018; Liu et al., 2020). All these approaches separately represent the syn- tactic information of the two input sentences as two coarse-grained embedding vectors, ignoring …

Content truncated to 500 words