Neural Quadratic Assignment Programming for Sentence Matching

Anonymous

Abstract: Studies have shown that both the syntactic structures and words' semantics are important for sentence matching. Existing studies usually model the syntactic structures and word semantics separately, resulting in matching models that overlook the relations and dependencies between syntactic structures and semantic meanings. How to jointly model the syntactic and semantic information has become a challenging problem in sentence matching. To address the issue, we formalize sentence matching as a problem of assigning the word of one sentence to that of another sentence, with the costs determined by the differences between the corresponding syntactic structures and word embedding similarities. The proposed method, referred to as neural quadratic assignment programming for sentence matching (NQAP-SM), represents the syntactic structures and semantic matching signals as an association graph.Solving the relaxed quadratic assignment programming (QAP) on this association graph achieves the final matching score. Experimental results on three public datasets demonstrated that NQAP-SM can outperform the state-of-the-art baselines in an effective and efficient way. The analysis also showed that NQAP-SM can match sentences in an interpretable way.

Paper Type: long