Divide and Conquer: Text Semantic Matching with Disentangled Keywords and Intents

Findings of the Association for Computational Linguistics: ACL 2022, pages 3622 - 3632 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Divide and Conquer: Text Semantic Matching with Disentangled Keywords and Intents Yicheng Zou1,2, Hongwei Liu2, Tao Gui1∗, Junzhe Wang2, Qi Zhang2∗, Meng Tang3, Haixiang Li3, Daniel Wang3 1Institute of Modern Languages and Linguistics, Fudan University, Shanghai, China 2School of Computer Science, Fudan University, Shanghai, China 2Shanghai Collaborative Innovation Center of Intelligent Visual Computing, Shanghai, China 3IPS, Tencent PCG, Beijing, China {yczou18,tgui,qz}@fudan.edu.cn, daniellwang@tencent.com Abstract Text semantic matching is a fundamental task that has been widely used in various scenarios, such as community question answering, infor- mation retrieval, and recommendation. Most state-of-the-art matching models, e.g., BERT, directly perform text comparison by processing each word uniformly. However, a query sen- tence generally comprises content that calls for different levels of matching granularity. Specif- ically, keywords represent factual information such as action, entity, and event that should be strictly matched, while intents convey abstract concepts and ideas that can be paraphrased into various expressions. In this work, we propose a simple yet effective training strategy for text se- mantic matching in a divide-and-conquer man- ner by disentangling keywords from intents. Our approach can be easily combined with pre- trained language models (PLM) without influ- encing their inference efficiency, achieving sta- ble performance improvements against a wide range of PLMs on three benchmarks. 1 Introduction Text semantic matching aims to predict a matching category or a matching score reflecting the seman- tic similarity given a pair of text sequences, which is a fundamental task employed in a wide range of applications (Huang et al., 2013; Hu et al., 2014; Palangi et al., 2016; Cer et al., 2017; Rücklé et al., 2020; Pang et al., 2021). Recently, pre-trained lan- guage models (PLM) show remarkable capability of representation learning and have accelerated the research of text semantic matching (Devlin et al., 2019; Liu et al., 2019; Lan et al., 2019). They typi- cally exploit large-scale corpora and well-designed self-supervised learning objectives to better learn semantic representations, achieving state-of-the- art performances or even surpassing the level of non-expert humans on general semantic matching benchmarks (Wang et al., 2019b,a). ∗Corresponding authors. Y A: What does a civil engineer do? B: How does civil engineering work? N A: What is the best game engine? B: What is game engine? N A: How do I repair my gas boiler? B: How do I repair boiler? Y A: Why does my nose bleed? B: What causes nose bleeds? Figure 1: Examples of sentence pairs sampled from the QQP dataset. The keywords are highlighted, while the other words constitute abstract intents. Y and N repre- sent whether the pair is matched or not. The original matching problem can be decomposed into two sub- problems: keyword matching and intent matching. A semantically equivalent pair generally means the key- word and intent are matched simultaneously. Most existing PLMs aim to establish a foun- dation for various downstream tasks (Bommasani et al., 2021) and focus …

Content truncated to 500 words