Abstract
Text matching is the core of natural language processing (NLP) system. It's considered as a touchstone of the NLP, and it aims to find whether text pairs are equal in semantics. However, the semantic gap in text matching is still an open problem to solve. Inspired by successes of cycle-consistent adversarial network (CycleGAN) in image domain transformation, we propose an enhanced text matching method based on the CycleGAN combined with the Transformer network. Based on the proposed method, the text semantics in a source domain is transferred to a similar or different target domain, and the semantic distance between text pairs is decreased. Meanwhile, we demonstrate our method in paraphrase identification and question answer matching. The matching degree is computed by a standard text matching model to evaluate the transforming influence on narrowing the text semantic gap. The experiments show that our method achieves text domain adaptation, and the effects on different matching models are remarkable.
Highlights
Text matching is a core issue in many natural language processing (NLP) tasks, such as paraphrase identification (PI) [1], question answering (QA) [2], information retrieval (IR) [3] and many other tasks
We focus on PI and QA matching
In this paper, we present that text matching performance can be enhanced by semantic transformation of text pairs from a new perspective
Summary
Text matching is a core issue in many natural language processing (NLP) tasks, such as paraphrase identification (PI) [1], question answering (QA) [2], information retrieval (IR) [3] and many other tasks. To address the problem of semantic gap, most prior works proposed various deep matching methods These methods depend on effective text representations, in which each word is represented as a distributed embedding vector via a deep neural network (DNN). The experiments on the task of paraphrase identification show that our method can achieve the similar semantic domain transformation, i.e. a question is transferred to a target question, and the matching degree is improved. For discriminator DX and DY , we freeze the generators, and minimize the loss functions LD, i.e. we should maximize the probability DX (x) and DY (y) for the original text, and minimize the probability DX (x) and DY (y) for the transferred text. G1 and G2 in our semantic transformation model is used to transfer the text semantic domain
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
