Adaptive Feature Discrimination and Denoising for Asymmetric Text Matching

Proceedings of the 29th International Conference on Computational Linguistics, pages 1146–1156 October 12–17, 2022. 1146 Adaptive Feature Discrimination and Denoising for Asymmetric Text Matching Yan Li 1, Chenliang Li2, Junjun Guo1, ∗ 1 Kunming University of Science and Technology 2 Wuhan University {liyayli,guojjgb}@163.com cllee@whu.edu.cn Abstract Asymmetric text matching has becoming increasingly indispensable for many down- stream tasks (e.g., IR and NLP). Here, asym- metry means that the documents involved for matching hold different amounts of in- formation, e.g., a short query against a rel- atively longer document. The existing solu- tions mainly focus on modeling the feature interactions between asymmetric texts, but rarely go one step further to recognize dis- criminative features and perform feature de- noising to enhance relevance learning. In this paper, we propose a novel adaptive fea- ture discrimination and denoising model for asymmetric text matching, called ADDAX. For each asymmetric text pair, ADDAX is devised to explicitly distinguish discrimina- tive features and filter out irrelevant features in a context-aware fashion. Concretely, a matching-adapted gating siamese cell (MAGS) is firstly devised to identify discriminative fea- tures and produce the corresponding hybrid representations for a text pair. Afterwards, we introduce a locality-constrained hashing de- noiser to perform feature-level denoising by learning a discriminative low-dimensional bi- nary codes for redundantly longer text. Exten- sive experiments on four real-world datasets from different downstream tasks demonstrate that the proposed ADDAX obtains substantial performance gain over 36 up-to-date state-of- the-art alternatives. 1 Introduction Given a pair of documents, text matching aims to precisely predict the semantic relations between them. An efficient and effective matching algo- rithm is now an indispensable asset in many in- formation retrieval, question answering and dia- logue systems. In these application scenarios, a text pair in matching (e.g., query-document and question-answer pair) ususlly has a large disparity ∗Corresponding author. Long Positive Document： Long Negative Document： Short Query： Power Station (Power Station), a Taiwanese pop rock concert group, was founded in 1994 and consists of two indigenous Taiwanese singers, You Qiuxing and Yan Zhilin…… Power Station refers to a factory that uses the chemical energy of coal, oil, natural gas or other fuels to produce electricity Who is included in the Power Station? Figure 1: An example of asymmetric text matching. in the quantity of information, a.k.a. asymmet- ric text matching. For example, a matching pair have 7.15 and 95.54 words respectively on average in InsuranceQA dataset (Feng et al., 2015) (i.e., the difference being about an order of magnitude). This asymmetry between a short query and a long document renders it as a nontrivial task. The existing solutions can be grouped into two categories, namely representation-based and interaction-based models (Khattab and Zaharia, 2020). The former category mainly utilizes con- volutional neural networks (CNN) and recurrent neural networks (RNN) to learn the latent repre- sentation of a document independently, including DSSM (Huang et al., 2013), SNRM (Zamani et al., 2018). On contrast, the latter category focuses on leveraging fine-grained interaction signals between them. It is widely recognized that exploiting …

Content truncated to 500 words