Matching Article Pairs with Graphical Decomposition and Convolutions

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 6284–6294 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 6284 Matching Article Pairs with Graphical Decomposition and Convolutions Bang Liu†, Di Niu†, Haojie Wei‡, Jinghong Lin‡, Yancheng He‡, Kunfeng Lai‡, Yu Xu‡ †University of Alberta, Edmonton, AB, Canada {bang3, dniu}@ualberta.ca ‡Platform and Content Group, Tencent, Shenzhen, China {fayewei, daphnelin, collinhe, calvinlai, henrysxu}@tencent.com Abstract Identifying the relationship between two arti- cles, e.g., whether two articles published from different sources describe the same break- ing news, is critical to many document un- derstanding tasks. Existing approaches for modeling and matching sentence pairs do not perform well in matching longer documents, which embody more complex interactions be- tween the enclosed entities than a sentence does. To model article pairs, we propose the Concept Interaction Graph to represent an ar- ticle as a graph of concepts. We then match a pair of articles by comparing the sentences that enclose the same concept vertex through a series of encoding techniques, and aggregate the matching signals through a graph convo- lutional network. To facilitate the evaluation of long article matching, we have created two datasets, each consisting of about 30K pairs of breaking news articles covering diverse top- ics in the open domain. Extensive evaluations of the proposed methods on the two datasets demonstrate significant improvements over a wide range of state-of-the-art methods for nat- ural language matching. 1 Introduction Identifying the relationship between a pair of arti- cles is an essential natural language understand- ing task, which is critical to news systems and search engines. For example, a news system needs to cluster various articles on the Internet report- ing the same breaking news (probably in different ways of wording and narratives), remove redun- dancy and form storylines (Shahaf et al., 2013; Liu et al., 2017; Zhou et al., 2015; Vossen et al., 2015; Bruggermann et al., 2016). The rich semantic and logic structures in longer documents have made it a different and more challenging task to match a pair of articles than to match a pair of sentences or a query-document pair in information retrieval. Traditional term-based matching approaches es- timate the semantic distance between a pair of text objects via unsupervised metrics, e.g., via TF- IDF vectors, BM25 (Robertson et al., 2009), LDA (Blei et al., 2003) and so forth. These methods have achieved success in query-document match- ing, information retrieval and search. In recent years, a wide variety of deep neural network mod- els have also been proposed for text matching (Hu et al., 2014; Qiu and Huang, 2015; Wan et al., 2016; Pang et al., 2016), which can capture the semantic dependencies (especially sequential de- pendencies) in natural language through layers of recurrent or convolutional neural networks. How- ever, existing deep models are mainly designed for matching sentence pairs, e.g., for paraphrase iden- tification, answer selection in question-answering, omitting the complex interactions among key- words, entities or sentences that are present in a …

Content truncated to 500 words