Optimal Partial Transport Based Sentence Selection for Long-form Document Matching

Proceedings of the 29th International Conference on Computational Linguistics, pages 2363–2373 October 12–17, 2022. 2363 Optimal Partial Transport based Sentence Selection for Long-form Document Matching Weijie Yu1, Liang Pang4, Jun Xu2,3,∗ , Bing Su2,3, Zhenhua Dong5 and Ji-Rong Wen2,3 1School of Information, Renmin University of China 2Gaoling School of Artificial Intelligence, Renmin University of China 3Beijing Key Laboratory of Big Data Management and Analysis Methods 4Institute of Computing Technology, Chinese Academy of Sciences 5Noah’s Ark Lab, Huawei {yuweijie, junxu, bingsu, jrwen}@ruc.edu.cn, pangliang@ict.ac.cn, dongzhenhua@huawei.com Abstract One typical approach to long-form document matching is first conducting alignment between cross-document sentence pairs and then aggre- gating all of the sentence-level matching sig- nals. However, this approach could be prob- lematic because the alignment between doc- uments is partial — despite two documents as a whole are well-matched, most of the sen- tences could still be dissimilar. Those dissim- ilar sentences lead to spurious sentence-level matching signals which may overwhelm the real ones, increasing the difficulties of learn- ing the matching function. Therefore, accu- rately selecting the key sentences for document matching is becoming a urgent problem. To ad- dress this issue, we propose a novel matching approach that equips existing document match- ing models with an Optimal Partial Transport (OPT) based component, namely OPT-Match, which selects the sentences that play a major role in matching. Enjoying the partial transport properties of OPT, the selected key sentences can not only effectively enhance the matching accuracy, but also be explained as the ratio- nales for the matching results. Extensive ex- periments on four publicly available datasets demonstrated that existing methods equipped with OPT-Match consistently outperformed the corresponding underlying methods. Evalua- tions also showed that the sentences selected by OPT-Match were consistent with human- provided rationales. 1 Introduction Long-form document matching, which identifies the semantic relationship between a source docu- ment and a target document, has become a funda- mental problem in both NLP and IR. Representa- tive tasks include cite recommendation (Jiang et al., 2019) and plagiarism detection (Foltýnek et al., 2020) etc. For example, in cite recommendation, ∗ Corresponding author Document 2: medicinal and aromatic plants are nowadays becoming the main constituents for medicines, colorings, preservatives, and fragrances… although it is of high priority, the effect of slowgrowth conservation on the chemical profile of prestored medicinal plants after conservation was not extensively investigated .this research aimed to analyze in vitro stored thymbra spicata l. var. … Document 1: ... the attraction of medicinal and aromatic plants as worthy farm crops has grown due to the demand created by consumer interest in these plants for culinary, medicinal, and other anthropogenic applications … an understanding of future opportunities in the medicinal and aromatic plant industry who 2003 is necessary to enable us growers to envision and invest in medicinal and aromatic crops that will meet market demands. Figure 1: A pair of matched long-form documents from S2ORC dataset. Documents 1 focuses on the future op- portunities in the medicinal and aromatic plant industry. Document 2 studies …

Content truncated to 500 words