Discourse Component to Sentence (DC2S): An Efficient Human-Aided Construction of Paraphrase and Sentence Similarity Dataset

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 6819–6826 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 6819 Discourse Component to Sentence (DC2S): An Efficient Human-Aided Construction of Paraphrase and Sentence Similarity Dataset Won Ik Cho1, Jong In Kim2, Young Ki Moon3,4, Nam Soo Kim1 Department of Electrical and Computer Engineering and INMC, Seoul National University1, Interdisciplinary Program in Cognitive Science, Seoul National University2, Department of Computer Engineering, Inha University3, Voithru Co., Ltd4 wicho@hi.snu.ac.kr, prows12@gmail.com, ykmoon0814@gmail.com, nkim@snu.ac.kr Abstract Assessing the similarity of sentences and detecting paraphrases is an essential task both in theory and practice, but achieving a reliable dataset requires high resource. In this paper, we propose a discourse component-based paraphrase generation for the directive utterances, which is efficient in terms of human-aided construction and content preservation. All discourse components are expressed in natural language phrases, and the phrases are created considering both speech act and topic so that the controlled construction of the sentence similarity dataset is available. Here, we investigate the validity of our scheme using the Korean language, a language with diverse paraphrasing due to frequent subject drop and scramblings. With 1,000 intent argument phrases and thus generated 10,000 utterances, we make up a sentence similarity dataset of practically sufficient size. It contains five sentence pair types, including paraphrase, and displays a total volume of about 550K. To emphasize the utility of the scheme and dataset, we measure the similarity matching performance via conventional natural language inference models, also suggesting the multi-lingual extensibility. Keywords: discourse component, paraphrasing, generation, sentence similarity test 1. Introduction How should the paraphrase be defined for the utterances with a relatively clear act? “Could you pass me over the salt?” is a request, but in some cultures, it might have a po- lite sibling such as “I can see salt on your side, so I wonder if it does not matter...”, though some unexpected nuances are inserted in. One may claim they are not precisely the same, but in real-life usage, paraphrasing does not neces- sarily aim to convey the identical possible world to the ad- dressee. How should the condition be defined for the above statement? How about other tricky sentences in real life? Checking sentence similarity and generating paraphrases has been studied a lot recently for their direct relationship with enriching the human language and its understanding (Fernando and Stevenson, 2008; Agirre et al., 2012; Agirre et al., 2013). The closest example is probably in question answering (QA) domain by matching the user query to the questions that are already in the database (Achananuparp et al., 2008), providing substantial information as an out- put. Similar approaches can be useful for recommendation systems or automatic counselors. In the non-task-oriented dialog between human and machine, sentence similarity datasets can help machines determine if the input is relevant to the conversation history. Also, the sentence relevance test has been of significant importance for some self-supervised pretraining of language models (Devlin et al., 2018). …

Content truncated to 500 words