Monolingual Phrase Alignment as Parse Forest Mapping

Proceedings of the The 12th Joint Conference on Lexical and Computational Semantics (*SEM 2023), pages 449–455 July 13-14, 2023 ©2023 Association for Computational Linguistics Monolingual Phrase Alignment as Parse Forest Mapping Sora Kadotani∗ NTT Human Informatics Laboratories, NTT Corporation, Japan sora.kadotani@ntt.com Yuki Arase Osaka University, Japan Artificial Intelligence Research Center, AIST, Japan arase@ist.osaka-u.ac.jp Abstract We tackle the problem of monolingual phrase alignment conforming to syntactic structures. The existing method formalises the problem as unordered tree mapping; hence, the align- ment quality is easily affected by syntactic am- biguities. We address this problem by expand- ing the method to align parse forests rather than 1-best trees, where syntactic structures and phrase alignment are simultaneously iden- tified. The proposed method achieves efficient alignment by mapping forests on a packed structure. The experimental results indicated that our method improves the phrase alignment quality of the state-of-the-art method by align- ing forests rather than 1-best trees. 1 Introduction Monolingual phrase alignment, which identifies se- mantically corresponding phrase pairs in sentences, is a fundamental technique useful for paraphrase recognition (Das and Smith, 2009), textual entail- ment recognition (MacCartney et al., 2008; Heil- man and Smith, 2010), question answering (Wang and Manning, 2010), and interpreting semantic tex- tual similarity (Agirre et al., 2015; Li and Srikumar, 2016). Its ability to declare overlapping informa- tion across sentences is also useful for summari- sation (Brook Weiss et al., 2021) and for interac- tive document exploration (Shapira et al., 2017; Hirsch et al., 2021). There are two approaches to phrase alignment: one aligns chunks of arbitrary spans (e.g., n-grams) (Yao et al., 2013; Ouyang and McKeown, 2019; Lan et al., 2021) while the other targets on syntactic phrases (Arase and Tsujii, 2017, 2020). In this study, we take the latter ap- proach to identify phrasal paraphrases conforming to syntactic structures that allow modelling sen- tences based on syntax (Socher et al., 2013; Tai et al., 2015). ∗This work was completed at Osaka University. [Person name] made this statement at the Karachi Shipyard when talking with reporters [Person name] gave this statement while speaking to media persons at Karachi Shipyard Figure 1: Phrase alignment example by the proposed method (corresponding nodes are colour-coded). Our method aligns a source (top) and target (bottom) parse forests simultaneously determining their structures. The current state-of-the-art syntactic phrase alignment (Arase and Tsujii, 2020) has formulated the phrase alignment as the unordered tree map- ping problem between trees of source and target sentences.1 Their method realised an efficient align- ment with a solid theoretical background by adopt- ing the constrained tree edit distance algorithm (Zhang, 1996), which aligns syntactic trees. How- ever, their experiments were limited to using man- ually assigned gold syntactic trees, disregarding the effects of syntactic ambiguities that cause parse errors in practical parsers. We address this problem by expanding the method proposed by Arase and Tsujii (2020) to align parse forests. Specifically, our method con- siders the likelihood of both syntactic structures and phrase alignment, i.e., it simultaneously identi- fies the syntactic …

Content truncated to 500 words