Automatic Generation of High Quality CCGbanks for Parser Domain Adaptation

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 129–139 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 129 Automatic Generation of High Quality CCGbanks for Parser Domain Adaptation Masashi Yoshikawa1 yoshikawa.masashi.yh8@ is.naist.jp Hiroshi Noji2 hiroshi.noji@aist.go.jp Koji Mineshima3 mineshima.koji@ocha.ac.jp Daisuke Bekki3 bekki@is.ocha.ac.jp 1Nara Institute of Science and Technology, Nara, Japan 2Artificial Intelligence Research Center, AIST, Tokyo, Japan 3Ochanomizu University, Tokyo, Japan Abstract We propose a new domain adaptation method for Combinatory Categorial Grammar (CCG) parsing, based on the idea of automatic gen- eration of CCG corpora exploiting cheaper re- sources of dependency trees. Our solution is conceptually simple, and not relying on a spe- cific parser architecture, making it applicable to the current best-performing parsers. We conduct extensive parsing experiments with detailed discussion; on top of existing bench- mark datasets on (1) biomedical texts and (2) question sentences, we create experimen- tal datasets of (3) speech conversation and (4) math problems. When applied to the proposed method, an off-the-shelf CCG parser shows significant performance gains, improving from 90.7% to 96.6% on speech conversation, and from 88.5% to 96.8% on math problems. 1 Introduction The recent advancement of Combinatory Cate- gorial Grammar (CCG; Steedman (2000)) pars- ing (Lee et al., 2016; Yoshikawa et al., 2017), combined with formal semantics, has enabled high-performing natural language inference sys- tems (Abzianidze, 2017; Mart´ınez-G´omez et al., 2017). We are interested in transferring the suc- cess to a range of applications, e.g., inference sys- tems on scientific papers and speech conversation. To achieve the goal, it is urgent to enhance the CCG parsing accuracy on new domains, i.e., solv- ing a notorious problem of domain adaptation of a statistical parser, which has long been addressed in the literature. Especially in CCG parsing, prior work (Rimell and Clark, 2008; Lewis et al., 2016) has taken advantage of highly informative cate- gories, which determine the most part of sentence structure once correctly assigned to words. is demonstrated that the annotation of only pre-It terminal categories is sufficient to adapt a CCG parser to new domains. However, the solution is limited to a specific parser’s architecture, making non-trivial the application of the method to the current state-of-the-art parsers (Lee et al., 2016; Yoshikawa et al., 2017; Stanojevi´c and Steedman, 2019), which require full parse annotation. Addi- tionally, some ambiguities remain unresolved with mere supertags, especially in languages other than English (as discussed in Yoshikawa et al. (2017)), to which the method is not portable. Distributional embeddings are proven to be powerful tools for solving the issue of domain adaption, with their unlimited applications in NLP, not to mention syntactic parsing (Lewis and Steed- man, 2014b; Mitchell and Steedman, 2015; Pe- ters et al., 2018). Among others, Joshi et al. (2018) reports huge performance boosts in con- stituency parsing using contextualized word em- beddings (Peters et al., 2018), which is orthogo- nal to our work, and the combination shows huge gains. Including Joshi et al. (2018), there are stud- ies to …

Content truncated to 500 words