IHS-RD-Belarus at SemEval-2016 Task 9: Transition-based Chinese Semantic Dependency Parsing with Online Reordering and Bootstrapping.

Proceedings of SemEval-2016, pages 1207–1211, San Diego, California, June 16-17, 2016. c⃝2016 Association for Computational Linguistics IHS-RD-Belarus at SemEval-2016 Task 9: Transition-based Chinese Semantic Dependency Parsing with Online Reordering and Bootstrapping. Artsiom Artsymenia Palina Dounar Maria Yermakovich IHS Inc. / IHS Global Belarus Minsk, Belarus IHS Inc. / IHS Global Belarus Minsk, Belarus IHS Inc. / IHS Global Belarus Minsk, Belarus {Artsiom.Artsymenia,Polina.Dovnar,Maria.Yermakovich}@ihs.com Abstract This paper is a description of our system de- veloped for SemEval-2016 Task 9: Chinese Semantic Dependency Parsing. We have built a transition-based dependency parser with on- line reordering, which is not limited to a tree structure and can produce 99.7% of the neces- sary dependencies while maintaining linear algorithm complexity. To improve parsing quality we used additional techniques such as pre- and post-processing of the dependency graph, bootstrapping and a rich feature set with additional semantic features. 1 Introduction Dependency parsing is one of the core tasks in nat- ural language processing, as it provides useful in- formation for other NLP tasks. Traditional syntac- tic parsing usually represents a sentence as a tree- shape structure and this restriction is essential for most of efficient algorithms developed in the past years. Semantic dependency parsing, on the other hand, deals with acyclic graphs, where words may have multiple incoming dependencies. It signifi- cantly complicates the task and requires develop- ment of the new algorithms or special adoption of the old ones. There are two main approaches to dependency parsing (Nivre and McDonald, 2008). The first one is a graph-based approach, for example, spanning tree algorithms (McDonald et al., 2005), where the goal is to find the highest scoring tree from a com- plete graph of dependencies between words in the sentence. The second approach is transition-based, which, instead of searching for global optimum, greedily finds local optimum with a chain of ac- tions that lead to a parsing tree. The main ad- vantage of the transition-based parsers is that they are in general faster than graph-based ones because of the linear complexity of the algorithm (Nivre, 2003). As we move from syntactic parsing to semantic parsing, and instead of projective trees have to deal with acyclic graphs, exact inference becomes NP- hard, so some strong independence assumptions or heuristics are needed. A lot of modifications have been proposed for transition-based parsers to sup- port non-projective structures. Nivre and Nilsson (2005) proposed a pseudo-projective parsing tech- nique which consists in modifying the input into a projective dependency tree with extended labels, performing the projective parsing and then apply- ing an approximate back transformation. Attardi (2006) introduced additional actions that add de- pendencies between the roots of non-adjacent sub- trees. Both techniques maintain linear algorithm complexity at the expense of incomplete coverage of all possible dependency trees. Complete cover- age of all non-projective trees is achieved by Nivre (2009) with a technique called on-line reordering but it increases the worst-case complexity from linear to quadratic. 2 System Description The core of our system is a transition-based de- pendency parser with …

Content truncated to 500 words