Improve Parsing Performance by Self-Learning

Improve Parsing Performance by Self-Learning Yu-Ming Hsieh, Duen-Chi Yang, and Keh-Jiann Chen Chinese Knowledge and Information Processing (CKIP) Institute of Information Science, Academia Sinica, Taipei {morris, ydc, kchen}iis@sinica.edu.tw Abstract There are many methods to improve performances of statistical parsers. Among them, resolving structural ambiguities is a major task. In our approach, the parser produces a set of n-best trees based on a feature-extended PCFG grammar and then selects the best tree structure based on association strengths of dependency word-pairs. However, there is no sufficiently large Treebank producing reliable statistical distributions of all word-pairs. This paper aims to provide a self-learning method to resolve the problems. The word association strengths were automatically extracted and learned by parsing a giga-word corpus. Although the automatically learned word associations were not perfect, the built structure evaluation model improved the bracketed f-score from 83.09% to 86.59%. We believe that the above iterative learning processes can improve parsing performances automatically by learning word-dependence knowledge continuously from web. 1. Introduction How to solve structural ambiguity is an important task in building a high-performance statistical parser, particularly for Chinese. Since Chinese is analytic language, words play different grammatical functions without inflections. A great deal of ambiguous structures will be produced by parsers if no structure evaluator is applied. There are three main steps in our approach aim to disambiguate the structures. The first step is to have parser produce n-best structures. Secondly, we extract word-to-word association from large corpora and build semantic information. The last one is to build a structural evaluator to find the best tree structure from n-best. Formerly, there were some approaches proposed to resolve structure ambiguities. For instances, � to add on lexical dependencies. Collins (1999) solves structural ambiguity by extracting lexical dependencies from Penn WSJ Treebank and applying dependencies to the statistic model. Lexical dependency (or Word-to-word association, WA) is one type of semantic information. It is a current trend to add on semantic related information in traditional parsers. Some incorporated word-to-word association in their parsing models, such as the Dependency Parsing in Chen et al. (2004). They take advantage of statistic information of word dependency in the parsing process to produce dependency structures. However, word association methods suffer low coverage for lacking very large tree-annotated training corpora, while checking dependency relation between word pairs. � to add on word semantic knowledge. CiLin and HowNet information are used in the statistic model in the experiment of Xiong et al. (2005). Their results prove to solve common parsing mistakes efficiently. � to use re-annotation method in grammar rule. Johnson (1998) thinks that re-annotating each node with the category of its parent category in Treebank is able to improve parsing performance. Klein et al. (2003) proposes internal/external/tag-splitting annotation strategies to obtain better results. � to build evaluator. Some people re-rank the structure values and find out the best parse (Collins, 2000; Charniak et al., 2005). At first hand, their parser produces a set of candidate parses for each sentence. Later, the reranker finds out the …

Content truncated to 500 words