Improving PP Attachment Disambiguation in a Rule-based Parser

Improving PP Attachment Disambiguation in a Rule-based Parser  Yoon-Hyung Roh a, Ki-Young Lee a, and Young-Gil Kim a aNatural Language Processing Research Team, Electronics and Telecommunications Research Institute, 161 Gajeong-dong, Yuseong-gu, Daejeon, 305-350, Korea {yhroh, leeky, kimyk}@etri.re.kr Abstract. This paper deals with how to enhance the performance of a rule-based parser using statistical Information. PP (Prepositional Phrase) attachment ambiguity is one of the main ambiguities found in parsing. We therefore conducted some experiments on extracting statistical information for PP attachment from a corpus, and on applying such information to a rule-based parser. Two types of information are used: supervised learning data and unsupervised learning data. In this paper, we show how we apply these types of information and to what degree they contribute to the PP attachment as well as to the overall parsing performance. The final results show a 5.42% performance improvement in PP attachment, with an 8.7% error reduction ratio in the overall parsing performance. Keywords: PP attachment, rule-based parsing, syntactic ambiguity, statistical information, lexical dependency  The work reported in this paper was supported by the IT R&D program of MKE, “Development of Machine Translation Technology for Korean/Chinese/English Spoken Language and Business Documents”. Copyright 2011 by Yoon-Hyung Roh, Ki-Young Lee and Young-Gil Kim 1 Introduction For a rule-based parser using CFG as its grammar, syntactic ambiguity is a challenging problem as it becomes more difficult to resolve conflicts between rules as the number of rules increases. Certain methods to solve this problem have been suggested, and have generally introduced statistical approaches (Sagae and Lavie, 2003; Foth and Menzel, 2006). Many recent parsing technologies have also taken statistical approaches, as we can get more linguistic data, such as the Penn Treebank (Collins, 1999; Collins, 2000; Charniak, 2000). These technologies have shown encouraging performances. But practically speaking, statistical parsing has efficiency and scalability problems. As a statistical parser generally uses a significant number of parameters, and the search space is large, the parsing speed is relatively low and is not adequate for a real-time application. In our previous research, we conducted some experiments on applying statistical information to a rule-based parser (Roh, 2009). The parsing performance was affected mainly in the attachment performance and by the attachment information. Therefore, we do not need to apply full statistical information. The scalability problem indicates the difficulty in incorporating other types of knowledge for use in parsing. Usually, a practical parser uses various types of additional knowledge for parsing, such as lexical patterns or semantic patterns. However, it is not easy to incorporate other knowledge into a parsing system based on probability. It is also difficult to tune the parser minutely with respect to each sentence. Thus, we want to use a rule-based parser as a base parsing system, and selective statistical information for ambiguity resolution. Our parsing system conducts bottom-up chart parsing using ACFG (Augmented Context Free Grammar) rules, which have many syntactic and semantic features and constraints for 25th Pacific Asia Conference on Language, Information and Computation, pages 559–566 …

Content truncated to 500 words