Linguistically-Motivated Grammar Extraction, Generalization and Adaptation

R. Dale et al. (Eds.): IJCNLP 2005, LNAI 3651, pp. 177 – 187, 2005. © Springer-Verlag Berlin Heidelberg 2005 Linguistically-Motivated Grammar Extraction, Generalization and Adaptation Yu-Ming Hsieh, Duen-Chi Yang, and Keh-Jiann Chen Institute of Information Science, Academia Sinica, Taipei {morris, ydc, kchen}@iis.sinica.edu.tw Abstract. In order to obtain a high precision and high coverage grammar, we proposed a model to measure grammar coverage and designed a PCFG parser to measure efficiency of the grammar. To generalize grammars, a grammar binari- zation method was proposed to increase the coverage of a probabilistic context- free grammar. In the mean time linguistically-motivated feature constraints were added into grammar rules to maintain precision of the grammar. The gen- eralized grammar increases grammar coverage from 93% to 99% and bracket- ing F-score from 87% to 91% in parsing Chinese sentences. To cope with error propagations due to word segmentation and part-of-speech tagging errors, we also proposed a grammar blending method to adapt to such errors. The blended grammar can reduce about 20~30% of parsing errors due to error assignment of pos made by a word segmentation system. Keywords: Grammar Coverage, Ambiguity, Sentence Parsing, Grammar Extraction. 1 Introduction Treebanks provide instances of phrasal structures and their statistical distributions. However none of treebanks provide sufficient amount of samples which cover all types of phrasal structures, in particular, for the languages without inflectional markers, such as Chinese. It results that grammars directly extracted from treebanks suffer low cover- age rate and low precision [7]. However arbitrarily generalizing applicable rule patterns may cause over-generation and increase ambiguities. It may not improve parsing per- formance [7]. Therefore a new approach of grammar binarization was proposed in this paper. The binarized grammars were derived from probabilistic context-free grammars (PCFG) by rule binarization. The approach was motivated by the linguistic fact that adjuncts could be arbitrarily occurred or not occurred in a phrase. The binarized gram- mars have better coverage than the original grammars directly extracted from treebank. However they also suffer problems of over-generation and structure-ambiguity. Con- temporary grammar formalisms, such as GPSG, LFG, HPSG, take phrase structure rules as backbone for phrase structure representation and adding feature constraints to elimi- nate illegal or non-logical structures. In order to achieve higher coverage, the backbone grammar rules (syntactic grammar) are allowed to be over-generation and the feature constraints (semantic grammar for world knowledge) eliminate superfluous structures 178 Y.-M. Hsieh, D.-C. Yang, and K.-J. Chen and increase the precision of grammar representation. Recently, probabilistic prefer- ences for grammar rules were incorporated to resolve structure-ambiguities and had great improvements on parsing performances [2, 6, 10]. Regarding feature constrains, it was shown that contexture information of categories of neighboring nodes, mother nodes, or head words are useful for improving grammar precision and parsing perform- ances [1, 2, 7, 10, 12]. However tradeoffs between grammar coverage and grammar precision are always inevitable. Excessive grammatical constraints will reduce grammar coverage and hence reduce parsing performances. On the other hand, loosely con- strained grammars cause structure-ambiguities and also reduce …

Content truncated to 500 words