Enhancing Unlexicalized Parsing Performance Using a Wide Coverage Lexicon, Fuzzy Tag-Set Mapping, and EM-HMM-Based Lexical Probabilities

Proceedings of the 12th Conference of the European Chapter of the ACL, pages 327–335, Athens, Greece, 30 March – 3 April 2009. c⃝2009 Association for Computational Linguistics Enhancing Unlexicalized Parsing Performance using a Wide Coverage Lexicon, Fuzzy Tag-set Mapping, and EM-HMM-based Lexical Probabilities Yoav Goldberg1∗ Reut Tsarfaty2† Meni Adler1‡ Michael Elhadad1 1Department of Computer Science, Ben Gurion University of the Negev {yoavg|adlerm|elhadad}@cs.bgu.ac.il 2Institute for Logic, Language and Computation, University of Amsterdam R.Tsarfaty@uva.nl Abstract We present a framework for interfacing a PCFG parser with lexical information from an external resource following a dif- ferent tagging scheme than the treebank. This is achieved by defining a stochas- tic mapping layer between the two re- sources. Lexical probabilities for rare events are estimated in a semi-supervised manner from a lexicon and large unanno- tated corpora. We show that this solu- tion greatly enhances the performance of an unlexicalized Hebrew PCFG parser, re- sulting in state-of-the-art Hebrew parsing results both when a segmentation oracle is assumed, and in a real-word parsing sce- nario of parsing unsegmented tokens. 1 Introduction The intuition behind unlexicalized parsers is that the lexicon is mostly separated from the syntax: specific lexical items are mostly irrelevant for ac- curate parsing, and can be mediated through the use of POS tags and morphological hints. This same intuition also resonates in highly lexicalized formalism such as CCG: while the lexicon cate- gories are very fine grained and syntactic in na- ture, once the lexical category for a lexical item is determined, the specific lexical form is not taken into any further consideration. Despite this apparent separation between the lexical and the syntactic levels, both are usually es- timated solely from a single treebank. Thus, while ∗Supported by the Lynn and William Frankel Center for Computer Sciences, Ben Gurion University †Funded by the Dutch Science Foundation (NWO), grant number 017.001.271. ‡Post-doctoral fellow, Deutsche Telekom labs at Ben Gu- rion University PCFGs can be accurate, they suffer from vocabu- lary coverage problems: treebanks are small and lexicons induced from them are limited. The reason for this treebank-centric view in PCFG learning is 3-fold: the English treebank is fairly large and English morphology is fairly sim- ple, so that in English, the treebank does provide mostly adequate lexical coverage1; Lexicons enu- merate analyses, but don’t provide probabilities for them; and, most importantly, the treebank and the external lexicon are likely to follow different annotation schemas, reflecting different linguistic perspectives. On a different vein of research, current POS tag- ging technology deals with much larger quantities of training data than treebanks can provide, and lexicon-based unsupervised approaches to POS tagging are practically unlimited in the amount of training data they can use. POS taggers rely on richer knowledge than lexical estimates de- rived from the treebank, have evolved sophisti- cated strategies to handle OOV and can provide distributions p(t|w, context) instead of “best tag” only. Can these two worlds be combined? We pro- pose that parsing performance can be greatly im- proved by using a wide …

Content truncated to 500 words