Lexicon Acquisition with a large-coverage unification-based grammar

Lexicon acquisition with a large-coverage unification-based grammar Frederik Fouvry Computational Linguistics Saarland University PO Box 15 11 50 D-66041 Saarbrticken, Germany fouvry@coli.uni—sb.de Abstract We describe how unknown lexical en- tries are processed in a unification-based framework with large-coverage gram- mars and how from their usage lexi- cal entries are extracted. To keep the time and space usage during parsing within bounds, information from exter- nal sources like Part of Speech (PoS) taggers and morphological analysers is taken into account when information is constructed for unknown words. 1 Introduction For Natural Language Processing (NLP) in gen- eral, and processing with linguistically rich frame- works more specifically, unknown words are a problem. The following gives an idea of the extent of the problem. In an evaluation of a large-scale grammar for unrestricted text on a newspaper cor- pus, we found that the number of failed parses due to unknown words accounted for around 89% of the total number of unsuccessful analyses. Even though this figure does not say anything about the grammar (these failures may be hiding many oth- ers), it shows the importance of the problem. For unification-based implementations, which often refer to linguistic theories and are therefore rich in information, one approach to deal with un- known words has been proposed several times: to exploit the syntactic context of completed analyses to collect information about a new word. A few implementations have been developed to demon- strate the feasibility of the technique, but to our knowledge it has not been applied yet to large- coverage grammars. In this note we discuss how we are applying it to such a grammar for unre- stricted text. Starting from this "standard" tech- nique, we extend it and integrate PoS and mor- phological information, originating from external resources. We will first describe the method of learning in- formation from the syntactic context. Then we dis- cuss the current results of our implementation, and how the external resources are put to use. Finally an evaluation scheme is presented and some issues we intend to investigate next. 2 Acquiring new information In a unification-based framework, information is percolated throughout the parse tree via the re- entrancies. Information that is underspecified in a lexical entry very often becomes more specific when it is used in a parse tree. Take the follow- ing example. The lexical entry for the French verb form etaient ("were") specifies that its sub- ject should be a plural noun phrase (and in the third person). When it is combined with the femi- nine plural noun phrase les conditions ("the condi- tions") via U in (1), the information about the sub- ject of etaient will also include the gender value. 87 HEAD�[verb] SUBJECT M [HEAD [noun] AGR� PERSON [third] NUMBER [plural] HEAD [noun] PERSON [MEd] �AGR�NUMBER [phirail GENDER [feminine] Normally this increase in information is not used for anything outside the current analysis. With un- known words however, this property can be used to find out how they can be used. When a word …

Content truncated to 500 words