Premium
THE DESIGN OF A STATISTICAL ALGORITHM FOR RESOLVING STRUCTURAL AMBIGUITY IN “V NP 1 usde NP 0 ”
Author(s) -
Li Wenjie,
Wong KamFai
Publication year - 2003
Publication title -
computational intelligence
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 0.353
H-Index - 52
eISSN - 1467-8640
pISSN - 0824-7935
DOI - 10.1111/1467-8640.00214
Subject(s) - noun phrase , computer science , determiner phrase , natural language processing , artificial intelligence , phrase , ambiguity , verb phrase , parsing , endocentric and exocentric , verb , sentence , phrase search , algorithm , noun , information retrieval , search analytics , web search query , programming language , search engine
The existence of structural ambiguity in modifying clauses renders noun phrase (NP) extraction from running Chinese texts complicated. It is shown from previous experiments that nearly 33% of the errors in an NP extractor were actually caused by the use of clause modifiers. For example, consider the sequence “V + NP 1 + ( of ) + NP 0 .” It can be interpreted as two alternatives, a verb phrase (i.e., [V[NP 1 ++ NP 0 ] NP ] VP ) or a noun phrase (i.e., [[V NP 1 ] VP ++ NP 0 ] NP ). To resolve this ambiguity, syntactical, contextual, and semantics‐based approaches are investigated in this article. The conclusion is that the problem can be overcome only when the semantic knowledge about words is adopted. Therefore, a structural disambiguation algorithm based on lexical association is proposed. The algorithm uses the semantic class relation between a word pair derived from a standard Chinese thesaurus, , to work out whether a noun phrase or a verb phrase has a stronger lexical association within the collocation. This can, in turn, determine the intended phrase structure. With the proposed algorithm, the best accuracy and coverage are 79% and 100%, respectively. The experiment also shows that the backed‐off model is more effective for this purpose. With this disambiguation algorithm, parsing performance can be significantly improved.