Structural Ambiguity and Conceptual Information Retrieval

Structual Ambiguity and Conceptural Information Retrieval Mathis Huey-chyun Chen and Jason J. S. Chang Department of Computer Science National Tsing Hua University Hsinchu 30043, Taiwan dr804311, jschang )@cs.nthu.edu.tw telephone: +886-35-731069 fax: +886-35-723694 Abstract Many researches found lexical preferences to be critical in resolving attachment ambiguity [WFB 90][FBK 82][ MP 80]. Most notably, information from Verb-Obj-Prep-Noun structures (VOPN) has been used to show that LA is very effective in the resolution of PP- attachment ambiguity [HR 93]. We investigated extensions to the lexical association strategy. The extensions include using conceptual association and acquiring the association information from different kind of lexical relations not limited to relations in VOPN structures. We refer to this approach as DeepAttach. Thus, it is possible to take information from all kinds of syntactical structures as long as they are alternations of a common deep structure [PU93] related to that implied by the intended attachment. A collection of sense-disambiguated sentences serves as the source of conceptual relations. No pre-processing is done to find the conceptual relations in these sentences. Instead, information retrieval technique is used to retrieve conceptually most relevant sentences using the words from the ambiguous structure as query. The prepositional phrase is then attached in favor of the constituent that has more conceptual presence in the ranked retrieved sentences. An experiment was implemented to embody the idea. The result shows that 75% of PP's in 260 VOPN structures can be attached correctly, when simple lexical relevance was considered. 1. Introduction Many researches found lexical preferences to be critical in resolving attachment ambiguity. Jenson and Binot [JB 87] propose to use dictionary definition for disambiguation. [HR 93] describe how co-occurrence of verbs and nouns with prepositions in corpus can be used as an indication of lexical preference. However, dictionary text especially the definitions are typically uneven in their coverage and inconsistent in the usage of words, while lexical co- occurrence suffers from sparseness of data. Resnik and Hearst (1994) find that using class information did not yield improved performance over lexical association due to multiplicity of classes for a word and lack of disambiguation. Our proposal is to use sense-disambiguated, sizable body of text as the source of lexical preference. Thus, for example, in the sentence 'He woke up to find the house on fire,' one of the word senses of the noun 'fire' occurs frequently in the context of the words which are conceptually similar to 'house.' This is evidence of a conceptual association of the direct object `house' with the prepositional object 'fire.' Unlike in previous proposals [HR 93][RH 93], these co-occurrences of concepts need not in the Verb-Object-Preposition relation. 115 2. The DeepAttach Algorithm For simplicity, the verb, direct object, preposition, and head of prepositional object of each testing sentence are fed to the prepositional phrase disambiguator. Category of Longman LEXICON [LO 92] is used as our semantic tag set. Given the sentence 'He woke up to find the house on fire,' the word 'fire' has four different semantic tags, and the other …

Content truncated to 500 words