Study On Attributes Computation In Conceptual Intension

Nowadays semantic processing is becoming a research hotspot in computationallinguistics. With the development of semantic processing, Chinese information pro-cessing is stepping up from lexicon, syntax to semantics. However, Chinese processinggets into trouble when analyzing syntax. The reason is that Chinese is short of mor-phological changes, that is, lexicons do not correspond to syntactic component strictly.So Chinese is actually not suitable for the syntactic theory of Indo-European language.As a result, Chinese processing may pay more attention to concept instead of syntax.According to the analysis of conceptual structure, this paper employs conceptual graphto represent semantics. The research starts from some instances of indexing conceptu-al graph manually, and the method and step of indexing conceptual are concluded. Onthis basis, this paper studies the approach of attribute acquisition, and the applicationof conceptual graph in information retrieval.The contributions of this paper reside in the following aspects.1. This paper proposes a revised conceptual graph based on conceptual inten-sional structure for representing Chinese semantics. Then the method of indexingconceptual graph manually is summed up. The conceptual graph can be employed ininformation retrieval system for indexing queries and documents, which overcomesthe shortage that traditional boolean model and vector model break semantic integrity.This is the basis and premise of this paper.2. This paper proposes two approaches to measure semantic similarity betweenwords. The first approach is based on large scale corpus, taking statistics of contextualpattern for computing similarity. This approach utilizes the massive data from Web,showing word similarity statistically. The second approach takes advantage of defini-tion in machine readable dictionary, which is the summary of expert knowledge andstatic data source. The merits of this approach are thus reliable and effective. Experi- mental results show that the two approaches are practicable, especially the second one.Similarity measurement is a foundation work of following works.3. This paper proposes a bootstrapping method for attributes extraction, acquiringattributes from definition in dictionary automatically. Firstly, some seeds are providedfor booting training, then extracting templets are generated iteratively. In order toimprove the extracting ability of templets, word similarity measurement is introducedinto sequence alignment algorithm, and synonym set is used to generalize templets.Experimental results show this method achieves good precision and recall.4. This paper proposes two methods for attribute names extension and a methodfor attribute names validation based on Web. These works extend the results of at-tributes extraction so as to construct a richer attribute knowledge base. The extendingmethods include hyponyms based and coordinated component based ways, which areboth summary and utilization of language phenomenon. This paper suggests a revisedPMI-IR algorithm for validating the candidate attribute names. Experimental resultsshow that these works enrich attributes names set effectively.5. This paper proposes an IR-oriented method for indexing conceptual graphautomatically. Firstly, the documents are processed using attributes extracting methodto construct a conceptual skeleton graph, transforming indexing conceptual graph intofilling skeleton graph. This paper introduces"Demand Focus"in order to describeuser goals detailedly, compared with traditional category of user goals. This paper alsosuggests a word segmenting method based on search engine, which is more suitableto recognize named entity in queries than traditional methods. Experimental resultsshow that the method for automatic indexing achieves good performance.6. This paper proposes an IR-oriented method for measuring similarity betweenconceptual graphs. Compared with the previous works, this paper analyzes differenceamong three kinds of nodes, suggesting three corresponding approaches to measurenodes similarity. On this basis, similarity of whole conceptual graph is computedrecursively. Experimental results show that retrieval model based on conceptual graphis effective, and Demand Focus helps to improve the quality of search results.