Research On Bilingual Mapping Technology Of Linking Open Schema

Schema information is quite critical in the domain of the Semantic Web and monolingual schema information has already shown its value in many research and application areas. Although cross-lingual schema link also has wide applications, the research on it is just starting and current cross-lingual schema mapping technology cannot be applied directly.Based on those facts, this thesis is focused on the research on bilingual mapping technology of linking open schema. Three aspects are involved:1. Methods to estimate similarity between categories from different languages. Four kinds of methods have been proposed:dictionary-based method, structure-based method, property-based method and instance based method. Dictionary-based method is used to fill the language gap and other three methods estimate the similarity based on different description of categories.2. Category mapping algorithm based on machine learning. Three different models have been tried:logistic regression, decision tree and multilayer perceptron. Results of those models have been compared and discussed.3. Methods to evaluate the mapping results. It is hard to evaluate the results because of the lacking of a golden-standard. This thesis uses random sampling and manual voting to do the precision analysis of the results. In addition, comparative experiment has also been conducted to evaluate the improvement in both precision and recall.As a conclusion, this thesis has proposed a bilingual mapping algorithm of linking open schema. Property-based method and instance-based method has utilized instance information to a smaller granularities. This can solve the instance mismatch problem in bilingual mapping environment. Those two methods are main innovations of this thesis.Apart from research on mapping algorithm, this thesis also discussed the construction method of bilingual linking open schema including schema information acquisition, construction and mapping. Finally thesis has built and published two domain datasets in e-commerce domain and online encyclopedia domain.