Research On Cross-lingual Fine-grained Sentiment Analysis

Sentiment analysis technology is to recognise and analyse the subjectiveinformation in text. The recent reseach is on coarse-grained classification problemson sentence level or document level. Based on these reseaches, reseachers’ worksfocus on the fine-grained information extraction, such as the holder or target ofopinion sentences. The recent extraction technologies, especially sequence labellingmethod, needs mass of fined-grained corpuses, which are lacking in nature languageprocessing domain. Thus we propose a cross-lingual method to transferfined-grained corpus from one language(source language) to another(target language)by machine translation system in order to enrich corpus in target language.According to the survey of recent fined-grained sentment analysis andcross-lingual technology, we summarize the main challenge in this problem. Thenwe propose a cross-lingual method based on syntactical sub-structure. The workcontains three parts:1. Propose a syntactical sub-structure based projection method. Apply machinetranslation system and phrase alignment information to translate and project sourcelanguage fined-grained corpus.2. Propose a multi-kernel support vector machine based fined-grainedsentiment analysis approach. It contains a syntactic based tree kernel, a semanticbased polynomial kernel and a word similarity based pivot function.3. Propose a transfer self-training method to filter cross-lingual corpus andimprove classifier.The contribution of work is:1. Propose a syntactical sub-structure to solve the mis-order orrer of machinetranslation system and realize the re-use of fined-grained sentiment corpus.2. Propose a multi-kernel support vector machine. Tree kernel solve theproblem of mis-order in machine translation. The pivot function improve precisionof classifier.3. Propose a transfer self-training method. Utilize syntactic and semantic kernelfunction in different procedures of self-training and get global optimum.The experiment in public data reveal a better perfomance than sequencelabelling method.