Improving Dependency Parsing Using Sentence Splitting

Parsing is an important task in natural language processing, which connects lexical and semantic analysis. Compared to phrase-structure grammar, dependency grammar has drawn great attraction for its simplicity and easy analyzing. Based on dependency grammar, we can get the dependency relations between words in a sentence. Syntactical dependency parsing has been greatly used in machine translation, automatic summarization, document classification, question answering, and so on.In this paper, several statistical dependency parsing algorithms were introduced firstly. Almost all of the existing dependency methods suffer from long sentences. Punctuations can be used to improve the result. Sub-sentences are parsed and so does the new sentence composed by head words of sub-sentences. The results are merged into the final result. However, errors in the results of sub-sentences will be enlarged in the remaining process of the method. Dependencies between head words are also parsed badly.So, we proposed a new method to improve dependency parsing with punctuations. The parsing results of sub-sentences are used to modify the result of the original long sentence. Dependency parsing on the long sentence was applied firstly. Then the splitting process was done and sub-sentences were trained and tested. Finally, we merge long-sentence parsing result and corresponding sub-sentences parsing results to generate a final result. Experiments show that a great improvement was achieved with our method.Not all of the punctuations are suitable for splitting sentence in our task. So, to get even better result, conditional random field model was introduced to make decision whether the punctuation should be used to split the sentence or not. This refined method also further improves the parsing result in our experiments.Now, there are no good tools for us to visualize the result of dependency parsing for further analysis. So, in this thesis, we design and realized a visualizing tool to meet the demand. Dependency trees can be shown with different styles and errors are highlighted. Advanced searching and statistical analysis are also implemented. The tool can even be used to annotate dependency tree corpus. The tool can provide great help for parsing researchers.