FST-Based Natural Language Processing Method for Opinion Extraction

摘要

This paper proposes a rule-based and Finite State Transducers (FST) based NLP method for extracting information from massive text. The method differs from n-gram based popular method which relies on probability statistics and machine learning. In our method, the rules are grammars of a language, summarized by people. FST is the implementation tool of rules. It can process natural language and generate a syntax tree for each sentence. To support applying the rules, we tokenize and generate the stem of words, and find many word features which are recorded in a dictionary. After generating a syntax tree, we extract useful information on many aspects, such as subject-verb-object matches and opinion matches. We evaluate our system on the accuracy rate of the syntax trees, and show that the result is satisfactory.