Research On Approximate Text Analysis Based Opinion Mining

In this paper, in-depth research on promoting the effectiveness of sentiment analysis in opinion mining by means of partial parsng is described. The achievements of the paper are as follows:(1) A partial-parsing-based method and its incremental implementation are proposed. The new parsing method is named Super Parsing, whose purpose is to seek all the possible interpretation for a given text (or text segment) by relaxing the constraints on reducing. It allows non-adjacent constituents to be merged, and allows one constituent to join multiple reduction relations. To optimize the enumerating process of eonsitutent combinations, the Candidate List Algorithm (CLA) is proposed. Approximate Text Analysis (ATA) is the incremental implementation of Super Parsing. It consists of two parts: Analyzing Component and Global Reduction Component. Analyzing Component takes the buffer queue as core data structure, which converts the Super Parsing problem into the Breadth-first Searching problem. While Global Reduction Component is the incremental implementation of CLA for constituent reduction.(2) A novell sentiment classification algorithm and its software implementation are proposed. The new algorithm is named ATA-based Sentiment Classification (ATA-SC). ATA-SC considers the semantic relationship between entity words and sentiment-relevant words, so its recognizing ability is better than traditional methods based on single subject hypothesis. And ATAFilter is the ATA-based sentiment classification module. This module has been intergated into the Mail Filtering Software VIHunter, and achieved resonable effect in both testing and application.(3) A new opinion extraction task and its solution are proposed. The new task is named Opinion Instance Extraction (OIE). It keeps the association between opinion storage and the source text, so that more context information can be utilized in future mining task. To solve the task, the algorithm Featurecentered Opinion Instance Extraction (FC-OIE) is proposed. It takes two steps: (â…°) for each feature instance (FI), to find the most semantically related subject instance (SI) with SARPO approach and make a "SI-FI" pair; (â…±) to determine the sentiment of each "SI-FI" pair by casting ATA-SC on the text segment near SI and/or FI.(4) A system for opinion instance extraction and retrieving is proposed and implemented, named Opinion Searching Systen (OSS). The system extracts opinion instances with FC-OIE from web reviews, and feeds them back via human-machine interface to users according to their searching interest.