Product Feature Extraction Algorithm Based On Tree Structure

Sentiment analysis is the computing process for the opinions, sentiment and subjectivity of the text. Sentiment analysis of feature-based customer reviews is the research hot in the field of sentiment analysis in recent years. It contains product feature extraction, opinion words extraction and opinion classification. In order to extract product features more precisely and flexibly, we propose a new frame based on tree structure. The frame combines the underlying algorithm and tree structure information, which can increase the performance of the product feature extraction and has high parameter flexiblity.Product feature extraction algorithm based on tree structure uses the existing product feature extraction algorithm as the underlying algorithm and guides the extraction through the tree structure information in the online review net. Firstly, the algorithm uses the association analysis as the underlying algorithm. After preprocessing, we generate the transaction files from the consumer reviews and apply the association rule mining algorithm to mine the candidate frequent features. After applying a series of prune operations, we get the final product features. Then the algorithm introduces the tree structure information to guide the product feature extraction. Online review websites usually classifies the product into a tree structure for convenience. Among the product tree, the none-leafs node present the product category and the leaf nodes present the product. As the deep goes, the product category is becoming more concrete and more similar between the adjacent products. We find that the consumer tend to use the similar words to describe the similar products and use the different words to describe the differnent ones. Product feature extraction algorithm based on tree structure guides the extraction process by assigning different weight to the adjacent product nodes. The scale of the weight differs according to the distance betweent the nodes. At last, the algorithm outputs the final product features after the feature ranking. Experiment result shows that, the prposed algorithm increases the recall rate significantly while remains the precison rate. Meanwhile, we find that the different parameter set can result in different performance. So we can set a series of reasonable parameters to get the prefer results in practice.