Customer Review Feature Extraction

Popular products often have thousands of reviews that contain far too much information for customers to digest. Our goal for the project is to implement a system that extracts opinions from these reviews and summarizes them in a concise form. This allows customers to quickly get an overview of a product and manufactures to efficiently process product feedbacks. In the past, we focused on the feature extraction directly on the word level. First, we would use association rule to extract all the adj.> rules and then use Pointwise Mutual Information to judge the polarization of the adjective. Processing directly on word level neglects the information contained on the sentence level. In this paper, we discuss how to extract features by first investigate on sentence level and then dive into the word level. Introduction Extracting features directly on word level neglects the information contained in the sentence level. For example, there might be a sentence such as “Today is a good day” appearing in the reviews. If we only perform the algorithm on the word level, we would consider as a feature of the product, but obviously it is not. Here we propose a method that first determines which sentences in a review might contain product features. If we are confident that a sentence does have feature, we would further process the sentence using the association rule and PMI to extract the features. To judge whether a sentence has any features, we used several Machine Learning methods. We will compare the performance of each method. Algorithms and Implementation Naïve Bayesian Naïve Bayesian technique is a powerful method for classification. In our problem, the two classes are Review Sentences with product feature (C1) and Review Sentences without product feature (C0). Pr C1 = 1{Ci = C1} training Size Pr(C0) = 1 – Pr(C1) The features we used for training are individual words within each sentence. We denote each sentence as Si and each word as Wj. The probability of the word Si appearing in a sentence of class Ck is given by: P Wj Ck = 1 Ci = Ck #{Wj in Si} Si 1 Ci = Ck #{Wj in Si} Si Wj Since some word might never appear in sentences of a particular class, we do not want to have a sentence with 0 probability of being in a particular class. We deploy a slightly modified version of Laplace Smoothing where λ = 0.1 instead of 1. P Wj Ck = 1 Ci = Ck ∗ #{Wj in Si} Si + 0.1 1 Ci = Ck ∗ #{Wj in Si} Si Wj + 0.1 ∗ |W| Finally, we have the following Naïve Bayesian classifier to determine the probability of a sentence being in a particular class, using the words in that sentence as features: P C1 Si = P(C1) P(Wj in Si|C1) P(Ck) P(Wj in Si|Ck) Ck P(C0|Si) = 1 – P(C1|Si) To avoid underflow, we can alternatively use log for calculation: log P C1 Si − log(P C0 Si) = logP C1 − logP(C0)