Building Sentiment Analysis Pipeline: A Case Study on Amazon Reviews

This paper presents our efforts in building a frame-work of automatic sentiment analysis for online user product reviews. The framework consists of web scraping to acquire a corpus of publicly available user product reviews, construction of a reference dictionary for removal of noisy entries such as typos and special characters from the reviews, extraction of Bag-of-Words (BoW) representations, building BoW-based sentiment models for cosine similarity classification, neural network-based classification, and performance evaluation and linguistic analysis modules. In a case study on Amazon user reviews, we analyze the impact of polarity salience- and exponential-based weighting of unigram and bigram BoW vectors and salience weight-driven dimensionality reduction on sentiment prediction accuracy and show that both strategies provide a dramatic reduction of equal error rates. Finally, we use the salience weighting matrix established in the classification experiments to rank-order discriminability of dictionary entries and their bigrams and report our findings in the form of word clouds. The dataset collected for this study, as well as the experimental framework, are made publicly available.