Sentiment Analysis for Review Rating Prediction in a Travel Journal

Abstract

This paper presents sentiment analysis to predict numerical rating of text reviews in a web-based travel journal application. The application allows users to record and provide text reviews on tourist spots visited. Text reviews undergo parts-of-speech (POS) tagging, rule-based phrase chunking and dependency parsing to extract opinion phrases in noun-adjective and noun-verb pairs from the original text. Each pair is further classified to one of the four categories: accommodation, food, entertainment and tourist attraction using the noun against a curated bag-of-words (BOW) to ensure that only relevant statements are included in the scoring. Word Sense Disambiguation is performed to correctly identify the word sense that matches the meaning of the sentence using WordNet. SentiWordNet, a lexical resource for sentiment analysis, was used to determine polarity score representing the emotional intensity of the review. The system predicted star rating was compared with the actual author rating in Google Maps and with human annotator ratings who are asked to label the text reviews. The predicted rating scored low mean absolute error (MAE) between the system and human rating which means that the rating predicted is closer to human interpretation of the text reviews. Overall rating prediction accuracy is 82%.