CONTEXTUAL RECOMMENDATION SYSTEM FOR LOCAL BUSINESSES

Abstract

Going through several reviews could be laborious when this has to be done for multiple restaurants. One could instead read a graphical representation of what is great at the restaurant. Currently on Yelp, the food recommendations are only based on the total number of mentions of the food item in the reviews. Higher mentions, irrespective of the context, get an up-vote toward recommended items. Including context from reviews and tips could greatly improve the list of recommended items. In this project, we combine Named Entity Recognition and Sentiment Analysis of reviews. Based on the sentiment of the reviews we aim to suggest the best dishes of a restaurant or the best restaurant offering a dish. We have leveraged various feature engineering methods to produce state-of-the-art results. We established that if chosen, the appropriate feature vectors can significantly improve the classification performance. Fine-tuning BERT and bi-directional LSTM are producing better results than the machine learning models and if trained for more epochs can eventually prove to be the best classifier models.