Aspect Extraction on Restaurant Reviews using Domain-Specific Word Embedding

Abstract

Reviews on the internet can be an important part of a business and can influence owners or consumers for their decision making. Easy access to information in the form of opinions, experiences, and feedback from others can be used as a reference for taking an action. For businesses in the food and beverage sector, consumers usually provide reviews with negative or positive sentiments based on several aspects of the related business. The taste of the food, atmosphere, price, service are examples of aspects that are commonly written in a review. In this work, aspect extraction on consumer reviews of restaurants in Indonesia is going to carried out. Reviews on the internet usually contains words that are informal and very domain specific. This is where Domain Specific Word embedding can be used to reduce the amount of out-of-vocabulary word (OOV) and give the model more context of the review text given. The model used is Deep Learning with Recurrent Neural Network architecture, using Domain Specific Embedding as Word Embedding, and several attempts to reduce out of vocabulary in the model. The model used is able to reduce OOV from 17.16% (based on previous research) to 3.62%, with an evaluation of the F1-Score model of 79.54% using the Bi-LSTM model.