Aspect Based Sentiment Analysis for Hotel Services in Afaan Oromo Text Using Deep Learning

Aspect-based sentiment analysis (ABSA) also known as phrase level, is a specific level of sentiment analysis task which tries to find the polarity of various aspects of an entity. It identifies aspects with in the given opinion and predicts the polarity corresponding to each aspect with in the text. In this research work, we proposed an aspect-based sentiment analysis for hotel services in Afaan Oromo texts using deep learning approach. Implementing this technique for hotel industries will have many significances. Tracking users’ opinions on the services they provide will help the organization understand their value in a competitive environment and to understand their customer opinion about the services. As there is no publicly available dataset used for this task, we have prepared an aspect-based sentiment analysis dataset from scratch by manually collecting the opinions from hotel customers by distributing a questionnaire. We have collected 1155 datasets for this research work through manually distributed questionnaires. The proposed solution has four steps: manual data collection, preprocessing, aspect category classification, and sentiment polarity classification. The classification process compares five deep learning models, CNN, LSTM, GRU, BiLSTM, and CNN-BiLSTM. In aspect category classification, we have six categories of aspects: Nyaata (Food), Dhugaatii (Drink), Bakka (Location), Gatii (Price), Keessummeessu (Reception), and Ciisicha (Room). In sentiment polarity classification, the polarity of each aspects is classified into either Negative or Positive classes. The result shows that the BiLSTM performed better than all the other models in both aspect category classification and sentiment polarity classification tasks. The BiLSTM model achieved an accuracy of 93.4% and an F1 score of 93% in the aspect category classification tasks. In the polarity classification task, BiLSTM achieved an accuracy of 87% and an F1 score of 88.9%. BiLSTM has been selected as an outperforming model to deploy our solution.