Aspect Based Multi-Class Sentiment Dataset for Bilingual eWOM of Commercial Food Products

Abstract

Aspect-Based Sentiment Analysis for product review opinion analysis is commonly utilized by the commercial food products manufacturing businesses to drive decisions regarding products. However, the general consumers are not facilitated with decision making ready end-user applications which generates insights to arrive at the purchase decision at the time of purchase due to the unavailability of products’ attribute-wise analysis-ready data. Although Electronic Word of Mouth (eWOM) platforms are comprised of opinions with a diversity of languages and expression formats, themselves do not generate any value to make comparable decision making. Hence, there is an existing gap of impactful information retrieval by the consumer to aid the purchase. Therefore, creating a publicly available analysis-ready dataset for the commercial food product domain contributes significantly to the Sri Lankan consumers and Government organizations. Through our research work, a manually annotated bilingual eWOM opinion text dataset for selected commercial food products categories has been delivered in which the opinions expressed in the Sinhala language have been translated into English language and each opinion has been manually rated into five levels by two domain experts. Two product attributes, “Price of the product”, “Safeness of product” have been considered as aspects to conduct the Aspect-Based Sentiment Analysis. This study describes the sub-tasks performed to conduct the Aspect- Based Sentiment Analysis on the dataset along with the basic statistical evaluation of the dataset. We have presented results on the performance of the dataset by utilizing an existing Long Short-Term Memory Model.