Hybrid Data Driven and Rule Based Sentiment Analysis on Greek Text

Sentiment analysis is a developing field dealing with the detection of sentiments or opinions expressed in a written text. In an age where online services are subject to immediate feedback by users and clients, sentiment analysis on bodies of online reviews is of particular interest to service-oriented businesses such as hospitality establishments. Methods based on Machine Learning are very widely used for this purpose. While some standardized methodologies exist, the sheer volume of data on which the analysis is to be performed in order to train a Deep Learning network, makes it important to come up with designs that are sufficiently complex so as to accurately detect sentiment but also simple enough so as to be scalable and efficient in terms of performance. In this paper, we examine trade-offs of accuracy versus efficiency by performing sentiment analysis on an annotated body of review texts collected from online hotel reserving resources. The reviews we use to train our Deep Learning models are preprocessed by a tool which uses linguistic rules to add sentiment tags to words and expressions in the text. The tags express sentiment polarity, i.e. whether the expression is positive or negative. We propose four different DL network designs, which receive as training input either the review texts, the review texts plus some information on the tag annotation, or the annotation of the text alone, and we present the trade-offs that each setup offers. (C) 2020 The Authors. Published by Elsevier B.V.