A Hybrid Method based on Rules and Deep Learning for Enhancing Single-Word and Multi-Word Aspects Extraction from French Reviews

In recent years, Aspect Based Sentiment Analysis (ABSA) has gained significant importance, particularly for enterprises operating in the commercial domain. These enterprises tend to analyze the customers’ opinions concerning the different aspects of their products. The primary objective of ABSA is to first identify the aspects (such as battery) associated with a given product (such as a smartphone) and then assign a sentiment polarity to each aspect. In this paper, we focus on the Aspects Extraction (AE) task, specifically for the French language. Previous research studies have mainly focused on the extraction of single-word aspects without giving significant attention to the multi-word aspects. To address this issue, we propose a hybrid method that combines linguistic knowledge-based methods with deep learning-based methods to identify both single-word aspects and multi-word aspects. Firstly, we combined a set of rules with a deep learning-based model to extract the candidate aspects. Subsequently, we introduced a new filtering algorithm to detect the single-word aspect terms. Finally, we created a set of 52 patterns to extract the multi-word aspect terms. To evaluate the performance of the proposed hybrid method, we collected a dataset of 2400 French mobile phone comments from the Amazon website. The final outcome proves the encouraging results of the proposed hybrid method for both mobile phones (F-measure value: 87.27% for single-word aspects and 82.38% for multi-word aspects) and restaurants (F-measure value: 78.79% for single-word aspects and 76.04% for multi-word aspects) domains. By highlighting the practical implications of these results, our hybrid method offers a promising outlook for Aspect Based Sentiment Analysis task, opening new avenues for businesses and future research.