A hybrid method for enhancing aspect sentiment classification in the french language: combining rule-based and learning methods

The aspect-based sentiment analysis (ABSA) task has become increasingly important to the scientific community with the growth of internet content. ABSA is a subfield of natural language processing (NLP) and it is principally divided into two sub-tasks: aspect extraction (AE) and the aspect sentiment classification (ASC). To accomplish this task, researchers first extract the aspect terms (e.g., camera, design, etc) of an entity (e.g., phone) and then classify them into positive, negative, or neutral. In this study, we focus only on the ASC task for the French language and we introduce a hybrid method that combines the rule-based method with the machine learning-based method. Firstly, we created a set of rules to annotate the dataset. These rules were constructed based on the opinion lexicon, dependency relations parser, and contextual information. Then, we introduced a new neural network architecture called CBGC (CNN-Bi-GRU-CRF), which combines deep learning models CNN (convolutional neural network) and Bi-GRU (bidirectional gated recurrent unit) with the classical machine learning model CRF (conditional random field). This architecture was trained on the annotated dataset to classify aspect terms into positive or negative. To evaluate the performance of the proposed hybrid method, we collected a dataset of 6000 French mobile phone comments from the Amazon website. The obtained results were very encouraging, with an F-measure value of 94.16% for the mobile phone domain and 84.38% for the SemEval-2016 restaurant dataset.