Sentiment analysis in Algerian dialects considering Code-switching

Abstract

Abstract : Arabic sentiment analysis in the Algerian Dialect considering code-switching is a difficult classification task that can allow us to understand the sentiment between human-generated content in social media. Although, the code-mixed type of content has been on the rise lately, with the continuous rise of social media use among all population’s categories, research around it has been rare due to the difficulty of the problem, and the scarcity of the needed data. Up till now, most of the solutions that have been adapted to deal with the problem need manually annotated data-sets, or pre-trained models with non-contextual word embeddings, due to the hard nature of the written Algerian dialects on social media, which makes their ability to deal with the problem at hand limited. In our work, we present an architecture that aggregates the most suited BERT-Based models for our study of the Algerian dialect, using aggregation methods that help us solve the problem of code-switching in text. The proposed architecture has been evaluated on the code-switched sub-data-set of the CERIST data-set, and on a data-set that has been manually collected, and the results achieved in both show that our architecture achieved good results. Being tested on the CERIST, the manually collected data-sets, our architecture achieved an accuracy of 77% and 64% respectively.