Sentiment Analysis with Neural Models for Hungarian

Abstract

Sentiment analysis is a powerful tool to gain insight into the emotional polarity of opinionated texts. Computerized applications can contribute to the establishment of next- generation models that can provide us with data of unprecedented quantity and quality. However, these models often require substantial amount of resources in order to meet the desired performance expectations. Therefore, numerous research efforts are targeted to achieve high-quality results while lowering the resource needs by improving the structure and function of the models used. From a cognitive perspective, it is important to understand the mental state of users when they engage in activities that potentially reflect their feelings and emotions. With the emergence of the widespread use of digital solutions, users post opinionated texts on social media, which can be used as a valuable source to detect their underlying sentiments. Therefore, these platforms offer an unparalleled opportunity to perform sentiment analysis. In recent years, natural language processing tasks, like sentiment analysis, can be solved with high performance, if a pre-trained language model is fine-tuned. Herein we present the first neural transformer-based sentiment analysis model for Hungarian, which achieved state-of-the-art performance. Several limitation factors can occur during fine-tuning, such as the lack of training corpora with appropriate size or the complete absence of usable training material. In our experiment, we use data augmentation methods, specifically machine translation and cross-lingual transfer, to increase the size of our training corpora. Here, we demonstrate our experimentation with 9 different language models. Our work provides evidence for the increased efficiency of the trained models if translation text is added to the training corpora. Furthermore, using the augmentation technique, we could further increase the performance of our models. Consequently, our findings represent an important milestone in the advancement of sentence-level and aspect- based sentiment analysis in the Hungarian language.