Enhanced Class Balancing Techniques for Classification of Text in Low Resource Languages

Abstract

The rise of Web 2.0 has empowered individuals to express their ideas on a wide range of social media and online retail platforms. People are increasingly depending on the reviews posted on various platforms before deciding to try a product or service for themselves. The emotions conveyed by individuals regarding a product or service through social media or e-commerce platforms indicate whether they had a positive or negative experience. As a result, sentiment analysis datasets sourced from these platforms have unequal distribution of different sentiment classes leading to the class imbalance problem. Due to this imbalanced class distribution, the learning model s performance will be biased towards the larger class. In our work, we addressed the class imbalance problem in two different domains- Aspect-based Sentiment Analysis and Offensive Content Identification. newlineAspect-based Sentiment Analysis (ABSA) enables a more fine-grained analysis by finding and categorising sentiments for each aspect separately. A more in-depth understanding of the opinions stated in the written reviews is provided by such a fine-grained analysis. The majority of existing research in the field of sentiment analysis is for English language text. However, there is limited research conducted for low-resource languages such as Hindi. A large population of people around the world speak Hindi (Official language of India), yet there is relatively limited data available for developing conversational AI systems. newlineABSA involves two tasks: extracting aspect terms and determining the sentiments expressed towards those terms. In this research, our focus is on sentiment classification of predetermined aspect terms in the Hindi language. We proposed two distinct models, LSTM-based and BERT-based, for aspect-based sentiment classification of Hindi text reviews. In LSTM-based model, we used two LSTM networks for capturing context of words surrounding the aspect term. In BERT-based model, we have finetuned the pre-trained BERT model using one additional laye