Breaking the Curse of Class Imbalance: Bangla Text Classification

Abstract

This article addresses the class imbalance issue in a low-resource language called Bengali. As a use-case, we choose one of the most fundamental NLP tasks, i.e., text classification, where we utilize three benchmark text corpora: fake-news dataset, sentiment analysis dataset, and song lyrics dataset. Each of them contains a critical class imbalance. We attempt to tackle the problem by applying several strategies that include data augmentation with synthetic samples via text and embedding generation in order to augment the proportion of the minority samples. Moreover, we apply ensembling of deep learning models by subsetting the majority samples. Additionally, we enforce the focal loss function for class-imbalanced data classification. We also apply the outlier detection technique, data resampling, and hidden feature extraction to improve the minority-f1 score. All of our experimentations are entirely focused on textual content analysis, which results in a more than 90% minority f1 score for each of the three tasks. It is an excellent outcome on such highly class-imbalanced datasets.

Sentiment Analysis in Low-Resource Settings: A Comprehensive Review of Approaches, Languages, and Data Sources

“…Similarly, different low-resource languages suffer from the problem of class imbalance including the Bengali language. [74] tries to resolve the problem by applying word embedding strategies as one of their methods in the sentiment classification task. Three DL models were used for the experiment and evaluation.…”

“…This omission is often attributed to the utilization of online open-source data repositories, where the availability of already generated and preprocessed data obviates the need for explicit preprocessing steps. Studies like [51], [67], [62], [74], highlight the critical role of stemming and lemmatization, particularly in low-resource sentiment analysis scenarios. The absence of adequate tools for stemming and lemmatization in certain languages within NLTK necessitates the development of tailored algorithms for these specific languages.…”

“…Nonetheless, challenges persist, particularly in accurately representing all languages within tokenizer frameworks due to variations in language characters and other linguistic factors. Data cleaning [95], [35], [96], [49], [62], [65], [58], [68], [67], [89], [74], [75], [76], [77], [88], [94], [100], [103], [104], [107], [80], [106], [81], [108], [109], [82], [53], [110], [111], [83], [92], [113], [114], [115], [116], [117] Stemming [51], [67], [104], [108],…”

“…Lemmatization [62], [74], [104], [108] Part-of-speech (POS) [95], [67], [76], [108] Tokenization [47], [48], [51], [67], [74], [94], [100], [104],…”

“…Additionally, one-hot encoding is mentioned in a few like [35] and [74], suggesting its presence in certain studies despite it various limitations. Embedding techniques, including both traditional embeddings and contextual embeddings, emerge as popular choices, with numerous studies highlighting their effectiveness.…”