Weakly-Supervised Text Classification for Estonian Sentiment Analysis
Name
Andreas Pung
Abstract
Text Classification is one of the most fundamental tasks in Natural Language Processing. Hand-labelling texts is costly and might need specialised domain knowledge – this is where unsupervised and weakly-supervised approaches could be useful. In this Master’s Thesis, the weakly-supervised text classification paradigm is used to classify the sentiment of Estonian texts. In this paradigm, the weak labels are created using labelling functions (Ratner et al., 2016). The aim of this thesis is to assess the applicability of weakly-supervised models trained with around 40× larger dataset in contrast to hand-labelling a smaller amount of texts to train a fully-supervised classifier. The compared models are fully and weakly-supervised BERT (Devlin et al., 2019); weakly-supervised COSINE (Yu et al., 2021) and WeaSEL (Cachay et al., 2021). Human evaluation is performed on texts where the models disagreed the most. As a result, we find that the fully-supervised models have the best performance. The best-performing weakly-supervised model trained on the larger dataset had an average classification accuracy of 7.29% worse (7.05% worse weighted F1-score) than the fully-supervised BERT model. The lower performance of weakly-supervised models might be caused by the low quality of labelling functions – developing them further might lead to better results.
Graduation Thesis language
English
Graduation Thesis type
Master - Computer Science
Supervisor(s)
Kairit Sirts
Defence year
2022