Construction and Evaluation of Sentiment Datasets for Low-Resource Languages: The Case of Uzbek

To our knowledge, the majority of human language processing technologies for low-resource languages don't have well-established linguistic resources for the development of sentiment analysis applications. Therefore, it is in dire need of such tools and resources to overcome the NLP barriers, so that, low-resource languages can deliver more benefits. In this paper, we fill that gap by providing its first annotated corpora for Uzbek language polarity classification. Our methodology considers collecting a medium-size manually annotated dataset and a larger-size dataset automatically translated from existing resources. Then, we use these datasets to train what, to our knowledge, are the first sentiment analysis models on the Uzbek language, using both traditional machine learning techniques and recent deep learning models. Both sets of techniques achieve similar accuracy (the best model on the manually annotated test set is a convolutional neural network with 88.89% accuracy, and on the translated set, a logistic regression with 89.56% accuracy); with the accuracy of the deep learning models being limited by the quality of available pre-trained word embeddings.