Sentiment Mining of Patient Experience Data Using T5 Architectures
Abstract
The article proposes an algorithm for three-class sentiment analysis of patient reviews of medical institutions in the Uzbek language. The work begins with the formation of a balanced, manually labeled corpus of compact (up to five sentences) user reviews, cleared of personal data. The total volume of the dataset collected by the authors is 12,000 reviews, which were collected from various open platforms. Two equal-sized T5 architectures are considered as a model core: the universal, instruction-tuned Flan-T5 and the domain-specific pretrained Clinical/Med-T5. Both models are retrained using the parameter-efficient LoRA adapters method, which allows keeping the training and inference processes on a single consumer GPU. The experimental part includes a comparison of models, where each of them shows fairly positive ratings. At the same time, the article also provides information on the Uzbek language, its grammar and nature. In addition, the study also provides information on existing similar solutions for both Uzbek and other low-resource Turkic languages.