<![CDATA[This study aims to evaluate the effectiveness of the Zero-Shot Learning (ZSL) approach using the DeepSeek-R1-Distill-Qwen-1.5B model in performing sentiment classification on Indonesian-language reviews of the DeepSeek AI application from the Google Play Store. A total of 2,000 unlabeled user reviews were collected and processed through instructional prompts to guide the model in classifying sentiments into three categories: positive, negative, and neutral. The model operates without fine-tuning and relies entirely on Zero-Shot Learning using Indonesian-language prompts. Out of 2,000 reviews, 1,348 were successfully classified with valid sentiment labels. Of these, 1,131 reviews (83.9%) were labeled as positive, 211 reviews (15.7%) as negative, and only 6 reviews (0.4%) as neutral. Evaluation results indicated an overall accuracy of 77.67%. The F1-Score for the positive class reached 86.66%, while the negative and neutral classes scored 33.56% and 16.66%, respectively, highlighting the performance disparity between dominant and underrepresented sentiment categories. These findings demonstrate that the DeepSeek-R1 model has strong potential in detecting positive sentiment in Indonesian without requiring additional training. However, its performance on negative and neutral sentiments remains limited, revealing the challenge of handling low-resource and imbalanced data in Zero-Shot settings. Future research should explore improved prompt engineering or multilingual adaptation to address the current limitations and enhance classification consistency across all sentiment categories]]>
Copyrights © 2025
