Multi-Aspect Sentiment Analysis on Turkish Food Delivery Reviews: A Multi-Task BERTurk Approach

Online food-delivery reviews provide valuable feedback for service monitoring and also constitute an important resource for natural language processing research. However, public benchmarks for aspect-based sentiment analysis (ABSA) in Turkish remain scarce. We present a systematic evaluation of the 676,214-comment Yemek Sepeti dataset, in which each review carries three 1-10 ratings for Speed, Service and Taste, under three task formulations: regression (T1), three-class ordinal (T2) and binary polarity (T3). We contrast Term Frequency- Inverse Document Frequency (TF-IDF) baselines with BERTurk single-task and multi-task (MTL) fine-tuning. A shared-encoder MTL with a class-weighted loss raises T2 Macro-F1 to 0.674 / 0.697 / 0.731 and Quadratic Weighted Kappa (QWK) to 0.752 / 0.783 / 0.818, reduces T1 MAE from 1.77 to 1.26, and reaches 92.1% accuracy on T3 Taste. Cross-seed variance is on the order of 10⁻³; a ConvBERTurk ablation confirms backbone independence. Error analysis surfaces three findings: the middle ordinal class is the hardest, performance halves under cross-aspect disagreement, and the corpus contains text-rating contradictions.