Authors - Lord Francis B. Navarro, Chris Jordan G. Aliac, Larmie S. Feliscuzo Abstract - This study benchmarks three Transformer-based encoder models for the sentiment classification stage of an aspect-based sentiment analysis pipeline applied to tourist reviews of the Chocolate Hills Complex in Bohol, Philippines. The work is motivated by the need for tourism analytics that remain usable under the computing constraints of Philippine local government units. A corpus of 5,885 Google Maps and TripAdvisor reviews was cleaned to 3,288 English textual reviews and transformed, through LLM-assisted silver-standard annotation, into 7,555 aspect-sentiment pairs across six tourism aspects and three sentiment classes. Three models — RoBERTa, DistilBERT, and TinyBERT — were finetuned for aspect-conditioned sentiment analysis and compared with TF-IDF baselines. Classification was evaluated on a held-out test set; deployment efficiency was tested on CPU-only hardware using latency, memory footprint, and parameter count. RoBERTa achieved the highest accuracy and macro-F1 but required substantially more memory and higher latency. TinyBERT achieved the lowest latency and memory use while maintaining usable macro-F1, making it the most deployment-practical option under the tested conditions. The results suggest that model selection for local tourism analytics should consider both predictive performance and operational feasibility.
Open Zoom