Skip to content
Conference Open access

Aspect Count-Guided Adaptive Top-K Correction for Multi-Aspect Sentiment Analysis in Hotel Reviews

Sep 2026 · 2026 8th International Conference on Cybernetics and Intelligent Systems (ICORIS) · pp. 1-6 · 0 citations · 28 references

Abstract

Aspect-Based Sentiment Analysis is important for extracting fine-grained customer insights from hotel reviews, where a single review may contain multiple aspects with different sentiment polarities. Existing multi-label aspect detection methods commonly rely on calibrated probability thresholds or fixed Top-K selection, which do not explicitly use the predicted aspect count to guide inference when aspect density varies. This study proposes an Aspect Count-Guided Adaptive Top-K Correction mechanism as an inference-stage stability and fallback strategy. The framework employs a multi-task BERT-BiLSTM architecture with three prediction heads for aspect detection, aspect count prediction, and aspect-level sentiment classification. During inference, calibrated thresholding first determines the active aspect set. When the number of detected aspects is lower than the predicted count, Adaptive Top-K selectively adds the highest-probability inactive aspects until the predicted aspect density is reached. Experiments compare Threshold Only, Fixed Top-K, and Adaptive Top-K across Location, Service, Price, and Other aspects. Threshold Only and Adaptive Top-K achieve identical best performance, with micro-F1 of 0.9388, macro-F1 of 0.9382, and recall of 1.0000. Adaptive Top-K therefore does not outperform the optimized threshold baseline, but provides a count-aware fallback mechanism expected to be useful when thresholding under-detects active aspects, while avoiding the degradation observed with Fixed Top-K.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.