Tourism Mining: Aligning Sentiment & Topic Representations of Sentence Transformer Embeddings for Tourism Opinion Mining

ACL ARR 2026 January Submission9186 Authors

Keywords: social media, opinion mining, contrastive learning, dense retrieval, stance detection, sentiment analysis, topic modeling, sentence transformers

Abstract:

We introduce **TourCSE**, a Sentence Transformer tailored for tourism opinion mining, leveraging training data from topic modeling with negative filtering using GISTEmbed. TourCSE enables identification of key aspects in large-scale exploratory analyses.

Beyond improved results, we investigate the sources of these improvements through ablation studies, including Low-Rank Adaptation, dropout augmentation, sentiment sampling, and negative filtering. Our experiments demonstrate the necessity of negative filtering, in combination with either topic modeling or sentiment sampling, to substantially improve topic–sentiment alignment. We also compare TourCSE with other available models in terms of both visualization and retrieval performance, confirming the value of domain-specific models. Finally, we advocate reconsidering the neutral sentiment in current benchmarks.

Paper Type: Long

Research Area: Computational Social Science, Cultural Analytics, and NLP for Social Good

Research Area Keywords: NLP tools for social analysis, contrastive learning, dense retrieval, stance detection, model bias/unfairness mitigation

Contribution Types: Model analysis & interpretability, NLP engineering experiment, Publicly available software and/or pre-trained models

Languages Studied: English

Submission Number: 9186