Visual Modality Dominance in Multimodal Hospitality Review Sentiment Analysis: A Comprehensive Baseline Evaluation on Tamil-English CHMR-1000
Abstract: Multimodal sentiment analysis of Tamil-English hospitality video reviews requires concurrent feature representations from text, audio, visual, and contextual signals. This paper presents a systematic four-modality feature extraction framework and comprehensive baseline classification study on CHMR-1000, a 906-record Tamil-English hospitality video review corpus spanning nine sub-domains with five concurrent modalities. XLM-RoBERTa text embeddings (768 dimensions), librosa MFCC audio features (120 dimensions), CLIP ViT-B/32 visual embeddings (512 dimensions), and YouTube engagement metadata (20 dimensions) are extracted and evaluated using ten machine learning classifiers across eleven unimodal and multimodal feature combinations with SMOTE class-imbalance correction on 87 training and 43 test gold-standard records. The key finding is that visual CLIP features achieve the highest unimodal classification accuracy (76.7%), surpassing XLM-RoBERTa text embeddings (74.4%) by 2.3 percentage points, demonstrating that language-independent visual sentiment signals captured from video keyframes are competitive with state-of-the-art cross-lingual text models. Text features achieve the highest Macro F1 (0.4119 via LDA), indicating complementary modality strengths with direct implications for intelligent hospitality AI system design.
Keywords: Multimodal sentiment analysis; Tamil-English NLP; CLIP visual features; MFCC audio features; XLM-RoBERTa.
How to Cite?: A. Martina Betsy, Dr. Sumathy Kingslin, "Visual Modality Dominance in Multimodal Hospitality Review Sentiment Analysis: A Comprehensive Baseline Evaluation on Tamil-English CHMR-1000", Volume 15 Issue 11, November 2026, International Journal of Science and Research (IJSR), Pages: 809-813, https://www.ijsr.net/getabstract.php?paperid=SR261119130858, DOI: https://dx.doi.org/10.21275/SR261119130858