SentimentLens: Reconciling Sentiment and Ratings via Dual-Modality in the Hospitality Sector

Abstract: Online travel platforms generate vast volumes of user-generated hotel reviews, offering rich opportunities to understand traveler experiences at scale. However, transforming unstructured textual feedback into structured, actionable insights remains a challenging task. This paper presents SentimentLens, a scalable analysis system based on Aspect-Based Sentiment Analysis that performs knowledge extraction from unstructured hotel reviews and organizes them into interpretable service categories. SentimentLens integrates aspect term extraction, aspect sentiment classification, semantic category assignment, and multi-level analytical modules to support region-level, hotel-level, and category-level evaluation. The system is designed to operate across different geographic contexts and hospitality settings. To demonstrate its practical utility, we apply SentimentLens to a large real-world dataset of over 10,000 publicly available hotel reviews. Through extensive analysis, the framework reveals how traveler sentiment varies across regions, service categories, and hotel archetypes. We further implement a cross-modal reconciliation of textual sentiment and numerical ratings to identify latent operational conflicts, structural inconsistencies in service quality, and high-impact improvement opportunities using importance--performance and entropy-based analyses. The results show that SentimentLens effectively transforms large-scale unstructured reviews into actionable intelligence, supporting data-driven decision-making for hospitality management and tourism policy. While demonstrated using a national case study, the proposed system is generalizable to other destinations and review-driven service domains.

• The paper introduces a self-hosted framework that combines Instruct-DeBERTa aspect-based sentiment analysis with rating and trip-type data across 10,000+ reviews, 100 Sri Lankan hotels, and 46,565 aspect mentions.
• The analysis finds Staff and Location are the strongest and most consistent service categories, while Booking Process, Room Quality, and Food and Dining show weaker or more variable sentiment, particularly in Northern Province.
• The paper demonstrates that high overall ratings can conceal category-level problems, identifying gaps such as Northern Province’s 0.81 normalized rating versus 0.40 Room Quality sentiment and supporting targeted investment decisions.

SentimentLens is an end-to-end analytical framework that fuses aspect-based sentiment analysis (ABSA) of hotel review text with structured rating and trip-type metadata to produce region-level, hotel-level, and category-level decision support for the hospitality sector (2606.00084). The system is demonstrated on a corpus of more than 10,000 publicly available reviews from 100 hotels across all nine provinces of Sri Lanka, yielding 46,565 aspect-level sentiment mentions. Its central premise is that star ratings summarize overall satisfaction but conceal category-specific weaknesses, while textual reviews expose those weaknesses but are difficult to aggregate systematically; only a joint, cross-modal treatment reveals both.

Motivation and positioning

The authors argue that prior ABSA research in hospitality is dominated by isolated modeling tasks—aspect extraction, sentiment classification, transformer comparisons—that stop short of integration into managerial workflows. Existing studies rarely perform systematic regional or cross-hotel aggregation, and unified pipelines that combine textual sentiment, numerical ratings, and multi-level aggregation remain scarce. SentimentLens addresses this gap by coupling ABSA outputs with rating analytics through an explicit reconciliation layer that flags discrepancies between the two modalities.

System architecture

The text pipeline uses Instruct-DeBERTa, a hybrid instruction-based model that jointly performs Aspect Term Extraction (ATE) and Aspect Sentiment Classification (ASC), handling both explicit and implicit aspects within a single framework. Extracted aspect terms are mapped to six standardized service categories—Facilities, Food and Dining, Room Quality, Staff, Location, and Booking Process—via centroid-based embedding prototypes built from GIST Sentence Transformer embeddings of curated category keywords, with assignment by cosine similarity above a confidence threshold. Notably, the entire pipeline is self-hosted locally without third-party LLM APIs, which the authors present as enabling scalable, cost-free, privacy-preserving analysis.

The rating pipeline analyzes star-rating distributions and traveler trip types with province-wise comparisons, one-way ANOVA, Tukey HSD post-hoc tests, and chi-square tests of independence. A fusion layer then reconciles the two modalities: conflicts are defined as hotel–category pairs where the absolute difference between the min–max normalized overall rating $R_{\mathrm{norm}}$ and the category-level sentiment score $S_a$ exceeds a threshold $\tau$.

Several derived indices support interpretation: a Province Competitiveness Index weighting category sentiment by mention frequency; sentiment entropy per category as a measure of experiential uncertainty; K-Means clustering of hotels on standardized category-sentiment vectors for archetype discovery (with $k=3$ selected via the elbow method); and an opportunity score defined as the gap between normalized category importance and normalized sentiment.

Sentiment-based findings

Staff dominates guest discourse (12,685 mentions) and achieves the highest average sentiment (0.864), followed closely by Location (0.836). Booking Process receives the lowest average sentiment (0.556) despite being least discussed (1,398 mentions). Province-level rankings place Eastern (0.821), Central (0.817), and Western (0.809) Provinces at the top and Northern Province last (0.580). No single province leads every dimension: Central Province excels in Location (0.908), Facilities (0.823), and Booking Process (0.680), while Eastern Province leads in Staff (0.907) and Food and Dining (0.801).

Stability analysis reinforces these patterns. Staff exhibits the lowest sentiment entropy (0.308) and the tightest cross-province distribution, making it the most predictable national strength; Location follows (entropy 0.450). Conversely, Booking Process has the highest entropy (0.746), with Room Quality and Food and Dining also high (0.649 each), marking infrastructure-related categories as both weaker and less consistent. Co-occurrence analysis shows Staff strongly linked with Food and Dining (5,793 co-mentions) and Room Quality (4,508 co-mentions), suggesting bundled intervention opportunities since improvements in one category may spill over into others.

Hotel-level clustering yields three interpretable archetypes: premium (strong across all categories, including Booking Process at 0.753), mid-range dependable, and low-performing or inconsistent properties whose mean Room Quality sentiment falls slightly below zero (−0.022). Low-performing hotels concentrate in Northern and Sabaragamuwa Provinces. The importance–performance analysis identifies Room Quality and Food and Dining in Northern, North Central, and Sabaragamuwa Provinces as the highest-priority opportunity gaps.

Rating-based findings

Ratings are heavily skewed toward the upper scale, yet provincial differences are statistically significant (ANOVA F = 50.934, p < 0.001). Central Province records the highest average rating (4.79) and Northern Province the lowest (4.24); Tukey HSD confirms Northern Province is significantly lower-rated than most other provinces, while top performers do not significantly differ among themselves. By contrast, ratings do not differ significantly across trip types (F = 1.199, p = 0.309), indicating that satisfaction is driven more by regional and hotel-specific factors than by traveler purpose. However, province and trip type are strongly associated ($\chi^2 \approx$ 623.72, p < 0.001): leisure provinces attract families and couples, whereas Western Province concentrates business travelers. The implication is that tourism strategy should be segmented geographically even though average satisfaction levels are uniform across segments.

Cross-modal reconciliation

The fusion layer substantiates the paper's core claim that rating-only analysis masks operational weaknesses. In Northern Province, a normalized rating of 0.81 coexists with Room Quality sentiment of only 0.40—a gap of 0.41—and Facilities sentiment of 0.46 (gap 0.35); Sabaragamuwa Province shows a 0.24 gap for Food and Dining against a normalized rating of 0.85. These latent conflicts demonstrate that high aggregate ratings can conceal recurring category-specific dissatisfaction that only aspect-level text analysis exposes. This is the paper's strongest empirical argument for dual-modality integration, and it aligns with the broader finding that the two modalities agree on regional rankings while diverging at finer granularity.

Limitations and open questions

The paper concedes several constraints. The case study is confined to Sri Lanka and to reviews scraped via Apify from Google Reviews and TripAdvisor, with the latest 100 reviews per hotel—a sampling scheme that may bias toward recent conditions and platform-specific populations. The conflict detection threshold $\tau$ is empirically selected rather than validated, leaving sensitivity of the conflict set unquantified. Category assignment depends on curated keyword prototypes and a cosine-similarity threshold, so aspects falling below the threshold default to "other," and the taxonomy of six categories may not capture all service dimensions. Sentiment scores are mapped coarsely to {−1, 0, +1}, discarding intensity information. The authors also note that generalization to other regions, languages, and domains (restaurants, airlines, attractions) is asserted rather than demonstrated, and multilingual operation remains future work. Open questions include whether the identified archetypes and opportunity rankings replicate under real-time data streams, whether predictive modeling can be layered onto the reconciliation framework, and how robust the conflict mechanism is to alternative thresholds and category taxonomies.

Conclusion

SentimentLens contributes a unified, self-hosted pipeline that links ABSA-derived aspect sentiment with structured rating analytics through explicit cross-modal conflict identification, importance–performance analysis, entropy-based stability measures, and hotel archetype clustering. Applied to over 10,000 Sri Lankan hotel reviews, it shows that Staff and Location are stable national strengths while Room Quality, Facilities, and Booking Process are weak, inconsistent, and regionally variable—with Northern Province flagged as the most under-served market. The principal empirical contribution is demonstrating that high overall ratings can mask category-level weaknesses detectable only through joint modality analysis, providing a concrete basis for targeted hospitality investment decisions.