Technical Summary: SentimentLens

Problem Statement

Online travel platforms generate vast volumes of user-generated hotel reviews, offering rich data for understanding traveler experiences. However, a significant challenge remains in transforming unstructured textual feedback into structured, actionable insights. While structured metadata (e.g., star ratings) provides a quick summary of overall satisfaction, it often obscures the specific reasons behind positive or negative experiences. Conversely, textual reviews contain detailed explanations but are difficult to analyze systematically at scale. Existing Aspect-Based Sentiment Analysis (ABSA) approaches often stop at prediction without integrating results into broader workflows for regional comparison or decision-making. Furthermore, relying on a single modality (either text or ratings) risks missing patterns that only emerge when both signals are combined. There is a lack of unified, end-to-end frameworks that integrate textual sentiment with numerical ratings to enable systematic regional and cross-hotel aggregation for managerial decision support.

Methodology

The paper proposes SentimentLens, a scalable information integration framework designed to reconcile textual sentiment and numerical ratings through a dual-modality approach. The system operates via the following pipeline:

1. Data Collection and Preprocessing:

   • Data is scraped from Google Reviews and TripAdvisor using Apify, covering over 10,000 reviews from 100 hotels across nine provinces in Sri Lanka.
   • The dataset includes structured attributes (hotel name, location, star rating, trip type) and unstructured free-text reviews.
   • Preprocessing involves cleaning, removing duplicates, and filtering records with missing metadata.

2. Text-Based Pipeline (ABSA):

   • Aspect Extraction and Sentiment Classification: The system utilizes Instruct-DeBERTa, a hybrid instruction-based model, to jointly perform Aspect Term Extraction (ATE) and Aspect Sentiment Classification (ASC). This model identifies explicit and implicit aspects and predicts sentiment polarity (positive, neutral, negative) without relying on third-party LLM APIs.
   • Semantic Categorization: Extracted aspect terms are mapped to six standardized service categories (Facilities, Food and Dining, Room Quality, Staff, Location, Booking Process) using centroid-based embedding prototypes derived from GIST Sentence Transformers.
   • Aggregation: The pipeline generates structured triples (aspect, category, sentiment) which are aggregated at the province, hotel, and category levels.

3. Rating-Based Pipeline:

   • Structured star ratings and trip types are analyzed separately.
   • Statistical methods, including one-way ANOVA and Tukey's HSD tests, are used to assess significant differences in ratings across provinces.
   • Chi-square tests of independence analyze the association between province and trip type.

4. Cross-Modal Reconciliation and Analysis:

   • Conflict Identification: The system identifies "latent conflicts" where high overall ratings mask category-specific weaknesses. This is calculated by comparing normalized rating signals against aggregated aspect-level sentiment scores.
   • Metrics: The framework computes Category Importance (frequency of mention), Average Sentiment, Province Competitiveness Index (PCI), and Sentiment Entropy to measure uncertainty in guest reactions.
   • Clustering: K-Means clustering is applied to hotel-level sentiment vectors to identify distinct hotel archetypes (Premium, Mid-tier, Struggling).
   • Opportunity Gap Analysis: An opportunity score is defined by combining category importance with performance gaps to highlight high-impact improvement areas.

Key Contributions

The paper outlines four primary contributions:

1. Unified Framework: The proposal of SentimentLens, a cross-modal reconciliation framework that integrates hotel review text and structured ratings to extract actionable insights at scale.
2. Systematic Alignment: A novel approach for aligning aspect-level sentiment with numerical rating signals, enabling the identification of inconsistencies, latent service gaps, and hidden performance issues that are invisible in single-modality analysis.
3. End-to-End Pipeline: The development of a complete analytical pipeline combining ABSA-based extraction, semantic category alignment, and multi-level aggregation (province, hotel, category) for interpretable analysis.
4. Empirical Validation: A large-scale real-world case study using over 10,000 reviews from Sri Lanka, demonstrating that cross-modal integration improves interpretability and reveals insights not observable from text or ratings alone.

Results

The application of SentimentLens to the Sri Lankan hotel dataset yielded several key findings:

• Global Sentiment Patterns: "Staff" emerged as the strongest performing category globally, while "Booking Process" and "Room Quality" showed weaker sentiment despite high discussion frequency.
• Regional Disparities: Eastern, Central, Western, and Southern provinces demonstrated the strongest overall sentiment and ratings. The Northern Province recorded the weakest performance across both modalities, reflecting issues in service consistency and infrastructure.
• Category Stability: "Staff" and "Location" were found to be the most stable and consistently positive dimensions across provinces. In contrast, "Room Quality," "Facilities," and "Booking Process" exhibited high variability and entropy, indicating inconsistent service delivery.
• Hotel Archetypes: Clustering analysis identified three distinct hotel archetypes: Premium (high satisfaction across all categories), Mid-tier (balanced but moderate performance), and Struggling (low performance, particularly in Room Quality). Struggling hotels were concentrated in the Northern and Sabaragamuwa provinces.
• Cross-Modal Conflicts: The system successfully identified latent conflicts. For example, in the Northern Province, a high normalized rating (0.81) masked a significantly lower sentiment score for "Room Quality" (0.40), revealing a specific operational weakness hidden by the aggregate rating.
• Opportunity Gaps: The analysis highlighted "Room Quality" and "Food and Dining" in underperforming regions (Northern, Sabaragamuwa, North Central) as the highest priority areas for improvement.

Significance and Claims

The paper claims that SentimentLens effectively transforms large-scale unstructured reviews into actionable intelligence, supporting data-driven decision-making for hospitality management and tourism policy. By reconciling textual sentiment with numerical ratings, the system addresses the limitation of relying on a single modality, which can obscure specific service failures.

The authors emphasize that while the system is demonstrated using a national case study in Sri Lanka, it is generalizable to other destinations and review-driven service domains (e.g., restaurants, attractions, airlines). The framework provides a more comprehensive basis for identifying targeted improvement opportunities, particularly in regions where strong overall ratings may mask specific category-level weaknesses. The study concludes that integrating these modalities is essential for uncovering structural inconsistencies in service quality and guiding strategic investments in the hospitality sector.