💬 NLP

XAI-enhanced Comparative Opinion Mining via Aspect-based Scoring and Semantic Reasoning

This paper introduces XCom, an interpretable transformer-based model that combines aspect-based rating prediction with semantic reasoning and Shapley additive explanations to enhance trust and performance in comparative opinion mining.

Original authors: Ngoc-Quang Le, T. Thanh-Lam Nguyen, Quoc-Trung Phu, Thi-Phuong Le, Duy-Cat Can, Hoang-Quynh Le

Published 2026-03-03

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ngoc-Quang Le, T. Thanh-Lam Nguyen, Quoc-Trung Phu, Thi-Phuong Le, Duy-Cat Can, Hoang-Quynh Le

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are standing in a massive beer aisle, overwhelmed by hundreds of choices. You want to know if Beer A is better than Beer B, but you don't have time to read every single review. You turn to the internet, hoping to find a clear answer.

Here's the problem: Most computer programs that try to help you are like magic 8-balls. They look at thousands of reviews and spit out an answer like, "Beer A is better!" But they can't tell you why. They are "black boxes"—you get the result, but the reasoning is hidden inside a locked room. This makes it hard to trust them.

This paper introduces XCom, a new tool designed to be a transparent, super-smart shopping assistant that not only tells you which beer is better but also shows you its "workings" so you can trust the advice.

Here is how XCom works, broken down into simple steps:

1. The "Same-User" Rule (The Fair Judge)

Imagine you are comparing two judges.

Judge A is a "strict" critic. If they say a beer is "Good," it's actually amazing.
Judge B is a "lenient" critic. If they say a beer is "Amazing," it might just be okay.

If you compare Judge A's "Good" beer against Judge B's "Amazing" beer, you might get the wrong answer because their standards are different.

XCom's Solution: It only compares reviews written by the same person. It's like asking one person, "Is your favorite movie better than your second favorite?" By keeping the "judge" the same, we remove the confusion of different personal styles and get a fair comparison.

2. The "Surgical" Approach (Aspect-Based Mining)

Reviews are messy. A user might say, "The beer tastes great, but the bottle is ugly."
Old computers might get confused and try to compare the taste of Beer A with the bottle of Beer B.

XCom's Solution: It acts like a surgical team with four specialized doctors:

Doctor Taste: Only looks at sentences about flavor.
Doctor Smell: Only looks at sentences about aroma.
Doctor Look: Only looks at appearance.
Doctor Mouthfeel: Only looks at how it feels to drink.

It slices the reviews into neat little pieces, ensuring that when it compares "Taste," it's only comparing "Taste."

3. The "Double-Check" System (Two Brains)

Once the reviews are sliced up, XCom uses two different "brains" to decide the winner for each category:

Brain 1 (The Scorekeeper): It looks for specific "score words" (like sweet, bitter, smooth) and calculates a mathematical score.
Brain 2 (The Storyteller): It reads the whole sentence to understand the vibe and context, even if there are no obvious score words.

It combines the results of both brains to make a final decision. This makes it much harder to fool the system.

4. The "Flashlight" (Explainable AI)

This is the most important part. When XCom says, "Beer A has a better taste," it doesn't just stop there. It turns on a flashlight (using a technique called SHAP) to show you exactly which words made it decide that.

Example: If the text says, "The taste is wonderful and balanced," XCom highlights those words in Red (meaning: "These words made me say 'Better'").
The Twist: If the text says, "The beautiful taste," XCom might highlight "beautiful" in Blue (meaning: "This word is usually for looks, not taste, so I'm ignoring it or thinking it's a mistake").

This transparency lets you see the logic. You can say, "Ah, I see why it chose Beer A; it liked the word 'balanced'."

Why Does This Matter?

Trust: You aren't just taking a computer's word for it; you can see the evidence.
Accuracy: By focusing on one person's opinion at a time and separating different parts of the product (taste vs. look), it avoids common mistakes.
Efficiency: It's a small, fast computer program, not a massive, energy-hungry giant. It does the job quickly without needing a supercomputer.

In short: XCom is like a honest, detail-oriented friend who reads all the reviews for you, compares them fairly, and then points to the exact sentences that convinced them, so you can make the best choice with confidence.

1. Problem Statement

The paper addresses the challenge of Comparative Opinion Mining (COM), specifically the task of inferring preferences between two products based on user reviews.

Implicit vs. Explicit: While explicit comparisons (e.g., "Product A is better than B") are easy to detect, most user reviews contain implicit comparisons distributed across separate texts. Inferring these requires contextual reasoning.
Same-User Constraint: To mitigate noise from individual user biases (e.g., rating leniency, writing style), the study focuses on comparing two reviews written by the same user. This isolates the preference signal against a stable baseline.
The Black-Box Issue: Existing transformer-based models for COM often lack transparency. They provide a comparative conclusion without explaining why or which aspects drove the decision, hindering user trust and adoption in high-stakes decision-making.

2. Methodology: The XCom Framework

The authors propose XCom, a hybrid architecture that integrates aspect-based sentiment analysis with semantic reasoning and Explainable AI (XAI). The model operates in three main phases:

A. Aspect-Based Pre-processing

Input: A pair of reviews from the same user.
Process: The model uses four distinct BERT-based classifiers (one for each aspect: appearance, aroma, palate, taste) to filter and extract sentences relevant to specific aspects.
Output: Structured tuples of sentence pairs $(s_1, s_2)$ for each aspect, where $s_1$ and $s_2$ are the extracted sentences from the first and second reviews, respectively.

B. Comparative Opinion Classification (Dual-Path Architecture)

XCom employs two parallel classifiers to generate a final prediction:

Score-Based Classifier:
- Calculates a rating score for each sentence.
- Adjective-based: Averages sentiment scores of adjectives in the sentence.
- Fallback: If no adjectives are found, it uses TF-IDF features fed into an XGBoost regressor to predict a score.
- Aggregation: The aspect rating for a review is the minimum score among its sentences (a conservative approach).
- Classification: The concatenated scores of both reviews are fed into a Transformer encoder to predict a comparative class (Better, Worse, or Similar).
Semantic-Based Classifier:
- Directly processes the raw sentence pairs using a Transformer encoder to capture semantic relationships and context that pure scoring might miss.
- Outputs a probability distribution over the three comparative classes.
Fusion: The final prediction is derived by summing the probability distributions from both classifiers ( $p_{final} = p_{score} + p_{semantic}$ ) and selecting the class with the highest probability.

C. SHAP-Based Explanation (XAI)

Mechanism: XCom integrates SHapley Additive exPlanations (SHAP) to attribute feature importance.
Process: It treats input tokens as features. By systematically masking tokens (replacing them with [MASK]) and observing the change in model output, it calculates Shapley values.
Goal: To visualize which specific words (tokens) positively or negatively influenced the comparative decision, transforming the model from a "black box" to a "white box."

3. Key Contributions

XCom Architecture: A novel dual-path model combining aspect-based rating prediction (quantitative) and semantic reasoning (qualitative) specifically designed for implicit, same-user comparative mining.
Explainability Integration: The first application of SHAP to comparative opinion mining, providing token-level insights into why a model prefers one product over another.
Same-User Focus: A rigorous approach to eliminating inter-user variance (bias, style) by strictly comparing reviews from the same individual.
Efficiency: The model achieves high performance using BERT-based encoders (~110M parameters), significantly outperforming massive Large Language Models (LLMs) while requiring far less computational resources (10 hours training vs. 40–80 hours for LLMs).

4. Experimental Results

Dataset: The model was evaluated on the SUDO dataset (beer reviews), which includes four aspects: appearance, aroma, palate, and taste.
Performance:
- XCom achieved a Macro-averaged F1 score of 58.56%, outperforming the second-best baseline (Finetuned-T5) by nearly 3 points.
- It significantly outperformed general-purpose LLMs (Llama-3, Gemini, ChatGPT), which scored between 25% and 49% F1, demonstrating that specialized architectures are superior for fine-grained comparative tasks.
- Per-Aspect Performance: Best results were seen in Appearance and Taste (clearer linguistic cues), while Palate was the most challenging due to subtle expressions.
Explainability Validation:
- SHAP visualizations correctly identified context-dependent sentiment (e.g., "beautiful" was negative for Taste but positive for Appearance).
- Faithfulness Test: Removing top-k tokens identified by SHAP caused an immediate drop in model performance, confirming that the explanations accurately reflect the model's decision logic.

5. Significance and Limitations

Significance:

Trustworthy AI: By providing interpretable reasons for comparative decisions, XCom bridges the gap between high-performance AI and user trust.
Resource Efficiency: It proves that specialized, smaller models can outperform massive LLMs in niche domains without the heavy computational cost.
Decision Support: The ability to highlight specific influential tokens helps users understand the rationale behind product rankings.

Limitations & Future Work:

Cascading Errors: The modular design (pre-processing $\to$ scoring $\to$ classification) risks error propagation if the aspect classification fails. Future work suggests a joint learning framework.
Interpretability for Non-Experts: SHAP visualizations are technical; future work aims to make these more readable for lay users.
Cross-Review Alignment: Current SHAP highlights important terms within a review but does not explicitly align contrasting terms between the two reviews. A graph-based approach is proposed to fix this.
Domain Generalization: The model was tested primarily on beer reviews; generalization to other product domains requires further validation.
Informal Language: The model struggles with slang and informal writing styles common in social media.

In conclusion, XCom represents a significant step forward in making comparative opinion mining both accurate and interpretable, offering a robust framework for analyzing implicit user preferences in a transparent manner.