XAI-enhanced Comparative Opinion Mining via Aspect-based Scoring and Semantic Reasoning
Authors: Ngoc-Quang Le, T. Thanh-Lam Nguyen, Quoc-Trung Phu, Thi-Phuong Le, Duy-Cat Can, Hoang-Quynh Le
Organizations: VNU University of Engineering and Technology, 144 Xuan Thuy, Cau Giay, Hanoi, Vietnam · School of Computing and Information Systems, Singapore Management University, 81 Victoria Street, Singapore 188065, Singapore · Centre Hospitalier Universitaire Vaudois, Rue du Bugnon 46, 1011 Lausanne, Switzerland · University of Lausanne, Quartier UNIL-Centre, 1015 Lausanne, Switzerland
Comparative opinion mining involves comparing products from different reviews. However, transformer-based models designed for this task often lack transparency, which can adversely hinder the development of trust in users. In this paper, we propose XCom, an enhanced transformer-based model separated into two principal modules, i.e., (i) aspect-based rating prediction and (ii) semantic analysis for comparative opinion mining. XCom also incorporates a Shapley additive explanations module to provide interpretable insights into the model's deliberative decisions. Empirically, XCom achieves leading performances compared to other baselines, which demonstrates its effectiveness in providing meaningful explanations, making it a more reliable tool for comparative opinion mining. Source code is available at: https://anonymous.4open.science/r/XCom.
Figures & tables
Figure 1: Illustration of explicit vs. implicit comparative opinions in user reviews. Explicit: Direct comparison using terms like “ is richer in taste than ”. Implicit: Two separate reviews are from the same user. Preference is implied, and not stated, which is harder to detect. Insufficient information in the cross-user setting: Two reviews are from different users. Comparisons are unreliable due to subjective differences in user standards (e.g., leniency, writing style), lacking a common baseline for inference.
Figure 2: XCom architecture.
Figure 3: Distribution of aspect and comparative labels.
Models
Micro-averaged
Macro-averaged
Prec.
Rec.
F1
Prec.
Rec.
F1
XCom
60.51 ±0.8
56.40±0.7
58.38 ±0.9
61.92 ±1.0
56.11±0.7
58.56 ±0.8
Feature-based baselines
FastText+SVM
52.57±1.0
49.15±0.8
50.08±1.1
52.83±0.9
49.76±0.8
51.05±1.0
FastText+XGBoost
52.03±1.1
42.28±1.0
46.65±0.8
53.65±1.0
41.38±0.7
46.22±1.1
Transformer-based baselines
Table 1: Model overall performance and comparison.
XCom
Micro-averaged
Macro-averaged
Prec.
Rec.
F1
Prec.
Rec.
F1
(A) Per-aspect performance
Appearance
61.28±0.9
60.59±0.8
60.93±0.7
63.58±1.0
59.98±0.9
61.18±0.8
Aroma
60.93±0.8
55.27±0.9
57.96±1.1
61.43±0.9
53.57±0.8
56.72±0.7
Palate
56.25±1.0
44.75±1.1
49.85±0.9
64.56±1.1
42.79±0.7
48.23±0.8
Taste
61.37±0.7
59.73±0.8
60.54±0.6
61.81±1.0
60.71±0.9
60.98±0.8
Table 2: Detailed model results of XCom .
Figure 4: The SHAP-based explanation examples for taste and appearance aspects. In our SHAP-based visualization, red tokens indicate a positive impact to the model’s output, blue tokens indicate negative impact. The arrow length represents the absolute SHAP value. The following highlighted text provides a more intuitive explanation by marking key tokens in the original text. More intense colors signify stronger influence.
Figure 5: SHAP faithfulness evaluation: Macro-F1 drop when removing top-k vs. bottom-k adjectives.
Figure 6: Ablation study showing the F1 drop when removing each model component.
Error type
Cause
Representative examples
Over-extraction of aspect-related sentence
Model generalizes beyond the intended scope of aspect-related content.
“This beer is very easy to drink” — General statement, not truly about a specific aspect.
Under-extraction of aspect-related sentence
Aspect reference is expressed indirectly or contextually, not explicitly.
“I like to get it from time to time when I want something with style that goes down smooth and tasty ” — Missed implicit aspect.
Aspect confusion
A term may refer to multiple aspect categories.
“ Hops is in the background however noticeable” — Confused between aroma and taste.
Aspect rating overestimation
Lexical complexity and implicit contrast between co-occurring aspects.
“The lace is amazing and makes this beer almost as good to watch as to taste (well, not quite but you get the idea)” — Model overestimates sentiment for taste.
Incorrect rating prediction
Failure to correctly interpret contrastive expressions.
“It’s sweet but not cloying . In fact, the sweetness comes in and out…” — Model predicts low score, missing contrast from “not cloying”.
Out-of-vocabulary word
The word is missing from the rating dictionary.
“The nose is accentuated with a slight hint of clove” — “Accentuated” absent from dictionary, causing mis-scoring.
Table 3: Error analysis and representative cases of XCom .