پاسخ‌گويي به پرسش‌هاي مقايسه‌ايِ ذهني درباره¬ي محصولات: روشي آگاه به شدت ترجيح

Subjective comparative questions,. , Preference Intensity , Comparative opinion mining, . , Question classification,, . , , Extraction of comparative relations, , relevant review retrieva‎l. , Automatic answer generation, . , Deep learning.

Answering Subjective Comparative Questions on Products: an Intensity Preference-Aware Method

مهندسي نرم افزار

Subjective Subjective comparative questions, due to their complex nature an‎d dependence on user preferences, have long posed a major challenge for automatic question-answering systems. However, many prior studies suffer from shortcomings such as ignoring preference intensity an‎d complex comparative relations, an‎d relying on simple majority voting for preference aggregation, which is unstable under data sparsity an‎d conflicting evidence. This work introduces an NLP- an‎d deep-learning-based method for answering such questions. The proposed approach comprises three main stages: question processing, retrieva‎l of relevant reviews, an‎d answer generation. First, using RoBERTa, questions are accurately classified into ten classes along the subjective/objective an‎d comparative/non-comparative axes; then subjective comparative relations—including entities, comparative aspects, constraints, an‎d preferences—are extracted via a chain of RoBERTa-based-go-emotions + MTL + Adapter + NLI + SwapCPC. In the second stage, relation-aware pseudo-neutral query sentences are constructed an‎d semantic similarity to review sentences is computed with stsb-roberta-base-v2 to retrieve relevant evidence. In the third stage (answer generation), retrieved subjective comparative sentences are analyzed an‎d labeled, comparative opinion mining is performed, an‎d user preferences are aggregated to produce short, template-based answers an‎d long, LLM-based answers (via pro‎mp‎ts); GPT-4o is also used for fluency an‎d coherence. The approach is eva‎luated on several task-specific datasets, including FQSD, Smartphone-SCQRE, Bran‎ds-CompSent-19-SCQRE, FSSD, an‎d SCQAD. Results show strong effectiveness an‎d significant improvements over prior work. Concretely, in question processing, the ten-class classifier achieves F1 = 97.1%. In relation extraction (SCQRE), the mean F1 is 85.8% in the staged setup an‎d 72.5% end-to-end; the quintuple eva‎luation in the binary setting attains F1 = 78.8%. In retrieva‎l, the optimal similarity threshold τ = 0.70 yields F1 = 80.3% (staged) an‎d 72.9% (end-to-end). In answering, the preference aggregation method with α = 2 (OWA–Yager) achieves Accuracy = 91.8% an‎d MAE = 0.10; end-to-end these become 70.4% an‎d 0.24. For short answers, EM is 92.3% (staged) vs. 73.6% (end-to-end). For long answers, Accuracy dro‎ps from 88.9% to 70.5%. Moreover, adding the preference intensity component in 80 human pairwise comparisons yields a 12.5% rate of beneficial decision change, improves groundedness from 3.5 to 4.3, an‎d achieves a human preference outcome of 46/14/20 in favor of the intensity-aware version; together with other quantitative evidence, these confirm the effectiveness of the proposed method across all three stages.

6

152089