Listwise Preference Optimization with Element-wise Confusions for Aspect Sentiment Quad Prediction

TL;DR

<3-5 sentence high-level summary> ASQP requires extracting a four-element sentiment quadruple (a, c, o, s) per sentence, but existing marker-based supervised fine-tuning struggles with inter-element dependencies and interpretability. The authors propose a reasoning-based generation framework that outputs both the quadruple and a synchronized natural-language rationale, coupled with Element-wise Confusable Candidates and a listwise preference optimization (E4L) to enforce structural validity and relational coherence. They build confusable candidates using syntactic distance and semantic similarity, training with a listwise objective that ranks the gold output above all confusions. Across four benchmark datasets, E4L consistently surpasses non-generative, generative, and collaborative baselines, achieving higher quad F1 and more coherent explanations, demonstrating improved robustness and interpretability in ASQP tasks.

Abstract

Aspect sentiment quad prediction (ASQP) is inherently challenging to predict a structured quadruple with four core sentiment elements, including aspect term (a), aspect category (c), opinion term (o), and sentiment polarity (s). Prior methods relying on marker-based prediction struggle with modeling the intricate relationships among elements and experience sharp performance declines when predicting higher-order elements (e.g., c and s) under standard supervised fine-tuning. To address these limitations, we employ reasoning-based generation to output both the quadruple and a natural language rationale under element prefixes within a unified template, encouraging explicit relational reasoning and interpretability. To further enhance element-wise alignment, we introduce a listwise preference optimization framework for improving structural validity and relational coherence. Specifically, we generate element-wise confusable candidates via syntactic and semantic proximity, then train the model with listwise objectives to prefer the gold candidates over closely competing alternatives. Extensive experiments on four benchmark datasets demonstrate that our framework effectively improves quadruple prediction accuracy and explanation consistency.

Figures (7)

• Figure 1: An example for the ASQP task. Previous methods remain maker-based prediction, where the extracted quads may be structurally valid but semantically incoherent, reflected as incorrect relationships in rationales.
• Figure 2: An overview of our proposed method (E4L). We first curate element-wise candidates via syntactic and semantic proximity, then compose them into listwise candidates for preference optimization.
• Figure 3: (a) Performance degradation as output structure complexity increases under standard supervised fine-tuning (SFT). A sharp decline is observed when expanding prediction targets from pairs (a, o) to triplets (a, o, c). (b) Improvement in target token probability for the challenging higher-order (e.g., $c$) element achieved by our method, compared to SFT-only with our reasoning-based template.
• Figure 4: Error analysis on ASQP-Rest16. Figure (a) demonstrates the proportion of coarse-grained error types. Figures (b) and (c) provide the fine-grained error breakdown for the two main error types.
• Figure 5: Heatmaps for beta at best F1 score under different learning rate (lr) and $\lambda$ settings on ASQP-Rest15 and ACOS-Rest16 datasets.