PaperlayerView exact revision on arXiv

Source-linked AI summary

STAR-OPD: Structured Aspect-Cascade-Aware On-Policy Reward Distillation for ABSA Quadruple Extraction

Tong Sun, Mingyang Ma, Jiayang Yu

arXiv:2608.20831v1cs.CLcs.AI

TL;DR

Distilling strong but costly chain-of-thought models for ABSA quadruple extraction remains difficult because teacher-only trajectories provide limited supervision for student-induced structural states. STAR-OPD trains on student rollouts with cascade-aware rewards, and consistently improves over baselines across two benchmarks while reducing target hallucination from 9.75% to 7.22% with Qwen3-4B.

  • Problem

    Off-policy distillation trains on teacher-generated trajectories, providing limited supervision for student-induced structural states in ABSA quadruple extraction, while costly CoT models motivate smaller deployable systems.

    Source: §1 · p. 1, §4.2 · p. 4, §6 · p. 8

  • Method

    STAR-OPD optimizes student rollouts using cascade-aware rewards for target–aspect binding consistency, target grounding, and aspect disambiguation.

    Source: §1 · p. 2

  • Results

    Across E-ABSA20K and SemEval-2014, STAR-OPD consistently improves over off-policy and general on-policy baselines; with Qwen3-4B, target hallucination drops from 9.75% to 7.22%.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The findings suggest that structured-extraction distillation should be designed around inference-time student states rather than teacher-only trajectories.

    Source: §6 · p. 8

    STAR-OPD’s generality beyond ABSA quadruple extraction, especially e-commerce reviews, remains to be validated.

    Source: §7 · p. 9

Abstract

from arXiv · show

Aspect-based sentiment analysis (ABSA) quadruple extraction requires jointly predicting target, aspect, opinion, and sentiment over reviews that often contain multiple fine-grained sentiment tuples. While large chain-of-thought (CoT) models perform well on this task, distilling them into smaller deployable models remains difficult. We identify a task-specific failure mode in distilled ABSA extraction: student errors at the target-aspect interface create structurally invalid states, such as broken target-aspect bindings and hallucinated targets, which then corrupt downstream predictions. Conventional off-policy distillation is poorly suited to this setting because it trains only on teacher-generated trajectories and provides little supervision on the student-induced structural states that dominate inference. To address this mismatch, we propose STAR-OPD (STructured Aspect-cascade-aware On-Policy Reward Distillation), which builds on generic on-policy distillation and instantiates it for ABSA quadruple extraction with cascade-aware, set-structured rewards. STAR-OPD trains on student rollouts and applies set-structured rewards that directly target binding consistency, target grounding, and fine-grained aspect disambiguation. Experiments on E-ABSA20K and SemEval-2014 show that STAR-OPD consistently outperforms off-policy and general on-policy baselines, reduces target hallucination, and substantially improves performance on structurally hard cases. With Qwen3-4B, STAR-OPD substantially narrows the student-teacher gap while improving inference efficiency, highlighting the importance of on-policy structural correction for distilled ABSA extraction.

1 Introduction

STAR-OPD addresses structured aspect cascade in ABSA quadruple extraction, where target–aspect errors create invalid tuples and off-policy distillation misses student-induced states. It trains on student rollouts with cascade-aware rewards, and the introduction reports improvements over off-policy and general on-policy baselines.

1 Introduction

  • ABSA quadruple extraction predicts target, aspect, opinion, and sentiment across reviews with multiple tuples, fine-grained aspect taxonomies, and targets that may be products or product-parts.Strong CoT LLM performance motivates distillation, while their inference cost is prohibitive for high-throughput e-commerce applications.Source: §1 · p. 1
  • Structured aspect cascade arises when target–aspect errors break bindings or hallucinate non-grounded targets, invalidating tuples even when local sentiment words remain plausible.A pilot SeqKD study found a substantial drop from target-only to target–aspect correctness.Source: §1 · pp. 1–2
  • Off-policy distillation trains only on teacher-generated, mostly valid trajectories, leaving little direct supervision for student-induced structural errors encountered during inference.Student predictions can include incorrect bindings and hallucinated targets that alter tuple structure.Source: §1 · p. 2
  • 9.75% to 7.22% target hallucination with Qwen3-4B; STAR-OPD improves over off-policy and strong general on-policy baselines on E-ABSA20K and SemEval-2014.It yields the largest gains on structurally hard reviews and improves deployment efficiency.Source: §1 · p. 2
  • STAR-OPD adapts generic on-policy distillation to ABSA quadruple extraction, using student rollouts and cascade-aware rewards for binding consistency, target grounding, and aspect disambiguation.Lightweight filtering and sampling act as stabilizers.Source: §1 · p. 2

2 Related Work

ABSA has advanced to quadruple extraction, but prior work does not address structural dependencies and distillation-specific error propagation in multi-quadruple reviews. STAR-OPD applies on-policy distillation to this setting with cascade-aware, set-structured rewards.

Aspect-Based Sentiment Analysis

  • ABSA progressed from aspect classification to triplet and quadruple extraction, while generative methods perform well but leave multi-quadruple structural dependencies and distillation errors unstudied.Prior work includes aspect-level classification, triplet extraction (Peng et al., 2020), quadruple extraction, and generative and LLM-based methods (Scaria et al., 2024; Wang et al., 2023; Hu et al., 2022).Source: §2 · p. 2

LLM Knowledge Distillation

  • Generic on-policy distillation optimizes sequence-level objectives; this work adapts it to ABSA with structural rewards for binding, grounding, and aspect disambiguation.The rewards provide task-specific structural credit assignment rather than proposing a new general on-policy distillation principle.Source: §2 · p. 2

Error Propagation in Structured Prediction

  • Exposure bias and structured-prediction research identify amplified upstream errors; in quadruple-extraction distillation, target–aspect interface errors bottleneck and corrupt downstream fields.Prior work includes Bengio et al. (2015) on exposure bias and Finkel et al. (2006) on structured-prediction error propagation.Source: §2 · p. 2

3 Problem Analysis

ABSA quadruple extraction requires jointly predicting interdependent targets, aspects, opinions, and sentiments, with the target–aspect interface shaping downstream interpretation. Distilled-model failures concentrate at this interface, while off-policy training leaves student-induced structural errors insufficiently corrected.

3.1 Task Structure and Failure Mode

  • Each review can contain multiple sentiment quadruples, whose target, fine-grained aspect, opinion span, and polarity fields are semantically interdependent.The target may be a generic product tag or a review-grounded mention; aspects come from a closed taxonomy, and polarity is positive, neutral, or negative.Source: §3.1 · pp. 2–3
  • For the smaller model vs. teacher, 50.6% vs. 81.1% target–aspect correctness and 40.0% vs. 76.3% full-quadruple correctness show structural corruption dominates failures.The diagnostic comparison found only a 6.6-point gap in target identification alone; the authors characterize the main bottleneck as maintaining valid target–aspect structure, not generic sequence noise.Source: §3.1 · p. 3
  • The structured aspect cascade begins with broken target–aspect bindings or hallucinated targets, causing downstream fields to rely on an incorrect tuple interpretation.In a SeqKD pilot, 8.1% of correctly identified target–aspect pairs still received an incorrect sentiment label.Source: §3.1 · p. 3

3.2 Why Off-Policy Distillation Is Insufficient

  • Off-policy distillation trains on mostly valid teacher trajectories, leaving student-generated structural errors largely unseen and without direct correction during training.Because early target–aspect mistakes change how remaining fields are interpreted, the proposed solution needs student rollouts and tuple-structure-level feedback.Source: §3.2 · p. 3

4 Method

STAR-OPD adapts on-policy distillation to ABSA quadruple extraction by training on student rollouts and using structured rewards to correct target–aspect binding and target-grounding errors. Its cascade-aware rewards align to unordered tuple sets and also provide fine-grained aspect-category supervision.

4.1 Overview

  • STAR-OPD prepares filtered pseudo-labels with a high-quality CoT teacher, then trains the student on its own rollouts using reverse-KL distillation, structure-aware rewards, difficulty-aware sampling, and replay.Difficulty-aware sampling and replay serve as lightweight stabilizers.Source: §4.1 · p. 3

4.2 On-Policy Cascade-Aware Distillation

  • Training on student-sampled outputs exposes the model to inference-like structural errors, while reward-weighted updates, reverse-KL teacher alignment, and reference regularization guide and stabilize correction.Higher-reward rollouts receive stronger teacher-aligned updates; lower-reward rollouts contribute less.Source: §4.2 · p. 4

4.3 Cascade-Aware Rewards

  • Rewards are computed after optimal bipartite matching, making them invariant to generation order; the binding reward credits a tuple only when its target and aspect are jointly correct.The set-level base reward supplies an overall tuple-quality signal, while binding targets the target–aspect interface.Source: §4.3 · pp. 4–5
  • The hallucination reward penalizes ungrounded non-PRODUCT targets, while teacher-scored aspect-category supervision provides a softer signal for semantically adjacent labels.The hallucination formulation avoids rewarding empty predictions and assigns full reward only to non-empty sets whose predictions are all grounded. Teacher scores are computed over the aspect taxonomy under the student rollout context.Source: §4.3 · p. 5
  • Additional format-validity rewards and input-adaptive weighting emphasize binding consistency and hallucination suppression for reviews identified as likely product-part cases.Binding and hallucination rewards are highlighted as the most important components because they address the target–aspect interface and tuple-invalidating hallucinated targets.Source: §4.3 · p. 5

5 Experiments

Experiments evaluate overall extraction quality, structural failure modes, and performance on ambiguity-sensitive cases across E-ABSA20K and SemEval-2014.

5.1 Setup

  • The setup uses E-ABSA20K as the primary benchmark and SemEval-2014 Restaurant and Laptop for cross-domain validation, comparing teacher-supervised, off-policy, and generic on-policy students.Baselines include Direct-SFT, SeqKD (Kim and Rush, 2016), GKD (Agarwal et al., 2024), MiniLLM, and G-OPD (Yang et al., 2026); student methods use Qwen3-4B unless otherwise specified.Source: §5.1 · p. 6
  • Quad-F1 is the primary metric, with T-F1, TA-F1, TAS-F1, and T-Hall reported for structural diagnosis.Quad-F1 uses Hungarian matching over target, aspect, and sentiment, followed by fuzzy character-level opinion evaluation using CharF1 at τ=0.5.Source: §5.1 · p. 6

5.2 Main Results

  • STAR-OPD beats off-policy and generic on-policy baselines, including MiniLLM and G-OPD, on E-ABSA20K while narrowing Qwen3-4B’s gap to the 32B teacher.On SemEval-2014, gains over generic on-policy baselines are smaller and statistical significance less consistent; the trend also transfers to the 1.7B student while retaining the smaller model’s efficiency advantage.Source: §5.2 · p. 6, Table 1 · p. 7

5.3 Reducing Structured Cascade

  • STAR-OPD improves every structural stage versus SeqKD, with the largest gain at target–aspect binding, and cuts T-Hall from 9.75% to 7.22%.The T-F1-to-TA-F1 drop identifies the target–aspect interface as the main bottleneck, with further declines through TAS-F1 and Quad-F1 reflecting downstream effects.Source: §5.3 · pp. 6–7, Figure 3 · p. 7

5.4 On-Policy and Reward Ablations

  • Removing on-policy training causes the largest ablation drop, while plain on-policy distillation without reward shaping also degrades performance.Student rollouts expose structurally invalid states absent from teacher-generated supervision; explicit structural credit assignment is needed beyond that exposure alone.Source: Table 2 · p. 7, §5.4 · pp. 7–8
  • A collapsed scalar reward underperforms component-specific rewards, and removing Rbind causes the largest drop among reward terms.Removing Rhall, Rcat, Rfmt, or Radap also degrades performance; greedy rather than Hungarian matching likewise performs worse, supporting order-invariant set-level alignment.Source: Table 2 · p. 7, §5.4 · pp. 7–8

5.5 Performance on Hard Cases

  • STAR-OPD achieves the best student performance on every E-ABSA20K-Hard subset, with the largest gains on product-part cases and improvements on contrastive and implicit-opinion reviews.GKD and G-OPD also improve over off-policy distillation on hard subsets but remain less effective on the most structure-sensitive cases.Source: Table 3 · p. 8, §5.5 · p. 8
  • In a product-part example, STAR-OPD restores the grounded target and fine-grained aspect where SeqKD drifts in binding and MiniLLM predicts a plausible but non-grounded target.The example illustrates why gains are largest when target granularity and aspect assignment are tightly coupled.Source: Figure 4 · p. 8, §5.5 · p. 8

6 Conclusion

STAR-OPD addresses brittle ABSA quadruple distillation by correcting student-induced structural states, especially failures at the target–aspect interface. Its results support designing structured-extraction distillation around student inference states, with finer-grained process-style rewards left for future work.

6 Conclusion

  • The central distillation challenge is recovery from student-induced structural states, particularly broken target–aspect bindings and hallucinated targets that degrade downstream tuple correctness.The conclusion says these interface failures, rather than teacher imitation alone, are the most damaging.Source: §6 · p. 8
  • Across E-ABSA20K and SemEval-2014, STAR-OPD consistently improves distilled students, reduces target hallucination, and yields the largest gains on structurally hard reviews.The conclusion attributes these gains to cascade-aware rewards for binding consistency and target grounding.Source: §6 · p. 8
  • For structured extraction, distillation should target student inference-time states rather than teacher-only trajectories; STAR-OPD rewards structured outcomes, while finer-grained process-style rewards remain future work.The authors describe outcome-level rewards as natural for unordered quadruple extraction.Source: §6 · p. 8

7 Limitations · A Task and Dataset Details · A.1 Full Task Definition

STAR-OPD is evaluated for ABSA quadruple extraction, but its generality and deployment safety remain limited by task scope, post-generation rewards, and dependence on teacher and pseudo-label quality. The task combines target, aspect, opinion, and sentiment predictions, with the target–aspect interface identified as a structural bottleneck.

7 Limitations

  • The study evaluates STAR-OPD on E-ABSA20K and SemEval-2014, but its focus on ABSA quadruple extraction, especially e-commerce reviews, leaves generality to other structured extraction settings unvalidated.Hard-subset analyses are diagnostic evidence of failure modes, not the sole basis for broad statistical claims.Source: §7 · pp. 8–9
  • Because rewards are computed on structured outcomes after generation, rather than at token or process level, they may limit intervention earlier during decoding.Source: §7 · p. 9
  • The method depends on teacher quality and pseudo-label filtering, while deployment may amplify annotation biases or residual target–aspect binding and grounding errors.Such errors could misstate user opinions or product attributes, so the method is framed as a research and assistive analysis tool, not a fully autonomous high-stakes system.Source: §7 · p. 9

A.1 Full Task Definition

  • Each ABSA quadruple contains a target, aspect, opinion span, and sentiment label, and a review may require predicting multiple such quadruples.Source: §A.1 · p. 10
  • Targets require a verbatim review entity, aspects come from a domain taxonomy of up to 42 categories, opinions are minimal contiguous spans, and sentiment is positive, neutral, or negative.Source: §A.1 · p. 10
  • Quadruple extraction combines target entity identification, aspect taxonomy classification, opinion span extraction, and sentiment polarity classification, making it structurally more challenging than simpler ABSA formulations.Source: §A.1 · p. 10
  • The target–aspect interface is a key bottleneck linking target grounding to downstream opinion and sentiment interpretation, making extraction vulnerable to structural error propagation.Source: §A.1 · p. 10

A.2 Dataset Statistics · A.3 Construction of E-ABSA20K-Hard

Experiments use E-ABSA20K and SemEval-2014, with E-ABSA20K-Hard constructed to diagnose structurally challenging quadruple extraction cases. The hard subset targets ambiguity and binding difficulties but is used diagnostically, not as the primary basis for significance claims.

A.2 Dataset Statistics

  • E-ABSA20K is a 20K-review benchmark spanning four e-commerce domains, and SemEval-2014 provides cross-domain validation (Sun et al., 2026; Pontiki et al., 2014).Source: §A.2 · p. 10
  • E-ABSA20K reviews average 6.0 sentiment quadruples, increasing opportunities for target confusion, aspect ambiguity, and structural binding errors.Source: §A.2 · p. 10

A.3 Construction of E-ABSA20K-Hard

  • E-ABSA20K-Hard includes reviews with at least one challenging phenomenon that increases structural ambiguity in quadruple extraction.These cases amplify aspect ambiguity, target–aspect binding difficulty, and target hallucination in distilled models.Source: §A.3 · p. 10, §A.2 · p. 10
  • Hard-case categories include fine-grained product-part targets that require preserving target identity and colloquial sarcasm that weakens surface lexical cues.Source: §A.3 · p. 11
  • The subset also includes contrastive sentences requiring aspect-sensitive resolution of clause-level sentiment reversals and implicit opinions requiring contextual polarity inference.Source: §A.3 · p. 11
  • These categories support diagnostic evaluation and hard-case analysis rather than serving as method requirements or the primary basis for significance claims.The corresponding subsets are smaller and partially overlapping.Source: §A.3 · p. 11

B Evaluation Details … D.1 CoT Teacher Fine-Tuning

The appendix defines structure-sensitive set evaluation and target-grounding diagnostics, while the analysis finds that target–aspect errors and hallucinated targets can corrupt or invalidate downstream quadruples.

B.1 Field-Level Matching Metrics

  • T-F1, TA-F1, TAS-F1, and Quad-F1 progressively score target, target–aspect, target–aspect–sentiment, and full quadruple correctness, with Quad-F1 primary.TA-F1 directly diagnoses target–aspect binding quality.Source: §B.1 · p. 11

B.2 Hungarian Matching for Set Prediction

  • One-to-one Hungarian matching aligns predicted and gold quadruples independently of generation order, excluding opinions from alignment but enforcing them in Quad-F1.This avoids unstable alignments from minor opinion-boundary variation while retaining opinion correctness in final evaluation.Source: §B.2 · p. 11

B.3 Opinion Matching

  • Opinion spans use normalized character-level F1, and an aligned opinion counts as correct when its CharF1 exceeds τ = 0.5.Character-level scoring is intended to tolerate minor boundary deviations.Source: §B.3 · pp. 11–12

B.4 Target Hallucination Rate (T-Hall)

  • T-Hall measures predicted non-PRODUCT targets whose entity span is absent from the review; lower rates indicate better grounding.PRODUCT and PRODUCT PART entity spans are checked by exact substring matching after lowercasing and whitespace normalization.Source: §B.4 · p. 12

C Additional Analysis of Structural Failure Modes

  • 9.75% T-Hall for SeqKD versus 6.67% for the teacher accompanies a failure mode where fabricated targets make entire quadruples spurious.Broken target–aspect bindings can also degrade downstream sentiment and opinion predictions; the analysis identifies these structural errors, rather than generic sequence noise, as the central failure pattern.Source: C Additional Analysis of Structural… · p. 12

D.1 CoT Teacher Fine-Tuning

  • The Qwen3-32B teacher is CoT-fine-tuned on manual training data, but only final quadruples are distilled; test annotations are excluded from adaptation and training.Teacher reasoning traces are not distilled, though teacher distributions may be queried for reward computation; development annotations support model selection.Source: §D.1 · pp. 12–13

D.2 Agreement with Expert Annotations … E.3 Reward-to-Weight Transformation

STAR-OPD validates its CoT teacher for pseudo-labeling, filters structurally invalid supervision, and uses weighted, cascade-aware rewards to guide student updates. The supplied sections describe the reward design and training safeguards, and report that explicit reward decomposition outperforms a collapsed scalar variant.

D.2 Agreement with Expert Annotations

  • The CoT teacher reaches Cohen’s κ = 0.92 against adjudicated expert labels, substantially exceeding a smaller direct model.Human–human agreement on the held-out E-ABSA20K-Hard subset was κ = 0.89; the authors treat teacher agreement as sufficient for pseudo-label generation.Source: §D.2 · p. 13, Table 4 · p. 13

D.3 Pseudo-Label Generation and Filtering

  • Pseudo-label filtering retains only schema-complete quadruples with valid aspect and sentiment labels, grounded non-PRODUCT targets, and duplicates removed.Filtering is intended to remove structurally invalid supervision; student-induced structural errors remain addressed by cascade-aware rewards during on-policy training.Source: §D.3 · p. 13
  • Filtering retains 70% of teacher-generated training instances overall, with domainwise rates of 66%–73%, and uses no validation or test annotations.Pseudo-labels are generated only for the student training split; E-ABSA20K-Hard is reserved for diagnostic evaluation.Source: §D.3 · p. 13

E.1 Reward Weights and Input-Adaptive Weighting

  • STAR-OPD weights binding and hallucination rewards most heavily, gives category supervision moderate weight, and uses format validity as a small stabilizing term.For likely PRODUCT PART reviews identified from teacher pseudo-labels, it increases emphasis on binding consistency and hallucination suppression; default weights are shared across domains.Source: §E.1 · pp. 13–14
  • Reward weights are selected on validation data with a constrained local search, and moderate ±0.05 perturbations around final values leave performance relatively stable.The search varies one coefficient at a time while monitoring Quad-F1, TA-F1, and T-Hall, rather than exhaustively testing combinations.Source: §E.1 · pp. 13–14

E.2 Collapsed Scalar Reward Variant

  • The collapsed scalar-reward variant performs worse than STAR-OPD, supporting explicit decomposition for structural credit assignment.It retains the structural signals for binding, hallucination suppression, and category support but removes component-specific and input-adaptive weighting; the rest of the on-policy setup is unchanged.Source: §E.2 · p. 14

E.3 Reward-to-Weight Transformation

  • STAR-OPD converts batch-normalized rollout rewards into bounded weights that preserve rollout-quality ordering and favor better-structured student outputs.The experiments use τw = 2.0, αmin = 0.2, αmax = 3.0, and ϵ = 10^-6; the passage reports greater stability than using raw rewards directly.Source: §E.3 · pp. 14–15

E.4 Training Dynamics of Reward Components … F Training Algorithm

The appendix examines reward alignment and describes sampling, replay, curriculum, optimization, runtime, and inference-efficiency choices for STAR-OPD. Its reward dynamics track intended structural metrics, while the deployment comparison reports faster student inference alongside a narrowed performance gap.

E.4 Training Dynamics of Reward Components

  • Rbind rises alongside validation TA-F1, while Rhall rises early and remains high as validation T-Hall decreases.The authors observe no obvious late-stage reward inflation or instability with metric degradation, but note that this analysis does not formally exclude shortcut optimization; the pattern is consistent with STAR-OPD’s gains on structurally difficult subsets.Source: §E.4 · p. 15

E.5 Difficulty-Aware Sampling and Replay

  • Harder instances are sampled more frequently using difficulty indicators derived from teacher pseudo-labels and lightweight heuristics, without diagnostic evaluation labels.The default sampling parameter is γ = 0.5.Source: §E.5 · p. 15
  • A FIFO buffer stores up to 4,096 low-reward student rollouts, and 20% of each batch is drawn from it when available to revisit unresolved structural failures.Rollouts enter when normalized reward is below −0.5; the other 80% of the batch comes from the training set.Source: §E.5 · p. 15

E.6 Curriculum

  • During the first 30% of training, difficulty-aware sampling probability increases linearly from 0.3 to 0.7, then stays at 0.7 to balance early stability with hard-case coverage.The curriculum is intended to reduce early optimization noise while ensuring sufficient exposure to structurally difficult cases.Source: §E.6 · p. 15

E.7 Optimization Hyperparameters

  • Student distillation runs for 2K updates with Qwen3-4B by default, effective batch size 32, and maximum rollout length of 1024 new tokens.Optimization uses AdamW at a peak learning rate of 1 × 10^-5, weight decay 0.01, cosine decay with 5% warmup, β = 0.1, temperature 0.8, top-p 0.95, and seed 42; Rcat teacher scoring is limited to the sampled rollout’s aspect step.Source: §E.7 · pp. 15–16

E.8 Hardware and Runtime

  • Teacher fine-tuning takes approximately 14 hours and on-policy student distillation about 24 hours on 4×A100 80G GPUs; reported evaluation averages three runs when applicable and uses paired bootstrap testing at p < 0.05.Teacher inference is served by vLLM (Kwon et al., 2023); inference-time efficiency is reported separately.Source: §E.8 · p. 16

E.9 Inference Efficiency

  • 2.18× faster than the Qwen3-32B teacher, the STAR-OPD Qwen3-4B student narrows the performance gap, while the 1.7B student is faster still.Latency and throughput are measured on E-ABSA20K Test under the same decoding configuration; Table 6 reports the measures.Source: §E.9 · p. 16

G Additional Qualitative Cases · H Full Prompt Template · I Aspect Taxonomy Example

The appendices show qualitative cases where STAR-OPD better preserves grounded targets and fine-grained structure, provide the extraction prompt, and give women’s bags as an example of the E-ABSA20K aspect taxonomy.

G Additional Qualitative Cases

  • The examples cover all four E-ABSA20K-Hard phenomena: product-part mentions, colloquial sarcasm, contrastive sentences, and implicit opinion expressions.Targets use the paper’s normalized schema for implicit product-level, explicit product, and explicit product-part targets.Source: G Additional Qualitative Cases · p. 16
  • Across product-part, sarcastic, contrastive, and implicit-opinion cases, STAR-OPD more reliably preserves grounded targets, fine-grained bindings, and complete tuples than SeqKD and MiniLLM.SeqKD tends to drift structurally after departing from teacher states, while MiniLLM often backs off to coarser structures or produces plausible but ungrounded targets.Source: G Additional Qualitative Cases · p. 16
  • Product-part and contrastive examples expose binding drift, target collapse, and cross-tuple error propagation; sarcasm and implicit opinions test fine-grained structure under ambiguous evidence.The authors say these cases mirror the quantitative hard-case trends and show gains beyond local sentiment prediction.Source: G Additional Qualitative Cases · pp. 16–17, Figure 6 · p. 17, Figure 7 · p. 17

H Full Prompt Template

  • The prompt requests (target, aspect_category, opinion, sentiment) quadruples, distinguishes general, specific-product, and product-part targets, and prohibits targets absent from the review.It also specifies predefined aspects, minimal exact opinion spans, and positive, neutral, or negative sentiment labels.Source: H Full Prompt Template · p. 17
  • Prompt rules require every applicable target–aspect pair, one quadruple per pair, shortest faithful opinions, correct negation and concession polarity, and target-specific multi-product assignment.The prompt also requires verifying that targets occur in the review and aspects come from the predefined list.Source: H Full Prompt Template · p. 17
  • The prompt’s reasoning checklist identifies targets, maps aspects, locates opinion spans, resolves polarity, and checks for invented targets and invalid aspects before emitting line-separated quadruples.The template supplies the review text and category list as input.Source: H Full Prompt Template · p. 17

I Aspect Taxonomy Example

Related papers

Loading 2608.20831v1…