Dependency Reasoning for Aspect Sentiment Quadruple Prediction

Shima Esfandyari; Seywan Soleymanbaigi; Fardin Akhlaghian Tab; Fatemeh Daneshfar; Mourad Oussalah

cs.CLOct 10, 2026

Dependency Reasoning for Aspect Sentiment Quadruple Prediction

Authors: Shima Esfandyari, Seywan Soleymanbaigi, Fardin Akhlaghian Tab, Fatemeh Daneshfar, Mourad Oussalah

Organizations: Department of Computer Engineering, University of Kurdistan, Sanandaj, Iran · University of Oulu, Finland

Abstract

Aspect Sentiment Quadruple Prediction (ASQP) is fundamentally a relational reasoning problem: identifying aspect-opinion-category-sentiment quadruples requires modeling structured dependencies, not merely sequential co-occurrence. Existing LLM-based approaches incorporate syntactic information indirectly, via retrieval-based demonstration selection or multi-stage training, rather than as unified, single-pass generation-time context. We present DRIFT, which linearizes dependency parse trees and injects them directly into LLM prompts as explicit relational context, enabling the model to reason over aspect-opinion dependencies rather than infer them from token proximity alone. DRIFT further integrates retrieval-augmented in-context learning and a syntax-aware LLM-based adjudication mechanism that complements exact-match evaluation under span-boundary variation. On SemEval Rest15, Rest16, and ACOS, DRIFT achieves the strongest performance among compared generative ASQP methods, reaching 54.47% and 64.10% F1 on Rest15 and Rest16, with consistent gains across short- and medium-distance aspect-opinion configurations; adjudication further shows exact-match metrics systematically underestimate semantically correct predictions.

Figures & tables

Figure 1: Comparison of surface-sequential reasoning in prior ASQP approaches and dependency-aware reasoning in DRIFT for aspect–opinion alignment. Note: the example sentence is quoted verbatim from the Rest15 dataset.

Figure 2: Overview of the DRIFT framework with dependency-guided prompting and retrieval augmentation.

Methods	Venue	Rest15			Rest16			Avg (F1)
Methods	Venue	Pre	Rec	F1	Pre	Rec	F1	Avg (F1)
GAS †	ACL21	45.31	46.70	45.98	54.54	57.62	56.04	51.01
Paraphrase	EMNLP21	46.16	47.72	46.93	56.63	59.30	57.93	52.43
DLO	EMNLP22	47.08	49.33	48.18	57.92	61.80	59.79	53.99
ILO	EMNLP22	47.08	50.38	49.05	57.58	61.17	59.32	54.19
MvP	ACL23	-	-	51.04	-	-	60.39	55.72

Table 1: Performance comparison (%);

\dagger

indicates results quoted from GenDA.

Judge Model	Rest15	Rest16
o3 (Default)	72.41	77.20
GPT-4o	63.97	71.58
GPT-4o-mini	61.94	68.43

Table 2: F1 (%) by LLM Judge Model.

Format	spaCy		Stanza
Format	Rest15	Rest16	Rest15	Rest16
spaCy (Default)	34.27	43.08	35.19	43.72
JSON	35.11	42.94	35.94	44.55
CoNLL-U	34.55	44.16	35.14	44.55
Plain Text	35.92	43.54	35.28	44.23
Hierarchical	35.85	43.94	35.51	45.34

Table 3: F1 (%) by dependency parser format.

Method	Backbone	Rest15				Rest16
Method	Backbone	0-shot	1-shot	5-shot	10-shot	0-shot	1-shot	5-shot	10-shot
THOR	Llama-3.1-8B	7.88	8.02	8.65	10.01	9.44	10.78	11.37	11.95
	Llama-3.1-70B	17.62	22.49	26.78	31.47	27.61	30.04	34.34	35.91
	GPT-4o-mini	18.18	21.52	24.29	29.36	29.83	31.50	32.84	34.05
	GPT-4o	30.79	33.37	35.12	36.81	37.23	39.22	42.05	43.57
LLMs for SA	Llama-3.1-8B	8.61	8.82	8.66	10.69	11.43	9.99	11.67	12.59

Table 4: F1 (%) of DRIFT and baseline methods under ICL.

Method	Backbone	Rest-ACOS				Laptop-ACOS
Method	Backbone	0-shot	1-shot	5-shot	10-shot	0-shot	1-shot	5-shot	10-shot
Zero-Shot ( Brown et al., 2020 )	GPT-4o-mini	31.28	-	-	-	11.18	-	-	-
Zero-Shot CoT ( Kojima et al., 2022 )	GPT-4o-mini	23.01	-	-	-	7.56	-	-	-
Few-Shot (N=5) ( Brown et al., 2020 )	GPT-4o-mini	-	-	32.76	-	-	-	13.69	-
Majority-Vote (N=5, K=8) ( Li et al., 2025 )	GPT-4o-mini	-	-	34.09	-	-	-	15.22	-
Retrieval-Augmented (N=5) ( Li et al., 2025 )	GPT-4o-mini	-	-	42.15	-	-	-	21.87	-

Table 5: Comparison of DRIFT with different In-Context Learning methods on ACOS benchmarks.

Panel of Figure 3 — Figure 3: F1 scores vs. number of shots in ICL.

Dep. Tree	Retrieval	PEFT		ICL
Dep. Tree	Retrieval	Rest15	Rest16	Rest15	Rest16
✗	✗	49.69	61.87	24.89	29.70
✗	✓	50.13	61.34	34.79	41.52
✓	✗	54.47	64.10	31.40	42.40
✓	✓	51.01	63.17	34.27	43.08

Table 6: Ablations in PEFT and ICL.

Dataset	Bucket	Vanilla F1	+Dep F1	$\Delta$
Rest15	1–2	37.57	39.69	+2.12
	3–4	27.04	27.62	+0.58
	$\geq$ 5	16.67	17.39	+0.72
Rest16	1–2	39.42	48.08	+8.66
	3–4	31.87	37.39	+5.52
	$\geq$ 5	21.85	17.74	$-$ 4.11

Table 7: Dependency-distance analysis.

Component	Time (s)	Std	Avg. Tokens	Complexity
Dependency Parsing	0.015	0.004	109.0	$\mathcal{O}(1)$
Few-shot Retrieval	0.014	0.004	81.7	$\mathcal{O}(n)$
Prompt + Prediction	1.306	0.923	590.1	$\mathcal{O}(1)$
Total	1.335	0.923	780.8	$\mathcal{O}(n)$

Table 8: Inference cost analysis of DRIFT.

Appendix figures & tables4 assets

Supplementary material from the paper’s appendix.

Appendix

Dataset	Rest15			Rest16
Dataset	Train	Valid	Test	Train	Valid	Test
N	834	209	537	1264	316	544
Positive	1005	252	453	1369	341	583
Neutral	34	14	37	62	23	40
Negative	315	81	305	558	143	176
Total	1354	347	795	1989	507	799

Appendix

Table 9: Dataset statistics for Rest15 and Rest16. N denotes the number of sentences. Positive, Neutral, and Negative represent the number of quadruples for each sentiment.

Dataset	Rest-ACOS			Laptop-ACOS
Dataset	Train	Dev	Test	Train	Dev	Test
N	1530	171	583	2934	326	816
Positive	1656	180	667	2583	279	716
Neutral	95	12	44	227	24	65
Negative	733	69	205	1362	137	380
Total	2484	261	916	4172	440	1161

Appendix

Table 10: Dataset statistics for Restaurant-ACOS and Laptop-ACOS. N denotes the number of sentences. Positive, Neutral, and Negative represent the number of quadruples for each sentiment.

Seed	Rest15 F1	Rest16 F1
1	54.20	64.10
2	54.13	64.05
3	54.33	63.89
4	54.47	63.95
5	54.28	64.03
Mean $\pm$ Std	54.28 $\pm$ 0.13	64.00 $\pm$ 0.08

Appendix

Table 11: Per-seed F1 (%) for the PEFT configuration with dependency-tree injection (no retrieval), across 5 random seeds.

Sentence (Query)	Reference	Prediction	Exact Match	Human	Reasoning
Great place to relax and enjoy your dinner.	[place, ambience general, positive, Great]	[place, ambience general, positive, Great] [place, ambience general, positive, enjoy]	✗	✓	Includes both Great and enjoy ; reference omits one.
The hot dogs are good, yes, but the reason to get over here is the fantastic pork croquette sandwich, perfect on its supermarket squishy bun.	[hot dogs, food quality, positive, good] [pork croquette sandwich, food quality, positive, fantastic] [bun, food quality, positive, perfect]	[hot dogs, food quality, positive, good] [pork croquette sandwich, food quality, positive, fantastic]	✗	✓	Reference wrongly assigns perfect to bun instead of the sandwich.
In the end you end up with a fair tab and NOTHING BUT A GREAT TIME!!!	[NULL, restaurant prices, positive, fair] [NULL, restaurant general, positive, GREAT]	[tab, restaurant prices, positive, fair] [NULL, restaurant general, positive, GREAT TIME]	✗	✓	Correctly identifies tab as aspect; reference misses it.
Which lets face it … at times it’s a good thing.	[NULL, service general, positive, good]	[NULL, restaurant miscellaneous, positive, good]	✗	✓	Reference assigns an ungrounded service general category; sentence never mentions service.
I’ve had my fair share of modern Japanese, and this spot delivers.	[modern Japanese, food quality, positive, delivers]	[spot, restaurant general, positive, delivers]	✗	✓	delivers refers to spot , not modern Japanese .

Appendix

Table 12: Qualitative examples where model predictions are semantically correct despite failing exact-match evaluation. ✓ = correct, ✗ = incorrect.

Aspect-Based Sentiment Analysis

Aspect-Based Sentiment Analysis (ABSA) identifies what people express opinions about and the sentiment or emotion attached to each aspect.

See all 51 papers

Apr 25, 2026

$\mathcal{S}^2$ IT: Stepwise Syntax Integration Tuning for Large Language Models in Aspect Sentiment Quad Prediction

Aspect Sentiment Quad Prediction (ASQP) has seen significant advancements, largely driven by the powerful semantic understanding and generative capabilities of large language models (LLMs). However, while syntactic structure information has been proven effective in previous extractive paradigms, it remains underutilized in the generative paradigm of LLMs due to their limited reasoning capabilities. In this paper, we propose S^2IT, a novel Stepwise Syntax Integration Tuning framework that progressively integrates syntactic structure knowledge into LLMs through a multi-step tuning process. The training process is divided into three steps. S^2IT decomposes the quadruple generation task into two stages: 1) Global Syntax-guided Extraction and 2) Local Syntax-guided Classification, integrating both global and local syntactic structure information. Finally, Fine-grained Structural Tuning enhances the model's understanding of syntactic structures through the prediction of element links and node classification. Experiments demonstrate that S^2IT significantly improves state-of-the-art performance across multiple datasets. Our implementation will be open-sourced at https://github.com/DMIRLAB-Group/S2IT.

Bingfeng Chen, Chenjie Qiu, Yifeng Xie +3

School of Computer Science, Guangdong University of Technology · Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) · Peng Cheng Laboratory +1

LLM Fine-Tuning Aspect-Based Sentiment Analysis Language Model Post-Training

Ask PDF

Mar 14, 2026

Generate Then Correct: Single Shot Global Correction for Aspect Sentiment Quad Prediction

Aspect-based sentiment analysis (ABSA) extracts aspect-level sentiment signals from user-generated text, supports product analytics, experience monitoring, and public-opinion tracking, and is central to fine-grained opinion mining. A key challenge in ABSA is aspect sentiment quad prediction (ASQP), which requires identifying four elements: the aspect term, the aspect category, the opinion term, and the sentiment polarity. However, existing studies usually linearize the unordered quad set into a fixed-order template and decode it left-to-right. With teacher forcing training, the resulting training-inference mismatch (exposure bias) lets early prefix errors propagate to later elements. The linearization order determines which elements appear earlier in the prefix, so this propagation becomes order-sensitive and is hard to repair in a single pass. To address this, we propose a method, Generate-then-Correct (G2C): a generator drafts quads and a corrector performs a single-shot, sequence-level global correction trained on LLM-synthesized drafts with common error patterns. On the Rest15 and Rest16 datasets, G2C outperforms strong baseline models.

Shidong He, Haoyu Wang, Wenjie Luo

School of Cyber Security and Computer, Hebei University, Baoding, China. · Faculty of Computer Science, Dalhousie University, Halifax, Canada.

Exposure Bias Aspect-Based Sentiment Analysis

Ask PDF

Mar 2, 2026

LLM-as-an-Annotator: Training Lightweight Models with LLM-Annotated Examples for Aspect Sentiment Tuple Prediction

Training models for Aspect-Based Sentiment Analysis (ABSA) tasks requires manually annotated data, which is expensive and time-consuming to obtain. This paper introduces LA-ABSA, a novel approach that leverages Large Language Model (LLM)-generated annotations to fine-tune lightweight models for complex ABSA tasks. We evaluate our approach on five datasets for Target Aspect Sentiment Detection (TASD) and Aspect Sentiment Quad Prediction (ASQP). Our approach outperformed previously reported augmentation strategies and achieved competitive performance with LLM-prompting in low-resource scenarios, while providing substantial energy efficiency benefits. For example, using 50 annotated examples for in-context learning (ICL) to guide the annotation of unlabeled data, LA-ABSA achieved an F1 score of 49.85 for ASQP on the SemEval Rest16 dataset, closely matching the performance of ICL prompting with Gemma-3-27B (51.10), while requiring significantly lower computational resources.

Nils Constantin Hellwig, Jakob Fehle, Udo Kruschwitz +1

Media Informatics Group, University of Regensburg, Regensburg, Germany · Information Science Group, University of Regensburg, Regensburg, Germany

Aspect-Based Sentiment Analysis LLM-Assisted Annotation

Ask PDF

Abstract

Aspect-Based Sentiment Analysis

Explore similar work

S2\mathcal{S}^2S2IT: Stepwise Syntax Integration Tuning for Large Language Models in Aspect Sentiment Quad Prediction

Generate Then Correct: Single Shot Global Correction for Aspect Sentiment Quad Prediction

LLM-as-an-Annotator: Training Lightweight Models with LLM-Annotated Examples for Aspect Sentiment Tuple Prediction

$\mathcal{S}^2$ IT: Stepwise Syntax Integration Tuning for Large Language Models in Aspect Sentiment Quad Prediction