nchellwig at SemEval-2026 Task 3: Self-Consistent Structured Generation (SCSG) for Dimensional Aspect-Based Sentiment Analysis using Large Language Models
Authors: Nils Constantin Hellwig, Jakob Fehle, Udo Kruschwitz, Christian Wolff
Organizations: Media Informatics Group, University of Regensburg, Regensburg, Germany · Information Science Group, University of Regensburg, Regensburg, Germany
We present Self-Consistent Structured Generation (SCSG) for Dimensional Aspect-Based Sentiment Analysis in SemEval-2026 Task 3 (Track A). SCSG enhances prediction reliability by executing a LoRA-adapted large language model multiple times per instance, retaining only tuples that achieve a majority consensus across runs. To mitigate the computational overhead of multiple forward passes, we leverage vLLM's PagedAttention mechanism for efficient key--value cache reuse. Evaluation across 6 languages and 8 language--domain combinations demonstrates that self-consistency with 15 executions yields statistically significant improvements over single-inference prompting, with our system (leveraging Gemma 3) ranking in the top seven across all settings, achieving second place on three out of four English subsets and first place on Tatar-Restaurant for DimASTE.
⟨ service , service#general , spotty , v=7.00,a=7.00⟩
DimASTE Output (Triplets)
⟨ Decor , nice , v=7.00,a=7.17⟩
Table 1: Example outputs for DimASQP and DimASTE tasks. DimASQP extracts quadruplets containing aspect term, aspect category, opinion term, and valence-arousal ( v,a ) dimensions, where v,a∈[1,9] , while DimASTE extracts triplets omitting the aspect category.
Table 2: Performance on the test set for DimASTE and DimASQP : Comparison of vanilla prompting (Baseline) and self-consistency (SC) with 5, 10, or 15 prompt executions. Results show continuous-level precision (cPrec), recall (cRec), and F1-score (cF1) in %. Bold values indicate the best performance for each language–domain pair and metric. Asterisks denote statistical significance of the improvement over the baseline, while daggers ( † ) and double daggers ( ‡ ) denote significance over 5 and 10 views, respectively, based on Holm-Bonferroni corrected p-values ( ∗:p<0.05,∗∗:p<0.01,∗∗∗:p<0.001 ).
Table 3: Top-6 leaderboard rankings for the two subtasks across all language–domain pairs. Our submission is highlighted in bold .
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 1: Prompt used for SCSG. The prompt comprises descriptions of the considered sentiment elements (4 for DimASTE, 5 for DimASQP), explanations regarding the range of valence and arousal, the desired output format, and the example text for which ABSA is to be performed.
Set
Domain
eng
jpn
rus
tat
ukr
zho
Train
Restaurant
2,284 / 3,659
—
1,240 / 2,487
1,240 / 2,487
1,240 / 2,487
6,050 / 8,523
Laptop
4,076 / 5,773
—
—
—
—
3,490 / 6,502
Hotel
—
1,600 / 2,846
—
—
—
—
Dev
Restaurant
200 / 408
—
48 / 102
48 / 102
48 / 102
300 / 761
Laptop
200 / 317
—
—
—
—
300 / 551
Hotel
—
200 / 364
—
—
—
—
Appendix
Table 4: Dataset statistics for DimASTE and DimASQP, reported as sentences / tuples per split, language, and domain. A dash (—) indicates that no data is available for the respective language–domain combination.
Input: “Decor is nice though service can be spotty.”
Predictions
Run
(Aspect Term, Sentiment Polarity) - Pair
Valence
Arousal
(Aspect Term, Sentiment Polarity) - Pair
Valence
Arousal
1
(Decor, nice)
6.92
7.13
(service, spotty)
5.53
6.03
2
(Decor, nice)
6.80
7.03
(service, be spotty)
5.60
6.10
3
(Decor, is nice)
6.67
6.90
(service, spotty)
5.40
5.90
4
(Decor, nice)
7.00
7.50
(service, spotty)
5.70
6.20
Appendix
Figure 2: Self-consistency majority voting for DimASTE over k=5 runs. Aspect-sentiment pairs (ignoring valence-arousal values) appearing in ≥τ=⌈k/2⌉ runs are aggregated by averaging their valence and arousal values. The aggregation section shows the explicit calculation. Light blue rows highlight matching (Decor, nice) variants; light green rows highlight matching (service, spotty) variants. For DimASQP, in addition to the aspect term and sentiment polarity, the aspect category is considered as well.
Subtask 2: DimASTE
Subtask 3: DimASQP
Language
Domain
BL
5 Views
10 Views
15 Views
BL
5 Views
10 Views
15 Views
English
Restaurant
77.93
78.35
78.15
78.45
75.17
75.30
75.39
75.26
English
Laptop
65.51
66.01
64.71
65.56
35.57
34.62
36.77
35.36
Japanese
Hotel
52.63
54.89
53.91
54.28
35.93
39.87
39.18
38.58
Russian
Restaurant
54.28
59.21
57.66
59.28
49.54
52.65
51.40
52.56
Tatar
Restaurant
52.72
52.83
52.99
53.54
38.65
44.65
43.95
43.56
Appendix
Table 5: Evaluation on the development set for DimASTE and DimASQP: cF1 scores across languages and domains comparing vanilla prompting (BL) and self-consistency (SC) using either 5, 10 or 15 prompt executions, with majority voting applied. Bold values indicate the best performance for each language–domain pair.
Training
Evaluation
Language
Domain
#
Time (s)
Time/1K (s)
#
w/ SC
w/o SC
w/o SC + w/o Batch.
Time (s)
Time/ex. (s)
Time (s)
Time/ex. (s)
Time (s)
Time/ex. (s)
English
Restaurant
2,284
3,860
1,690
1,000
320
0.321
37
0.037
1,151
1.151
English
Laptop
4,076
6,515
1,598
1,000
289
0.289
32
0.032
1,017
1.018
Japanese
Hotel
1,600
3,415
2,134
800
234
0.293
29
0.037
822
1.028
Russian
Restaurant
1,240
2,638
2,127
630
277
0.440
38
0.060
884
1.405
Appendix
Table 6: Training and evaluation duration across languages and domains for both DimASTE and DimASQP. We report duration with Self-Consistency (w/ SC, 15 views), without (w/o SC), and without vLLM’s batch processing (w/o Batch.).
Table 7: Evaluation on different LLMs (cF1). We compared G (Gemma-3-27B), M (Mistral-Small-24B), and Q (Qwen3-32B). Green cells ( x ) indicate performance surpassing the Rank 1 team on the official competition leaderboard for the respective language and domain combination. Overall, the results demonstrate that increasing the number of self-consistency views leads to consistent performance gains, with several configurations significantly outperforming the competition’s top-ranked entries.
In this paper, we present AILS-NTUA system for Track-A of SemEval-2026 Task 3 on Dimensional Aspect-Based Sentiment Analysis (DimABSA), which encompasses three complementary problems: Dimensional Aspect Sentiment Regression (DimASR), Dimensional Aspect Sentiment Triplet Extraction (DimASTE), and Dimensional Aspect Sentiment Quadruplet Prediction (DimASQP) within a multilingual and multi-domain framework. Our methodology combines fine-tuning of language-appropriate encoder backbones for continuous aspect-level sentiment prediction with language-specific instruction tuning of large language models using LoRA for structured triplet and quadruplet extraction. This unified yet task-adaptive design emphasizes parameter-efficient specialization across languages and domains, enabling reduced training and inference requirements while maintaining strong effectiveness. Empirical results demonstrate that the proposed models achieve competitive performance and consistently surpass the provided baselines across most evaluation settings.
Stavros Gazetas, Giorgos Filandrianos, Maria Lymperaiou +3
School of Electrical and Computer Engineering, AILS Laboratory National Technical University of Athens
This paper presents an approach to the SemEval-2026 Task 3: Dimensional Aspect-Based Sentiment Analysis. We investigate methods for moving beyond traditional categorical sentiment (e.g., positive or negative) to predict fine-grained, real-valued scores for sentiment "valence" (positivity) and "arousal" (intensity). We participate in two subtasks: predicting these scores for given aspects (Subtask 1) and extracting full sets of sentiment details, including aspects, categories, and opinions alongside their scores (Subtask 3). Our approach for the regression task involves a weighted ensemble of transformer-based encoder models. For the Russian language, we further enhance the input by using a large language model (LLM) to generate synthetic sentiment descriptions. For the extraction task, we fine-tune a decoder LLM to perform structured prediction, allowing the system to identify sentiment elements and estimate their numerical scores simultaneously.
We present our system for SemEval-2026 Task 3 on dimensional aspect-based sentiment regression. Our approach combines a hybrid RoBERTa encoder, which jointly predicts sentiment using regression and discretized classification heads, with large language models (LLMs) via prediction-level ensemble learning. The hybrid encoder improves prediction stability by combining continuous and discretized sentiment representations. We further explore in-context learning with LLMs and ridge-regression stacking to combine encoder and LLM predictions. Experimental results on the development set show that ensemble learning significantly improves performance over individual models, achieving substantial reductions in RMSE and improvements in correlation scores. Our findings demonstrate the complementary strengths of encoder-based and LLM-based approaches for dimensional sentiment analysis. Our development code and resources will be shared at https://github.com/aaronlifenghan/ABSentiment
A. J. W. de Vink, Filippos Karolos Ventirozos, Natalia Amat-Lefort +1
LIACS, Leiden University, NL · Manchester Metropolitan University, UK · Leiden University Medical Center, NL