LogSigma at SemEval-2026 Task 3: Uncertainty-Weighted Multitask Learning for Dimensional Aspect-Based Sentiment Analysis

Baraa Hikal; Jonas Becker; Bela Gipp

cs.CLMar 26, 2026

LogSigma at SemEval-2026 Task 3: Uncertainty-Weighted Multitask Learning for Dimensional Aspect-Based Sentiment Analysis

Authors: Baraa Hikal, Jonas Becker, Bela Gipp

Organizations: University of Göttingen, Germany · LKA NRW, Germany

Abstract

This paper describes LogSigma, our system for SemEval-2026 Task 3: Dimensional Aspect-Based Sentiment Analysis (DimABSA). Unlike traditional Aspect-Based Sentiment Analysis (ABSA), which predicts discrete sentiment labels, DimABSA requires predicting continuous Valence and Arousal (VA) scores on a 1-9 scale. A central challenge is that Valence and Arousal differ in prediction difficulty across languages and domains. We address this using learned homoscedastic uncertainty, where the model learns task-specific log-variance parameters to automatically balance each regression objective during training. Combined with language-specific encoders and multi-seed ensembling, LogSigma achieves 1st place on five datasets across both tracks. The learned variance weights vary substantially across languages due to differing Valence-Arousal difficulty profiles-from 0.66x for German to 2.18x for English-demonstrating that optimal task balancing is language-dependent and cannot be determined a priori.

Figures & tables

Figure 1: Overview of LogSigma . Left: Input text and aspect are encoded as [CLS] text [SEP] aspect [SEP] and passed through a language-specific pretrained encoder to obtain $\mathbf{h}\in\mathbb{R}^{d}$ . Middle: Dual fully-connected heads predict Valence and Arousal independently; during training (dashed region), learnable log-variance parameters $s=\log\sigma^{2}$ balance the MSE losses via precision weighting. Right: At inference, predictions from three seeds are averaged to produce the final V.VV#A.AA output.

Language	Encoder
English	twitter-roberta-large- topic-sentiment-latest $\dagger$
Japanese	bert-base-japanese-v3
Russian	rubert-base-cased
Tatar	rubert-base-cased $\ddagger$
Ukrainian	ukr-roberta-base
Chinese	chinese-roberta-wwm-ext

Table 1: Language-specific encoders. (

\dagger

) sentiment-pretrained; (

\ddagger

) related-language proxy; (

\star

) multilingual fallback. This strategy outperforms a single multilingual encoder by 8.2% on Track A and up to 28.5% on Track B (Section 5 ).

Track	Dataset	System	RMSE $\downarrow$	Rank
A	eng_laptop	LogSigma	1.241	1 st /33
		Kimi-K2 Thinking	2.189	32 nd /33
		Qwen-3 14B	2.809	33 rd /33
	eng_restaurant	LogSigma	1.104	1 st /37
		Kimi-K2 Thinking	2.146	36 th /37
		Qwen-3 14B	2.643	37 th /37

Table 2: Blind test results with baselines from the dataset papers ( Lee et al., 2026 ; Becker et al., 2026 ) . LogSigma outperforms all baselines across every dataset, ranking 1 st on five of seven submitted datasets. See Appendix O for a discussion of the performance gap.

Dataset	Encoder	w/o	w/	$\Delta$
jpn_hotel (A)	XLM-R-L	0.672	0.619	$-$ 7.9%
jpn_finance (A)	XLM-R-L	0.673	0.606	$-$ 9.9%
deu_pol (B)	GBERT-L	0.944	0.940	$-$ 0.5%
eng_env (B)	Tw-RoBERTa	1.180	1.086	$-$ 8.0%
pcm_pol (B)	Tw-RoBERTa	0.702	0.624	$-$ 11.2%

Table 3: Uncertainty weighting ablation on the dev set (RMSE

\downarrow

). Each row compares the same encoder with and without learned uncertainty. The “w/” column includes 3-seed ensembling. Gains are largest for Pidgin (

-

11.2%), where the V/A difficulty imbalance is greatest.

Figure 2: Learned task variance (

\sigma^{2}

) across Track B languages (seed 42). Higher

\sigma^{2}

indicates greater task-specific noise, causing the model to down-weight that objective. English and Pidgin show

\sigma^{2}_{V}>\sigma^{2}_{A}

(arousal receives higher weight), while German shows the opposite (

\sigma^{2}_{A}>\sigma^{2}_{V}

). English and Pidgin learn near-identical values despite per-language training, confirming their shared V/A difficulty profile.

Appendix figures & tables24 assets

Supplementary material from the paper’s appendix.

Appendix

Language	Domain	Size
English	Restaurant	2,507
English	Laptop	1,869
Japanese	Hotel	1,892
Japanese	Finance	1,245
Russian	Restaurant	1,987
Tatar	Restaurant	892

Appendix

Table 4: Track A dataset statistics (6 languages, 4 domains, 10 datasets). Sizes are total aspect-level instances across train and dev splits. Tatar is the smallest dataset (892), contributing to its high prediction difficulty.

Language	Domain	Size
German	Politics	1,245
English	Environmental	987
Nigerian Pidgin	Politics	1,156
Swahili	Politics	876
Chinese	Environmental	1,089

Appendix

Table 5: Track B dataset statistics (5 languages, stance detection). All datasets are smaller than Track A, with Swahili being the most resource-constrained (876 instances).

Hyperparameter	Value
Optimizer	AdamW
Encoder learning rate	$2\times 10^{-5}$
Uncertainty lr ( $s_{V}$ , $s_{A}$ )	$5\times 10^{-2}$
Weight decay	0.01
Batch size / grad. accum.	16 / 4 (eff. 64)
Max epochs (Track A / B)	25 / 15

Appendix

Table 6: Full training hyperparameters. All languages share the same configuration except English environmental (Track B), which uses a lower encoder learning rate of

1\times 10^{-5}

. The uncertainty parameters receive a

2{,}500\times

higher learning rate than the encoder to enable rapid adaptation to task-specific noise levels.

Language	Domain	Text (excerpt)	Aspect	V	A
Track A — Multi-Domain Reviews
English	Restaurant	I am so disappointed in the service…	service	3.38	5.75
English	Laptop	I got this laptop and am very impressed…	laptop	8.00	8.00
Japanese	Hotel	客室のトイレが古すぎて臭かったです…	トイレ (toilet)	2.62	6.50
Japanese	Hotel	施設のスタッフ…とても丁寧でした	スタッフ (staff)	8.00	7.83
Russian	Restaurant	Сервис вполне хороший.	Сервис (service)	6.50	5.00

Appendix

Table 7: Representative examples from Track A and Track B datasets. Valence (V) and Arousal (A) are gold-annotated scores on a 1–9 scale. The examples illustrate why V/A imbalance is non-trivial: high arousal can co-occur with either high valence (excited approval) or low valence (angry criticism).

Dataset	RMSE $\downarrow$	RMSE V	RMSE A	PCC V / PCC A
eng_restaurant	0.754	0.742	0.766	0.919 / 0.684
eng_laptop	0.660	0.600	0.719	0.934 / 0.611
jpn_hotel	0.688	0.690	0.687	0.909 / 0.575
jpn_finance	0.650	0.801	0.500	0.828 / 0.489
rus_restaurant	1.028	0.898	1.158	0.856 / 0.433
tat_restaurant	1.285	1.434	1.136	0.519 / 0.330

Appendix

Table 8: Track A full development set results (language-specific encoders + learned uncertainty + 3-seed ensemble). Chinese datasets are consistently easiest (0.44–0.58 RMSE), while low-resource languages are hardest: Tatar (1.29) and Ukrainian (1.23). Arousal RMSE exceeds Valence RMSE on 3 of 10 datasets (both English and Russian); the remaining 7 have RMSE V > RMSE A , including Japanese finance and all Chinese datasets—see Appendix G for analysis.

Dataset	Encoder	RMSE $\downarrow$	RMSE V	RMSE A	PCC V	PCC A
deu_politics	GBERT-Large	0.940	0.887	0.993	0.884	0.534
eng_environmental	Tw-RoBERTa-Large	1.111	1.267	0.956	0.784	0.487
pcm_politics	Tw-RoBERTa-Large	0.624	0.567	0.681	0.977	0.746
swa_politics	XLM-R-Large	1.234	n/a $\dagger$		0.788	0.621
zho_environmental	Zh-RoBERTa	0.443	0.471	0.415	0.898	0.463
Average (4 lang)		0.779	0.798	0.761	0.886	0.557

Appendix

Table 9: Track B full development set results (best configuration per language). All use learned uncertainty with 3-seed ensembling except Swahili (multilingual ensemble, no uncertainty). The 4-language average excludes Swahili for fair comparison. Pidgin achieves the best RMSE (0.62) despite being a low-resource language, benefiting from the English-pretrained encoder.

\dagger

Per-dimension RMSE not available for Swahili’s multilingual ensemble.

Language	Seed	$\boldsymbol{\sigma^{2}_{V}}$	$\boldsymbol{\sigma^{2}_{A}}$	$\boldsymbol{\sigma^{2}_{V}/\sigma^{2}_{A}}$
German	42	0.519	0.790	0.66
	12	0.800	0.944	0.85
	73	0.301	0.504	0.60
English	42	0.912	0.418	2.18
	12	0.621	0.582	1.07
	73	0.306	0.457	0.67

Appendix

Table 10: Full per-seed learned

\sigma^{2}

values (Track B). Higher

\sigma^{2}

indicates greater task-specific noise, causing the model to down-weight that objective. German is the only language where

\sigma^{2}_{A}>\sigma^{2}_{V}

across all seeds. English and Pidgin converge to within

\Delta\leq 0.001

on every seed. Chinese shows the highest cross-seed instability (

\sigma^{2}_{V}

: 0.39–0.89), suggesting sensitivity to initialization for this language.

Figure 3: English–Pidgin

\sigma^{2}

convergence across all three seeds. Despite per-language training on separate datasets, English and Pidgin learn near-identical variance parameters on every seed (

\Delta\leq 0.001

). This quantitatively confirms that the Twitter-RoBERTa encoder perceives these two languages as having identical V/A difficulty profiles, consistent with Nigerian Pidgin being an English-based creole.

Language	Seed	$\boldsymbol{\sigma^{2}_{V}}$	$\boldsymbol{\sigma^{2}_{A}}$	$\boldsymbol{\sigma^{2}_{V}/\sigma^{2}_{A}}$
English	21	0.615	0.288	2.13
	99	0.859	0.559	1.54
	42	0.910	0.418	2.18
Japanese	21	0.354	0.752	0.47
	99	0.443	0.525	0.84
	42	0.392	0.469	0.84

Appendix

Table 11: Track A learned

\sigma^{2}

values for four languages (seeds 21, 99, 42). English uses restaurant values; Japanese uses finance values (hotel is near-identical). English is the only language with

\sigma^{2}_{V}>\sigma^{2}_{A}

across all seeds, and the only Track A language using a sentiment-pretrained encoder (Twitter-RoBERTa). Japanese and Russian converge to near-identical ratios despite different encoders (BERT-Japanese vs. ruBERT), unrelated language families, and different review domains (finance/hotel vs. restaurant)—their shared factor is general-purpose encoder pretraining, suggesting that without sentiment pretraining, arousal is consistently the harder target. Russian and Tatar are trivially identical as they share the same ruBERT encoder. Ukrainian and Chinese values were not logged for this run.

Dataset	XLM-R-Large	XLM-R-Base	Afro-XLMR	mDeBERTa	XLM-Politics
deu_politics	1.143	1.389	1.161	1.580	1.449
eng_environmental	1.383	1.523	1.418	1.628	1.602
pcm_politics	0.985	1.300	0.941	1.431	1.284
swa_politics	1.464	1.437	1.498	1.539	1.566
zho_environmental	0.464	0.734	0.465	0.636	0.633
Average	1.088	1.277	1.097	1.363	1.307

Appendix

Table 12: Encoder architecture comparison on Track B development set (RMSE

\downarrow

, all without uncertainty weighting). XLM-RoBERTa-Large wins on 4 of 5 datasets. mDeBERTa-v3-base is the weakest encoder—near-zero arousal PCC on German (0.13) and English (0.18)—suggesting its pretraining does not transfer to VA regression. XLM-Twitter-Politics, despite being pretrained on political sentiment text, underperforms generic XLM-R-Large. XLM-R-Base wins only on Swahili, likely due to less overfitting on this low-resource language.

Figure 4: Arousal PCC per encoder across Track B languages. XLM-RoBERTa-Large and Afro-XLMR-Large achieve the highest arousal correlations on most datasets. mDeBERTa and XLM-Twitter-Politics fall below the near-zero threshold (dashed line) on German, demonstrating that their pretraining does not transfer to arousal regression. Pidgin is the only language where all encoders achieve reasonable arousal PCC (

>

0.29), likely because its English-based lexicon benefits all English-pretrained models.

Dataset	PCC V	PCC A	Gap ( $V-A$ )
eng_restaurant	0.919	0.684	+0.235
eng_laptop	0.934	0.611	+0.323
jpn_hotel	0.909	0.575	+0.334
jpn_finance	0.828	0.489	+0.339
rus_restaurant	0.856	0.433	+0.423
tat_restaurant	0.519	0.330	+0.189

Appendix

Table 13: Valence vs. Arousal prediction difficulty across all 15 datasets. Arousal is harder to predict than Valence in every dataset, with an average PCC gap of +0.29. The gap is largest for Chinese environmental (+0.44) and Russian restaurant (+0.42), and smallest for Chinese restaurant (+0.14). Swahili has the smallest Track B gap (+0.17). However, PCC-based ranking can disagree with RMSE-based ranking: for Japanese finance and all Chinese Track A datasets, RMSE V > RMSE A (Table 8 ), meaning the model makes larger absolute errors on Valence despite ranking Arousal samples less accurately. This discrepancy arises because PCC measures rank correlation while RMSE measures absolute error.

Dataset	5 Epochs	10 Epochs	Winner
eng_laptop	0.750	0.687	10ep
eng_restaurant	0.834	0.745	10ep
jpn_finance	0.673	0.770	5ep
jpn_hotel	0.672	0.686	5ep
rus_restaurant	1.002	0.947	10ep
tat_restaurant	1.306	1.332	5ep

Appendix

Table 14: Epochs ablation on Track A development set (XLM-RoBERTa-Large, RMSE

\downarrow

). Five epochs wins on 6 of 10 datasets—all Japanese, all Chinese, and Tatar. Ten epochs wins on English, Russian, and Ukrainian. This suggests that CJK and low-resource languages are more prone to overfitting. Our final system uses early stopping (patience 3) rather than a fixed epoch count to adapt per language.

Dataset	Random Init	Track A Init	$\boldsymbol{\Delta}$
deu_politics	1.143	1.069	$-$ 6.5%
eng_environmental	1.383	1.490	+7.8%
pcm_politics	0.985	0.901	$-$ 8.5%
swa_politics	1.464	1.362	$-$ 6.9%
zho_environmental	0.464	0.409	$-$ 12.0%

Appendix

Table 15: Cross-domain transfer from Track A to Track B, using XLM-RoBERTa-Large (RMSE

\downarrow

). Track A initialization improves 4 of 5 languages (up to

-

12% on Chinese). English environmental is the only dataset where transfer hurts (+7.8%), despite Chinese environmental benefiting—suggesting that language-specific factors beyond domain mismatch influence transferability. This experiment uses a single XLM-R model, not our final language-specific setup.

Figure 5: Nigerian Pidgin model progression (RMSE on dev set). Switching from multilingual XLM-R (0.98) to English-pretrained Twitter-RoBERTa yields

-

21.4% RMSE. Adding 3-seed ensembling and learned uncertainty yields a further

-

19.4% (0.77

\to

0.62), demonstrating that encoder selection, ensembling, and loss design each contribute substantially.

Figure 6: Translation vs. native model for Russian restaurant. Valence PCC is nearly identical (0.938 vs. 0.934), confirming that positive/negative sentiment transfers across languages. However, arousal PCC drops from 0.586 (native) to 0.557 (translated), and overall RMSE increases by 5.4%. This suggests that arousal is encoded in language-specific features—word choice, emphasis patterns, morphological intensity markers—that translation strips away.

Dataset	RegressLM	LogSigma	$\boldsymbol{\Delta}$
eng_laptop	1.438	0.660	$-$ 54.1%
eng_restaurant	1.334	0.754	$-$ 43.5%
jpn_hotel	1.333	0.688	$-$ 48.4%
jpn_finance	1.523	0.650	$-$ 57.3%
rus_restaurant	1.873	1.028	$-$ 45.1%
tat_restaurant	2.137	1.285	$-$ 39.9%

Appendix

Table 16: RegressLM (T5-Gemma, pretrained, 60M params) on Track A dev set (RMSE

\downarrow

), text + aspect input only.

\Delta

shows the relative RMSE reduction of LogSigma over RegressLM. The opinion-augmented variant is in Table 19 in Appendix K .

Dataset	PCC V	PCC A	RMSE
eng_laptop	0.69	0.32	1.560
eng_restaurant	0.78	0.56	1.294
jpn_finance	0.03	$-$ 0.03	1.610
jpn_hotel	$-$ 0.05	$-$ 0.11	1.969
rus_restaurant	0.32	0.23	1.997
tat_restaurant	0.41	0.24	1.951

Appendix

Table 17: From-scratch RegressLM (2E2D,

d_{\text{model}}

= 512, 31.2M params, lr = 5e-5, early stopped at epoch 46) on Track A dev set. Japanese datasets produce negative PCC on both dimensions, meaning predictions are anti-correlated with gold labels. Even the best language (English restaurant, PCC V = 0.78) falls far short of pretrained encoders (PCC V = 0.93 with LogSigma ).

Dataset	Aspect-only $\sigma$		Aspect+Opinion $\sigma$		$\Delta$ %
	V	A	V	A	V	A
eng_laptop	1.28	0.85	0.33	0.41	$-$ 74.5	$-$ 51.6
eng_restaurant	1.30	0.87	0.30	0.37	$-$ 76.8	$-$ 57.0
jpn_hotel	0.91	0.49	0.22	0.24	$-$ 76.0	$-$ 50.9
jpn_finance †	0.92	0.43	0.13	0.12	$-$ 85.8	$-$ 71.5
rus_restaurant	1.34	0.80	0.16	0.30	$-$ 88.2	$-$ 61.8

Appendix

Table 18: Within-group standard deviation of VA scores when instances are grouped by aspect alone versus aspect+opinion. Adding opinion reduces

\sigma_{V}

by 47–88% and

\sigma_{A}

by 38–72%, confirming that opinion is the dominant source of within-aspect VA variance (

\eta^{2}_{V}{=}0.98

\eta^{2}_{A}{=}0.95

). (

\dagger

) Synthetic DeepSeek-generated opinions; all other datasets use ground-truth labels.

Panel of Figure 7 — Figure 7: English system prompt and few-shot template for synthetic opinion generation with DeepSeek. Each of the 6 languages uses an equivalent native-language prompt. Examples cover positive, negative, and null opinion categories.

Dataset	$-$ Opinion	+Opinion	$\boldsymbol{\Delta}$
eng_laptop	1.438	1.279	$-$ 11.1%
eng_restaurant	1.334	1.188	$-$ 10.9%
jpn_hotel	1.333	1.224	$-$ 8.2%
jpn_finance ‡	1.523	1.379	$-$ 9.5%
rus_restaurant	1.873	1.545	$-$ 17.5%
tat_restaurant	2.137	2.205	+3.2%

Appendix

Table 19: Opinion ablation with RegressLM (T5-Gemma, pretrained, 60M params) on Track A dev set (RMSE

\downarrow

-

Opinion : text + aspect only (same baseline as Table 16 ); +Opinion : input includes opinion terms (Ground-Truth for 8 sets; Synthetic for 2 sets marked with

\ddagger

). Inclusion of opinion terms reduces average RMSE by 9.2%; see Table 16 for LogSigma comparison.

Dataset	RMSE V	RMSE A	RMSE	LogSigma
rus_restaurant	0.538	1.077	0.808	1.028
jpn_finance	0.835	0.455	0.645	0.650

Appendix

Table 20: Post-competition Qwen3-4B fine-tuning (2 regression heads, 5 epochs, MSE loss) on two Track A dev datasets (RMSE

\downarrow

). While Qwen3-4B delivers a substantial improvement in the Russian domain, it exhibits diminishing returns in Japanese finance, leading to our decision to prioritize more efficient encoder models.

Dataset	Track	Seed 1	Seed 2	Seed 3	Ensemble	$\boldsymbol{\Delta}$
eng_restaurant	A	1.277	1.211	1.214	1.153	$-$ 4.8%
eng_environmental	B	1.218	1.201	1.251	1.161	$-$ 3.3%

Appendix

Table 21: Single-seed vs. 3-seed ensemble (RMSE

\downarrow

\Delta

is the relative improvement from the best individual seed to the ensemble. Averaging predictions across three seeds consistently improves over even the best single run, smoothing variance from random initialization.

#	Language	Track	Domain	Best RMSE $\downarrow$
1	Chinese	A	Finance	0.373
2	Chinese	B	Environmental	0.409
3	Chinese	A	Laptop	0.520
4	Chinese	A	Restaurant	0.532
5	Japanese	A	Finance	0.606
6	Japanese	A	Hotel	0.619

Appendix

Table 22: All 15 dataset-language combinations ranked by best development RMSE across all configurations tested. Chinese is the easiest language across all domains and tracks (0.37–0.53). Tatar and Swahili are the hardest (

>

1.23), both lacking dedicated pretrained encoders. English environmental ranks 13 th —harder than German politics—likely because environmental stance text is distributionally different from the sentiment-rich data that Twitter-RoBERTa was pretrained on. Note: some entries use earlier configurations (e.g., XLM-R-Large) that outperformed the final language-specific system on individual datasets; final system results are in Tables 8 and 9 .

Multilingual Sentiment Analysis

See all 10 papers

NCL-BU at SemEval-2026 Task 3: Fine-tuning XLM-RoBERTa for Multilingual Dimensional Sentiment Regression
Apr 10, 2026Tong Wu, Nicolay Rusnachenko, Huizhi LiangMultilingual Sentiment Analysis Sentiment Analysis
Independent Researcher · Centre for Applied Creative Technologies (CFACT+), Bournemouth University, UK · School of Computing, Newcastle University, Newcastle upon Tyne, UK
AILS-NTUA at SemEval-2026 Task 3: Efficient Dimensional Aspect-Based Sentiment Analysis
Mar 5, 2026Stavros Gazetas, Giorgos Filandrianos, Maria Lymperaiou +3Aspect-Based Sentiment Analysis
School of Electrical and Computer Engineering, AILS Laboratory National Technical University of Athens
SemEval-2026 Task 3: Dimensional Aspect-Based Sentiment Analysis (DimABSA)
Apr 8, 2026Liang-Chih Yu, Jonas Becker, Shamsuddeen Hassan Muhammad +14Sentiment Analysis Aspect-Based Sentiment Analysis
Yuan Ze University · University of Göttingen · Imperial College London +10

Apr 10, 2026J/K move · Enter open · S save

NCL-BU at SemEval-2026 Task 3: Fine-tuning XLM-RoBERTa for Multilingual Dimensional Sentiment Regression

Tong Wu, Nicolay Rusnachenko, Huizhi Liang

Independent Researcher · Centre for Applied Creative Technologies (CFACT+), Bournemouth University, UK · School of Computing, Newcastle University, Newcastle upon Tyne, UK

Multilingual Sentiment Analysis Sentiment Analysis Aspect-Based Sentiment Analysis

Open paper AskPDF arXiv

Dimensional Aspect-Based Sentiment Analysis (DimABSA) extends traditional ABSA from categorical polarity labels to continuous valence-arousal (VA) regression. This paper describes a system developed for Track A, Subtask 1 (Dimensional Aspect Sentiment Regression), aiming to predict real-valued VA scores in the [1, 9] range for each given aspect in a text. A fine-tuning approach based on XLM-RoBERTa-base is adopted, with dual regression heads with sigmoid-scaled outputs for valence and arousal prediction. Separate models are trained for each language-domain pair (English and Chinese across restaurant, laptop, and finance domains), and training and development sets are merged for final test predictions. In development experiments, the fine-tuning approach is compared against several large language models under a few-shot prompting setting, demonstrating that task-specific fine-tuning outperforms these LLM-based methods across all evaluation datasets.

Abstract

Multilingual Sentiment Analysis

Explore similar work

NCL-BU at SemEval-2026 Task 3: Fine-tuning XLM-RoBERTa for Multilingual Dimensional Sentiment Regression

AILS-NTUA at SemEval-2026 Task 3: Efficient Dimensional Aspect-Based Sentiment Analysis

SemEval-2026 Task 3: Dimensional Aspect-Based Sentiment Analysis (DimABSA)