DimABSA: Dimensional Aspect Sentiment Analysis

• DimABSA is a reformulation of ABSA that uses continuous valence–arousal scores to capture nuanced sentiment beyond categorical labels.
• It operationalizes sentiment regression, triplet extraction, and quadruplet prediction across multiple languages and domains.
• DimABSA bridges structured information extraction and affective science, enabling precise sentiment analysis with detailed calibration.

Searching arXiv for the cited DimABSA papers to ground the article in current sources. arXiv search query: "Dimensional Aspect-Based Sentiment Analysis SemEval 2026 Task 3 DimABSA" Dimensional Aspect-Based Sentiment Analysis (DimABSA) is a reformulation of Aspect-Based Sentiment Analysis (ABSA) in which aspect-level sentiment is represented not by categorical polarity labels such as positive, negative, or neutral, but by continuous valence–arousal (VA) scores. In this formulation, valence represents the degree of positivity or negativity, while arousal represents the level of activation or intensity, with both dimensions defined on a real-valued scale in $[1,9]$ (Yu et al., 8 Apr 2026). DimABSA was consolidated as a shared-task framework in SemEval-2026 Task 3, where it was paired with Dimensional Stance Analysis (DimStance) and operationalized through regression, triplet extraction, and quadruplet prediction subtasks across multiple languages and domains (Yu et al., 8 Apr 2026). The framework is designed to retain the structured aspect-centric character of ABSA while introducing a finer-grained affective representation grounded in affective science (Lee et al., 30 Jan 2026).

1. Conceptual basis and relation to traditional ABSA

Traditional ABSA analyzes sentiment at the granularity of aspects and opinion spans rather than document-level polarity, but in most prior formulations it assigns coarse categorical labels to aspect instances (Gazetas et al., 5 Mar 2026). DimABSA replaces these categorical labels with continuous sentiment in the valence–arousal space, thereby modeling both pleasantness and activation for each aspect (Hikal et al., 26 Mar 2026). In the SemEval-2026 task definition, valence and arousal are both real-valued in $[1,9]$, where $1$ denotes extreme negative valence or low arousal, $9$ denotes extreme positive valence or high arousal, and $5$ denotes neutral valence or medium arousal (Yu et al., 8 Apr 2026).

The main motivation is increased expressivity. The DimABSA resource paper states that categorical polarity is too coarse to distinguish differences in strength and activation, as in contrasts such as “good” versus “excellent,” or modifier-sensitive expressions such as “a little slow” versus “extremely slow” (Lee et al., 30 Jan 2026). The SemEval task paper likewise emphasizes that VA captures nuanced distinctions beyond positive, negative, and neutral, including modifier effects in opinion terms and fine-grained aspect-linked sentiment variation (Yu et al., 8 Apr 2026). This means that DimABSA preserves the localized structure of ABSA—aspect terms, opinion spans, and optionally categories—while replacing label discretization with continuous affective geometry.

The framework is also positioned as a bridge between ABSA and dimensional affect modeling. The resource paper describes three subtasks that couple standard ABSA elements with continuous VA scores, explicitly presenting them as a bridge from traditional ABSA to dimensional ABSA (Lee et al., 30 Jan 2026). A plausible implication is that DimABSA is not only a task reformulation but also a methodological interface between structured information extraction and continuous affect regression.

2. Task formulation and output schema

SemEval-2026 Task 3 is organized into two tracks: Track A for DimABSA and Track B for DimStance (Yu et al., 8 Apr 2026). Track A contains three subtasks, while Track B contains only the regression subtask.

Track A Subtask 1 is Dimensional Aspect Sentiment Regression (DimASR). Its input is a sentence $x$ together with one or more provided aspect terms $a$, and its output is a pair of VA scores for each aspect, $(V,A)\in [1,9]\times [1,9]$ (Yu et al., 8 Apr 2026). The task is formalized by the prediction function

$f(x,a)\rightarrow (\hat{V},\hat{A}),$

with $\hat{V},\hat{A}\in [1,9]$ (Yu et al., 8 Apr 2026).

Track A Subtask 2 is Dimensional Aspect Sentiment Triplet Extraction (DimASTE). Given a sentence $[1,9]$0, the system must output all triplets $[1,9]$1, where $[1,9]$2 is an aspect term, $[1,9]$3 is an associated opinion term, and $[1,9]$4 (Yu et al., 8 Apr 2026). The structured output is written as $[1,9]$5 (Yu et al., 8 Apr 2026). The AILS-NTUA system paper further specifies a JSON-oriented representation with keys {Aspect, Opinion, Valence, Arousal} and notes that matches are case-sensitive and exact at the span level (Gazetas et al., 5 Mar 2026).

Track A Subtask 3 is Dimensional Aspect Sentiment Quadruplet Prediction (DimASQP). Its output consists of quadruplets $[1,9]$6, where $[1,9]$7 is an aspect category label of the form ENTITY#ATTRIBUTE, such as FOOD#QUALITY or SERVICE#GENERAL (Yu et al., 8 Apr 2026). No holder entity is annotated in this task; the quadruplets are explicitly $[1,9]$8 (Yu et al., 8 Apr 2026).

Track B reformulates stance detection as dimensional regression. In DimStance, stance targets are treated as aspects, and the task uses the DimASR formulation: the input is an utterance or post $[1,9]$9 with one or more stance targets $1$0, and the output is a VA pair for each target (Yu et al., 8 Apr 2026). The task paper describes this as a way to extend ABSA beyond consumer reviews to public-issue discourse, including political, energy, and climate issues (Yu et al., 8 Apr 2026).

Examples supplied in the task documentation illustrate the intended outputs. For DimASR, “The food was excellent” with the provided aspect yields 8.00#8.12; for DimASTE, “Service at the bar was a little slow” yields (Service, a little slow, 4.10#4.30); and for DimASQP, “Their sodas are usually expired and flat” yields (sodas, DRINKS#QUALITY, usually expired, 1.90#7.20) and (sodas, DRINKS#QUALITY, flat, 2.40#6.80) (Yu et al., 8 Apr 2026).

3. Datasets, annotation, and multilingual scope

The DimABSA resource introduced the first multilingual, multidomain DimABSA dataset, containing 76,958 aspect instances across 42,590 sentences, spanning six languages and four domains (Lee et al., 30 Jan 2026). In the SemEval-2026 task formulation, Track A covers English, Japanese, Russian, Tatar, Ukrainian, and Chinese across the restaurant, laptop, hotel, and finance domains, while finance is available only for Subtask 1 (Yu et al., 8 Apr 2026). Track B covers five languages—English, German, Chinese, Nigerian Pidgin, and Swahili—across environmental protection and politics, with 11,746 stance targets across 7,365 texts (Yu et al., 8 Apr 2026).

The resource paper enumerates ten language–domain pairs for Track A. Restaurant data are available in English, Russian, Tatar, Ukrainian, and Chinese; laptop data in English and Chinese; hotel data in Japanese; and finance data in Japanese and Chinese (Lee et al., 30 Jan 2026). The SemEval paper specifies representative sources: ACOS and Yelp for English restaurant, ACOS and Amazon Reviews 2023 for English laptop, Rakuten Travel for Japanese hotel, EDINET for Japanese finance, SIGHAN-2024 plus Google Reviews and PTT for Chinese restaurant, Mobile01 for Chinese laptop, and MOPS for Chinese finance (Yu et al., 8 Apr 2026).

Annotation combines categorical tuple extraction with continuous VA scoring. For full-subtask datasets in restaurant, laptop, and hotel, annotators create quadruplets $1$1; for finance and all DimStance datasets, they annotate $1$2 pairs because only the regression formulation applies (Yu et al., 8 Apr 2026). The SemEval task paper states that Track A uses two independent annotators to extract categorical tuples, with a third adjudicator resolving disagreements, while Track B uses LLMs to extract candidate stance targets and then validates them via majority voting by five annotators (Yu et al., 8 Apr 2026).

VA rating uses five annotators per item, and the final VA score is the average (Yu et al., 8 Apr 2026). The resource paper adds that the dataset uses the Self-Assessment Manikin (SAM) protocol with VA emojis to calibrate annotators on the 1–9 scale, and that outlier ratings beyond mean $1$3 standard deviations are discarded before averaging (Lee et al., 30 Jan 2026). Inter-annotator agreement for categorical tuples is reported as F1, while continuous agreement is reported as RMSE against the annotator mean (Yu et al., 8 Apr 2026).

Several structural properties of the datasets are noteworthy. The resource paper reports that restaurant has 18 aspect categories, laptop has 148, and hotel has 47, making laptop markedly more long-tailed (Lee et al., 30 Jan 2026). It also reports that Chinese and Japanese show more compact VA distributions, and that finance has constrained arousal because of its formal genre (Lee et al., 30 Jan 2026). The AILS-NTUA system paper notes that selected training partitions contain NULL labels for implicit aspects or opinions, increasing extraction difficulty, and that distribution shifts in review length, tuple density, and category distributions correlate with development–test performance gaps (Gazetas et al., 5 Mar 2026).

4. Evaluation metrics and exact-match constraints

DimABSA uses different official metrics for regression and structured extraction. For DimASR and DimStance, the official metric is RMSE over VA:

$1$4

where $1$5 is the number of instances, $1$6 are predictions, and $1$7 are gold values (Yu et al., 8 Apr 2026).

For DimASTE and DimASQP, the SemEval task introduces continuous F1 (cF1) to jointly evaluate exact categorical matching and VA prediction error (Yu et al., 8 Apr 2026). Standard F1 is insufficient because these subtasks combine discrete extraction with continuous regression (Lee et al., 30 Jan 2026). The cF1 definition assigns a continuous true positive to each matched prediction:

$1$8

where $1$9 is the set of predictions whose categorical elements exactly match a gold tuple in the same sentence (Yu et al., 8 Apr 2026). For triplets, the exact match condition is on $9$0; for quadruplets, it is on $9$1 (Yu et al., 8 Apr 2026).

The normalized Euclidean distance is

$9$2

This normalization ensures $9$3 on the $9$4 scale (Yu et al., 8 Apr 2026). cRecall is the sum of cTP over matched predictions divided by the number of gold tuples; cPrecision is the same sum divided by the number of predicted tuples; and cF1 is their harmonic mean (Yu et al., 8 Apr 2026).

Two properties of cF1 are central. First, exact categorical match is a gating condition: partial span overlap does not count (Yu et al., 8 Apr 2026). Second, when VA prediction is perfect, cPrecision and cRecall reduce to standard precision and recall (Yu et al., 8 Apr 2026). The appendix example in the task paper gives a total cTP of 1.375 over 4 predictions and 3 gold tuples, yielding $9$5, $9$6, and $9$7 (Yu et al., 8 Apr 2026).

This metric design makes structured extraction bottlenecks highly consequential. The SemEval analysis notes that structured extraction accuracy can bottleneck cF1 even when VA predictions are good, because categorical exact-match gating determines cTP (Yu et al., 8 Apr 2026). A plausible implication is that systems must optimize extraction fidelity and VA calibration jointly rather than treating continuous prediction as an independent post-processing step.

5. Modeling approaches and system designs

The SemEval-2026 task paper reports official baselines for Track A and Track B. For Track A, the baselines are Kimi K2 Thinking with one-shot prompting and Qwen3-14B fine-tuned per dataset with QLoRA; for Track B, the baselines are mBERT with full fine-tuning and Mistral-3-14B with QLoRA (Yu et al., 8 Apr 2026). The top-performing systems exhibit several recurring design patterns: language-specific or language-adapted backbones, parameter-efficient tuning, calibration, uncertainty-aware multitask objectives, retrieval augmentation, and ensembling (Yu et al., 8 Apr 2026).

The AILS-NTUA system for Track A adopts a unified but task-adaptive design. For DimASR it uses language-appropriate encoder backbones, such as DeBERTa-v3 base for English, Chinese RoBERTa wwm-ext for Chinese, ruBERT base cased for Russian, and XLM-R base for Ukrainian and Tatar (Gazetas et al., 5 Mar 2026). Inputs are serialized as aspect-conditioned sentence pairs—for example, Aspect: {a}. Sentence: {x}.—and the encoder output is pooled with attention, followed by two linear heads for valence and arousal (Gazetas et al., 5 Mar 2026). Training uses a composite objective combining MSE, Concordance Correlation Coefficient, and a VA-guided hinge triplet regularizer (Gazetas et al., 5 Mar 2026). For DimASTE and DimASQP, AILS-NTUA uses instruction-tuned LLMs with LoRA, constrained JSON outputs, greedy decoding, post-processing for schema validity, clipping to $9$8, and exact case-sensitive span preservation (Gazetas et al., 5 Mar 2026).

LogSigma focuses on the regression-only setting and targets the difficulty imbalance between valence and arousal. It uses language-specific encoders, aspect-aware [CLS] text [[SEP](https://www.emergentmind.com/topics/semantic-entropy-production-sep-metric)] aspect [SEP] inputs, and dual regression heads (Hikal et al., 26 Mar 2026). Its distinguishing feature is learned homoscedastic uncertainty, with trainable log-variances $9$9 and $5$0 used to weight the valence and arousal losses:

$5$1

According to the paper, this allows the model to automatically balance valence and arousal according to dataset-specific difficulty profiles, and the learned variance ratios differ substantially by language (Hikal et al., 26 Mar 2026). The same work also uses multi-seed ensembling and reports that averaging three seeds reduces RMSE by roughly 3–5% versus the best single run (Hikal et al., 26 Mar 2026).

Among the SemEval winners, PAI is described as using Qwen3-32B fine-tuned with LoRA together with distributional adaptation that aligns predicted VA with the training distribution while preserving valence–arousal correlation, followed by post-hoc calibration via the Sinkhorn algorithm (Yu et al., 8 Apr 2026). TeleAI uses a single multilingual, multi-domain Qwen2.5-7B model fine-tuned with LoRA, Smooth L1 loss with R-Drop consistency, embedding-level PGD adversarial training, and post-hoc linear calibration (Yu et al., 8 Apr 2026). Takoyaki uses retrieval-based in-context learning with BM25 variants to retrieve similar training examples for Gemini 3.0 Pro generation, combined with an agreement-based ensemble and LLM-mined correction rules (Yu et al., 8 Apr 2026). For DimStance, YangS_team uses mDeBERTa-v3-base with aspect-aware marker encoding and a 5-fold ensemble, while CYUT uses a geometry-informed multi-task framework built on Qwen2-7B with LoRA and auxiliary geometry-derived signals from VA annotations (Yu et al., 8 Apr 2026).

An earlier Chinese shared-task system, ZZU-NLP at SIGHAN-2024, proposed Coarse-to-Fine In-context Learning on top of Baichuan2-7B for DimABSA in Chinese restaurant reviews (Zhu et al., 2024). Its second stage retrieves top-3 similar labeled training instances based on cosine similarity between Chinese BERT opinion embeddings, augments prompts with related opinion words and their average VA scores, and filters retrieved examples by valence polarity consistency (Zhu et al., 2024). This predates the SemEval-2026 framework but illustrates retrieval-centric refinement for continuous aspect sentiment prediction in a narrower monolingual setting.

6. Empirical findings, challenges, and limitations

The SemEval-2026 shared task attracted more than 400 participants, resulting in 112 final submissions and 42 system description papers (Yu et al., 8 Apr 2026). Across Track A and Track B, the task paper identifies several broad performance patterns. DimASR RMSE tends to be lower on Chinese and Japanese datasets, with the highest errors on low-resource Tatar; DimASTE achieves the highest cF1 on English and the lowest on Tatar; DimASQP is harder than DimASTE because it adds domain-dependent aspect category classification; and DimStance achieves the lowest RMSE on Chinese and the highest on Swahili (Yu et al., 8 Apr 2026).

Selected leaderboard results illustrate the range. In Track A Subtask 1, LogSigma achieved RMSE 1.1035 on English restaurant and 1.2408 on English laptop, compared with baseline Kimi scores of 2.1461 and 2.1893 and baseline Qwen scores of 2.6427 and 2.8089 (Yu et al., 8 Apr 2026). TeleAI reached 0.5561 on Japanese hotel and 0.6581 on Japanese finance; ICT-NLP and TeleAI were nearly tied on Chinese restaurant with 0.9256 and 0.9265; and HUS@NLP-VNU achieved 0.4841 on Chinese finance (Yu et al., 8 Apr 2026). In Track A Subtask 2, Takoyaki obtained cF1 0.7021 on English restaurant and 0.6366 on English laptop, while PAI reached 0.5793 on Russian restaurant and 0.5787 on Ukrainian restaurant (Yu et al., 8 Apr 2026). In Track A Subtask 3, Takoyaki led English restaurant with 0.6514, PAI led Russian restaurant with 0.5599, and NYCU Speech Lab led Chinese restaurant with 0.5521 (Yu et al., 8 Apr 2026). In Track B, LogSigma achieved 1.4734 on English environmental, 1.3417 on German politics, and 1.7959 on Swahili politics, while YangS_team achieved 0.5468 on Chinese environmental and CYUT achieved 1.1024 on Nigerian Pidgin politics (Yu et al., 8 Apr 2026).

The resource and system papers converge on several sources of difficulty. Low-resource languages such as Tatar and Swahili are repeatedly identified as challenging (Yu et al., 8 Apr 2026, Hikal et al., 26 Mar 2026). Arousal is consistently harder than valence: the DimABSA resource reports higher disagreement for arousal than for valence (Lee et al., 30 Jan 2026), AILS-NTUA reports that valence tends to correlate better than arousal across languages (Gazetas et al., 5 Mar 2026), and LogSigma reports average PCC values of $5$2 for valence and $5$3 for arousal across 15 datasets (Hikal et al., 26 Mar 2026). Long-tailed aspect categories, especially in laptop and hotel, reduce DimASQP performance (Yu et al., 8 Apr 2026, Lee et al., 30 Jan 2026). Exact-match evaluation also sharply penalizes paraphrase-like span differences, malformed structured outputs, and translation-induced span drift (Gazetas et al., 5 Mar 2026).

The papers also note limitations of cross-lingual and cross-cultural comparability. The SemEval task paper states that cultural interpretations of valence and arousal can vary, and therefore results should be interpreted within language–community–domain contexts, even though the annotation protocol uses five native-speaker annotators, consistent 1–9 scales, and shared guidelines (Yu et al., 8 Apr 2026). The resource paper similarly cautions that cultural differences in the interpretation of VA may affect cross-lingual comparability (Lee et al., 30 Jan 2026). The LogSigma paper observes that translation from Russian to English preserves valence PCC better than arousal PCC, suggesting that arousal depends more strongly on language-specific cues such as emphasis, morphology, and intensity markers (Hikal et al., 26 Mar 2026).

Quantitative comparisons to traditional categorical ABSA are limited. The SemEval task paper explicitly states that evaluation is conducted within the dimensional paradigm using RMSE and cF1, and that quantitative comparisons to categorical ABSA baselines are not reported (Yu et al., 8 Apr 2026). The resource paper does report an auxiliary categorical conversion by partitioning valence into positive, neutral, and negative intervals, concluding that categorical ABSA is easier than dimensional ABSA (Lee et al., 30 Jan 2026). This suggests that the added expressiveness of DimABSA comes with measurable additional modeling difficulty.

7. Resources, reproducibility, and future directions

The official SemEval-2026 task resources are available through a GitHub repository containing datasets, annotation guidelines, aspect category lists, the official cF1 evaluation script, a starter kit, and beginner resources (Yu et al., 8 Apr 2026). The competition ran on Codabench with development and evaluation phases, and the last run in the evaluation phase was used for ranking (Yu et al., 8 Apr 2026). Reproduction requires adherence to the specified input–output formats: DimASR expects per-aspect VA predictions, DimASTE outputs structured $5$4 tuples, and DimASQP outputs $5$5 tuples (Yu et al., 8 Apr 2026).

The AILS-NTUA paper provides substantial reproduction detail for Track A, including backbone assignments, learning rates, batch sizes, LoRA target modules, quantization choices, and decoding procedures (Gazetas et al., 5 Mar 2026). LogSigma similarly reports full hyperparameters for its uncertainty-weighted regression setup, including encoder learning rate, log-variance learning rate, warmup, early stopping, and seed values (Hikal et al., 26 Mar 2026). Such reporting reflects a broader pattern in DimABSA research: because cF1 depends on both structured exactness and continuous calibration, implementation-level details in preprocessing, schema validation, clipping, case sensitivity, and deduplication are operationally important rather than incidental (Gazetas et al., 5 Mar 2026).

Several future directions are explicitly proposed. The SemEval task paper calls for expanded language coverage, evaluation of measurement invariance across languages and cultures, improved handling of long-tailed aspect categories, and further refinement of calibration and uncertainty modeling, especially in low-resource and stance settings (Yu et al., 8 Apr 2026). The resource paper suggests adding modalities such as images and audio, incorporating dialog contexts, exploring improved metrics that account for annotation uncertainty, and investigating structured decoders adapted for VA (Lee et al., 30 Jan 2026). LogSigma points to heteroscedastic uncertainty and transfer of learned variance profiles to new languages as promising extensions (Hikal et al., 26 Mar 2026). AILS-NTUA highlights multi-task and multi-language adapter sharing, more robust span canonicalization and VA calibration, and better handling of implicit aspects and opinions (Gazetas et al., 5 Mar 2026).

Taken together, these directions indicate that DimABSA is evolving along three axes simultaneously: richer affect representation, broader multilingual and multidomain generalization, and tighter integration of extraction, calibration, and uncertainty modeling. The available evidence suggests that progress in the area depends less on any single architectural template than on how effectively systems reconcile exact structured outputs with continuous affect prediction under heterogeneous linguistic and domain conditions (Yu et al., 8 Apr 2026).