Which tasks does the ABSA system cover?

All three tasks of SemEval-2026 Task III Track A: valence-arousal regression and triplet and quadruplet extraction.

How does it avoid text generation and backbone tuning?

It decomposes the tasks into typed questions answered with rubric scores, label probabilities, and yes/no judgments, then aligns them with the annotation scheme through 488 coefficients fitted on CPU.

What regression result does the system report?

On valence-arousal regression over ten corpora in six languages the system reaches 1.0645 RMSE, the lowest aggregate error of any participating system, the authors report.

What extraction scores does it reach?

52.09 and 44.06 continuous F1 on triplet and quadruplet extraction, above the fine-tuned Llama-3.3-70B and GPT-OSS-120B baselines.

Where does the accuracy come from, per the ablations?

Supervised calibration roughly halves the raw regression error, exact valence-arousal would add only 4.5 F1 to extraction, and the learned combination of span-boundary evidence carries the extraction systems.

What model answers these typed decisions?

Jev, a frozen model that answers typed questions with rubric scores, label probabilities, and yes/no judgments; only the 488 alignment coefficients are fitted.

Where is the code for this ABSA system?

Code is available at github.com/ZhangYiqun018/jev-dimabsa.

CommunityJev Practices & Patterns

Decide, Don't Generate: Competitive Dimensional ABSA with Jev's Typed Decisions (arXiv 2609.35293)

A system that decomposes all three tasks of SemEval-2026 Task III Track A into typed Jev decisions aligned with 488 CPU-fitted coefficients. It reaches 1.0645 RMSE on valence-arousal regression and 52.09/44.06 continuous F1 on extraction, above fine-tuned Llama-3.3-70B and GPT-OSS-120B baselines.

See highlights

Open arxiv.org ZhangYiqun018/jev-dimabsa

Stars: 7
Language: Python
Format: Article
Published by: Community
Last verified: Oct 2026

ZhangYiqun018/jev-dimabsa

User repository on GitHub

View on GitHub

TypeSafe Jev (System One) on DimABSA, SemEval-2026 Task 3 Track A, all three subtasks: V/A regression (lowest 10-corpus micro RMSE among full-coverage teams), triplet extraction, quadruplet extraction. Pure-Jev inference with BM25 examples and train-fitted calibration.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

Tagscommunity research arxiv classification calibration multilingual

Works forJev for Classification Jev for Scoring Jev for Extraction

Used inJev for Software & DevTools Jev for Media & Entertainment

Languages: Python

Highlights

Decomposes all three tasks of SemEval-2026 Task III Track A into typed Jev decisions.
Aligns decisions with the annotation scheme through 488 coefficients fitted on CPU, with no backbone tuning.
Reaches 1.0645 RMSE on valence-arousal regression over ten corpora in six languages, the lowest of participating systems.
Triplet and quadruplet extraction reach 52.09 and 44.06 continuous F1.
Supervised calibration roughly halves the raw regression error in the ablations.

Watch out

arXiv preprint posted September 28, 2026 (cs.CL) by Yiqun Zhang, Peidong Wang, Zihan Wang, and Shi Feng; 14 pages. The headline numbers against the Llama-3.3-70B and GPT-OSS-120B baselines come from the authors' own evaluation of the shared task, and roughly half the regression gain comes from supervised calibration fitted on task data.

Questions about this entry

Short answers drawn from the linked project or page.

Related terms

Glossary definitions related to this entry.

More Jev resources like this

Emo-Jev: Probabilistic Reasoning for Emotion Classification with Jev (arXiv 2610.08829)

A training-free framework with two implementations that turn emotion classification into Jev judgments: one decomposes tasks into atomic judgments, the other aggregates complementary judgment paths. Across eight datasets standard Jev averages 62.93% macro-F1 versus 67.28% for the strongest LLM baseline.