Arctic-ABSA: Smarter, multilingual aspect sentiment with reasoning
Ever wish reviews told you exactly how people feel about the camera, battery, or price? That's aspect-based sentiment analysis (ABSA).
Arctic-ABSA is a new family of reasoning-infused models built for real-world ABSA. Trained on a large blend of public and carefully generated synthetic data—about 20× the classic SemEval14 size—they:
- Recognize five sentiments: positive, negative, neutral, mixed, and unknown.
- Judge both per-aspect opinions and overall text sentiment.
- Work across multiple languages with one model.
By injecting step-by-step reasoning (including a new pretraining method for encoder models), Arctic-ABSA generalizes better and stays accurate when the data shifts.
Results: a 395M-parameter encoder and an 8B-parameter decoder set new state-of-the-art ABSA scores, beating leading general LLMs (e.g., GPT-4o, Claude 3.5 Sonnet) by up to 10 percentage points. One multilingual model maintains 87–91% accuracy across six languages without hurting English.
Plus, the team releases ABSA-mix, a large benchmark spanning 17 public datasets and 92 domains.
Paper: https://arxiv.org/abs/2601.03940v1
Paper: https://arxiv.org/abs/2601.03940v1
Register: https://www.AiFeta.com
NLP SentimentAnalysis ABSA AI LLM Multilingual Reasoning Benchmark Research
