TeleAI at SemEval-2026 Task 3: Large Language Models for Dimensional Aspect-Based Sentiment Analysis
Year: 2026
Type: conference-paper
Abstract: We present TeleAI's system for SemEval-2026 Task 3 (Track A), focusing on Subtask 1 (Di-mASR): given a text and aspect terms, the system predicts continuous Valence-Arousal (VA) scores under the official RM SE V A metric (Yu et al., 2026).Our regression framework uses Qwen2.5-7B as a feature extractor with LoRA-based parameter-efficient finetuning, multilingual/multi-domain mixed training, R-Drop-style consistency regularization, embedding-level PGD adversarial training, output interval mapping, and post-hoc linear calibration.Experimental observations and ablation trends indicate that mixed training and robustness/consistency regularization substantially reduce RM SE V A , yielding competitive performance across multiple language-domain settings.We also tried lightweight direct-JSON prompting for Subtasks 2 and 3, but those runs were less competitive, so we focus on Subtask 1.
Institution: China Telecom
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7166806618