TeamLasse at SemEval-2026 Task 3: A Hybrid Generative-Discriminative Framework for Dimensional Aspect-Based Sentiment Analysis
Year: 2026
Type: conference-paper
Abstract: In this paper, we present our system for SemEval-2026 Task 3 Track A: Dimensional Aspect-Based Sentiment Analysis (DimABSA).The core objective is to extract structural sentiment elements-such as aspects, opinions, and categories-from text and predict their corresponding continuous Valence-Arousal (VA) scores.The primary challenge lies in simultaneously handling structural extraction and continuous numerical regression across highly imbalanced datasets encompassing multiple languages and domains.To address this complexity, we propose a decoupled, two-stage hybrid generative-discriminative framework.A generative Large Language Model first extracts structured sentiment tuples, while an encoder-based language model performs the continuous VA regression.To foster cross-lingual and crossdomain generalization, we train our models using a targeted data balancing mechanism.Our results demonstrate that this approach is highly effective for structural extraction in data-scarce scenarios, notably securing 2nd place in the underrepresented Japanese Hotel domain.However, Subtask 1 results highlight a structural limitation of our pipeline: because this subtask provides only the aspect term, our regression model-optimized for full aspect-opinion tuples-experiences a performance drop when the explicit opinion term is missing during inference.
Authors: Lasse Strothe, Shaghayegh Kolli, Jana Diesner
Institution: Technical University of Munich
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
Funder: Technische Universität München
OpenAlex ID: W7166881725