The Classics at SemEval-2026 Task 3: Combining Transformer Models and LLM-Generated Annotations for Dimensional Aspect-Based Sentiment Analysis

Technical Summary: The Classics at SemEval-2026 Task 3

Problem Statement
This paper addresses the challenges of Dimensional Aspect-Based Sentiment Analysis (DimABSA), a paradigm shift from traditional categorical sentiment classification (e.g., positive/negative) to predicting fine-grained, continuous real-valued scores. Specifically, the authors tackle SemEval-2026 Task 3, which requires mapping sentiment onto a two-dimensional space of valence (positivity) and arousal (intensity). The task is divided into two subtasks:

1. Subtask 1 (Dimensional Aspect Sentiment Regression): Predicting valence and arousal scores for a given text and a specific aspect.
2. Subtask 3 (Dimensional Aspect Sentiment Quad Prediction): An end-to-end extraction task requiring the simultaneous identification of aspects, categories, opinion phrases, and their corresponding continuous valence/arousal scores.

The authors note that traditional models often oversimplify complex human emotions and lack the opacity required for nuanced decision-making, necessitating methods capable of continuous regression and structured extraction.

Methodology
The authors propose two distinct pipelines tailored to the specific requirements of each subtask:

• Subtask 1: Discriminative Ensemble with LLM Augmentation

  • Encoder Ensemble: The core approach involves fine-tuning a weighted ensemble of transformer-based encoder models (RoBERTa-Base, RoBERTa-Large, and DeBERTaV3-Large for English; XLM-RoBERTa-Base and XLM-RoBERTa-Large for Russian). A linear regression head is applied to the [CLS] token to output the two continuous variables.
  • LLM-Based Augmentation (Russian Track): To address the difficulty of extracting subtle sentiment nuances in Russian, the authors introduce a zero-shot data augmentation technique. An instruction-tuned Large Language Model (LLM) generates synthetic descriptive rationales (focusing on pleasantness and intensity) for each training instance. These rationales are concatenated with the original input text to provide explicit descriptive context for the encoder models.
  • Ensemble Strategy: Predictions are aggregated using a weighted average, with weights empirically optimized based on development set performance.

• Subtask 3: Unified Generative Framework

  • Decoder LLM: Instead of a multi-step pipeline (which risks error propagation), the authors employ a unified generative approach. A decoder-based LLM (Llama-3.1-8B) is fine-tuned using Low-Rank Adaptation (LoRA) to perform structured prediction.
  • Joint Estimation: The model is trained to process input text and directly generate a structured sequence containing the full quadruplet (Aspect, Category, Opinion, Valence-Arousal) in a single forward pass. This allows the model to learn the relationship between descriptive textual phrases and their corresponding emotional intensity simultaneously.

Key Results

• Subtask 1 Performance:
  • English Track: The weighted ensemble achieved competitive results, with the final system outperforming organizer baselines (e.g., Kimi-K2 Thinking and Qwen-3 14B) in both Restaurant and Laptop domains. For instance, in the Restaurant domain, the system achieved an RMSE of 1.23 compared to the baseline's 2.14.
  • Russian Track: The LLM-based augmentation proved effective, reducing the RMSE of XLM-RoBERTa models from the 1.43–1.45 range to 1.40–1.41.
• Subtask 3 Performance:
  • The end-to-end generative LLM approach achieved a Continuous F1-score (cF1) of 0.3072 on the Laptop domain test set.
  • Ablation Findings: Contrary to the initial hypothesis that encoders are superior for numerical regression, the hybrid pipeline (using an encoder to predict scores for spans extracted by the LLM) underperformed (cF1 0.3024). This suggests the decoder LLM successfully captures the relationship between text and emotional intensity without a separate regression module.
  • Sensitivity: The study highlighted high sensitivity to prompt phrasing and learning rates; alternative prompts or higher learning rates (2e-4) significantly degraded performance.

Significance and Claims
The paper claims that its primary contribution lies in demonstrating the efficacy of unified generative frameworks for complex sentiment extraction tasks, challenging the conventional reliance on separate extraction and regression modules.

• LLM Capabilities: The results indicate that instruction-tuned decoders can effectively handle continuous numerical regression when trained jointly with extraction tasks, a finding that contradicts the assumption that encoders are strictly more stable for regression.
• Contextual Augmentation: The work validates that using LLMs to generate synthetic descriptive context can significantly enhance smaller encoder models, particularly in languages with complex sentiment nuances like Russian.
• Future Directions: The authors suggest that extending the purely generative approach to standard dimensional regression (Subtask 1) represents a promising direction, though they were unable to fully explore this within the shared task timeline. They also note limitations regarding the scalability of their methods to larger models and the sensitivity of results to hyperparameter and prompt design.

The system secured 7th place out of 18 in Subtask 3 and placed in the top 16 across various domains in Subtask 1, demonstrating competitive efficacy in fine-grained sentiment analysis.