QuadAI at SemEval-2026 Task 3: Ensemble Learning of Hybrid RoBERTa and LLMs for Dimensional Aspect-Based Sentiment Analysis

The QuadAI system for SemEval-2026 Task 3 achieves superior performance in dimensional aspect-based sentiment regression by employing an ensemble learning framework that combines a hybrid RoBERTa encoder with large language models, leveraging the complementary strengths of both architectures to significantly reduce RMSE and improve correlation scores.

The Gist

Imagine you are trying to understand how a person is feeling just by reading their text message. You don't just want to know if they are "happy" or "sad." You want to know how happy they are (from a tiny smile to a screaming joy) and how intense that feeling is (from a calm contentment to a frantic excitement).

This is what the QuadAI team at SemEval-2026 tried to solve. They built a system to measure these complex emotions, specifically looking at two scales: Valence (Positive vs. Negative) and Arousal (Calm vs. Intense).

Here is how they did it, explained with simple analogies:

1. The Two Experts: The "Mathematician" and the "Storyteller"

The team didn't rely on just one type of AI. They used two different "experts" and asked them to work together.

• Expert A: The Hybrid RoBERTa (The "Mathematician")
  Think of this as a very smart, fast calculator that has been trained on millions of sentences.

  • The Problem: Usually, these calculators either try to guess a specific number (Regression) or pick a bucket from a list (Classification). Guessing a number can be wobbly and unstable. Picking a bucket is stable but lacks precision.
  • The Solution: The team built a "Hybrid" calculator. It does both at the same time. It guesses the exact number and picks the bucket. Then, it takes the average of both guesses.
  • The Analogy: Imagine you are guessing the weight of a watermelon. One person guesses "5.2 kg" (Regression), and another says "It's in the 5-to-6kg bucket" (Classification). By averaging these two methods, you get a much more stable and reliable guess than if you used just one method.

• Expert B: The Large Language Model (The "Storyteller")
  This is the modern, giant AI (like the ones you might chat with) that understands context and nuance very well.

  • The Trick: Instead of just asking the AI to guess, the team gave it a "cheat sheet" of similar examples before it answered. This is called In-Context Learning.
  • The Quality Control: Before using these examples, they had a "Triple-LLM" panel (three different AIs) act as judges. If the judges agreed an example was weird or wrong, they threw it out. This ensured the AI was only learning from high-quality stories.

2. The Coach: Ensemble Learning

Having two experts is great, but what if they disagree? The team acted as a Coach to combine their answers. This is called Ensemble Learning.

• The Strategy: They didn't just let the two experts shout their answers and pick the loudest one. They used a "Ridge Regression" method, which is like a smart coach who listens to both experts but knows that one is usually better at math and the other is better at nuance.
• The Result: The Coach blends their answers together. In their tests, this "Coach" approach was significantly better than listening to either expert alone. It reduced errors and made the predictions much more accurate.

3. The Results: How Did They Do?

The team tested their system on two types of data: reviews about Laptops and reviews about Restaurants.

• The "Mathematician" (Hybrid RoBERTa): Did a great job on its own, beating standard models by a wide margin.
• The "Coach" (Ensemble): When the Coach combined the Mathematician and the Storyteller, the results were even better. The errors dropped significantly, and the system's ability to predict the right emotional intensity improved.

4. The Catch (Limitations)

Because of a tight deadline and some unexpected issues, the team couldn't submit their best possible version (the one with the Coach and the Storyteller) to the final competition. They only submitted the "Mathematician" version.

Even with this "lightweight" version, they did very well, ranking in the top half of all teams. They are now planning to test their full "Coach + Storyteller" system on new data and even try it on other languages like Chinese.

Summary

In short, the QuadAI team built a sentiment analyzer that works like a dream team:

1. A Hybrid Calculator that is stable and precise.
2. A Storyteller AI that understands context and nuance.
3. A Smart Coach that blends their opinions to get the most accurate emotional reading possible.

They proved that when you combine different types of AI, you get a result that is much stronger than the sum of its parts.

Technical Summary

Here is a detailed technical summary of the paper "QuadAI at SemEval-2026 Task 3: Ensemble Learning of Hybrid RoBERTa and LLMs for Dimensional Aspect-Based Sentiment Analysis."

1. Problem Definition

The paper addresses Dimensional Aspect-Based Sentiment Regression (DimABSA), specifically focusing on Track-A1 (Valence-Arousal Prediction). Unlike traditional Aspect-Based Sentiment Analysis (ABSA) which classifies sentiment into discrete categories (e.g., positive/negative), this task requires predicting continuous scores for two psychological dimensions:

• Valence: The degree of positivity or negativity.
• Arousal: The intensity of the emotion (from low to high).

The challenge lies in modeling the complex, continuous nature of human emotion while handling data scarcity and domain-specific variations (e.g., Laptop vs. Restaurant reviews). The authors participated in the SemEval-2026 shared task, aiming to minimize the Root Mean Square Error (RMSE) and maximize Pearson correlation ($\rho$) between predicted and ground-truth scores.

2. Methodology

The proposed system, QuadAI, employs a three-stage architecture combining encoder-based models, Large Language Models (LLMs), and ensemble learning.

A. Hybrid RoBERTa Encoder

The core of the system is a hybrid encoder that leverages the strengths of both regression and classification to improve prediction stability.

• Architecture: Uses a pre-trained RoBERTa encoder.
• Dual Heads:
  1. Regression Head: Directly predicts continuous Valence and Arousal scores.
  2. Discretized Classification Head: The continuous target space is split into $n$ bins (e.g., 31 bins). The model predicts a probability distribution over these bins using cross-entropy loss. The final classification output is the expected value of the bins ($\sum p_i c_i$).
• Fusion: The final prediction is a weighted average of the regression output ($\hat{y}_{reg}$) and the classification expected value ($\hat{y}_{cls}$):
  $$\hat{y} = w \cdot \hat{y}_{reg} + (1-w) \cdot \hat{y}_{cls}$$
  This approach aims to balance the smoothness of regression with the stability and confidence of discretized classification.

B. Large Language Models (LLMs) & In-Context Learning (ICL)

The authors explored using LLMs (specifically Gemini, Claude, and GPT-5.2) via in-context learning.

• Prompt Engineering: They compared zero-shot prompting against few-shot prompting using randomly selected examples versus semantically similar examples (selected via sentence embedding similarity).
• Data Cleaning (Tri-LLM Consensus): To ensure high-quality in-context examples, a "Triple-LLM" workflow was implemented:
  1. Training data is clustered using HDBScan based on Valence/Arousal dimensions.
  2. Three different LLMs are prompted to identify outliers within each cluster.
  3. Instances where all three LLMs agree are outliers are removed from the candidate pool.
• Best Performer: Gemini was selected as the primary LLM for the final prompting stage due to superior performance.

C. Ensemble Learning (Prediction-Level Fusion)

The system combines the Hybrid RoBERTa and LLM predictions using Model Stacking (Late Fusion).

• Combiner: A Ridge Regression model acts as the meta-learner to combine predictions.
• Inputs: The combiner takes the outputs of the base models ($\hat{y}_1, \hat{y}_2$) and optional external features (VADER lexicon scores: compound, positive, negative, neutral).
• Training: The combiner is trained using Out-of-Fold (OOF) predictions to prevent label leakage.
• Mathematical Formulation:
  $$\hat{y}_{final} = \mathbf{w}^T [\hat{y}_1, \hat{y}_2, f] + b$$
  Where $f$ represents the VADER features.

3. Key Contributions

1. Hybrid Encoder Design: Introduction of a RoBERTa-based model that jointly optimizes regression and discretized classification heads, demonstrating that averaging these outputs yields more stable and accurate predictions than either method alone.
2. LLM Data Cleaning Pipeline: A novel consensus-based filtering mechanism using three distinct LLMs to curate high-quality in-context learning examples, reducing noise in the prompt context.
3. Ensemble Strategy: Demonstration that combining encoder-based models (efficient, domain-adaptive) with LLMs (strong reasoning, few-shot capability) via Ridge Stacking significantly outperforms individual models.
4. Comprehensive Evaluation: Extensive ablation studies on hyperparameters (number of bins, loss weights) and feature inclusion (VADER) across multiple domains (Laptop, Restaurant, Environmental Protection).

4. Experimental Results

The system was evaluated on the SemEval-2026 Development Sets.

• Hybrid RoBERTa Performance:
  • On the Laptop domain, the Hybrid model achieved an RMSE of 0.7361, outperforming the Regression-only model (0.7836) and the Bin-only model (0.7898).
  • On the Restaurant domain, the Hybrid model showed a massive improvement, reducing MSE from 0.8176 (Regression) to 0.4919, and RMSE from 0.9042 to 0.7013.
• LLM Performance:
  • The LLM (Gemini) alone achieved an RMSE of 0.695 on the Laptop set, outperforming the Hybrid RoBERTa (0.7361) and improving the mean Pearson correlation to 0.757.
• Ensemble Performance:
  • The Ensemble (Hybrid + LLM) achieved the best results.
  • Laptop Domain: The ensemble reduced RMSE to 0.6344 (compared to 0.7361 for Hybrid and 0.695 for LLM).
  • Feature Analysis: The inclusion of VADER features did not improve performance (slightly degraded scores), suggesting that the neural/LLM components already captured the necessary sentiment signals, making the lexicon features redundant or noisy.
• Submission Rankings:
  • Due to submission constraints, only the Hybrid RoBERTa (without LLM/Ensemble) was submitted.
  • Laptop: Ranked 16th/30 (Score: 1.4062 vs. Best: 1.2408).
  • Restaurant: Ranked 22nd/33 (Score: 1.3632 vs. Best: 1.1035).
  • Despite not using the full ensemble, the Hybrid model significantly outperformed the baseline and bottom-tier teams.

5. Significance and Future Work

• Significance: The paper validates that dimensional sentiment analysis benefits from a multi-faceted approach. It proves that combining the structural stability of discretized classification with the fluidity of regression, and further augmenting this with LLMs via ensemble learning, creates a robust system for continuous emotion prediction.
• Limitations: The full ensemble (LLM + Hybrid) was not submitted to the test set due to time constraints; results are based on development set performance. The system was primarily tested on English data.
• Future Directions:
  • Applying the full ensemble pipeline to the test set.
  • Exploring generalizability on non-English languages (e.g., Chinese).
  • Implementing automated hyperparameter tuning (e.g., Optuna) for the hybrid configuration.
  • Investigating more complex ensemble methods beyond simple ridge stacking.

In conclusion, QuadAI demonstrates that a hybrid encoder-LLM ensemble is a highly effective strategy for dimensional aspect-based sentiment regression, offering a promising direction for future NLP tasks involving continuous emotional modeling.