EAOS-Llama-3.2-1B: Emotion-Aware Opinion Summarization

This repository contains the weights and evaluation code for EAOS-Llama-3.2-1B, the fine-tuned model presented in the ACL 2026 paper: "Bridging Cognition and Affect: Emotion-Aware Opinion Summarization using LLMs".

Model Overview

Standard opinion summarization systems primarily reduce affect to binary polarity, thereby losing the nuanced emotional perspectives embedded in user text. The Emotion-Aware Opinion Summarization (EAOS) framework addresses this limitation by grounding abstract summaries in Plutchik's Wheel of Emotions (Joy, Trust, Fear, Surprise, Sadness, Disgust, Anger, Anticipation).

This model is a parameter-efficient fine-tuned version of meta-llama/Llama-3.2-1B-Instruct trained on the synthetic EAOS-SUMM dataset. Despite a constrained 1B parameter count, it matches the zero-shot performance of 70B-scale foundation models on the emotion-aware summarization task.

• Base Architecture: meta-llama/Llama-3.2-1B-Instruct
• Task: Emotion-Aware Opinion Summarization
• Language: English
• Publication: ACL 2026

Usage and Inference

The model has internalized the Plutchik framework through instruction tuning. Therefore, it does not require complex zero-shot prompting definitions. Users must construct the input exactly as shown below.

Inference Pipeline (Python)

from transformers import pipeline

generator = pipeline("text-generation", model="cuteasduck/EAOS-Llama-3.2-1B", device_map="auto")

product_title = "Product Name"
reviews = """
Review 1:
Title: Example Title
Text: Example review text containing specific emotional cues.
""" 

prompt = f"""<|im_start|>system
You are a helpful assistant specialized in creating emotion-aware summaries of product reviews.<|im_end|>
<|im_start|>user
Product: {product_title}

Reviews:
{reviews}<|im_end|>
<|im_start|>assistant
"""

result = generator(
    prompt, 
    max_new_tokens=400, 
    temperature=0.2,
    top_p=0.95,
    top_k=25,
    do_sample=True,
    return_full_text=False
)

print(result[0]['generated_text'].strip())

Note on Decoding Parameters

The decoding parameters provided above (temperature=0.2, top_p=0.95, top_k=25) match the rigorous evaluation methodology described in our paper. These parameters guarantee deterministic, reproducible summaries for academic benchmarking. For production or downstream applications prioritizing stylistic fluidity and creativity, users may opt to increase the temperature (e.g., temperature=0.7) combined with an appropriate repetition penalty (e.g., repetition_penalty=1.15).

Citation

If you utilize this model or the EAOS framework in your research, please cite the corresponding ACL 2026 publication:

@inproceedings{EAOS-2026,
  title={Bridging Cognition and Affect: Emotion-Aware Opinion Summarization using LLMs},
  author={Arnav},
  booktitle={Proceedings of the Annual Meeting of the Association for Computational Linguistics (ACL)},
  year={2026}
}