Analisis Perbandingan Few-Shot Dan Iterative Optimization Prompting Pada Sistem Large Language Model Ensemble Untuk Deteksi Emosi Multilabel Berbahasa Inggris

Abstract

Multilabel emotion detection in English text remains a challenging task in Natural Language Processing due to the ability of a single text to express multiple emotions simultaneously and the presence of contextual ambiguity. This study aims to analyze the impact of few-shot prompting and compare it with iterative prompt optimization within a Large Language Model (LLM) ensemble framework for multilabel emotion detection. The proposed system adopts the LLM ensemble pipeline from SemEval-2025 Task 11 and employs five LLMs, namely ChatGPT-4o, DeepSeek-V3, Mistral-7B, Qwen-2.5-0.5B, and Gemma-2B. Lightweight local models are fine-tuned using the AdaLoRA technique, while three few-shot examples are automatically selected using semantic search based on sentence embeddings. Predictions from individual LLMs are then aggregated through a twostage ensemble approach involving neural networks, XGBoost, LightGBM, linear regression, and weighted voting. Evaluation is conducted on the SemEval-2025 Task 11 Track A dataset using the macro F1-score metric. Experimental results show that the few-shot prompting configuration achieves its highest macro F1- score of 0.694 using the weighted voting method, whereas the iterative prompt optimization approach employed by the PAI system attains a higher macro F1- score of 0.832. Although the overall performance of few-shot prompting is lower than that of iterative prompt optimization, the findings indicate that few-shot prompting provides stable and consistent performance across models and can serve as an alternative prompt engineering strategy for LLM ensemble-based multilabel emotion detection under limited computational resources.