LM-SODP: Language Model Self-Optimizing Discrete Prompt for Aspect Based Sentiment Analysis.

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Discrete prompts are standard for Large Language Models (LLMs) due to interpretability and compatibility.
• Optimizing prompts for fine-grained tasks like Aspect-Based Sentiment Analysis (ABSA) is difficult due to error propagation and manual design effort.

Purpose of the Study:

• To present LM-SODP, a Reinforcement Learning (RL) framework for automated discrete prompt and prediction order optimization in ABSA.
• To improve LLM performance on ABSA by enhancing task-specific information utilization and reducing output uncertainty.

Main Methods:

• Utilized a distilled GPT-2 model within an RL framework (LM-SODP).
• Optimized discrete prompts to reduce output entropy and improve task-specific information use.
• Developed an independent method to determine optimal execution sequences for ABSA subtasks.

Main Results:

• LM-SODP demonstrated stable performance improvements across various conditions and domains.
• The framework effectively guides LLMs with limited computational resources using optimized prompts.
• Achieved reduced error propagation and uncertainty in ABSA predictions.

Conclusions:

• LM-SODP offers an automated approach to prompt optimization for ABSA, overcoming limitations of fixed prediction orders.
• The method enhances LLM efficiency and effectiveness in fine-grained NLP tasks.
• Opens new possibilities for automated prompt token generation and improved human-AI interaction.