Parse Trees Guided LLM Prompt Compression.

Area of Science:

• Natural Language Processing
• Artificial Intelligence
• Computational Linguistics

Background:

• Large Language Models (LLMs) benefit from rich context, but long prompts increase computational cost and may exceed input limits.
• Existing prompt compression methods include generative approaches (prone to hallucination) and selective methods (lacking linguistic rules and global structure awareness).

Purpose of the Study:

• To introduce PartPrompt, a novel selective prompt compression method that addresses the limitations of current techniques.
• To improve the efficiency and effectiveness of providing context to LLMs through optimized prompt length.

Main Methods:

• PartPrompt employs linguistic rules to generate parse trees for sentences and calculates local information entropy.
• It organizes parse trees into a global structure reflecting hierarchical dependencies (sentences, paragraphs, sections).
• Root-ward and leaf-ward propagation adjust node values, followed by recursive pruning based on these values.

Main Results:

• PartPrompt achieves state-of-the-art performance across diverse datasets, metrics, compression ratios, and LLMs for inference.
• Ablation studies confirm the effectiveness of PartPrompt's design components.
• Additional experiments demonstrate superior coherence and effectiveness in extreme long prompt scenarios.

Conclusions:

• PartPrompt offers a linguistically informed and structurally aware approach to prompt compression.
• The method significantly enhances LLM performance while managing prompt length and maintaining coherence.
• PartPrompt represents a significant advancement in optimizing LLM input for various applications.