Enhancing Large Language Models for Fashion Smart Manufacturing via Dynamic Collaborative Routing-Based Retrieval...

Area of Science:

• Artificial Intelligence
• Machine Learning
• Information Retrieval

Background:

• Large language models (LLMs) in fashion manufacturing require external knowledge for reliable technical support.
• Specialized terminology and complex queries in fashion manufacturing challenge existing retrieval methods.
• Simple query-matching leads to irrelevant and redundant document chunks, hindering LLM performance.

Purpose of the Study:

• To propose a retrieval optimization framework for enhancing LLM performance in fashion manufacturing knowledge retrieval.
• To address the issue of contextually loose and redundant document retrieval.
• To improve the quality of technical support and decision-making assistance provided by LLMs.

Main Methods:

• A dynamic capsule routing network with an embedded semantic graph (SGDCR) framework is proposed.
• The framework employs a two-step process: filtering and reranking of retrieved documents.
• A capsule routing mechanism models semantic relations and contribution scores among documents for filtering, followed by deep semantic similarity matching for reranking.

Main Results:

• The SGDCR framework effectively filters irrelevant and redundant document chunks.
• Reranking integrates relevance and contribution scores for contextually coherent document prompts.
• Experimental results show superior performance compared to existing reranking approaches on QA datasets.

Conclusions:

• The proposed SGDCR framework significantly enhances the accuracy and efficiency of fashion manufacturing knowledge retrieval for LLMs.
• This approach improves process control and intelligent production efficiency.
• The method demonstrates effectiveness in both general open-domain QA and specialized fashion manufacturing QA systems.