Automatic Taxonomy Generation Based on Large Language Models in the Automotive Domain

The effective extraction and analysis of text-based data from social and niche media by automotive companies can provide more comprehensive and in-depth insights into user needs, thereby enhancing user experience and optimizing product offerings and services. A critical step in enabling effective analysis lies in structured content organized by categorical labels. However, most existing methodologies for constructing labeling systems and developing labeled datasets remain heavily dependent on manual analysis and processing by industry experts, rendering them inadequate to adapt to the rapid shifts in contemporary markets. Recent advancements in large language models (LLMs) offer promising avenues for achieving automated data extraction. This paper introduces a methodology for constructing large-scale labeling systems through prompt generation techniques at various stages using LLMs, coupled with pseudo-labeling to support deployment and application. The proposed approach features a two-stage framework, employing LLMs to execute LLM based label generation and text classification. Experimental results indicate that this method can autonomously generate accurate labeling systems and classification datasets, achieving an optimal bal-ance between classification accuracy and operational efficiency.