INFDok - Dokumentenvolltextserver des Fachbereichs Informatik

Kurzfassung auf Englisch:

Online job platforms provide access to a large number of employment opportunities, yet conventional title- and keyword-based search methods often fail to identify positions that genuinely match a user’s skills, experience, and qualifications. This thesis proposes a transparent and fully open-source job-matching system that combines LLM-based requirement extraction (LRE) with token-embedding-driven similarity search (TSS). The system processes unstructured job descriptions and extracts mandatory skills, experience requirements, and qualifications into a structured format. These requirements are then transformed into vector representations and compared with similarly structured user profiles using semantic similarity measures. The approach relies exclusively on lightweight open-source models and does not require task-specific fine-tuning. Four open-source Large Language Models were evaluated on a dataset of 209 real-world job descriptions using an automated LLM-as-a-judge framework. Qwen3-8B achieved the highest overall performance, with an aggregated score of 0.50. Although schema-guided generation successfully ensured syntactically valid and machine-readable outputs, the models continued to exhibit limitations in semantic extraction, particularly regarding completeness, categorization, and hallucinated requirements. A comparison with LinkedIn job recommendations showed that the proposed architecture provides transparent and traceable similarity scores. However, incomplete or inaccurate extraction frequently reduced the scores of otherwise relevant postings. The results demonstrate that lightweight opensource LLMs provide a viable foundation for accessible and interpretable semantic job matching, although further improvements in extraction quality are required before the system can compete with industrial-grade solutions.