# SkillGPT: a RESTful API service for skill extraction and standardization using a Large Language Model

**Authors:** Nan Li, Bo Kang, Tijl De Bie  
**arXiv:** [2304.11060](https://arxiv.org/abs/2304.11060)  
**Date:** 2023-04-17  
**Summary page:** https://www.summarizepaper.com/en/paper/2304.11060v2/

## Key points

- SkillGPT is a tool for skill extraction and standardization from job descriptions and user profiles
- It utilizes an open-source Large Language Model (LLM) called Llama
- SkillGPT performs its tasks through summarization and vector similarity search
- The design choices of SkillGPT were carefully considered and experimented with
- It supports various use cases including different document types, ESCO concept types, and languages
- Users can input free-styled documents in French, English or Dutch and choose the document type between job description and user profile
- The backbone LLM distills the skills contained in the document providing an output skill list in the same language as the original input
- Users can choose which ESCO concept type to standardize the free-style skill descriptions
- Extracted codes for the same content in different languages may differ somewhat
- SkillGPT has limitations such as potential loss of subtle skills when treating summarized text as a single document
- Future plans include addressing these limitations, qualitative & quantitative evaluations on SES & various downstream tasks, optimizing for smaller languages, and supporting all 25 European Union languages.

## Summary

We present SkillGPT, a tool for skill extraction and standardization (SES) from free-style job descriptions and user profiles. SkillGPT utilizes an open-source Large Language Model (LLM) as its backbone, specifically based on Llama which is free for academic use. Unlike previous methods that require supervision or rely on heavy data-preprocessing and feature engineering, SkillGPT performs its tasks in steps through summarization and vector similarity search. This approach balances speed with precision providing users with the convenience of conversational SES efficiently and reliably. The design choices of SkillGPT were carefully considered and experimented with. The justification of these choices as well as an ablation study will be presented in an extended manuscript. The main components of SkillGPT include tools such as summarization and vector similarity search which can be flexible to accommodate other options. SkillGPT supports various use cases including different document types (job description/user resume), ESCO concept types (Skill/Occupation/Occupation group), and languages (En/Fr/Nl). The current version allows for 18 possible use cases. Users can input free-styled documents in French, English or Dutch and choose the document type between job description and user profile. By clicking "Summarize," the backbone LLM distills the skills contained in the document providing an output skill list in the same language as the original input. Additionally users can choose which ESCO concept type to standardize the free-style skill descriptions; the corresponding most plausible ESCO terminologies will be returned however it should be noted that extracted codes for the same content in different languages may differ somewhat. While SkillGPT is efficient economical and often delivers truthful and plausible results it does have some limitations; treating summarized text as a single document might cause subtle skills to be lost since dominant qualities may overshadow them; furthermore there are many options for optimizing performance of LLMs that have not been thoroughly examined due to time limitations and rapidly evolving nature of LLM utilization. In future authors plan to address these limitations by considering qualitative & quantitative evaluations on SES & various downstream tasks in e-recruitment recommendation; they also aim to optimize SkillGPT for smaller languages & support full range of 25 European Union languages.