Advanced Job Classification Framework Combining BERT, CNN2D

This paper proposes a novel two-stage system for job title identification in online job advertisements, addressing challenges of small datasets and noisy text. The first stage employs Bidirectional Encoder Representations from Transformers (BERT) to classify job ads by sector, such as IT or agriculture, effectively capturing contextual semantics. The second stage uses the Euclidean Distance algorithm to match job titles within the predicted sector, achieving high accuracy without requiring extensive labeled data. Comparing BERT with models like TFIDF and Word2Vec, this approach demonstrates superior performance. Further refinements involve feature selection via CHI2 and the integration of CNN2D, boosting accuracy to 96%. This methodology is validated on real-world datasets and reveals insights into emerging job market demands.