Secure and Scalable Skill Extraction from Real-World Resumes using RoBERTa-Based NER with Blockchain-Enabled Data Integrity

It must be able to stand alone since abstracts are frequently given apart from Automated resume parsing is central to modern digital recruitment, but two issues persist in the literature: current systems are mostly validated on small, curated datasets that are not representative of the noise and heterogeneity of real resumes, and they do not provide any mechanism to ensure that candidate data extracted remains unaltered once collected. In this paper, we propose a secure and scalable framework that combines a RoBERTa-based Named Entity Recognition (NER) model and an Ethereum blockchain layer for tamper-evident data management. We formulate skill extraction as a BIO-tagged sequence-labeling task and fine-tune the model on a manually annotated real-world corpus of 4,473 resumes collected in JSON format, under a three-stage training schedule with mixed-precision optimization. Skill profiles are extracted and hashed with SHA-256 and anchored to an Ethereum smart contract that allows employers to verify the integrity of a candidate's skill profile without exposing the underlying resume text. On the held-out validation split, the proposed model achieves an entity-level Precision, Recall, F1-score and token-level Accuracy 0.640, 0.666, 0.653 and 0.910 respectively on the seqeval evaluation protocol. This exceeds the rule-based (F1 = 0.37), Conditional Random Field (F1 = 0.50), BiLSTM-CRF (F1 = 0.56) and BERT based (F1 = 0.62) baselines evaluated on the same dataset and under the same experimental conditions. The collapse in performance on small samples and substantial recovery at full scale is confirmed by a dataset-scale ablation, highlighting the value of large, realistic corpora for training on this task. We demonstrate the addition of blockchain verification to be about 2.5 ms of end-to-end latency per profile (store-and-verify cycle, ~160.7K gas for storage and ~27.0K gas for verification, over 200 simulated profiles on a local, Ethereum compatible development network) and the benefits of full tamper-evidence, transparent traceability and controlled access to candidate skill records. The results show that the combination of contextual transformer-based extraction and decentralized integrity verification provide an accurate and trustworthy pipeline for large-scale, real-world recruitment systems.