A Decision Support System Based on Transformer-Driven Sentiment Analysis of Social Media Data

The growing availability of social media data offers new opportunities for decision support systems (DSS) in large-scale human resource screening. This study proposes a technology-driven DSS architecture integrating transformer-based sentiment analysis to support early-stage candidate profiling. Its novelty lies in combining IndoBERT-based sentence embeddings with a structured DSS layer that aggregates tweet-level sentiment into risk-aware recommendations, rather than treating sentiment classification as a standalone output. Using a quantitative experimental design, 5,000 public posts from 100 users were processed through an NLP pipeline incorporating mean-pooled embeddings, feature engineering, principal component analysis, and Support Vector Machine classification. The model achieved 69.1% accuracy, with weighted precision, recall, and F1-score of 0.694, 0.691, and 0.691, outperforming baseline models by 6.5–15.0 percentage points. Sentiment outputs are treated as probabilistic behavioral signals within an advisory DSS framework, not direct indicators of candidate suitability. Preliminary validation on 50 cases showed moderate correlations (ρ = 0.52–0.61) with conventional assessments. The system remains non-automated, incorporating confidence thresholds, uncertainty handling, and mandatory human oversight. Limitations include moderate accuracy, reliance on text-only data, and linguistic ambiguity.