Augmenting Predictive Models of Turnover with Natural Language Insights

This study evaluates whether open-ended employee comments improve turnover prediction beyond structured HR and satisfaction data. Using recent data from a consulting company, four modeling strategies are compared: (i) structured data only; (ii) structured data plus text represented with TFIDF; (iii) a hybrid transformer model that combines BERT text embeddings with structured inputs; and (iv) a narrative variant that renders structured fields as short sentences. The structuredonly baseline attains 0.74 accuracy. Adding TF-IDF text lifts accuracy to 0.79, and the hybrid transformer reaches 0.84 accuracy with 0.88 F1 on the “Out” class; the narrative variant attains 0.80 accuracy. An ablation indicates that content-preserving text representations (TF-IDF, transformer embeddings) provide most gains, whereas high-level proxies (sentiment and zeroshot topics alone) do not consistently outperform structured-only <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$(0.73)$</tex>. These findings suggest that appropriately modeled freetext materially enhances identification of at-risk employees while remaining practical for HR use. To the best of current knowledge, this is the first study to systematically quantify the incremental value of open-ended employee text for turnover prediction against a structured-only baseline.