Presentation Information
[1Yin-B-62]Extraction of Bidding Decision Factors in a Competitive Bidding Career Service Based on BERT Feature Attribution
〇Yoichi Kitahara1 (1. Livesense Inc.)
Keywords:
BERT,feature attribution,explainable AI,XAI
This study proposes an analysis method based on ModernBERT feature attribution to extract bidding factors for job seekers in a competitive-bidding career service. First, ModernBERT is fine-tuned to predict whether a company would bid on a candidate (sending a scout message with a salary offer) based on their resume text data. Next, to identify the tokens contributing to the prediction, multiple feature attribution methods are applied, including LIME, Integrated Gradients, AttnLRP, DecompX, GAF (Generalized Attention Flow), and LibraGrad. Comparative evaluation shows that DecompX, GAF, and LibraGrad are particularly effective for this task.