DeepCT Deep Contextualized Term Weighting for First-Stage Retrieval

A supervised deep contextualized term-weighting framework for first-stage bag-of-words retrieval that maps each word's first-subword contextualized BERT embedding T_{t, c} in text c to a scalar term importance prediction y^_{t, c} = w^T T_{t, c} + b by minimizing per-token Mean Squared Error loss_MSE = sum_c sum_t (y_{t, c} - y^_{t, c})^2 against relevance-derived recall targets: in DeepCT-Index, ground-truth passage term importance is defined as Query Term Recall QTR(t, d) = |Q_{d, t}| / |Q_d| in [0, 1] (the fraction of queries relevant to passage d that contain term t) and query-independent offline predictions are scaled to integer pseudo-frequencies TF_DeepCT(t, d) = round(y^_{t, d} * N) (with N = 100) that replace raw term frequencies TF(t, d) in standard inverted index postings; in DeepCT-Query, ground-truth query term importance is defined as Term Recall TR(t, q) = |D_{q, t}| / |D_q| in [0, 1] to weight bag-of-words and sequential dependency queries.