The paper derives and tests a refined logQ-correction loss that drops the positive item from the sampled denominator and scales each positive example by one minus its estimated model probability.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.IR 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Correcting the LogQ Correction: Revisiting Sampled Softmax for Large-Scale Retrieval
The paper derives and tests a refined logQ-correction loss that drops the positive item from the sampled denominator and scales each positive example by one minus its estimated model probability.