linkedin / detext

DeText: A Deep Neural Text Understanding Framework for Ranking and Classification Tasks
BSD 2-Clause "Simplified" License
1.26k stars 134 forks source link

Support shallow tower #68

Closed StarWang closed 3 years ago

StarWang commented 3 years ago

Description

Support shallow tower. Final score becomes final_score = MLP score + shallow tower score (linear).

This is to reserve the original feature space of sparse features (MLP introduce non-linearity).