Events

Reinforcement learning to align NLG learning objectives with desired model behaviors

Speaker: Shiyue Zhang

Location: 60 Fifth Avenue, Room 204

Date: Thursday, March 23, 2023

We have witnessed impressive progress in natural language generation (NLG) recently. However, reliability issues still remain. On the one hand, NLG models sometimes still show unwanted behaviors, for example, models may generate incoherent and repetitive text following a prompt. On the other hand, we face a dilemma when evaluating generation systems: human evaluation is desirable yet expensive and not reproducible, while automatic metrics are cheap and reproducible yet not always trustworthy. In this talk, I will present my past works that use reinforcement learning to align NLG learning objectives with desired model behaviors and combine human and automatic evaluations to find a good trade-off between both worlds.