This repository is the official implementation of Counterfactual Maximum Likelihood Estimation for Training Deep Network.

We propose two different training methods for Counterfactual Maximum Likelihood Estimation (CMLE), Implicit CMLE and Explicit CMLE. The models we use in the real-world experiments are illustrated below:

We perform experiments on two different tasks, natural language inference (NLI) and image captioning (IC), that require different environments, datasets and models. To check each of the tasks, see README in ./NLI for the natural language inference experiments and see README in ./IC for image captioning experiments.