As we've already discussed, RL training methods applied to the seq2seq problem can potentially improve the final model. The main reasons are:
- Better handling of multiple target sequences. For example, hi could be replied with hi, hello, not interested, or something else. The RL point of view is to treat our decoder as a process of selecting actions when every action is a token to be generated, which fits better to the problem.
- Optimizing the BLEU score directly instead of cross-entropy loss. Using the BLEU score for the generated sequence as a gradient scale, we can push our model toward the successful sequences and decrease the probability of unsuccessful ones.
- By repeating the decoding process, we can generate more episodes to train on, which will lead to better gradient estimation.
- Additionally, using the self-critical sequence training approach, we can get the baseline almost for free, without increasing the complexity of our model, which...