baseline in the code and paper

[Victor0118]

In the paper, the baseline is not detailed. In the code,

https://github.com/nyu-dl/QueryReformulator/blob/master/run.py#L556

r = reward - reward_last - bl, according to my understanding, reward is the recall by querying with the reformulated query text. reward_last is the recall by querying with the raw query text. But I did not understand what bl means. I know it is the output of NN for each query. But I did not think it means baseline. I think reward_last is the baseline. Correct? If true, what is bl?

[rodrigonogueira4]

bl is the baseline. reward_last is the recall by querying with the original query and it is used as a heuristic to slightly speed up training. However, the model should reach an equally good recall if you just use r = reward - bl.