generating P(T) for Anti-LM

[kuhanw]

Hi all,

I am trying to understand the implementation of the anti-LM model, in particular the meaning of this line:

line 128: all_prob_t = model_step(dummy_encoder_inputs, cand['dec_inp'], dptr, target_weights, bucket_id)

where dummy_encoder_inputs is dummy_encoder_inputs = [np.array([data_utils.PAD_ID]) for _ in range(len(encoder_inputs))].

in tf_chatbot_seq2seq_antilm/lib/seq2seq_model_utils.py.

This is presumably the probability of the target (P(T)) from the paper https://arxiv.org/pdf/1510.03055.pdf, but how does feeding in an encoder input sequence of PAD give you the probability of T?

Anyone have any ideas?

Cheers,

Kuhan

[zhongpeixiang]

Probably training a separate language model on the dataset containing target responses would be an idea to calculate P(T)? Or using dummy encoder_inputs means given no input sentence, what's the probability of target response T.

[rtygbwwwerr]

@kuhanw I think it might make sense. Using PAD as the initial input of the encoder results in the same initial state of the decoder corresponding different p(T|S). Naturally, the first several output words of the decoder are more influenced by p(T|S) rather than U(T), that is consistent with the original thought of JiweiLi's paper. Commonly, the decoder is considered as a language model, and I think input empty is a simple way to implement anti-MMI without external Model. I hope it could be explicated by the author.