Cog-TAIL

Conditional Generation of Text with Generative Adversarial Imitation Learning

Preprocessing

1. Introduction

Not important.

2. Top3600 Words

The following is detail of pickle files.

• Pre_reviews : It consists of preprocessed Amazon electronic products review files. The original can be found at here.
• Top3600_dist : We use only 3600 words that appear frequently in learning. This file contains the distribution of them.
• Top3600_AtoB, Top3600_BtoA : Each word in the Top3600 has a unique index, which matches one-to-one with the result of the BERT encoding. This index can be used to make the first word, to see what word it was in a batch, and so on. AtoB converts from a unique index to a BERT index, while BtoA does the opposite.
• Top3600_first_pos, Top3600_first_neg : The cumulative distribution of the first word for each emotion. Use it to pick the first word of a sentence.

3. Batch File Links

Each batch file has 5 different keys.

• states : BERT embedded sentence sequence, but just one index.
• actions : The word just after the sentence used to create states.
• codes : The sentiment of the sentence.
• action_ids : Top3600 index of the action.
• prev_action_ids : Top3600 index of the previous action.

Batch file created differently depending on the index used when creating states.

• CLS
• The Word just before SEP
• SEP