This repo contains the code for our work published in AAAI 2021:

Keyword-Guided Neural Conversational Model

The paper is available here

This repo is mainly adapted from https://github.com/squareRoot3/Target-Guided-Conversation.

Datasets

• ConvAI2: https://www.dropbox.com/s/1fw2gwpuyud2bkq/convai2.zip?dl=0
• Reddit/Casual: https://www.dropbox.com/s/hhesm3z6judpsxa/casual.zip?dl=0

Preprocessing

• Preprocess conversations and build word vocabulary.
• Use TF-IDF-based keyword extraction techniques (based on here) to extract keywords from each utterance.
• Build the keyword vocabulary.
• Based on the word and keyword vocabularies, preprocess conceptnet and build the concept vocabulary.

Steps:

1. Download the datasets.
2. Run python train_keyword_prediction.py --config ./configs/keyword_prediction_config.json to train a keyword prediction model.
3. Run python train_retrieval.py --config ./configs/retrieval_config.json to train a response retrieval model.

Note that this repo contains a portion of experimental code that is not used in the paper.