fagan2888 / kg-goal-oriented-dialog

This code is written in Python and implements a goal-oriented dialog system which takes as input a conversation history as well as the underlying database, and predicts the best next utterance.

kg-goal-oriented-dialog

This code is written in Python and implements a goal-oriented dialog system which takes as input a conversation history as well as the underlying database, and predicts the best next utterance.

Get started

Prepare the data

Build data

python build_all_data.py -train_path $path_to_train_data -dev_path $path_to_dev_data -db_path $path_to_database -out_dir $path_to_out_dir [-test_path $path_to_test_data]

Dump pretrained word embeddings

python build_pretrained_w2v.py -emb $path_to_pretrained_word_embeddings -vocab_path $path_to_vocabulary -emb_size $embedding_size -out_path $path_to_output_file

Train the model

python train.py -vocab_size $vocab_size -mf $path_to_saved_model -data_dir $path_to_data_dir -pre_w2v  $pretrained_word_embeddings -hidden_size $hidden_size -word_emb_size $word_emb_size -model_name $model_name -eval_every $evaluate_model_every_steps -exp $experiment_index

Evaluate the model

python test.py -data_dir $path_to_data_dir -mf $path_to_saved_model -act $action_type

To compute evaluation metrics, set $action_type as evaluate; To interpret the model (e.g., sample test dialogs with intermediate model attentions), set $action_type$ as interpret.