asappresearch / constrained-dialogue-generation

Constrained dialogue generation

This codebase contains the code for constrained dialogue generation. We include files to run the approach as well as the public datasets we run experiments on.

Pre-process data

• ABCD
  • Download and extract the dataset https://github.com/asappresearch/abcd/blob/master/data/abcd_v1.1.json.gz into data/ABCD
• MultiWoz
  • Download and extract the dataset https://github.com/lexmen318/MultiWOZ-coref/blob/main/MultiWOZ2_3.zip into data/MultiWoz
  • Run the preprocessing script: cd data/MultiWoz and python preprocess_multiwoz.py --output-file multiwoz_processed.json
• TaskMaster-3
  • Download and extract the dataset svn checkout https://github.com/google-research-datasets/Taskmaster/trunk/TM-3-2020/data into data/TaskMaster
  • Run the preprocessing script: cd data/TaskMaster and python preprocess_taskmaster.py --output-file taskmaster_processed.json

Requirements

• Python 3.7.6 (this version is verified to run the code)
• pip install -r requirements.txt

Fine-tune models

• Example for ABCD
  • Train the customer model.

    • cd finetune
      python main.py --do-train --local-rank -1 --config-file configs/abcd_customer.ini
      

  • Evaluate the customer model.

    • python main.py --nodo-train --do-eval --local-rank -1 --config-file configs/abcd_customer.ini
      

  • Similarly, train and evaluate an agent model using these commands:

    • python main.py --do-train --local-rank -1 --config-file configs/abcd_agent.ini
      python main.py --nodo-train --do-eval --local-rank -1 --config-file configs/abcd_agent.ini
      

Build datastores

• Example for ABCD
  • Create the train datastore.

    • cd datastore
      model_path="Enter path to the customer model here"
      python knn_datastore.py \
             --build-datastore \
             --model_path "${model_path}" \
             --data-path ../data/ABCD/abcd_v1.1.json \
             --output-dir ../data/ABCD/DATASTORE \
             --split train \
             --finetuned \
             --fp16
      

  • Create the test datastore.

    • model_path="Enter path to the customer model here"
      python knn_datastore.py \
             --build-datastore \
             --model_path "${model_path}" \
             --data-path ../data/ABCD/abcd_v1.1.json \
             --output-dir ../data/ABCD/DATASTORE \
             --split test \
             --finetuned \
             --fp16
      

Run approaches

• Example for ABCD

  •   cd approaches
      model_path="Enter path to the customer model here"
      agent_model_path="Enter path to the agent model here"
      bash run_individual.sh \
           --run-approaches=wfirst,finetuned,prompt,dbs,cgmh,retrieve,windowfop \
           --MODEL-TYPE=finetuned \
           --MODEL-PATH="${model_path}" \
           --AGENT-MODEL-PATH="${agent_model_path}" \
           --config-file=../finetune/configs/abcd.ini \
           --data-dir=../data/ABCD/ \
           --save-dir=abcd_results
    

• Get results table

  • Run the jupyter notebook in approaches/get-latex-results-table.ipynb with the appropriate result directories.

• Plot graphs for simulated conversations

  • python plot.py --save_dir <results directory> --eval_type simulated

** You can follow a similar set of steps for the other datasets with the corresponding config files. **