COOKING UP A CONVERSATION WITH SOLOIST - FINETUNING AN END-TO-END MODEL FOR A NEW DOMAIN

Introduction

• For documentation of SOLOIST, please refer to the repository of Soloist's authors
• In this repository we show case a new domain: recipe recommendation, where the bot is asked to recommend a suitable recipe based on:
  • ingredients: user's preferable main ingredient
  • type: main dish, salad, soup, beverage, dessert
  • is_vegan: binary
  • difficulty: easy, medium, difficult
  • time: time (in minutes) to prepare and cook
• The bot could provide serveral pieces of information
  • recipe's name
  • ingredients: list of all ingredients needed
  • equipment: e.g oven, big salad bowl...
  • instruction: step-by-step how to prepare and cook
  • substitute: if it is possible to substitute some ingredients with an alternative
  • and further information related to each recipe: type, vegan, difficulty, time...
• For the detailed report, please refer to this link
• For a short version of report, please refer to this link

Installation

Require python 3.6.

• Please use the below commands to clone the repo and install required package.

git clone https://github.com/Yen444/soloist.git

• We recommend creating a virtual environment using conda

conda create -n myenv python=3.6
# OR
conda create -p /path/to/myenv python=3.6

• Activate virtual environment

conda activate /path/to/myenv

• Install dependencies

Navigate to the root directory of soloist, where you see requirements_soloist_recipe.txt

pip install -r requirements_soloist_recipe.txt

• Download pretrained model

# gtg_pretrained 
https://drive.google.com/file/d/1BNhY_GCx5f_Ubv_8mx6PHa6mFDAG7ujh/view?usp=drive_link
# finetuned_models
https://drive.google.com/drive/folders/1VjnxouEe04yrokzllFpevXi7Jw-h_JDK?usp=sharing

• Copy gtg_pretrained and finetuned_models to the same directory where you see soloist_train.py, that is:

soloist/soloist/gtg_pretrained
soloist/soloist/finetuned_models

Pipeline

Data format

    {
        "history": [
            "user : I'm in the mood for a dessert. Can you suggest something sweet? "
        ],
        "kb": "kb : more than three",
        "belief": "belief : type = dessert ; ingredients = not mentioned ",
        "reply": "system : Sure! I have a few recipes for dessert. Do you have any preferences or restrictions?",
        "dp": "dp : request ( ingredients ) "
    }

We use json to represent a training example. As shown in the example, it contains the following fields:

• history - The context from session beginning to current turn
• belief - The belief state of the user (slot - value pair).
• kb - Database query results. If not blank, inference is slower but better.
• reply - The target system respose. It can be a template, an api call or natural language.
• dp - The system action or dialogue policy.

Training

• For baseline model

python soloist_train_experiment.py --output_dir $OUTPUT --model_type=gpt2 --model_name_or_path $MODEL_NAME --do_train --train_data_file $TRAIN_FILE --eval_data_file $EVAL_FILE --add_special_action_tokens=$SPECIAL_TOKEN_FILE --per_gpu_train_batch_size 1 --num_train_epochs $EPOCHS --learning_rate 5e-5 --overwrite_cache --max_seq 100 --overwrite_output_dir --max_turn 15 --num_candidates 1 --mc_loss_efficient 0.33 --add_response_prediction --add_same_belief_response_prediction --add_belief_prediction --save_steps 6000 [--add_kb_to_context][--evaluate_during_training] [--add_dp_to_response]

• output_dir: Path of the saving model
• model_name_or_path: Initial checkpoint
• train_data_file: Path to training file (in soloist-json format)
• eval_data_file: Path to validation file (in soloist-json format)
• add_special_action_tokens: Path to txt file that contains all special tokens (used for delexicalization)
• num_train_epochs: Number of training epochs, should be tuned using validation set. For 40 training dialogues we fine-tuned with 9 epochs.
• learning_rate: Learning rate; 5e-5, 1e-5, or 1e-4.
• num_candidates: number of candidate; recommend 1.
• mc_loss_efficient: contrastive loss coefficient; 0.1 to 1.
• add_belief_prediction: if add contrastive loss item for belief span.
• add_response_prediction: if add contrastive loss item for response prediction.
• add_same_belief_response_prediction: if add contrastive loss item for belief span and response.
• Optional arguments
  • add_dp_to_response: if output dialogue policy before natural language response.
  • add_kb_to_context: if add database states to context.
  • evaluate_during_training: if compute perplexity for each step.

Generation

python soloist_decode_experiment.py model_type=gpt2 --model_name_or_path $OUTPUT --num_samples $NS --input_file=$TEST_FILE --top_k $TOP_K --temperature $TEMP --output_file $GENERATE --max_turn 15 [--add_kb_to_context]

• model_name_or_path : Path of the saved model.
• num_samples : Number of samples; 1 or 5 for reranking.
• top_k : Top k sampling, 3
• temperature : Temperature sampling; 0.7 - 1.5
• input_file : Path to test file in soloist-json format.
• output_file : Path to save results.
• Optional argument
  • add_kb_to_context : If model is aware of database state.

Interaction

We provide an demo interface to chat with finetuned models.

Start the backend server:

# under soloist/examples/recipe
python recipe_server.py

Start serving frontend page:

npm install
npm run serve 

Open localhost:8080, you will see the following page. Note that the backend port should be consistent with the port used in html/compoents/chat.vue.