TRIP

This repository contains code and data for the paper Target-constrained Bidirectional Planning for Generation of Target-oriented Proactive Dialogue accepted by ACM Transactions on Information Systems (TOIS).

Overview

We propose a novel target-constrained bidirectional planning (TRIP) approach for target-oriented proactive dialogue systems. It plans an appropriate dialogue path by looking ahead and looking back, bidirectionally generating a dialogue path consisting of a sequence of <action, topic> pairs using two Transformer decoders. They are expected to supervise each other and converge on consistent actions and topics by minimizing the decision gap and contrastive generation of targets.

Moreover, we propose a target-constrained decoding algorithm with a bidirectional agreement to better control the inference process of dialogue planning.

Subsequently, we adopt the planned dialogue paths to guide dialogue generation in a pipeline manner, where we explore two variants: prompt-based generation and plan-controlled generation.

Requirements

The required packages are listed in requirements.txt. Suppose you use Anaconda to manage the Python dependencies, you can install them by running:

conda create -n trip python=3.9.7
conda activate trip
pip install -r requirements.txt

Datasets

Please download the repurposed datasets from the following OneDrive links:

• DuRecDial (Chinese)
• DuRecDial 2.0 (English)

Quickstart

Training TRIP

python main_plan.py --mode train --lang "en" \
    --train_data "data/DuRecDial2_en/sample_train.jsonl" \
    --dev_data "data/DuRecDial2_en/sample_dev.jsonl" \
    --cache_dir "caches/DuRecDial2_en/plan" \
    --log_dir "logs/DuRecDial2_en/plan" \
    --num_epochs 10 \
    --batch_size 6

Planned Path Generation

python main_plan.py --mode test --lang "en" \
    --test_data "data/DuRecDial2_en/sample_test_seen.jsonl" \
    --cache_dir "caches/DuRecDial2_en/plan" \
    --log_dir "logs/DuRecDial2_en/plan" \
    --output_dir "outputs/DuRecDial2_en/plan" \
    --test_batch_size 1 \
    --max_dec_len 80 \
    --beam_size 3

Note: The output file best_model_test_seen.jsonl or best_model_test_unseen.jsonl will be saved in the --output_dir.

Dialogue Model Fine-tuning

# for prompt-based generation
python main_dial.py --mode train --lang "en" \
    --train_data "data/DuRecDial2_en/sample_train.jsonl" \
    --dev_data "data/DuRecDial2_en/sample_dev.jsonl" \
    --cache_dir "caches/DuRecDial2_en/dial" \
    --log_dir "logs/DuRecDial2_en/dial" \
    --num_epochs 5 \
    --batch_size 8 \
    --use_control false

Note: If you adopt the plan-controlled generation, please set --use_control true during training.

Dialogue Generation

python main_dial.py --mode test --lang "en" \
    --test_data "data/DuRecDial2_en/sample_test_seen.jsonl" \
    --plan_data "outputs/DuRecDial2_en/plan/best_model_test_seen.jsonl" \
    --cache_dir "caches/DuRecDial2_en/dial" \
    --log_dir "logs/DuRecDial2_en/dial" \
    --output_dir "outputs/DuRecDial2_en/dial" \
    --turn_type_size 6 \
    --test_batch_size 4 \
    --max_dec_len 100

Note: The --plan_data is the file path to the planned dialogue paths generated by TRIP, --output_dir is the directory to save the generated utterances.

Evaluation

To evaluate the performance dialogue planning, please run:

python eval/eval_planning.py --eval_file <path_to_eval> --gold_file <path_to_gold_data>

Note: The --eval_file is the file path to the generated dialogue paths, e.g., outputs/DuRecDial2_en/plan/best_model_test_seen.jsonl. The --gold_file is the file path to the ground-truth dialogue paths, e.g., data/DuRecDial2_en/sample_test_seen.jsonl.

To evaluate the performance of dialogue generation, please run:

python eval/eval_dialogue.py --lang <"zh"|"en"> \
    --eval_file <path_to_eval> \
    --gold_file <path_to_gold_data>

Note: The --eval_file is the file path to the generated dialogue utterances, e.g., outputs/DuRecDial2_en/dial/best_model_test_seen.jsonl. The --gold_file is the file path to the ground-truth dialogue utterances, e.g., data/DuRecDial2_en/sample_test_seen.jsonl.

Citation

If you find this repo helpful, please kindly cite our work as:

@article{wang2024target,
  title={Target-constrained Bidirectional Planning for Generation of Target-oriented Proactive Dialogue},
  author={Wang, Jian and Lin, Dongding and Li, Wenjie},
  journal={ACM Transactions on Information Systems},
  volume={42},
  number={5},
  pages={1--27},
  year={2024},
  publisher={ACM New York, NY}
}