search

microsoft / task_oriented_dialogue_as_dataflow_synthesis

Code to reproduce experiments in the paper "Task-Oriented Dialogue as Dataflow Synthesis" (TACL 2020).
MIT License
308 stars 66 forks source link

readme

Task-Oriented Dialogue as Dataflow Synthesis

License: MIT

This repository contains tools and instructions for reproducing the experiments in the paper Task-Oriented Dialogue as Dataflow Synthesis (TACL 2020). If you use any source code or data included in this toolkit in your work, please cite the following paper.

@article{SMDataflow2020,
  author = {{Semantic Machines} and Andreas, Jacob and Bufe, John and Burkett, David and Chen, Charles and Clausman, Josh and Crawford, Jean and Crim, Kate and DeLoach, Jordan and Dorner, Leah and Eisner, Jason and Fang, Hao and Guo, Alan and Hall, David and Hayes, Kristin and Hill, Kellie and Ho, Diana and Iwaszuk, Wendy and Jha, Smriti and Klein, Dan and Krishnamurthy, Jayant and Lanman, Theo and Liang, Percy and Lin, Christopher H. and Lintsbakh, Ilya and McGovern, Andy and Nisnevich, Aleksandr and Pauls, Adam and Petters, Dmitrij and Read, Brent and Roth, Dan and Roy, Subhro and Rusak, Jesse and Short, Beth and Slomin, Div and Snyder, Ben and Striplin, Stephon and Su, Yu and Tellman, Zachary and Thomson, Sam and Vorobev, Andrei and Witoszko, Izabela and Wolfe, Jason and Wray, Abby and Zhang, Yuchen and Zotov, Alexander},
  title = {Task-Oriented Dialogue as Dataflow Synthesis},
  journal = {Transactions of the Association for Computational Linguistics},
  volume = {8},
  pages = {556--571},
  year = {2020},
  month = sep,
  url = {https://doi.org/10.1162/tacl_a_00333},
  abstract = {We describe an approach to task-oriented dialogue in which dialogue state is represented as a dataflow graph. A dialogue agent maps each user utterance to a program that extends this graph. Programs include metacomputation operators for reference and revision that reuse dataflow fragments from previous turns. Our graph-based state enables the expression and manipulation of complex user intents, and explicit metacomputation makes these intents easier for learned models to predict. We introduce a new dataset, SMCalFlow, featuring complex dialogues about events, weather, places, and people. Experiments show that dataflow graphs and metacomputation substantially improve representability and predictability in these natural dialogues. Additional experiments on the MultiWOZ dataset show that our dataflow representation enables an otherwise off-the-shelf sequence-to-sequence model to match the best existing task-specific state tracking model. The SMCalFlow dataset, code for replicating experiments, and a public leaderboard are available at \url{https://www.microsoft.com/en-us/research/project/dataflow-based-dialogue-semantic-machines}.},
}

Understand SMCalFlow Programs

Please read this document to understand the syntax of SMCalFlow programs, and read this document to understand their semantics.

Install

# (Recommended) Create a virtual environment
virtualenv --python=python3 env
source env/bin/activate

# Install the sm-dataflow package and its core dependencies
pip install git+https://github.com/microsoft/task_oriented_dialogue_as_dataflow_synthesis.git

# Download the spaCy model for tokenization
python -m spacy download en_core_web_md-2.2.0 --direct

# Install OpenNMT-py and PyTorch for training and running the models
pip install OpenNMT-py==1.0.0 torch==1.4.0

SMCalFlow Experiments

Follow the steps below to reproduce the results reported in the paper (Table 2).

NOTE: We highly recommend following the instructions for the leaderboard to report your results for consistency. If you use your own evaluation script, please pay attention to the notes in Step 2 and Step 7.

  1. Download and unzip the SMCalFlow 1.0 dataset.

    dataflow_dialogues_dir="output/dataflow_dialogues"
mkdir -p "${dataflow_dialogues_dir}"

cd "${dataflow_dialogues_dir}"
# Download the dataset `smcalflow.full.data.tgz` or `smcalflow.inlined.data.tgz`
# The `PATH_TO_DATA_TGZ` is the path to the tgz file of the corresponding dataset.
tar -xvzf PATH_TO_DATA_TGZ
    • Both SMCalFlow 1.0 and SMCalFlow 2.0 can be found under the datasets folder.
    • The dataset is distributed under the CC BY-SA 4.0 license.

  2. Compute data statistics:

    dataflow_dialogues_stats_dir="output/dataflow_dialogues_stats"
mkdir -p "${dataflow_dialogues_stats_dir}"
python -m dataflow.analysis.compute_data_statistics \
    --dataflow_dialogues_dir ${dataflow_dialogues_dir} \
    --subset train valid \
    --outdir ${dataflow_dialogues_stats_dir}

    • Basic statistics

      num_dialogues num_turns num_kept_turns num_skipped_turns num_refer_turns num_revise_turns
      train 32,647 133,821 121,200 12,621 33,011 9,315
      valid 3,649 14,757 13,499 1,258 3,544 1,052
      test 5,211 22,012 21,224 7,88 8,965 3,315
      all 41,517 170,590 155,923 14,667 45,520 13,682
      • We currently do not release the test set, but we report the data statistics here.
      • NOTE: There are a small number of turns (num_skipped_turns in the table) whose sole purpose is to establish dialogue context and should not be directly trained or tested on. The dataset statistics reported in the paper are based on non-skipped turns only.

  3. Prepare text data for the OpenNMT toolkit.

    onmt_text_data_dir="output/onmt_text_data"
mkdir -p "${onmt_text_data_dir}"
for subset in "train" "valid"; do
    python -m dataflow.onmt_helpers.create_onmt_text_data \
        --dialogues_jsonl ${dataflow_dialogues_dir}/${subset}.dataflow_dialogues.jsonl \
        --num_context_turns 2 \
        --include_program \
        --include_described_entities \
        --onmt_text_data_outbase ${onmt_text_data_dir}/${subset}
done
    • We use --include_program to add the gold program of the context turns.
    • We use --include_described_entities to add the entities (e.g., entity@123456) described in the generation outcome for the context turns. These entities mentioned in the context turns can appear in the "inlined" programs for the current turn, and thus, we include them in the source sequence so that the seq2seq model can produce such tokens via a copy mechanism.
    • You can vary the number of context turns by changing --num_context_turns.

  4. Compute statistics for the created OpenNMT text data.

    onmt_data_stats_dir="output/onmt_data_stats"
mkdir -p "${onmt_data_stats_dir}"
python -m dataflow.onmt_helpers.compute_onmt_data_stats \
    --text_data_dir ${onmt_text_data_dir} \
    --suffix src src_tok tgt \
    --subset train valid \
    --outdir ${onmt_data_stats_dir}

  5. Train OpenNMT models. You can also skip this step and instead download the trained model from the table below.

    onmt_binarized_data_dir="output/onmt_binarized_data"
mkdir -p "${onmt_binarized_data_dir}"

src_tok_max_ntokens=$(jq '."100"' ${onmt_data_stats_dir}/train.src_tok.ntokens_stats.json)
tgt_max_ntokens=$(jq '."100"' ${onmt_data_stats_dir}/train.tgt.ntokens_stats.json)

# create OpenNMT binarized data
onmt_preprocess \
    --dynamic_dict \
    --train_src ${onmt_text_data_dir}/train.src_tok \
    --train_tgt ${onmt_text_data_dir}/train.tgt \
    --valid_src ${onmt_text_data_dir}/valid.src_tok \
    --valid_tgt ${onmt_text_data_dir}/valid.tgt \
    --src_seq_length ${src_tok_max_ntokens} \
    --tgt_seq_length ${tgt_max_ntokens} \
    --src_words_min_frequency 0 \
    --tgt_words_min_frequency 0 \
    --save_data ${onmt_binarized_data_dir}/data

# extract pretrained Glove 840B embeddings (https://nlp.stanford.edu/projects/glove/)
glove_840b_dir="output/glove_840b"
mkdir -p "${glove_840b_dir}"
wget -O ${glove_840b_dir}/glove.840B.300d.zip http://nlp.stanford.edu/data/glove.840B.300d.zip
unzip ${glove_840b_dir}/glove.840B.300d.zip -d ${glove_840b_dir}

onmt_embeddings_dir="output/onmt_embeddings"
mkdir -p "${onmt_embeddings_dir}"
python -m dataflow.onmt_helpers.embeddings_to_torch \
    -emb_file_both ${glove_840b_dir}/glove.840B.300d.txt \
    -dict_file ${onmt_binarized_data_dir}/data.vocab.pt \
    -output_file ${onmt_embeddings_dir}/embeddings

# train OpenNMT models
onmt_models_dir="output/onmt_models"
mkdir -p "${onmt_models_dir}"

batch_size=64
train_num_datapoints=$(jq '.train' ${onmt_data_stats_dir}/nexamples.json)
# validate approximately at each epoch
valid_steps=$(python3 -c "from math import ceil; print(ceil(${train_num_datapoints}/${batch_size}))")

onmt_train \
    --encoder_type brnn \
    --decoder_type rnn \
    --rnn_type LSTM \
    --global_attention general \
    --global_attention_function softmax \
    --generator_function softmax \
    --copy_attn_type general \
    --copy_attn \
    --seed 1 \
    --optim adam \
    --learning_rate 0.001 \
    --early_stopping 2 \
    --batch_size ${batch_size} \
    --valid_batch_size 8 \
    --valid_steps ${valid_steps} \
    --save_checkpoint_steps ${valid_steps} \
    --data ${onmt_binarized_data_dir}/data \
    --pre_word_vecs_enc ${onmt_embeddings_dir}/embeddings.enc.pt \
    --pre_word_vecs_dec ${onmt_embeddings_dir}/embeddings.dec.pt \
    --word_vec_size 300 \
    --attention_dropout 0 \
    --dropout 0.5 \
    --layers ??? \
    --rnn_size ??? \
    --gpu_ranks 0 \
    --world_size 1 \
    --save_model ${onmt_models_dir}/checkpoint

    • Hyperparameters for models reported in the Table 2 in the paper.

      --layers --rnn_size model
      dataflow 2 384 link
      inline 3 384 link

  6. Make predictions using a trained OpenNMT model. You need to replace the checkpoint_last.pt in the following script with the final model you get from the previous step.

    onmt_translate_outdir="output/onmt_translate_output"
mkdir -p "${onmt_translate_outdir}"

onmt_model_pt="${onmt_models_dir}/checkpoint_last.pt"
nbest=5
tgt_max_ntokens=$(jq '."100"' ${onmt_data_stats_dir}/train.tgt.ntokens_stats.json)

# predict programs using a trained OpenNMT model
onmt_translate \
    --model ${onmt_model_pt} \
    --max_length ${tgt_max_ntokens} \
    --src ${onmt_text_data_dir}/valid.src_tok \
    --replace_unk \
    --n_best ${nbest} \
    --batch_size 8 \
    --beam_size 10 \
    --gpu 0 \
    --report_time \
    --output ${onmt_translate_outdir}/valid.nbest

  7. Compute the exact-match accuracy (taking into account whether the program_execution_oracle.refer_are_correct is true).

    evaluation_outdir="output/evaluation_output"
mkdir -p "${evaluation_outdir}"

# create the prediction report
python -m dataflow.onmt_helpers.create_onmt_prediction_report \
    --dialogues_jsonl ${dataflow_dialogues_dir}/valid.dataflow_dialogues.jsonl \
    --datum_id_jsonl ${onmt_text_data_dir}/valid.datum_id \
    --src_txt ${onmt_text_data_dir}/valid.src_tok \
    --ref_txt ${onmt_text_data_dir}/valid.tgt \
    --nbest_txt ${onmt_translate_outdir}/valid.nbest \
    --nbest ${nbest} \
    --outbase ${evaluation_outdir}/valid

# evaluate the predictions (all turns)
python -m dataflow.onmt_helpers.evaluate_onmt_predictions \
    --prediction_report_tsv ${evaluation_outdir}/valid.prediction_report.tsv \
    --scores_json ${evaluation_outdir}/valid.all.scores.json

# evaluate the predictions (refer turns)
python -m dataflow.onmt_helpers.evaluate_onmt_predictions \
    --prediction_report_tsv ${evaluation_outdir}/valid.prediction_report.tsv \
    --datum_ids_json ${dataflow_dialogues_stats_dir}/valid.refer_turn_ids.jsonl \
    --scores_json ${evaluation_outdir}/valid.refer_turns.scores.json

# evaluate the predictions (revise turns)
python -m dataflow.onmt_helpers.evaluate_onmt_predictions \
    --prediction_report_tsv ${evaluation_outdir}/valid.prediction_report.tsv \
    --datum_ids_json ${dataflow_dialogues_stats_dir}/valid.revise_turn_ids.jsonl \
    --scores_json ${evaluation_outdir}/valid.revise_turns.scores.json
    • NOTE: The numbers reported using the scripts above should match those reported in Table 2 in the paper. The leaderboard has used a slightly different evaluation script that canonicalizes both the gold and predicted programs, and thus, the accuracy would be slightly higher (e.g., 0.665 vs. 0.668 on the test set). To obtain the leaderboard results, please add --use_leaderboard_metric when running python -m dataflow.onmt_helpers.create_onmt_prediction_report to create the report.

  8. Calculate the statistical significance for two different experiments.

    analysis_outdir="output/analysis_output"
mkdir -p "${analysis_outdir}"
python -m dataflow.analysis.calculate_statistical_significance \
    --exp0_prediction_report_tsv ${exp0_evaluation_outdir}/valid.prediction_report.tsv \
    --exp1_prediction_report_tsv ${exp1_evaluation_outdir}/valid.prediction_report.tsv \
    --scores_json ${analysis_outdir}/exp0_vs_exp1.valid.scores.json
    • The exp0_evaluation_outdir and exp1_evaluation_outdir are the evaluation_outdir in Step 7 for corresponding experiments.
    • You can also provide --datum_ids_jsonl to carry out the significance test on a subset of turns.

MultiWOZ Experiments

  1. Download the MultiWoZ dataset and convert it to dataflow programs.

    # creates TRADE-processed dialogues
raw_trade_dialogues_dir="output/trade_dialogues"
mkdir -p "${raw_trade_dialogues_dir}"
python -m dataflow.multiwoz.trade_dst.create_data \
    --use_multiwoz_2_1 \
    --output_dir ${raw_trade_dialogues_dir}

# patch TRADE dialogues
patched_trade_dialogues_dir="output/patched_trade_dialogues"
mkdir -p "${patched_trade_dialogues_dir}"
for subset in "train" "dev" "test"; do
    python -m dataflow.multiwoz.patch_trade_dialogues \
        --trade_data_file ${raw_trade_dialogues_dir}/${subset}_dials.json \
        --outbase ${patched_trade_dialogues_dir}/${subset}
done
ln -sr ${patched_trade_dialogues_dir}/dev_dials.json ${patched_trade_dialogues_dir}/valid_dials.json

# create dataflow programs
dataflow_dialogues_dir="output/dataflow_dialogues"
mkdir -p "${dataflow_dialogues_dir}"
for subset in "train" "valid" "test"; do
    python -m dataflow.multiwoz.create_programs \
        --trade_data_file ${patched_trade_dialogues_dir}/${subset}_dials.json \
        --outbase ${dataflow_dialogues_dir}/${subset}
done
    • To create programs that inline refer calls, add --no_refer when running the dataflow.multiwoz.create_programs command.
    • To create programs that inline both refer and revise calls, add --no_refer --no_revise.

  2. Prepare text data for the OpenNMT toolkit.

    onmt_text_data_dir="output/onmt_text_data"
mkdir -p "${onmt_text_data_dir}"
for subset in "train" "valid" "test"; do
    python -m dataflow.onmt_helpers.create_onmt_text_data \
        --dialogues_jsonl ${dataflow_dialogues_dir}/${subset}.dataflow_dialogues.jsonl \
        --num_context_turns 2 \
        --include_agent_utterance \
        --onmt_text_data_outbase ${onmt_text_data_dir}/${subset}
done
    • We use --include_agent_utterance following the setup in TRADE (Wu et al., 2019).
    • You can vary the number of context turns by changing --num_context_turns.

  3. Compute statistics for the created OpenNMT text data.

    onmt_data_stats_dir="output/onmt_data_stats"
mkdir -p "${onmt_data_stats_dir}"
python -m dataflow.onmt_helpers.compute_onmt_data_stats \
    --text_data_dir ${onmt_text_data_dir} \
    --suffix src src_tok tgt \
    --subset train valid test \
    --outdir ${onmt_data_stats_dir}

  4. Train OpenNMT models. You can also skip this step and instead download the trained models from the table below.

    onmt_binarized_data_dir="output/onmt_binarized_data"
mkdir -p "${onmt_binarized_data_dir}"

# create OpenNMT binarized data
src_tok_max_ntokens=$(jq '."100"' ${onmt_data_stats_dir}/train.src_tok.ntokens_stats.json)
tgt_max_ntokens=$(jq '."100"' ${onmt_data_stats_dir}/train.tgt.ntokens_stats.json)

onmt_preprocess \
    --dynamic_dict \
    --train_src ${onmt_text_data_dir}/train.src_tok \
    --train_tgt ${onmt_text_data_dir}/train.tgt \
    --valid_src ${onmt_text_data_dir}/valid.src_tok \
    --valid_tgt ${onmt_text_data_dir}/valid.tgt \
    --src_seq_length ${src_tok_max_ntokens} \
    --tgt_seq_length ${tgt_max_ntokens} \
    --src_words_min_frequency 0 \
    --tgt_words_min_frequency 0 \
    --save_data ${onmt_binarized_data_dir}/data

# extract pretrained Glove 6B embeddings
glove_6b_dir="output/glove_6b"
mkdir -p "${glove_6b_dir}"
wget -O ${glove_6b_dir}/glove.6B.zip http://nlp.stanford.edu/data/glove.6B.zip
unzip ${glove_6b_dir}/glove.6B.zip -d ${glove_6b_dir}

onmt_embeddings_dir="output/onmt_embeddings"
mkdir -p "${onmt_embeddings_dir}"
python -m dataflow.onmt_helpers.embeddings_to_torch \
    -emb_file_both ${glove_6b_dir}/glove.6B.300d.txt \
    -dict_file ${onmt_binarized_data_dir}/data.vocab.pt \
    -output_file ${onmt_embeddings_dir}/embeddings

# train OpenNMT models
onmt_models_dir="output/onmt_models"
mkdir -p "${onmt_models_dir}"

batch_size=64
train_num_datapoints=$(jq '.train' ${onmt_data_stats_dir}/nexamples.json)
# approximately validate at each epoch
valid_steps=$(python3 -c "from math import ceil; print(ceil(${train_num_datapoints}/${batch_size}))")

onmt_train \
    --encoder_type brnn \
    --decoder_type rnn \
    --rnn_type LSTM \
    --global_attention general \
    --global_attention_function softmax \
    --generator_function softmax \
    --copy_attn_type general \
    --copy_attn \
    --seed 1 \
    --optim adam \
    --learning_rate 0.001 \
    --early_stopping 2 \
    --batch_size ${batch_size} \
    --valid_batch_size 8 \
    --valid_steps ${valid_steps} \
    --save_checkpoint_steps ${valid_steps} \
    --data ${onmt_binarized_data_dir}/data \
    --pre_word_vecs_enc ${onmt_embeddings_dir}/embeddings.enc.pt \
    --pre_word_vecs_dec ${onmt_embeddings_dir}/embeddings.dec.pt \
    --word_vec_size 300 \
    --attention_dropout 0 \
    --dropout ??? \
    --layers ??? \
    --rnn_size ??? \
    --gpu_ranks 0 \
    --world_size 1 \
    --save_model ${onmt_models_dir}/checkpoint

    • Hyperparameters for models reported in the Table 3 in the paper.

      --dropout --layers --rnn_size model
      dataflow (--num_context_turns 2) 0.7 2 384 link
      inline refer (--num_context_turns 4) 0.3 3 320 link
      inline both (--num_context_turns 10) 0.7 2 320 link

  5. Make predictions using a trained OpenNMT model. You need to replace the checkpoint_last.pt in the following script with the actual model you get from the previous step.

    onmt_translate_outdir="output/onmt_translate_output"
mkdir -p "${onmt_translate_outdir}"

onmt_model_pt="${onmt_models_dir}/checkpoint_last.pt"
nbest=5
tgt_max_ntokens=$(jq '."100"' ${onmt_data_stats_dir}/train.tgt.ntokens_stats.json)

# predict programs on the test set using a trained OpenNMT model
onmt_translate \
    --model ${onmt_model_pt} \
    --max_length ${tgt_max_ntokens} \
    --src ${onmt_text_data_dir}/test.src_tok \
    --replace_unk \
    --n_best ${nbest} \
    --batch_size 8 \
    --beam_size 10 \
    --gpu 0 \
    --report_time \
    --output ${onmt_translate_outdir}/test.nbest

  6. Compute the exact-match accuracy of the program predictions.

    evaluation_outdir="output/evaluation_output"
mkdir -p "${evaluation_outdir}"

# create the prediction report
python -m dataflow.onmt_helpers.create_onmt_prediction_report \
    --dialogues_jsonl ${dataflow_dialogues_dir}/test.dataflow_dialogues.jsonl \
    --datum_id_jsonl ${onmt_text_data_dir}/test.datum_id \
    --src_txt ${onmt_text_data_dir}/test.src_tok \
    --ref_txt ${onmt_text_data_dir}/test.tgt \
    --nbest_txt ${onmt_translate_outdir}/test.nbest \
    --nbest ${nbest} \
    --outbase ${evaluation_outdir}/test

# evaluate the predictions
python -m dataflow.onmt_helpers.evaluate_onmt_predictions \
    --prediction_report_tsv ${evaluation_outdir}/test.prediction_report.tsv \
    --scores_json ${evaluation_outdir}/test.scores.json

  7. Evaluate the belief state predictions. 

    belief_state_tracker_eval_dir="output/belief_state_tracker_eval"
mkdir -p "${belief_state_tracker_eval_dir}"

# creates the gold file from TRADE-preprocessed dialogues (after patch)
python -m dataflow.multiwoz.create_belief_state_tracker_data \
    --trade_data_file ${patched_trade_dialogues_dir}/test_dials.json \
    --belief_state_tracker_data_file ${belief_state_tracker_eval_dir}/test.belief_state_tracker_data.jsonl

# creates the hypo file from predicted programs
python -m dataflow.multiwoz.execute_programs \
    --dialogues_file ${evaluation_outdir}/test.dataflow_dialogues.jsonl \
    --cheating_mode never \
    --outbase ${belief_state_tracker_eval_dir}/test.hypo

python -m dataflow.multiwoz.create_belief_state_prediction_report \
    --input_data_file ${belief_state_tracker_eval_dir}/test.hypo.execution_results.jsonl \
    --format dataflow \
    --remove_none \
    --gold_data_file ${belief_state_tracker_eval_dir}/test.belief_state_tracker_data.jsonl \
    --outbase ${belief_state_tracker_eval_dir}/test

# evaluates belief state predictions
python -m dataflow.multiwoz.evaluate_belief_state_predictions \
    --prediction_report_jsonl ${belief_state_tracker_eval_dir}/test.prediction_report.jsonl \
    --outbase ${belief_state_tracker_eval_dir}/test
    • The scores are reported in ${belief_state_tracker_eval_dir}/test.scores.json.

  8. Calculate the statistical significance for two different experiments.

    analysis_outdir="output/analysis_output"
mkdir -p "${analysis_outdir}"
python -m dataflow.analysis.calculate_statistical_significance \
    --exp0_prediction_report_tsv ${exp0_evaluation_outdir}/test.prediction_report.tsv \
    --exp1_prediction_report_tsv ${exp1_evaluation_outdir}/test.prediction_report.tsv \
    --scores_json ${analysis_outdir}/exp0_vs_exp1.test.scores.json
    • The exp0_evaluation_outdir and exp1_evaluation_outdir are the belief_state_tracker_eval_dir in Step 7 for corresponding experiments.