ICASSP2017: End-to-end joint learning of natural language understanding and dialogue manager
Switch branches/tags
Nothing to show
Clone or download
Fetching latest commit…
Cannot retrieve the latest commit at this time.
Permalink
Type Name Latest commit message Commit time
Failed to load latest commit information.
data
log
model
.gitignore
AgentActClassifyingModel.py
BaselineModel.py
DataSetCSV.py
DataSetCSVagentActPred.py
DataSetCSVjoint.py
DataSetCSVslotTagging.py
JointModel_multitask_jointraining.py
LICENSE
PipelineLstmModel.py
README.md
SlotTaggingModel_multitask.py
analyze_log.py
csv_stats.py
iob2csv.py
keras
prepare_data.py
test_models.sh
train_models.sh
utils.py

README.md

END-TO-END JOINT LEARNING OF NATURAL LANGUAGE UNDERSTANDING AND DIALOGUE MANAGER

This repository releases the source code for our paper END-TO-END JOINT LEARNING OF NATURAL LANGUAGE UNDERSTANDING AND DIALOGUE MANAGER. Please cite the following paper if you use this code as part of any published research.

[1] Xuesong Yang, Yun-Nung Chen, Dilek Hakkani-Tür, Paul Crook, Xiujun Li, Jianfeng Gao, and Li Deng. "End-to-end joint learning of natural language understanding and dialogue manager." In Acoustics, Speech and Signal Processing (ICASSP), 2017 IEEE International Conference on, pp. 5690-5694. IEEE, 2017.

@inproceedings{yang2017end,
  title={End-to-end joint learning of natural language understanding and dialogue manager},
  author={Yang, Xuesong and Chen, Yun-Nung and Hakkani-T{\"u}r, Dilek and Crook, Paul and Li, Xiujun and Gao, Jianfeng and Deng, Li},
  booktitle={Acoustics, Speech and Signal Processing (ICASSP), 2017 IEEE International Conference on},
  pages={5690--5694},
  year={2017},
  organization={IEEE}
}

License

The code is released under MIT License.

Data

We used DSTC4 data and split data into train/dev/test in the following:

Split Total Sub-Dialog IDs
train 14 001, 002, 003, 004, 006, 007, 008, 009, 010, 012, 013, 017, 019, 022
dev 6 011, 016, 020, 025, 026, 028
test 9 021, 023, 024, 030, 033, 035, 041, 047, 048

[Note]:

  1. We are only allowed to provide user utterance index, IOB data, and their corresponding speech act and attributes. User raw utterances are not allowed to release. Please contact the committee of DSTC4 or DSTC5 for the whole data.

Utterance index consists of two components: sub-dialog folder and utter_id.

e.g. 011_129 represents that the sub-dialog folder is 011, and utter_id is 129 in the corresponding your_DSTC_directory/011/label.json file.

  1. Although we only mentioned DSTC4 in our paper, users can also locate these sub-dialogs from DSTC5, since it provides all the same data in DSTC4, and plus two extra Chinese dialogs (055, 056).

Prerequisites

  1. pip install nltk
  2. pip install python-crfsuite
  3. pip install prettytable
  4. keras 1.2.0
  5. theano 0.9.0dev4

Executable scripts

  1. training models: $ bash train_models.sh
  2. testing models using selected parameters: $ bash test_models.sh

Auxiliary Label

  1. "null" label for system actions.

    From human annotations, "null" label is used to identify that there is not any system action that makes response to current user utterance. In other word, "null" is not supposed to be one of the system actions. During the testing process, if the posterior prob for each oneVSall binary classifier is less than its decision threshold, "null" is considered as the predicted label.

  2. "null" label for user intent.

    Similar explanation to the one for system actions.

Model Selection

  1. JointModel:

    • slotTagging
      • weights=./model/joint_4770/weights/ep=8_tagF1=0.438_intentF1=0.494th=0.221_NLUframeAcc=0.296_actF1=0.302frameAcc=0.047th=0.131.h5
    • userIntent
      • weights=./model/joint_4770/weights/ep=13_tagF1=0.425_intentF1=0.519th=0.342_NLUframeAcc=0.379_actF1=0.300frameAcc=0.035th=0.139.h5
      • threshold=0.342
    • agentAct
      • weights=./model/joint_4770/weights/ep=172_tagF1=0.418_intentF1=0.492th=0.387_NLUframeAcc=0.360_actF1=0.189frameAcc=0.212th=0.009.h5
      • threshold=0.009
  2. SlotTaggingModel:

    • slotTagging
      • weights=./model/slot_4768/weights/ep=14_tagF1=0.468frameAcc=0.757_intentF1=0.399frameAcc=0.329th=0.203.h5
    • userIntent
      • weights=./model/slot_4768/weights/ep=196_tagF1=0.448frameAcc=0.759_intentF1=0.496frameAcc=0.419th=0.391.h5
      • threshold=0.391
  3. AgentActModel:

    • agentAct
      • weights=./model/agentAct_4769/weights/ep=139_f1=0.228_frameAcc=0.202_th=0.154.h5
      • threshold=0.154
  4. BaselineModel: model_folder=./model/baseline_4771

Performance

Table 1 Perforamce of End2End Models for System Act Prediction.

Models Fscore Precision Recall Accuracy_Frame
Baseline(CRF+SVMs+SVMs) 0.3115 0.2992 0.3248 0.0771
Pipeline(biLSTM+biLSTM+biLSTM) 0.1989 0.1487 0.3001 0.1196
JointModel(biLSTM+biLSTM+biLSTM) 0.1904 0.1853 0.1957 0.2284
Oracle(SVMs) 0.3061 0.3020 0.3104 0.0765
Oracle(biLSTM) 0.2309 0.2224 0.2401 0.1967

Table 2 Perforamce of NLU Models

Models tagF tagP tagR tagAccFr intF intP intR intAccFr nluAccFr
Baseline 0.4050 0.6141 0.3021 0.7731 0.4975 0.5256 0.4724 0.3719 0.3313
Pipeline 0.4615 0.5463 0.3996 0.7684 0.4748 0.5219 0.4355 0.3996 0.3638
JointModel 0.4504 0.5335 0.3897 0.7649 0.4967 0.5222 0.4735 0.4220 0.3738