Code and Data for Dialogue Topic Segmenter

This repository maintains:

• [PART I] The source code of "Improving Unsupervised Dialogue Topic Segmentation with Utterance-Pair Coherence Scoring", SIGDIAL-21.
• [PART II - Continously Update] A dataset collection of Dialogue Topic Segmentation cleaned to a standardized format.

PART I:

This list of python scripts are together as the source codebase of our paper:

• For training:
  • train.py : contains the main code (Sec 3.2 in the paper) for the training process of the utterance-pair coherence scoring model grounded on BERT (Next Sentence Prediction).
  • data_utils.py : contains the main code (Sec 3.1 in the paper) for pseudo training data generation, which will be loaded to train the coherence scoring model.
  • model_utils.py : contains the class of coherence scoring model.
• For evaluation:
  • segment.py : contains the main code to conduct evaluation procedure based on the TextTiling segmentation framework.
  • neural_texttiling.py : contains the detailed implementation of TextTiling with different settings of text encoder (e.g., Bi-encoder, cross-encoder, coherence scoring etc)

👉🏼 Training/Testing Steps:

0. Instaill env requirements

pip install -r requirements.txt

1. Download DailyDial from this link and add the following three files to ./data/train/dailydialog/ :

+ dialogues_text.txt
+ dialogues_topic.txt
+ dialogue_act.txt

2. Execute the training command, for example:

python train.py -t ./data/train/dailydialog/ -e bert-base-uncased -s ./checkpoints -m 1 -r 10 -b 32

💡Notes:

• The current data loading/generation code (data_utils.py) is specifically implemented for DailyDialog, please adjust it accordingly to load your datasets if need.
• The training code only support bert-based language model supporting mode of Next Sentence Prediction, otherwise it will run into errors (e.g., loading roberta, sbert as text encoder).
• Checkpoint of each 1000 steps will be saved to your specified directory.
• Training log can be found in training_log.txt.

3. For evaluation, you can run command like:

python segment.py -t ./data/eval/dialseg_711.json -e ./checkpoints/cpt_0.pth -m CM

💡Instruction:

• The evaluation script supports the standarized data form defined in our Dialogue Topic Segmentation Data Hub. You can easily conduct test by loading any corpus under this directory (./data/eval/).
• The evaluation script provides a easy-to-use TextTiling-based pipeline where you can plug different bert-based text encoders (e.g., roberta, sbert, tod-bert, dse). For example:

  python segment.py -t ./data/eval/dialseg_711.json -e roberta-base -m SC
  python segment.py -t ./data/eval/dialseg_711.json -e sentence-transformers/all-mpnet-base-v2 -m SC
  python segment.py -t ./data/eval/dialseg_711.json -e TODBERT/TOD-BERT-JNT-V1 -m NSP
  python segment.py -t ./data/eval/dialseg_711.json -e aws-ai/dse-bert-base -m NSP
  

• The evaluation script supports three sentence pair scoring paradigms:
  • Sequence Classification (SC) : Encode each sentence individually and compute consine similarty as sentence-pair score.

    python segment.py -t ./data/eval/dialseg_711.json -e bert-base-uncased -m SC
    

  • Next Sentence Prediction (NSP) : Encode a pair of sentences together and use the next sentence probability as sentence-pair score.

    python segment.py -t ./data/eval/dialseg_711.json -e bert-base-uncased -m NSP
    

  • Coherence Modeing (CM) : Encode a pair of sentences together with trained coherence scoring model and use the coherence score as sentence-pair score.

    python segment.py -t ./data/eval/dialseg_711.json -e ./checkpoints/cpt_1.pth -m CM
    

• The default setting to obtain sentence representation is mean-pooling over all token hidden states, if you want to explore other options (e.g., CLS representation), replace line 17 -> line 18 in neural_texttiling.py .
• The default included evaluation metrics are: P_k, Windiff, F1. To add your own metrics, please adjust the code in neural_texttiling.TextTiling.
• Sometimes you may want to investigate actual segmentation prediction by case more than metric values, neural_texttiling.TextTiling returns segment_prediction as well.

👉🏼 Reference (If you use code in this database for your research, please include our paper in reference)

@inproceedings{xing-carenini-2021-improving,
    title = "Improving Unsupervised Dialogue Topic Segmentation with Utterance-Pair Coherence Scoring",
    author = "Xing, Linzi and Carenini, Giuseppe",
    booktitle = "Proceedings of the 22nd Annual Meeting of the Special Interest Group on Discourse and Dialogue",
    year = "2021",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2021.sigdial-1.18",
    pages = "167--177"}

PART II (Continously Update):

We also maintain a data hub for dialogue topic segmentation with corpora cleaned to a standardized format compatible with our evaluation code. We believe maintaining this data hub will also save efforts of other ongoing research as there is no need to write data loading code for each individual corpus anymore.

The current list of dataset for dialogue topic segmentation covers:

• DialSeg_711
• Doc2Dial
• Tiage
• AMI (meeting)
• ICSI (meeting)
• Committee (meeting)

Don't forget citing these papers if you use any of these dataset in your work!

For more details about this data hub and how to contribute to it, please refer this.