Skip to content

Latest commit

 

History

History

dialogue-cse

README.md

DialogueCSE

Code for the paper DialogueCSE: Dialogue-based Contrastive Learning of Sentence Embeddings

Training & Evaluation

Steps to train DialogueCSE:

  1. Put dialogues in data/session.txt. The format should be {session_id}\t{role}\t{text}\n.
  2. Run python data/data_generator.py to generate the training data.
  3. Run sh run_train.sh to train DialogueCSE.
  4. Run sh eval/batch_test_cmd.sh {ecd|jddc} to evaluate the model.

The argument bert_init_dir in run_train.sh refers to a pre-trained BERT model, the parameters of which could be either the original version or that with continued pre-training on the training data of DialogueCSE. The latter produces the best performance.

To conduct continued pre-training, please refer to the standard BERT codebase.

Datasets

Download evaluation data from Google Drive and move them to dialogue-cse/dataset

  1. JDDC: JDDC is an open-source dataset released by JD AI [1].
  2. ECD: ECD is released in [2] . We have been granted for releasing our evaluation data which is derived from the original ECD dataset.
  3. MDC: The license of MDC does not support secondary distribution and we are in communication with relevant parties.

[1] https://jddc.jd.com/2019/jddc

[2] Modeling Multi-turn Conversation with Deep Utterance Aggregation.