bsantraigi / Hierarchical-Transformer-Dialog

Development repository for implementations of Models/Experiments in "Hierarchical Transformer for Task Oriented Dialog Systems" - NAACL 2021 (Long Paper)

HIER Project

Error Analysis Steps

1. First run main_acts.py or main.py with test_split uncommented (inside run function). This will save the prediction json files in running folder.
2. Then run new_eval/new_eval.ipynb to obtain the stats file with inform and success rates.
3. Finally, the error analysis is done in error_reader.ipynb notebook.

main_acts.py (eval only) -> error_analysis -> new_eval -> error_reader.ipynb

DO NOT TRY TO MERGE MAIN_ACTS AND ERROR_ANALYSIS. In ErrorAnalysis.ipydb, batch_size needs to be 1, so it's slow.

DATASET: (From https://github.com/wenhuchen/HDSA-Dialog)

generated using create_delex_data.py in original multiwoz repository for multiwoz 2.1 version.
add preprocessed data - train.json, val.json and test.json into hdsa_data/hdsa_data/ folder
Add delex.json file into data folder, large file(~87MB)

wget --directory-prefix=data/  https://hdsa-dialog.s3-us-west-2.amazonaws.com/delex.json

MODELS:

set, hier, mat, set++, hier++

Number of layers: e1,e2,d

SET/SET++:

initially, positional embeddings wrt individual utterances are added. After e1 layers, positional embedding wrt whole dialog are addded.
For all e1+e2 layers, attention in between individual utterances.

HIER/HIER++:

For e1 layers, attention in between individual utterances. For e2 layers, cross attention between current and last utterance. For d layers in decoder, lower triangular mask to prevent from attending to future positions of response.

MAT:

For e2 layers, cross attention between current and last utterance. For d layers in decoder, lower triangular mask to prevent from attending to future positions of response.

HYPERPARAMETERS:

Number of layers: e1,e2,d Number of heads in each encoder/decoder

Reproducability:

For SET/HIER/MAT: python main.py

For SET++/HIER++: python main_acts.py