Adapting Pretrained Transformer to Lattices for Spoken Language Understanding

This repo contains source code of our ASRU 2019 paper "Adapting Pretrained Transformer to Lattices for Spoken Language Understanding"

Requirements

• Python >= 3.6

Required python packages are listed in requirements.txt.

Dataset

Unfortunately, we are not allowed to redistribute the dataset(ATIS). The dataset needs to be obtained from LDC

• https://catalog.ldc.upenn.edu/LDC93S5
• https://catalog.ldc.upenn.edu/LDC94S19
• https://catalog.ldc.upenn.edu/LDC95S26

Preprocess

Convert lattices to PLF format

We use the PLF format lattices, you can use this script to convert Kaldi lattices to PLF format

https://github.com/noisychannel/phrase_speech_translation/blob/master/asr_util/kaldi2FST.sh

Create dataset

python3 preproc-lattice.py [-h] dataset_file lattice_file out_file

• dataset_file: csv file with fields id, text, labels. The id field should match with the utterance ids.
• lattice_file: PLF lattice generated from the above script.
• out_file: output filename.

Training

Sample usage:

python3 run_openai_gpt_atis_lattice.py
    --train_dataset <train_csv_file>
    --eval_dataset <eval_csv_file>
    --model_name openai-gpt
    --output_dir <output_dir>
    --do_train --do_eval
    --task <intent/slot>
    --num_train_epochs 5
    --attn_bias
    --probabilistic_masks

• probabilistice_masks: Whether to use probabilistic_masks. Binary masks will be used if not set.
• linearize: linearize lattices.

Reference

Please cite the following paper

@inproceedings{
    huang2019adapting,
    title={Adapting Pretrained Transformer to Lattices for Spoken Language Understanding},
    author={Chao-Wei Huang and Yun-Nung Chen},
    booktitle={2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)},
    year={2019},
    organization={IEEE}
}