Leveraging Multilingual Self-Supervised Pretrained Models for Sequence-to-Sequence End-to-End Spoken Language Understanding

[Paper]

NLU

Training

1. Install "transformers<=4.19.0".
2. Go to dataset directory: cd nlu/slurp/.
3. Prepare data: ./prepare_data.py /path/to/slurp data.
4. Train a model: ./finetune.sh (tested on A6000 GPU).
5. Run evaluation: ./evaluate.py data/test.target output/test_generations.txt.

Alternatively, you can use the pretrained models hosted on Hugging Face Hub.

Pretrained Models

Dataset	Directory	Pretrained model
SLURP	nlu/slurp	akreal/mbart-large-50-finetuned-slurp
SLUE	nlu/slue	akreal/mbart-large-50-finetuned-slue
CATSLU	nlu/catslu	akreal/mbart-large-50-finetuned-catslu
MEDIA	nlu/media	akreal/mbart-large-50-finetuned-media
PortMEDIA-Dom	nlu/portmedia_dom	akreal/mbart-large-50-finetuned-portmedia-dom
PortMEDIA-Lang	nlu/portmedia_lang	akreal/mbart-large-50-finetuned-portmedia-lang

SLU

Training

1. Install the regular ESPnet version.
2. Copy the model configuration file from this repository: cp slu/slurp/train_asr_conformer_xlsr_mbart.yaml /path/to/espnet/egs2/slurp_entity/asr1/conf/
3. Run the recipe: ./run.sh --asr_config conf/train_asr_conformer_xlsr_mbart.yaml.
4. If you want to use pretrained Adaptor, download it from the link in the next section and run the recipe with it: ./run.sh --asr_config conf/train_asr_conformer_xlsr_mbart.yaml --pretrained_model downloads/conformer08x08h_d1024_xlsr_ts_lr5e-5_attcela_7kh_ave.pth:::decoder --asr_tag postdec-aed_7kh.

Pretrained Models

The following are SLU models trained with the Adaptor that is pretrained on 7k hours with PostDec-AED loss.

Dataset	Recipe	Pretrained model
SLURP	slurp_entity	Link
SLUE	slue-voxpopuli	Link
CATSLU	catslu_entity	Link
MEDIA	media	Link
PortMEDIA-Dom	portmedia_dom	Link
PortMEDIA-Lang	portmedia_lang	Link

Cross-lingual PortMEDIA-Lang SLU model finetuned from the MEDIA SLU model: Link.

Adaptor

Training

1. Install the custom version of ESPnet: git clone --branch v.202207 --depth 1 git@github.com:espnet/espnet.git /path/to/espnet-adaptor-pretrain
2. Copy the modifications: rsync -avh adaptor/espnet/ /path/to/espnet-adaptor-pretrain/
3. Follow ESPnet installation instructions.
4. Run the recipe: cd /path/to/espnet-adaptor-pretrain/egs2/commonvoice/asr1; ./run.sh

Pretrained Models

Loss	Configuration	Pretrained model
PreEnc MC	conf/train_adaptor_conformer_preenc-mc.yaml	Link
PreEnc CTC	conf/train_adaptor_conformer_preenc-ctc.yaml	Link
PostEnc MC	conf/train_adaptor_conformer_postenc-mc.yaml	Link
PostDec MC	conf/train_adaptor_conformer_postdec-mc.yaml	Link
PostDec AED	conf/train_adaptor_conformer_postdec-aed.yaml	Link
PreEnc CTC + PostDec AED	conf/train_adaptor_conformer_preenc-ctc_postdec-aed.yaml	Link
PreEnc CTC + PostDec AED + PostEnc MC	conf/train_adaptor_conformer_preenc-ctc_postenc-mc_postdec-aed.yaml	Link
PostEnc MC (1K hours English data)		Link
PostDec AED (7K hours data)		Link

Citation

@INPROCEEDINGS{10389655,
  author={Denisov, Pavel and Vu, Ngoc Thang},
  booktitle={2023 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)}, 
  title={Leveraging Multilingual Self-Supervised Pretrained Models for Sequence-to-Sequence End-to-End Spoken Language Understanding}, 
  year={2023},
  volume={},
  number={},
  pages={1-8},
  keywords={Training;Error analysis;Conferences;Predictive models;Benchmark testing;Filling;Data models;spoken language understanding;self-supervised learning;end-to-end;sequence-to-sequence;multilingual},
  doi={10.1109/ASRU57964.2023.10389655}
}