bert-base-ner-atc-en-atco2-1h

This model allow to perform named-entity recognition (NER) on air traffic control communications data. We solve this challenge by performing token classification (NER) with a BERT model. We fine-tune a pretrained BERT model on the ner task.

For instance, if you have the following transcripts/gold annotations:

• Utterance: lufthansa three two five cleared to land runway three four left

Could you tell what are the main entities in the communication? The desired output is shown below:

• Named-entity module output: [call] lufthansa three two five [/call] [cmd] cleared to land [/cmd] [val] runway three four left [/val]

This model is a fine-tuned version of bert-base-uncased on the atco2_corpus_1h.

It achieves the following results on the development set:

• Loss: 1.4282
• Precision: 0.6195
• Recall: 0.7071
• F1: 0.6604
• Accuracy: 0.8182

Paper: ATCO2 corpus: A Large-Scale Dataset for Research on Automatic Speech Recognition and Natural Language Understanding of Air Traffic Control Communications

Authors: Juan Zuluaga-Gomez, Karel Veselý, Igor Szöke, Petr Motlicek, Martin Kocour, Mickael Rigault, Khalid Choukri, Amrutha Prasad and others

Abstract: Personal assistants, automatic speech recognizers and dialogue understanding systems are becoming more critical in our interconnected digital world. A clear example is air traffic control (ATC) communications. ATC aims at guiding aircraft and controlling the airspace in a safe and optimal manner. These voice-based dialogues are carried between an air traffic controller (ATCO) and pilots via very-high frequency radio channels. In order to incorporate these novel technologies into ATC (low-resource domain), large-scale annotated datasets are required to develop the data-driven AI systems. Two examples are automatic speech recognition (ASR) and natural language understanding (NLU). In this paper, we introduce the ATCO2 corpus, a dataset that aims at fostering research on the challenging ATC field, which has lagged behind due to lack of annotated data. The ATCO2 corpus covers 1) data collection and pre-processing, 2) pseudo-annotations of speech data, and 3) extraction of ATC-related named entities. The ATCO2 corpus is split into three subsets. 1) ATCO2-test-set corpus contains 4 hours of ATC speech with manual transcripts and a subset with gold annotations for named-entity recognition (callsign, command, value). 2) The ATCO2-PL-set corpus consists of 5281 hours of unlabeled ATC data enriched with automatic transcripts from an in-domain speech recognizer, contextual information, speaker turn information, signal-to-noise ratio estimate and English language detection score per sample. Both available for purchase through ELDA at this http URL. 3) The ATCO2-test-set-1h corpus is a one-hour subset from the original test set corpus, that we are offering for free at this url: https://www.atco2.org/data. We expect the ATCO2 corpus will foster research on robust ASR and NLU not only in the field of ATC communications but also in the general research community.

Code — GitHub repository: https://github.com/idiap/atco2-corpus

Intended uses & limitations

This model was fine-tuned on air traffic control data. We don't expect that it keeps the same performance on some others datasets where BERT was pre-trained or fine-tuned.

Training and evaluation data

See Table 6 (page 18) in our paper: ATCO2 corpus: A Large-Scale Dataset for Research on Automatic Speech Recognition and Natural Language Understanding of Air Traffic Control Communications. We described there the data used to fine-tune our NER model.

• We use the ATCO2 corpus to fine-tune this model. You can download a free sample here: https://www.atco2.org/data
• However, do not worry, we have prepared a script in our repository for preparing this databases:
  • Dataset preparation folder: https://github.com/idiap/atco2-corpus/tree/main/data/databases/atco2_test_set_1h/data_prepare_atco2_corpus_other.sh
  • Get the data in the format required by HuggingFace: speaker_role/data_preparation/prepare_spkid_atco2_corpus_test_set_1h.sh

Writing your own inference script

The snippet of code:

from transformers import pipeline, AutoTokenizer, AutoModelForTokenClassification

tokenizer = AutoTokenizer.from_pretrained("Jzuluaga/bert-base-ner-atc-en-atco2-1h")
model = AutoModelForTokenClassification.from_pretrained("Jzuluaga/bert-base-ner-atc-en-atco2-1h")


##### Process text sample
from transformers import pipeline

nlp = pipeline('ner', model=model, tokenizer=tokenizer, aggregation_strategy="first")
nlp("lufthansa three two five cleared to land runway three four left")

# output:
[{'entity_group': 'callsign', 'score': 0.8753265, 
'word': 'lufthansa three two five', 
'start': 0, 'end': 24},
{'entity_group': 'command', 'score': 0.99988264, 
'word': 'cleared to land', 'start': 25, 'end': 40}, 
{'entity_group': 'value', 'score': 0.9999145, 
'word': 'runway three four left', 'start': 41, 'end': 63}]

Cite us

If you use this code for your research, please cite our paper with:

@article{zuluaga2022bertraffic,
  title={BERTraffic: BERT-based Joint Speaker Role and Speaker Change Detection for Air Traffic Control Communications},
  author={Zuluaga-Gomez, Juan and Sarfjoo, Seyyed Saeed and Prasad, Amrutha and others},
  journal={IEEE Spoken Language Technology Workshop (SLT), Doha, Qatar},
  year={2022}
  }

and,

@article{zuluaga2022how,
    title={How Does Pre-trained Wav2Vec2. 0 Perform on Domain Shifted ASR? An Extensive Benchmark on Air Traffic Control Communications},
    author={Zuluaga-Gomez, Juan and Prasad, Amrutha and Nigmatulina, Iuliia and Sarfjoo, Saeed and others},
    journal={IEEE Spoken Language Technology Workshop (SLT), Doha, Qatar},
    year={2022}
  }

and,

@article{zuluaga2022atco2,
  title={ATCO2 corpus: A Large-Scale Dataset for Research on Automatic Speech Recognition and Natural Language Understanding of Air Traffic Control Communications},
  author={Zuluaga-Gomez, Juan and Vesel{\`y}, Karel and Sz{\"o}ke, Igor and Motlicek, Petr and others},
  journal={arXiv preprint arXiv:2211.04054},
  year={2022}
}

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 32
• eval_batch_size: 16
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• lr_scheduler_warmup_steps: 500
• training_steps: 3000

Training results

Training Loss	Epoch	Step	Validation Loss	Precision	Recall	F1	Accuracy
No log	125.0	500	0.8692	0.6396	0.7172	0.6762	0.8307
0.2158	250.0	1000	1.0074	0.5702	0.6970	0.6273	0.8245
0.2158	375.0	1500	1.3560	0.6577	0.7374	0.6952	0.8119
0.0184	500.0	2000	1.3393	0.6182	0.6869	0.6507	0.8056
0.0184	625.0	2500	1.3528	0.6087	0.7071	0.6542	0.8213
0.0175	750.0	3000	1.4282	0.6195	0.7071	0.6604	0.8182

Framework versions

• Transformers 4.24.0
• Pytorch 1.13.0+cu117
• Datasets 2.7.0
• Tokenizers 0.13.2