eduzhai > Applied Sciences > Engineering >

Robust Prediction of Punctuation and Truecasing for Medical ASR

  • king
  • (0) Download
  • 20210505
  • Save

... pages left unread,continue reading

Document pages: 10 pages

Abstract: Automatic speech recognition (ASR) systems in the medical domain that focuson transcribing clinical dictations and doctor-patient conversations often posemany challenges due to the complexity of the domain. ASR output typicallyundergoes automatic punctuation to enable users to speak naturally, withouthaving to vocalise awkward and explicit punctuation commands, such as "period ", "add comma " or "exclamation point ", while truecasing enhances user readabilityand improves the performance of downstream NLP tasks. This paper proposes aconditional joint modeling framework for prediction of punctuation andtruecasing using pretrained masked language models such as BERT, BioBERT andRoBERTa. We also present techniques for domain and task specific adaptation byfine-tuning masked language models with medical domain data. Finally, weimprove the robustness of the model against common errors made in ASR byperforming data augmentation. Experiments performed on dictation andconversational style corpora show that our proposed model achieves ~5 absoluteimprovement on ground truth text and ~10 improvement on ASR outputs overbaseline models under F1 metric.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×