Robust Spoken Language Understanding with Unsupervised ASR-Error Adaptation

  • Advised by Prof. Kai Yu;

  • Published on IEEE International Conference on Acoustics, Speech, and Signal Processing, 2018. Paper;

  • Designed an adversarial adaptation method to learn ASR-error pattern for SLU without annotation on ASR hypotheses, and proposed a unsupervised task-adaptation (Bi-LM) architecture for robust SLU;

  • Exploited the share part of the parameters in feature encoders to enforce the tasks in a similar feature space, and transferred the transcript side slot tagging model to ASR hypotheses;

  • Implemented in PyTorch;

  • Achieved significant improvements over strong baselines, reached performance very close to the oracle system (Baseline: 81.90%/78.71%; Oracle: 84.65%/85.64%; Ours: 85.11%) on Chinese character level.