eduzhai > Applied Sciences > Engineering >

Deep F-measure Maximization for End-to-End Speech Understanding

  • king
  • (0) Download
  • 20210506
  • Save

... pages left unread,continue reading

Document pages: 5 pages

Abstract: Spoken language understanding (SLU) datasets, like many other machinelearning datasets, usually suffer from the label imbalance problem. Labelimbalance usually causes the learned model to replicate similar biases at theoutput which raises the issue of unfairness to the minority classes in thedataset. In this work, we approach the fairness problem by maximizing theF-measure instead of accuracy in neural network model training. We propose adifferentiable approximation to the F-measure and train the network with thisobjective using standard backpropagation. We perform experiments on twostandard fairness datasets, Adult, and Communities and Crime, and also onspeech-to-intent detection on the ATIS dataset and speech-to-image conceptclassification on the Speech-COCO dataset. In all four of these tasks,F-measure maximization results in improved micro-F1 scores, with absoluteimprovements of up to 8 absolute, as compared to models trained with thecross-entropy loss function. In the two multi-class SLU tasks, the proposedapproach significantly improves class coverage, i.e., the number of classeswith positive recall.

Please select stars to rate!

         

0 comments Sign in to leave a comment.

    Data loading, please wait...
×