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

TL;DR

This paper introduces a differentiable F-measure maximization method for training neural networks, improving fairness and class coverage in speech understanding and other tasks by addressing label imbalance.

Contribution

It proposes a novel differentiable approximation to the F-measure and demonstrates its effectiveness across multiple datasets and tasks.

Findings

Up to 8% absolute improvement in micro-F1 scores.

Significant increase in class coverage and positive recall.

Effective across speech and non-speech datasets.

Abstract

Spoken language understanding (SLU) datasets, like many other machine learning datasets, usually suffer from the label imbalance problem. Label imbalance usually causes the learned model to replicate similar biases at the output which raises the issue of unfairness to the minority classes in the dataset. In this work, we approach the fairness problem by maximizing the F-measure instead of accuracy in neural network model training. We propose a differentiable approximation to the F-measure and train the network with this objective using standard backpropagation. We perform experiments on two standard fairness datasets, Adult, and Communities and Crime, and also on speech-to-intent detection on the ATIS dataset and speech-to-image concept classification on the Speech-COCO dataset. In all four of these tasks, F-measure maximization results in improved micro-F1 scores, with absolute improvements of up to 8% absolute, as compared to models trained with the cross-entropy loss function. In the two multi-class SLU tasks, the proposed approach significantly improves class coverage, i.e., the number of classes with positive recall.