Joint Training for Selective Prediction

TL;DR

This work introduces a novel joint-training approach that simultaneously optimizes learned representations used by the classifier module and a learned deferral policy, which leads to better SP outcomes over two strong baselines, but also improves the performance of both modules.

Abstract

Classifier models are prevalent in natural language processing (NLP), often with high accuracy. Yet in real world settings, human-in-the-loop systems can foster trust in model outputs and even higher performance. Selective Prediction (SP) methods determine when to adopt a classifier's output versus defer to a human. Previous SP approaches have addressed how to improve softmax as a measure of model confidence, or have developed separate confidence estimators. One previous method involves learning a deferral model based on engineered features. We introduce a novel joint-training approach that simultaneously optimizes learned representations used by the classifier module and a learned deferral policy. Our results on four classification tasks demonstrate that joint training not only leads to better SP outcomes over two strong baselines, but also improves the performance of both modules.

Figures (3)

• Figure 1: The model architecture with three steps. CL is the classification model of the given task, DP is the Deferral Policy model. A is the reward function from equation (\ref{['eq:reward']}).
• Figure 2: Four-valued reward signal $A$. The desirable outcomes $a$ and $d$ should have higher values than the undesirable outcomes $c$ and $d$.
• Figure 3: Sensitivity of four metrics to deferral weight ($d$ in the reward signal $A$) on the BEETLE dataset: deferral rate (blue), SP accuracy (red), CL accuracy (brown) and DP accuracy (green).