Analysis and Prediction of NLP Models Via Task Embeddings

TL;DR

This paper introduces MetaEval, a collection of NLP tasks, and demonstrates how task embeddings can analyze, predict, and improve zero-shot transfer learning performance across diverse NLP tasks.

Contribution

It proposes a unified transformer model conditioned on task embeddings for analyzing and predicting task properties, enabling zero-shot inference without annotated examples.

Findings

Task embeddings reveal meaningful relationships among NLP tasks.

Predicted embeddings improve zero-shot performance on GLUE tasks.

MetaEval serves as a new benchmark for transfer learning research.

Abstract

Task embeddings are low-dimensional representations that are trained to capture task properties. In this paper, we propose MetaEval, a collection of $101$ NLP tasks. We fit a single transformer to all MetaEval tasks jointly while conditioning it on learned embeddings. The resulting task embeddings enable a novel analysis of the space of tasks. We then show that task aspects can be mapped to task embeddings for new tasks without using any annotated examples. Predicted embeddings can modulate the encoder for zero-shot inference and outperform a zero-shot baseline on GLUE tasks. The provided multitask setup can function as a benchmark for future transfer learning research.