Acquiring Bidirectionality via Large and Small Language Models

TL;DR

This paper introduces a method to enhance token representations by combining large unidirectional LMs with newly trained small backward LMs, significantly improving performance in token classification tasks, especially in rare domains and few-shot scenarios.

Contribution

The paper proposes training a small backward language model and concatenating its representations with existing models to improve token classification performance.

Findings

Over 10-point performance improvement in named entity recognition benchmarks.

Enhanced effectiveness in rare domain and few-shot learning settings.

Demonstrated that combining bidirectional and unidirectional models benefits downstream tasks.

Abstract

Using token representation from bidirectional language models (LMs) such as BERT is still a widely used approach for token-classification tasks. Even though there exist much larger unidirectional LMs such as Llama-2, they are rarely used to replace the token representation of bidirectional LMs. In this work, we hypothesize that their lack of bidirectionality is keeping them behind. To that end, we propose to newly train a small backward LM and concatenate its representations to those of existing LM for downstream tasks. Through experiments in named entity recognition, we demonstrate that introducing backward model improves the benchmark performance more than 10 points. Furthermore, we show that the proposed method is especially effective for rare domains and in few-shot learning settings.