Enhancing Linguistic Competence of Language Models through Pre-training with Language Learning Tasks

TL;DR

This paper introduces L2T, a pre-training framework that incorporates language learning tasks to explicitly enhance linguistic competence in language models, inspired by human language acquisition.

Contribution

It presents a novel pre-training method combining raw text with structured language learning tasks, improving linguistic skills and accelerating learning in language models.

Findings

L2T improves performance on linguistic competence benchmarks.

Pre-training with L2T accelerates linguistic skill acquisition.

Maintains competitive reasoning performance.

Abstract

Language models (LMs) are pre-trained on raw text datasets to generate text sequences token-by-token. While this approach facilitates the learning of world knowledge and reasoning, it does not explicitly optimize for linguistic competence. To bridge this gap, we propose L2T, a pre-training framework integrating Language Learning Tasks alongside standard next-token prediction. Inspired by human language acquisition, L2T transforms raw text into structured input-output pairs to provide explicit linguistic stimulation. Pre-training LMs on a mixture of raw text and L2T data not only improves overall performance on linguistic competence benchmarks but accelerates its acquisition, while maintaining competitive performance on general reasoning tasks.