The Rediscovery Hypothesis: Language Models Need to Meet Linguistics

TL;DR

This paper investigates whether linguistic knowledge is essential for language model performance, showing compressed models retain linguistic information and proposing an information-theoretic framework to relate language modeling to linguistic knowledge.

Contribution

It introduces an information-theoretic framework linking language modeling objectives with linguistic knowledge and demonstrates that compressed models still retain linguistic understanding.

Findings

Compressed models retain linguistic knowledge despite size reduction.

Proposed framework quantifies the impact of linguistic information on word prediction.

Experiments on synthetic and real NLP tasks support the rediscovery hypothesis.

Abstract

There is an ongoing debate in the NLP community whether modern language models contain linguistic knowledge, recovered through so-called probes. In this paper, we study whether linguistic knowledge is a necessary condition for the good performance of modern language models, which we call the \textit{rediscovery hypothesis}. In the first place, we show that language models that are significantly compressed but perform well on their pretraining objectives retain good scores when probed for linguistic structures. This result supports the rediscovery hypothesis and leads to the second contribution of our paper: an information-theoretic framework that relates language modeling objectives with linguistic information. This framework also provides a metric to measure the impact of linguistic information on the word prediction task. We reinforce our analytical results with various experiments, both on synthetic and on real NLP tasks in English.