Assessing the Ability of LSTMs to Learn Syntax-Sensitive Dependencies

TL;DR

Problem: whether LSTMs can learn syntax-sensitive dependencies such as English subject-verb agreement without explicit structural representations. Approach: evaluate LSTMs under strong supervision (number prediction and grammaticality judgments) and under language-modeling on a corpus with subject-verb agreement, and analyze error patterns. Key findings: strong supervision yields very high accuracy overall, but performance degrades on harder cases that require encoding structural information; language-modeling alone performs poorly on these cases and is highly sensitive to recent nonstructural nouns; explicit supervision is necessary for reliable learning of the dependency. Significance: indicates limits of purely predictive objectives for syntax, informs NLP applications that require long-distance agreement, and motivates exploring stronger architectures or direct supervision to capture hierarchical dependencies.

Abstract

The success of long short-term memory (LSTM) neural networks in language processing is typically attributed to their ability to capture long-distance statistical regularities. Linguistic regularities are often sensitive to syntactic structure; can such dependencies be captured by LSTMs, which do not have explicit structural representations? We begin addressing this question using number agreement in English subject-verb dependencies. We probe the architecture's grammatical competence both using training objectives with an explicit grammatical target (number prediction, grammaticality judgments) and using language models. In the strongly supervised settings, the LSTM achieved very high overall accuracy (less than 1% errors), but errors increased when sequential and structural information conflicted. The frequency of such errors rose sharply in the language-modeling setting. We conclude that LSTMs can capture a non-trivial amount of grammatical structure given targeted supervision, but stronger architectures may be required to further reduce errors; furthermore, the language modeling signal is insufficient for capturing syntax-sensitive dependencies, and should be supplemented with more direct supervision if such dependencies need to be captured.