Do LLMs Understand Syntactic Center Embedding? Should They?

Anonymous

TL;DR: We find that GPT-3.5 struggles to understand multiple center embeddings, while GPT-4 achieves superhuman levels of accuracy.

Abstract: We consider the case of syntactic center embedding, where an embedding phrase contains material on both sides of the embedded phrase. While a single center embedding is easily understandable for human language users, multiple center embeddings are generally uninterpretable. Despite this, it has been claimed that multiple embeddings are in fact grammatically acceptable. We construct sentences with center embeddings of varying levels, ranging from 1-4, and we find that GPT-3.5, like humans, interprets level 1 sentences correctly, but fails with higher levels. On the other hand, GPT-4 achieves superhuman accuracy levels, with nearly perfect results even with 3 or 4 levels of embeddings. We suggest that this raises relevant questions about the relation of LLMs to the human language faculty.

Paper Type: short

Research Area: Linguistic theories, Cognitive Modeling and Psycholinguistics

Contribution Types: Model analysis & interpretability, Theory

Languages Studied: English