Dependency syntax as the simplest theory of grammar

Area of Science:

• Linguistics
• Cognitive Science
• Artificial Intelligence

Background:

• The complexity of human language syntax has been a long-standing debate, with some theories suggesting it's unlearnable from input alone.
• The emergence of large language models (LLMs) with remarkable syntactic abilities challenges traditional views.
• Chomsky's transformational grammar, emphasizing phrase structure and transformations, has been a dominant framework.

Purpose of the Study:

• To propose a simplified view of syntax based on dependency rules.
• To challenge the necessity of phrase structure and transformation rules in explaining language.
• To reconcile human language acquisition and LLM learning with a unified syntactic framework.

Main Methods:

• Theoretical analysis of linguistic syntax.
• Comparison of dependency grammar with transformational grammar.
• Evaluation against human language processing data and crosslinguistic universals.
• Speculation on the learning mechanisms of LLMs.

Main Results:

• A dependency grammar model accounts for diverse syntactic phenomena and crosslinguistic word order universals.
• This simple view of syntax better explains specific human language data.
• The dependency grammar approach resolves the syntax learnability problem.

Conclusions:

• Human language syntax can be understood as a set of dependency rules, eliminating the need for complex phrase structure or transformation rules.
• Large language models likely learn syntax through pattern recognition, akin to child language acquisition.
• This perspective offers a more parsimonious and learnable model of syntax for both humans and AI.