The architecture of agreement: from strings to structures

In this paper we raise the question “why is agreement so common across natural languages?”. We will argue that the challenge of grammar inference in natural and artificial languages provides key insights into the ubiquity of agreement. By grammar inference, we mean the discovery of a procedure that (i) determines string well-formedness on the basis of exposure to unannotated expressions of a language; and (ii) allows for the construction of structural descriptions for well-formed strings. The idealized version of this problem results in the identification of the generator of a stringset, or -more realistically- a restriction of the class of possible generators. We argue that agreement plays a crucial role not only in flagging dependencies between expressions at the string level, but also, considering that agreement relations occur in restricted structural configurations, in restricting the class of structural descriptions compatible with a string. As such, agreement mediates between strings and structure, providing a parser with information to solve the grammar inference problem. We will furthermore argue that the mechanisms involved in grammar identification are not restricted to natural language acquisition and processing, but in fact extend to a class of problems that motivated much research in the theory of symbolic encoding of dynamical systems and machine learning.