Projects

The interface of syntax and semantics

The study of human language has traditionally distinguished semantics (meaning) from syntax (the order of words and other constituents). Word order is clearly relevant to meaning but the relation is not consistent across languages. A jaguar grabbing a man is quite different from a man grabbing a jaguar but in languages like Hixkaryana, spoken in Brazil, the order is the opposite of English:

toto    yahosiye    kamara

man   grabbed     jaguar

The jaguar grabbed the man

Linking syntax and semantics

There was a research program in the 1960s called Generative Semantics, which sought to generate syntactic structures directly from meaning. Some current theories, such as Head-Driven Phrase Structure Grammar (HPSG), retain the linkage between semantics and syntax by encoding information about syntactic categorization and semantic roles directly onto lexical items. The use of lexical coding, however, rules out a universal representation of meaning. 

Semantics in natural language generation

Large language models like AI ChatGPT have made remarkable progress in generating natural language texts in response to natural language queries. They create the illusion of semantic comprehension by mining an enormous corpus of existing texts to find statistical links to the query, enabling the program to pull up and remix content that is likely to be responsive. But, in addition to requiring training on a massive corpus, large language models do not represent the meaning underlying the text that they manipulate, leaving them prone to mistakes, hallucinations, and any biases found in the corpus on which they are trained.

The Theta Bridge Research Program

We have discovered a system of semantic notation that encodes fine-grained meaning in a universal format that is linked to but independent of the lexicon and that can be converted directly into semantically equivalent syntactic strings in multiple languages without the use of artificial intelligence or corpus training. This system makes it possible to realize the goal of Generative Semantics by reducing syntactic representation to a by-product of semantic processing. We are currently developing practical applications employing this system.

If you are interested in more information, please contact us.