CompCog: Generalization versus item-specificity in language processing and change

Language allows people to produce and understand sentences that have never been said before, making human language unique among the communication systems found in nature. Humans do this by recombining pieces of sentences that have been heard before in new ways to express new meanings. But among the multitude of ways that humans can make new sentences, there are still some constraints.

For example, speakers need to know what is allowed in their language and what is not in addition to knowing detailed statistics of how words and sentence structures are typically used. Deciding how to say a new sentence therefore requires searching through many ways to recombine known words and phrases to express a message in a way that is both clear and natural sounding.

Understanding how language users sort through these possibilities to express new ideas has been a challenge for a scientific understanding of the human mind. This project combines tools from linguistics, psychology, and computational approaches to advance an understanding of language processing in the mind. Advances in this area contribute to improved diagnosis and treatment of language disorders, improvements in second language instruction, and improvements in teaching other language-related skills such as programming.

This project supports education by training students in quantitative skills including experimental methods, programming, and statistics. Teaching these quantitative skills in the context of linguistics provides an opportunity to reach more diverse audiences, such as women and underrepresented minorities, who are better represented in linguistics compared to some other disciplines.

The project also creates publicly available corpora of naturally occurring instances of verb usages that are manually annotated. These corpora enable future research on a wide variety of questions in linguistics by researchers with fewer resources for data annotation.

This project focuses on how speakers use different types of knowledge when deciding what to say. For example, many verbs in English (such as “give” or “send”) can be used in two structures: (1) I gave the man a book, and (2) I gave a book to the man. Whether a speaker prefers one structure versus the other depends in part on the verb.

For example, speakers prefer to use “give” in structure (1) but “send” in structure (2). Preference also depends in part on other properties of the sentence. For example, speakers prefer to use structure (1) when the receiver is short (e.g., “I gave him a book”) and structure (2) when the receiver is long (e.g., “I gave a book to the old man who I met at the grocery store”).

These other properties are more generalizable in that they apply to all verbs, unlike the idiosyncratic verb-specific preferences. This project examines how speakers make use of both the generalizable constraints and the verb-specific preferences when deciding in real time which structure to use for a sentence. This project first builds computational models of how the mind integrates these two different types of knowledge.

Then, these models are tested using controlled experiments. Finally, further computational models of how language changes over time are built. Each generation of speakers must learn the existing language, but each generation also subtly re-shapes the language, creating a situation in which speaker preferences and the language itself are simultaneously co-evolving.

Computational models are used to investigate how human capacity for language and the structure of the languages become jointly optimized.

This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.