What do the successes of AI language systems tell us about how humans represent and use language?

What is it about?

Every week there seems to be another report of an amazing artificial intelligence (AI) system that can do something remarkable: have a conversation, create pictures from a description, identify objects in scenes, write a story, and so on. These advances are truly amazing, especially from the perspective of the slow progress of AI throughout the 20th century. As systems come closer and closer to levels of human performance (or even surpass them), what does this tell us about human psychology? Our article focuses on the representation of words and sentences. Natural Language Processing (NLP) systems have improved greatly in recent years, and they are able to pass some tests of their understanding of words, phrases, and sentences. Some have argued that we should take the representations of such systems as theories of psychological representation, that people may be using representations similar to those of successful AI systems. Our article reviews basic facts about the psychology of word meaning, including what a theory must explain in human behavior. We argue that NLP systems focus on only a small slice of human language behavior and cannot immediately explain other pieces of the pie. We look at extensions and new proposals that might make such systems more complete. Ultimately, a complete system of word meaning must break out of the circle of words that most NLP systems operate in. That is, their inputs and outputs are only text samples, whereas people use language to refer to and change the world as well as to express and change internal beliefs and desires. People build word meanings in part through those social and physical interactions. When AI systems incorporate these aspects of language, they will come much closer to using words the way we do.

Featured Image