The Rational Processing of Language Illusions

(Please email jiayi.lu [at] stanford.edu (Jiayi Lu) for Zoom link)

Abstract: Language processing is not always compositional, and the extracted linguistic representation does not always map to the sentence’s literal meaning. Language illusion is such a case and poses an interesting example for us to understand the nature of language processing. In this talk, I will present a unified rational account inspired by the noisy-channel framework in information theory (Shannon, 1948) to explain why language illusion arises. I will also present work from the evaluation of large language models to contribute to another discussion of whether LLMs have human-like language processing behavior by getting tricked by these illusions. 

Here are the illusion sentences I am going to focus on:

• Depth-charge illusion: No head injury is too trivial to be ignored.
• Comparative illusion: More people have been to Russia than I have.
• Negative polarity illusion: Many authors that few critics recommended have ever received acknowledgment for a best-selling novel.