The Rational Processing of Language Illusions

Abstract

This thesis presents a unified rational account for three so-called language illusions in English. Language illusions refer to sentences that are ungrammatical, semantically ill-formed, or implausible but are still considered acceptable by native speakers of that language. They pose a great challenge to the scientific inquiry into the nature of language processing, especially in the understanding of why there is a mismatch between the actual comprehension and the literal form and meaning of the sentence. My thesis shows that this pattern of language comprehension manifests a unique trait of human beings when they deal with uncertainty. The three language illusions are the depth-charge illusion (e.g., No head injury is too trivial to be ignored.), the comparative illusion (e.g., More people have been to Russia than I have.), and the negative polarity illusion (e.g., Many authors that few critics recommended have ever received acknowledgment for a best-selling novel). The problems for each of the language illusions can be traced back to meaning implausibility, semantic ill-formedness, and ungrammaticality, respectively. Numerous studies have investigated the effect of different linguistic factors on the degree of the illusion but none has provided a satisfactory explanation, let alone a unified one. The unified account I attempt here argues that humans rationally infer the intended message behind the perceived linguistic signal, by weighing the prior probability of that intended message and how likely the message is to be encapsulated in the perceived signal. The inference could cover the entire sentence or only part of it. This theory originates from the noisy-channel theory of language processing (e.g., Futrell et al., 2020; Gibson et al., 2013a; Levy, 2008) under the information theory (Shannon, 1948). Ultimately, this account exemplifies rationality in language comprehension. By synthesizing the literature on rationality, I summarize the rational behavior of language processing as a set of behaviors that optimize the information gain from perceived sentences by applying the Bayes rule to deal with uncertainty and making trade-offs between information accuracy and cognitive load. I provide empirical data from human experiments as support. I also investigate whether large language models can be “tricked” by these illusions, as a starting point to investigate whether their processing patterns exhibit traits of rationality.