Seminaria w Instytucie Podstaw Informatyki PAN

Streszczenie (autorskie):

Daniel Dennett speculated in Kinds of Minds (1996):
“Perhaps the kind of mind you get when you add language to it is so different from the kind of mind you can have without language that calling them both minds is a mistake.”

Recent work in AI can be seen as testing Dennett’s thesis by exploring the performance of AI systems with and without linguistic training. I will argue that the success of Large Language Models at inferential reasoning, limited though it is, supports Dennett’s radical view about the effect of language on thought. I suggest it is the abstractness and efficiency of linguistic encoding that lies behind the capacity of LLMs to perform inferences across a wide range of domains. In a slogan, language makes inference computationally tractable. I will assess what these results in AI indicate about the role of language in the workings of our own biological minds.

Streszczenie (autorskie):

In this talk, I discuss the problem of nominal properties of complementizer phrases (CPs) in both direct and indirect discourse. I use data from nominalization of CPs in Japanese to show that the availability and formal properties of such structures are not uniform across types of matrix verbs. Then I argue that these properties are regulated by a hierarchy of verbs recently developed by Susi Wurmbrand and her colleagues. Based on these findings, I propose new formal accounts of matrix verbs capturing the effects observed for nominalised CPs. I conclude by drawing more general prospects for deriving the reportative character of matrix verbs, regulated by the abovementioned hierarchy.

Streszczenie (autorskie):

Plain disjunctive sentences, such as The mystery box contains a blue ball or a yellow ball, typically imply that the speaker does not know which of the two disjuncts is true. This is known as an 'ignorance' inference. We can distinguish between two aspects of this inference: the negated universal upper bound part (i.e., the speaker is uncertain about each disjunct), which we call 'uncertainty', and the existential lower bound part (i.e., the speaker considers each disjunct possible), which we call 'possibility'. In the traditional approach, uncertainty is derived as an implicature, from which possibility follows. When disjunctions are embedded under a universal operator, such as It is certain that the mystery box contains a blue ball or a yellow ball, an inference analogous to possibility arises (i.e., it's possible that the mystery box contains a blue ball and it's possible that it contains a yellow ball). This inference is generally called a 'distributive' inference and the traditional approach derives it in the same way as in the unembedded case, from the corresponding negated universal inference (i.e. it is uncertain that the mystery box contains a blue ball and it is uncertain that it contains a yellow ball). In both the simple and the embedded cases, the traditional implicature approach predicts that the derived lower bound inference should never arise without the corresponding negated universal one. In this talk, we report on two experiments using a sentence-picture verification task based on the mystery box paradigm, the results of which challenge the traditional approach to ignorance and distributive inferences. Our findings show that possibility can arise without uncertainty, and that, in the same way, the distributive inference can arise without the corresponding negated universal one, thus calling for a reevaluation of the traditional view of disjunction and its inferences. We discuss how alternative theories can account for the observed patterns of inference derivation in a unified fashion and the open challenges that remain.