Research Projects

In general, my research aims to uncover how the human mind processes language in order to understand broader aspects of human cognition and consciousness. I delve into these questions using insights from Large Language Models (LLMs), particularly examining the phenomenon of emergence, where the properties of a system cannot be simply reduced to the sum of its parts (underlying principles). Does emergence in LLMs compare to similar processes in human language development? Are there specific linguistic phenomena that emerge similarly in both systems?

What is (syntactic) movement?

Philosopher C. McGinn (2019) described the current theory of physical motion as an "occult phenomenon" because its "driving force" is "inherently unobservable except by its effects." A similar issue arises with syntactic movement: although displacement is ubiquitous* in natural language (expressions are often heard in one position but interpreted in another), its mechanics remain "murky and controversial" (Branigan 2012). Since Chomsky (1957), using some sort of transformational rule (e.g., Move-α, Move/Attract, Internal Merge) has been a common approach to explaining displacement in natural language.

While most efforts within the Minimalist framework have focused on identifying a legitimate trigger behind displacement (e.g., EPP, Case, Labeling), I choose not to join this quest. Instead, adopting J. Norton's (2003) perspective that the principle of causation is an "a priori science that tries to legislate in advance how the world must be," I propose to tackle the "trigger question" for syntactic movement in a "cut-the-Gordian-knot" manner. I suggest that two movement-based constructions in English — Raising-to-Object and Tough-movement — do not rely on an independent movement transformation in their derivations.

*Though my own conviction is that (numerous) reports of its ubiquity are — if not greatly — at least to some degree exaggerated.

Raising-to-Object as an emergent phenomenon

I propose an account 🔗 of Raising-to-Object (RtO) constructions in English, in which their particular syntax and semantics are seen as a "by-product" of the interplay between two "third factor" constraints: s-selection and Full Interpretation. The combination of s-selection requirements and the need to satisfy Full Interpretation naturally leads to the "dislocation" of the infinitival subject into the matrix clause without the need for any formal trigger enforcing this operation.

This interaction of constraints gives rise to an outcome not predictable from the components of grammar alone — what Chomsky (2020) refers to as "the very strange raising-to-object phenomenon." Its "strange" nature is due to the mismatch between the properties of the system and those of its individual parts — a typical feature of all emergent phenomena that tend to defy our linear cause-and-effect thinking.

"This construction is tough to believe that anyone could analyze in terms of movement"

While I am fully sympathetic to the sentiment of this subheading, which features an example of a tough-construction (TC) borrowed from Culicover & Jackendoff (2006), two major brands of analysis of this construction in generative grammar seem to be much less so 😞. Instead, these analyses assume that TCs involve an A̅-movement step at a certain point in the derivation, whether by moving an invisible (null) wh-operator or the tough-subject itself. However, these approaches turn out to be problematic because TCs exhibit some exceptional properties that set them apart from canonical A̅-movement constructions.

I suggest that TCs do not involve A̅-movement 🔗. Rather, TCs constitute an emergent property – a "trade-off" between ease and clarity – two opposing tendencies "inherent to any communicative system" (Piantadosi et al. 2011; Zipf 1949). Williams (1983, 2003) argues that tough-predicates are semantically vague: there is "a huge variety of ways a thing could be easy/tough, and in TCs the infinitive specifies which way." The semantic underspecification of tough-predicates also manifests itself outside the scope of TCs: easy recipe, easy answer, easy question, easy prey.

Piantadosi et al. contend that ambiguity is a desirable property of natural language, as it allows for a communication system that is "short and simple" because unambiguous language is partly redundant with the context and ambiguity allows the re-use of words. Thus, given the semantic underspecification of tough, Sylvie is tough would be "short and simple" from the speaker’s perspective. However, in the absence of the relevant context, the intended meaning (in what way Sylvie is tough) may not be recovered from the signal with high probability by the hearer.

This new account not only eliminates "strange" movement operations and theory-internal formatives ("null operators") from the analysis of TCs but also offers an explanation for other properties of TCs that remain unaccounted for by movement approaches, in particular, the unavailability of subject gaps.

The puzzle of wager-class verbs

The so-called wager-class verbs (e.g., wager, allege, say, claim) in English are similar to believe-class verbs (e.g., believe, consider, understand) in that they both can take finite propositional complements. However, unlike the believe-class, for most speakers, wager-class verbs do not allow infinitival (RtO) propositional complements (We believed/*wagered Mary to be the most likely winner). While most existing accounts of this paradox concentrate on the peculiarities of syntactic relations between the wager-class predicate and the embedded infinitival subject, I, building on Pesetsky’s (2021) Exfoliation theory, attempt to derive the contrast through the lens of semantic relations between the wager-class verb and the infinitive 🔗 (to the exclusion of the embedded subject).

Page updated

Report abuse