Speaker: Paul Smolensky (Microsoft/Johns Hopkins University)
Title: “Generative Linguistics meets Generative AI”
When: Wednesday, April 22nd, 1pm-2:30pm (Day 1) + Thursday, April 23rd, 12:30-2pm (Day 2)
Where: 32-D461
Abstract: Do the unprecedented syntactic abilities of modern generative AI systems — Large Language Models (LLMs) — falsify traditional generative principles of linguistic knowledge, given that these principles rely heavily on symbolic computation, which appears to be absent from the neural network architectures of these LLMs? In Friday’s colloquium I will argue: no. The technical foundation of this argument is introduced in the tutorial lectures of Wednesday and Thursday.
Day 1: Explaining higher cognition: Symbol structures in neural patterns
This first tutorial shows how the compositional structure of symbolic representations like those employed standardly in generative linguistics can emerge from a type of compositional structure within activation patterns in neural networks: TPRs. TPRs constitute a novel type of compositional structure that is in general continuous — non-discrete — although special cases include standard discrete structures (strings, trees, propositions). Discrete TPRs can be described as standard symbol structures that can be operated upon by neural computations to compute sophisticated discrete compositional functions such as tree-adjoining and beta-reduction. Non-discrete TPRs enable new types of formal explanations for linguistic competence (and linguistic performance), discussed primarily in Thursday’s lecture. The internal numerical representations learned by modern AI systems, in a wide variety of settings, can be well approximated by TPRs, enabling compositional interpretation and control of otherwise opaque systems.
Day 2: Synergies between symbolic grammatical theory and neural network computation
This second tutorial identifies strong synergies between the grammar of symbol structures and neural network computation. The compositionally-structured neural representations presented in Wednesday’s tutorial, TPRs, can be described at an abstract level as symbol structures and at a lower level as numerical activation patterns. Well-formed representations can be characterized at both levels as those that maximize Harmony, giving rise to symbolic grammar formalizations based in optimization: Harmonic Grammar and Optimality Theory. TPRs enable non-discrete compositional representations in which symbols have gradient strength and can blend together within a structural position: this enables novel types of grammatical explanation exemplified here by an analysis of French liaison and of source-markedness-controlled vowel harmony. Liaison consonants’ complex alternation between presence and absence is explained from their underlying strength deficiency, and marked vowels’ failure to trigger harmony is explained by the relative surface weakness resulting from their markedness.