Titel: Understanding of, and Explanations with, Large Language Models

Professor Dr. Michael Franke vom Seminar für Sprachwissenschaft der Universität Tübingen erhält eine Momentum-Förderung der Volkswagenstiftung für ein Forschungsprojekt zu KI-gestützten Sprachmodellen. Franke will darin Risiken und Chancen dieser Modelle eruieren und dazu geisteswissenschaftliche Erkenntnisse mit Forschung aus Kognitionswissenschaft und künstlicher Intelligenz verbinden. Für das Projekt mit dem Titel „Understanding of, and Explanations with, Large Language Models“ erhält der Wissenschaftler eine Förderung in Höhe von rund 920.000 Euro über vier Jahre. Die „Momentum-Förderung für Erstberufene“ der Volkswagenstiftung soll Wissenschaftlerinnen und Wissenschaftlern die Möglichkeit geben, ihre Professur strategisch und inhaltlich weiterzuentwickeln.

Sprachmodelle wie ChatGPT, sogenannte Large Language Models (LLMs), werden rasant weiterentwickelt. Eine abwägende wissenschaftliche Analyse der Auswirkungen dieser Technologien ist dabei kaum möglich, zumal LLMs so komplex sind, dass sich ihre internen Mechanismen und der Output, den sie produzieren, nur unzureichend nachvollziehen lassen. Michael Franke will mit einem interdisziplinären Ansatz dazu beitragen, die Folgen von Anwendungen mit LLMs für die Gesellschaft transparenter zu machen.

Title: Linguistic Meaning and Bayesian Modelling (LMBayes)

The project addresses the mathematical modelling of logical conclusions and inferences as an essential part of our use of language. Human communication processes cannot be
modelled by pure logic alone; probability calculations and world knowledge have to be added. The integration of such probability calculations is currently being driven by Bayesian models that are combined with traditional logic. However, the use of Bayesian methods for modelling human communication in all its diversity and complexity poses difficult mathematical challenges to the field, which the project aims to address in a collaboration of linguistics, mathematics and computer science, thus advancing the field as a whole.

Title: CRC 1718: Common Ground - Cognition – Grammar – Communication

The chair is involved in two projects in the CRC 1718: Common Ground - Cognition – Grammar – Communication.

A1: Probabilistic Reasoning about Common Ground
In search for a cognitively plausible formal model of reasoning about uncertainty in common ground, this project addresses the problems of representational and of inferential complexity that naive approaches face. By drawing on tenets from Relevance Theory, common ground is treated as inferred, rather than as given (ex ante). The project aims to develop probabilistic models of this inference process by drawing inspiration from resource-bounded rationality, dual-processing accounts of reasoning and formal representations of inattentiveness. Novel experimental tasks are developed to test downstream predictions from our probabilistic models.

A7: Modeling Great Ape Signaling Behavior: Evolutionary Roots of Common Ground
Great ape communication is a critical puzzle piece for understanding the evolution of language, including communicative practices that rely on shared contextual information as a form of common ground. This project seeks to provide novel methodological tools – bespoke probabilistic models – for analyzing observational data from great ape communicative signaling to provide an empirical foundation for theorizing about the evolution of common ground-based communication. Models will be developed to infer contextual meaning of multi-modal signals in newly collected large data sets of orangutan and chimpanzee multimodal signaling behavior in the wild.

Title: CommuniCause
Duration: 01.01.2025 until 01.01.2028
Abstract:
Knowledge of causal processes is vital for all aspects of our lives, from mundane
consumer choices to high-impact socio-political decision making. Much of our
causal knowledge is acquired not from individual experience, but from cultural
transmission via language. While theorists have amassed a large body of
philosophical and psychological understanding about causation, surprisingly little
research has been devoted to a theoretical understanding of the processes that
underlie communication of causal information and the role of the linguistic signal
in this transmission.
The main objective of this project is to apply methods and tools from experimental
pragmatics to shed light on a wide range of puzzles about causal language and
cognition. We introduce a new pragmatic framework, the "CommuniCause"
approach, which makes new and empirically testable predictions and offers a
unified explanation for a number of disparate phenomena, both old and new.
CommuniCause brings together two important strands of research that have been
isolated until now, in a way that will benefit both communities. It draws on
established philosophical theorizing and recent computational models of
individual causal cognition, but is distinguished by its focus on linguistic factors:
Communicating Causality