Event details from: ACLC Seminar | Fausto Carcassi

Date
15 December 2023
Time
16:00 – 17:00
Room
1.15

Open-ending cognitive models of pragmatics

Collaborative project with Polina Tsvilodub and Michael Franke

Building explanatory computational models of pragmatic language production and interpretation has been a long-standing goal in cognitive science (e.g., Frank & Goodman, 2012; Degen, 2023; Franke & Jäger, 2016). Despite all the recent progress, current cognitive models are mostly limited to handling manually pre-specified utterances and interpretations. To address this limitation, we present a neuro-symbolic framework for modeling open-ended language use, ScAffolded Generative models for Explanation (SAGE). This framework combines LLMs and symbolic modules to build cognitive models. I will present two case studies. First, a model of language production inspired by Dale & Reiter (1995) applied to scene description (Tsvilodub & Franke, 2023). Second, a model of implicature interpretation that encodes Gricean-style reasoning (Grice, 1975). This model is tested on a novel dataset and a dataset from Hu et al. (2023). Modelling results are also compared to the data from a behavioural experiment. In both case studies, performance robustly exceeds a GPT-3.5-turbo baseline. In sum, our framework opens up avenues for open-ended yet cognitively motivated and explainable modeling of pragmatic language production and interpretation.

About the ACLC seminar series

The ACLC seminar series is a two weekly lecture series organized by the ACLC, the Amsterdam Center for Language and Communication.