Prediction plays an important role in verbal interactions. Independent of the neuroscientific research inspired by Predictive Coding and similar approaches, prediction is also at the heart of artificial systems achieving state-of-the-art performance in speech recognition, translation, and other NLP tasks. Current language models are primarily trained on prediction tasks and subsequently tuned to specific applications. We may characterize them as multilevel, generative models aimed at prediction, conceptually similar to current cognitive and neuroscientific approaches to speech production and comprehension.
The aim of this PhD project is to extend and apply current acoustic and language models to conversations and use them to model communicating agents’ predictive processes. Communicating agents are hypothesized to predict each other’s behavior at multiple levels, resulting in general “alignment” or “synchronisation”. Recent language models provide a novel opportunity for capturing this process. In the current project we will select applicable models that can generate predictions for upcoming utterances on the basis of interaction history and validate them against experimental data. We use the best individual models to construct dialogue-level models in order to test the general synchronization hypothesis as a basic feature of human communication.
előírt nyelvtudás: English or Hungarian további elvárások: basic knowledge of machine learning, basic knowledge of signal processing