Work Package 3 - Compositionality and contextuality
Research lines initiated by coordinating postdoc
Project 1 - The neural basis of the combinatoriality and contextuality of language
This project brings together insights from computational linguistics, psycholinguistics, neuroscience, and cognitive science aiming to reveal universal and fundamental neural mechanisms, such as flexible binding and memory retrieval, that underlie not only the human capacity to understand and produce language, but also reasoning, planning and visual cognition. Focus is mainly on one of the main questions of WP3, namely: "What are the primitive building blocks of language stored in lexical memory, and what aspects of an utterance are computed online by binding operations?"
The project consists of two parts, which both aim at modelling the neural instantiation of a critical component of the human language system.
The first part of the project, in close collaboration with Willem Zuidema, investigates the neural basis of the combinatoriality of language. Our main focus is on spike-time-dependent mechanisms of flexible, dynamic binding, and the development of a novel type of neural network models of language processing that incorporates effective spiking dynamics. Among others, we will explore the possibility of a neural switchboard that flexibly routes properly coded neural signals between remote brain areas. The project, as well as the closely related PhD project 13, aims at integrating formal, computational accounts of syntactic processing with neurally plausible implementations of binding, capitalizing on the various expertises of the members of work packages 2 and 3, which cross the disciplines of a.o. computational linguistics, psycholinguistics and neuroscience.
The second part of the project deals with the contextuality and context-dependency of language, which question we approach from the perspective of the human memory system. The basic idea is that generating a novel sentence involves both binding abstract (context-free) elements from a semantic memory (part 1), as well as reuse of larger, sentence fragments (multi-word constructions) that are stored in episodic memory, where they are embedded in the context of previously analysed sentences (part 2). Building on our expertise from computational linguistics, particularly Data Oriented Parsing, and the neuroscience of memory, we develop a computational model of the interaction between episodic and semantic memory in language, which we apply to the tasks of syntactic parsing and next word prediction.
It was shown that cognitively informed notion of building blocks leads to a more accurate model of language processing when tested on benchmarks. Paper currently under review. The cognitively inspired natural language parser developed by Dr. Borensztajn has been used by other researchers in the field.