Call for demos

If you are interested in giving a demonstration, please contact us for more information.

Jan Doenges - Learning grammatical constructions with a biologically plausible neural network

University of Mainz

Typically, biologically accurate neural network models owe their functioning to the cooperation of a large amount of rather simple units like neurons, columns, or cell assemblies. Usage-based models of construction grammar follow a similar ‘maximalist’ principle, as they reduce the complexity of single constructions by allowing information to be specified redundantly over a multitude of constructions. Accordingly, a biologically plausible neural network model can be constructed, in which units learn to categorize and link activation representing syntactic and semantic information, and thus parallel the behavior of grammatical constructions.
The characteristic connection pattern forming a construction is supposed to arise from a previously unspecified network with a vast amount of sub-threshold connections by reinforcing connections between units that are coherently active.
This modelling approach has been exemplified for a small network with spiking neurons and grammatical constructions according to Radical Construction Grammar, but further investigations with larger networks are still necessary.
Back to schedule - Back to top

Martin Loetzsch - Recruiting Language Universals: The Perspective Reversal Experiment

Sony CSL Paris

All human languages allow speakers to adopt perspectives other than their own in conceptualising the world, and to mark perspective explicitly through words or grammatical constructions. How can we explain this universal of human languages? We argue that speakers and hearers are able to configure and apply different strategies to deal with communicative tasks and retain their use if and only if this increases communicative success and expressive power while minimizing effort in terms of memory and information processing. Each strategy requires the recruitment of neural mechanisms, which are not specific to language. We present as a concrete case study a model in which embodied robot agents are capable of creating and learning concepts, and language to express them. The agents play language games, describing objects moving about in their immediate surroundings to each other. Because the agents are embodied and situated, the perspective from which they perceive and describe a scene influences their individual perception of it. We show that in this case, giving agents the capacity to project each other's visual perspectives increases their communicative success, as does allowing them to encode perspective in language. Because human beings are embodied agents as well, situated in the physical world, their communicative success will analogically increase when they use perspective-marking language. The model thus correctly predicts that perspective marking will be culturally selected as a universal strategy in the evolution of human language.
Back to schedule - Back to top

Andras Lorincz and Palotai Zsolt - Communication Tools for Persons with Special Communication Needs

Collegium Budapest

There is an ongoing project at the Eötvös Loránd University. We have been developing tools for severely handicapped, non-speaking but speech understanding children at the Hungarian Bliss Foundation. Main components of the project are (i) the development of special communication tools including a controllable pet, Aibo, (ii) the development of machine learning techniques for situational optimization, and (iii) the development of Body Sensor Network (BSN) to form ambient intelligence.

The project has two subprojects. The first one is concerned with human-computer interface development using (i) webcam based eye communication, (ii) webcam based head communication, (iii) webcam based facial expression detection, (iv) RF-MEMS sensor based body sensor network. The other subproject deals with the tracking of communication between children who use the new tools communicate with each other by sending images/symbols and image/symbol series.  

The demonstration shows the interlinked tools, including RF-MEMS tools, webcam, the Aibo pet and the present stage of the software developed for image/symbol based communication.
Back to schedule - Back to top

Remi van Trijp - Fluid Construction Grammar as a tool for Linguistic Research

Sony CSL Paris

There is a growing trend in linguistics to take human interaction into account when explaining the internals of language. Instead of studying the Saussurean sign (i.e. the pairing of form and meaning) isolated from the language user, the population dynamics determine whether a new construction gets conventionalised or not. This means that the language faculty shows extreme fluidity: grammatical conventions are only partially shared by members of the population and constructions may not exist at some point. A grammar formalism should therefore be able to capture this fluidity and cope with the noisy data that occurs in social interaction. More specifically, the formalism has to support grammars that are dynamic because they may change after every interaction, that are bi-directional because the language speaker can act both as a speaker and a hearer, and that are extremely flexible so that utterances that are incomplete or ungrammatical can still be parsed and interpreted to some extent.

This demonstration gives a brief introduction to Fluid Construction Grammar (FCG), a grammar formalism that has been designed by Luc Steels for doing experiments on the origins of language. FCG captures all the above-mentioned fluidity and can be implemented in situated embodied agents that are engaged in a series of language games, i.e. routinised communicative interactions. The demonstration especially focuses on how the syntax-lexicon continuum is represented in FCG and how different templates interact with each other to form a construction. It also shows how FCG can contribute to a more complete theory of language.

Back to schedule - Back to top