University of Twente Student Theses


Optimization of a control network for binding in combinatorial sentence structures

Meijer, K. (2019) Optimization of a control network for binding in combinatorial sentence structures.

[img] PDF
Abstract:The ability of human to understand and produce combinatorial structures in language is an important feat of cognition. It allows for the understanding of a huge set of sentences generated by the combination of a known sentence structure and known words. How these combinatorial structures can be instantiated in neural terms is still faced by some challenges. The neural blackboard architecture (NBA) aims to solve these challenges. In this architecture words are bound together in a temporary sentence structure that encodes the relations between the words. The question remains with what network the binding in the NBA can best be controlled. Therefore, the goal of this thesis was twofold. The first being the replication of the FFN trained by van der Velde and de Kamps (2010). The second being finding the best hidden layer size for this FFN. Accuracy as reported by Keras showed a decline for processing of the test sentences for the models with six or fewer hidden nodes. However, a model configuration that completely omitted the hidden layer scored an accuracy equal to the model with 12 hidden nodes. In conclusion it can be said that a model without a hidden layer is not adequate to perform the controlling task for binding in the NBA if the performance requirements of van der Velde and de Kamps (2010) are applied.
Item Type:Essay (Master)
1992, Enschede, Nederland
Faculty:BMS: Behavioural, Management and Social Sciences
Subject:77 psychology
Programme:Psychology MSc (66604)
Link to this item:
Export this item as:BibTeX
HTML Citation
Reference Manager


Repository Staff Only: item control page