[4Rin1] Interactive Session 2

[4Rin1-08] Evaluation of Holographic Reduced Representations in Question-Answering System using Spiking Neuron Model

〇Keiji Iwao¹, Bisser Raytchev¹, Toru Tamaki¹, Kazufumi Kaneda¹ (1. Univ. of Hiroshima)

Keywords:spiking neuron, question-answering, linguistic representation, brain computing

Symbolic reasoning is an important ability of the brain. Symbolic structures can be represented as distributed representations with binding operations. We aim to extend a conventional model for word or phrase level reasoning to sentence level reasoning and to find the optimal distributed representation for processing structural representations such as languages in a biological brain model under noise. We construct a spiking neuron model which can answer answer questions about a group of sentences. We employed Holographic Reduced Representations (HRRs) to prevent the increasing of vector dimensionality for binding operations. Experimental results show that our model can select an appropriate sentence for question-answering even in similar sentences. Experiments suggest that a noise-robust structural representations for languages can be realized by changing the role vector depending on the depth of the structure and by embedding a verb into a role vector and removing the predicate part.