Automatically assigned DDC number:

Manually assigned DDC number: 00635

Number of references: 0

Title: Proceedings of ICNN'94, Orlando, Florida

Subject: Proceedings of ICNN'94, Orlando, Florida

Description: ¾¾ We study the internal representations developed by recurrent networks learning to associate semantic role descriptions with simple sentences. Our analysis of the state-space of the hidden units makes explicit the structure of the representations and shows how it reflects the similarities between the sentences. We find that the representation scheme developed by learning is related to a formally defined one. word is input, the network must instantiate the semantic roles corresponding to the part of the sentence which was already presented. This task makes use of the schema-like structure model --- a schema being regarded as a set of bindings --- employed by Smolensky when introducing tensor product representations in [11]. We only consider non-recursive structures here. I. INTRODUCTION To tackle the temporal component of the task in a flexible way, we chose recurrent networks (RNs). To ease analysis, the RNs we use have three layers, with only one --- recurrent --- hi...

Contributor: The Pennsylvania State University CiteSeer Archives

Publisher: unknown

Date: 1998-10-19

Pubyear: unknown

Format: ps

Identifier: http://citeseer.ist.psu.edu/140506.html

Source: http://www.rfai.li.univ-tours.fr/webrfai/chercheurs/crucianu/postscript/icnn94.ps.gz

Language: en

Rights: unrestricted