The KBS Class Project

KBS-Media Lab
Lund University
[projects] Last update 1995.05.29

Knowledge-Based System for automatic CLASSification of building texts - KBS-CLASS.

Goal/Purpose:

The KBS-CLASS project aims at, with the aid of neural network technology, building a tool for automatic classification of building information that resides in textual form. The tool is meant to be used as an integrated part in a dynamic knowledge net of highly interconnected machines and people, in situations where information is needed quickly and is filtered out of news streams or retrieved from large databases. Its task will be to classify texts so that these may be retrieved via the vocabulary of a classification system, and to classify natural language search queries for searches in databases that already use a classification system. The network is trained with a building products database from AB Svensk Byggtjänst (the Swedish Building Centre) that is indexed with the BSAB classification system. A back-propagation neural network is used for predicting the classification of a given text and different input representations are tested and evaluated. We arrive at a tool that uses trigrams selected on a statistical basis as input to the neural network. Different sets of inputs are evaluated. Finally we outline future directions for further development of the tool and/or similar tools. The overall design and systems engineering has been made by Jörgen Modin while the programming has been done by Peter Andersson. The project has been supervised by Per Christiansson.

Relevance:

There will be a great need for such tools in the future when access to building knowledge in electronic form will be of key importance to an organization's success. A current need for filtering out information about installations from other text in building protocols has been identified in the COOCOM1 project.

Novelty:

To our knowledge an automatic classifier using neural network technology has not been conceived and built previously for the building sector. Furthermore the project points a the need for automated tools, agents, in the electronic and knowledge intense future of the building industry.

Methodology:

Descriptions of a buildings products database from Svensk Byggtjänst have been used as the raw material for the neural network. Different input and output representations have been tested based on statistics and different information retrieval methods. A drag-and drop classifier is now under construction.

Results:

The project has given valuable input to the KBS-CROSS project. A refeered journal article (Modin, 1994) ""KBS-Class: A neural network tool for automatic content recognition of building texts".

Project set-up:

Feedback Navigate Front Page [projects]