Semantic Map

Technische Beschreibung

Semantic MapsKohonen Maps are self organising neural networks that can map highdimensional input onto a planar map. When supplying semantic data, such as texts to a Kohonen Map, the mapping will locate texts that are semantically correlated close to each other. In netzspannung.org the abstracts and/or detailed descriptions of artistic and research projects are fed as input to a Kohonen Map . The Mapping process can be devided into four steps: text vectorization, neural net training, map extraction and visualization. Text vectorisation and net training demand excessive time and memory requirements, so they cannot be performed in real-time. The map extraction step then derives regions, labels and keywords from the neural network thus creating a knowledge map. The knowledge map can be visualized in various ways by the clientside visualisation interface.Text VectorisationFor text vectorisation all abstracts were automatically translated into a common language (english). Word frequencies, a stopword list, and random projection were then applied to select a pool of 1000 relevant words out of all abstracts.Each abstract is related to a wordvector, that assigns a relevance factor to each of the 1000 words. The IDFxTF algorithm assigns high relevance factors to terms that appear more frequently in the corresponding abstract than in others. Net TrainingThe Neural Net consists of a grid holding 10 by 10 neurons, where each neuron is represented by a random vector. The word vectors are provided to the network one by one, stimulating all neurons at the same time. The amount of stimulation is determined by the similarity of input vectors and neuron vectors. The winner-neuron which is stimulated most, is reinforced, as well as its neighbourhood. Both, power of reinforcement and neighbourhood size, decrease over time. Thus in the beginning of the training we can observe the formation of a coarse structure, whereas later on the fine tuning takes place.Map ExtractionWhen the neural net has been trained it implicitly contains the semantic map. By stimulating the net with input vectors the neural correlate of each artistic project is detected. The neurons are also grouped by similarity of their vectors, thus forming regions on the map. Each region is represented by a vector which has a high relevance fector for the term that is used as a region label. The individual projects are also labeled using keywords that are selected from their word vectors.VisualizationThe resulting Kohenen maps are not directly visualized, but rather form an abstract foundation for a variety of different knowledgemap visualisations. The visualisation modules provide different views on the map data and feature different modes of interaction. They are adapted to specific user tasks like orientation (overview), inspection (detailed view) and side by side comparison (split view).

Hardware / Software

Schockwave Player

Lizenz

• frei für nicht-kommerzielle Nutzung

Veröffentlichungen

• Novak, J.; Müller, B.; Schneider, M.; ”The netzspannung.org Knowledge Map System”, GMD, September 2001, to be published as GMD Technical Report

Sekundärliteratur

• Fleischmann, M; Strauss, W.; Novak, J.; Paal, S.; Müller, B.; Blome, G.; Peranovic, P.; Seibert, C.; Schneider, M.; netzspannung.org - an Internet media lab for knowledge discovery in mixed realities, In Proceedings of cast01//Living in Mixed Realities, International Conference on Communication on Art, Science and Technology, Schloss Birlinghoven, Fraunhofer-IMK, September 2001
  » http://netzspannung.…rg/cast01/proceedings [link 04]

Referenzen

• Lin, X.; Soergel, D.; Marchionini, G.; A self-organizing semantic map for information retrieval, in Proc. of 14th ACM/SIGIR Conf. Research and Development in Information Retrieval, 1991
  » http://www.acm.org/dl [link 05]
• Warren Sack, “Conversation Map: An Interface for Very Large-Scale Conversations,” Journal of Management Information Systems (to appear, Winter 2000-2001)