Treffer: Search mechanisms using a neural network model: comparison with the vector space model

This paper proposes a model to define an associative IR system. This model is based on the Neural Network theory, it takes into account the results obtained in the more frequently used models including classical search processes as a relevance feedback mechanism. Our model ensures that the search performance are at least the ones obtained in the vector space model. More importantly, our model improves the results due to the usage of associations between the search language terms. The conncectionist features have been introduced, firstly to improve the information modelling: the document indexing terms, the search language are representedin an homogeneous and simple way, maling it possible to take into account semantic and statistic dependencies between the terms. Thus, the query terms spread their activation to the indexing terms. Secondly because of perspectives of the learning capabilities, making possible an adaptation of the information representation to the user's goal.