Neural Net Learning Issues in Classification of Free Text Documents

In intelligent analysis of large amounts of text, not any single clue indicates reliably that a pattern of interest has been found. When using multiple clues, it is not known how these should be integrated into a decision. In the context of this investigation, we have been using neural nets as parameterized mappings that allow for fusion of higher level clues extracted from free text. By using higher level clues and features, we avoid very large networks. By using the dominant singular values computed by Latent Semantic Indexing (LSI) [Deerwester, et al., 90] and applying neural network algorithms for integrating these values and the outputs from other "sensors", we have obtained preliminary encouraging results with text classification.