?
Semantic Text Analysis Using Artificial Neural Networks Based on Neural-Like Elements with Temporal Signal Summation
Text as an image is analyzed in the human visual analyzer. In this case, the image is scanned along the points of the greatest informativity, which are the inflections of the contours of the equitextural areas, into which the image is roughly divided. In the case of text analysis, individual characters of the alphabet are analyzed in this way; but as reading skills are mastered, their groups (words) are perceived as whole objects, and then groups of words and phrases are perceived. Next, the text is analyzed as repetitive language elements of varying complexity. Dictionaries of levelforming elements of varying complexity are formed, the top of which is the level of acceptable compatibility of the root stems of words (names) in sentences of the text, that is, the semantic level. In the case of digitized text, the analysis is greatly simplified due to the absence of necessity to recognize alphabetic characters. The level of semantics represented by pairs of root stems is virtually a homogeneous directed semantic network. Re-ranking the weights of the network vertices corresponding to the root stems of individual names, as occurs in the hippocampus, makes it possible to move from the frequency characteristics of the network to their semantic weights: vertices associated with many other vertices that have large weights increase their weight to the detriment of other vertices. Such networks can be used to analyze texts that represent them: one can compare them with each other, classify and use to identify the most significant parts of texts (generate abstracts of texts), etc. Based on this approach, software technology TextAnalyst was implemented for the semantic analysis of texts. The frequency of occurrence of the root stems of words and pairs of root stems in sentences of the text is analyzed with an artificial neural network based on neurons with temporal summation of signals. The weights of the network vertices are re-ranked using a Hopfield-like iterative algorithm. The resulting technology was used for informational and analytical expert evaluation of texts, ranking of human capital assets parameters, extracting implicit information from texts (using the analysis of texts by V. Nabokov and J. Brodsky as an example), and classifying the results of genetic analysis (based on the analysis of signal genetic networks