Textual Energy of Associative Memories: Performant Applications of Enertex Algorithm in Text Summarization and Topic Segmentation

Hopfield [1, 2] took as a starting point physical systems like the magnetic Ising model �formalism resulting from statistical physics describing a system composed of units with two possible states named spins) to build a Neural Network �NN) with abilities of learning and recovery of patterns. The capacities and limitations of this Network, called associative memory, were well established in a theoretical frame in several studies [1, 2]: the patterns must be not correlated to obtain free error recovery, the system saturates quickly and only a little fraction of the patterns can be stored correctly. As soon as their number exceeds ≈ 0� 14N , any pattern is recognized. This situation strongly restricts the practical applications of Hopfield Network. However, in NLP, we think that it is possible to exploit this behavior. Vector Space Model �VSM) [3] represents the sentences of a document into vectors. These vectors can be studied as Hopfield NN. With a vocabulary of N terms of a document, it is possible to represent a sentence as a chain of N neurons actives �words are presents) or inactives �words are absents). A document with P sentences is formed of P chains in the vector space Ξ of