SYSTEM, METHOD, AND MEDIUM FOR LOW COMPLEXITY AND RICH SEMANTIC AUTOMATIC QUESTION ANSWERING

Given a data processing system, it is an objective of the present invention to reduce semantic modelling computational load for finding an answer to a query. The objective is solved by the system comprising: a storage medium (100) holding data representing a set of vectors (q1', q2', q3', q4', q5') and a set of answer strings (A1, A2, A3, A4, A5) wherein each vector can be pre-calculated by computation-expensive semantic modelling and then mapped (m1, m2, m3, m4, m5) to an answer string, and each vector represents a question; and a shallow similarity neural network (NN) is able to effectively processing the similarity from the discriminative feature that is specifying an input in the form of two vectors (qa', qb'), and then trained to provide similarity (s) as output; wherein the system comprises means for carrying out the steps: a) receive a query (210); b) generate a vector from the query (220); c) iteratively (230) feed the neural network with a pair of vectors including the vector generated from step b) and one vector from the set of vectors; d) based on similarity outputs from the neural network (NN) generate in step c), either reject the query if none of matched vector from the set of of vectors (q1', q2', q3', q4', q5') or find (240) at least one vector as similar vector from the set of vectors (q1', q2', q3', q4', 5'); e) based on the similar vector, use the map (m1, m2, m3, m4, m5) to find a corresponding answer string (260); f) respond to the query using the corresponding answer string (270). Because of the pre-calculated a set of vectors, system is not able to optimally jointly train them together to highlight the discriminative, the effective discriminative feature layer is introduced to optimize the performance, together with shallow similarity neural network that naturally requires small number of training corpus due to small number of parameters of similarity neural networks.