ENTERPRISE RELATIONSHIP EXTRACTION METHOD AND DEVICE, AND STORAGE MEDIUM

摘要

Disclosed are an enterprise relationship extraction method and device, and a storage medium. The method comprises: extracting sentences comprising pairs of related enterprise entities from a repository as training sample sentences to establish a sample database; extracting all training sample sentences, each of which comprises a pair of enterprise entities, and performing word segmentation, mapping each word to a word vector x_i, and mapping each sentence to a sentence vector S_i; using LSTM to calculate a first hidden-layer-state vector h_i and a second hidden-layer-state vector h_i' of the word vector x_i, performing splicing to obtain a comprehensive hidden-layer-state vector, and then obtaining a feature vector T_i; substituting the feature vector T_i into an average vector expression to calculate an average vector S; substituting the average vector S and a relation type of the pair of enterprise entities into a softmax classification function to calculate a weight a_i of each training sample sentence; and extracting each sentence comprising two enterprise entities, obtaining a feature vector T_i by means of a bi-LSTM, and inputting the vector into a trained RNN model to predict the relation between the two enterprises, so that labor costs are reduced, and the relation between two enterprise entities can be predicted more accurately.