Method and device for text-enhanced knowledge graph joint representation learning

1. A method for text-enhanced knowledge graph joint representation learning, at least comprising:

learning a structure vector representation based on entity objects and their relation linking in a knowledge graph and forming structure representation vectors;

discriminating credibility of reliable feature information and building an attention mechanism model, aggregating vectors of different sentences and obtaining association-discriminated text representation vectors; and

building a joint representation learning model, and

using a dynamic parameter-generating strategy to perform joint learning for the text representation vectors and the structure representation vectors based on the joint representation learning model,

wherein:

the reliable feature information is obtained by extraction based on convolutional neural network model learning, in which,

entity-related text description information is represented based on the convolutional neural network model learning, so as to provide semantic enhancement to the structure representation vectors, and

wherein representing the entity-related text description information based on the convolutional neural network model learning comprises:

associating entity-relation linking data in the knowledge graph with texts and automatically labeling entities that have identical names;

performing local semantic combination on word vector sequences in the texts based on context relation of word vector sets in the texts and extracting local semantic features of the sentences; and

training sequence vectors in the texts based on a convolutional neural network model, and optimizing vector semantics of sentence representation through iterative training based on a non-linear activation function,

and

mirroring data of the entities and relations in the knowledge graph to a low-dimension continuous vector space, so that a certain level of computational-sum-based association exists among entity representation vectors with linking relations; and

wherein discriminating credibility of reliable feature information comprises:

acquiring sentence sets with entity pair concurrence and extracting corresponding entity structure representation vectors and/or sentence neighboring word vectors; and

performing vector optimization on entities with specified names based on location feature information and word-order feature information of entity pairs in the sentences, so that a certain level of discrimination exists among representation vectors of the same entity in different said sentences; and

wherein similarity of sentences is computed and a threshold vale is assigned for degree of similarity, and

wherein when the similarity value of the text vector representation corresponding to the candidate set of specified vectors in the knowledge graph is smaller than said threshold value, it is determined that the vectors lack for text information, and are unreliable feature information, and

wherein when the similarity value of the text vector representation corresponding to the candidate set of specified vectors in the knowledge graph is not smaller than said threshold value, it is determined that the vectors do not lack for text information, and are reliable feature information, and

wherein a structure representation vector satisfying said threshold is selected as a text relation representation vector, and

wherein aggregating vectors of different sentences and obtaining association-discriminated text representation vectors further comprises:

selecting and performing information reengineering according to location features of the structure representation vectors in the knowledge graph so as to form the text representation vectors.

2. The method of claim 1 , wherein forming the text representation vectors further comprises:

establishing an attention weight matrix, and

computing attention weight values of the sentences based on structure entity vectors in the knowledge graph and relation representation vectors of the sentences using the attention weight matrix.

3. The method of claim 1 , wherein the method further comprises:

setting up a threshold mechanism to perform comprehensive weight addition on the structure representation vectors and the text representation vectors.

4. The method of claim 3 , wherein the method further comprises:

mining implicit association features among the vectors based on the convolutional neural network model, and

performing knowledge association prediction and/or data intelligence reasoning for missing relation in the knowledge, graph.

5. The method of claim 4 , wherein the dynamic parameter-generating strategy comprises:

a knowledge graph having a more complicated structure has structured vector representation with a greater weight, for knowledge graphs of sparse lands, textual vector representation takes a larger share.

6. The method of claim 3 , wherein the dynamic parameter-generating strategy is expressed as: h=hs·θh+ht·(1−θh), r=rs·θr+rt·(1−θr), t=ts·θt+tt·(1−θt), wherein the values of parameters θh, θr and θt are all at [0,1], which are obtained using a logistic sigmoid function.