Construction method, device, computing device, and storage medium for constructing patent knowledge database
The present disclosure discloses construction method, device, computing device, and storage medium for constructing patent knowledge database. The method comprises: obtaining patent resource data; analyzing the patent resource data to obtain explicit information and implicit information, the explicit information comprises attribute information of each entity in a preset entity set, the implicit information comprises an entity relationship of technical elements; fusing the implicit information to obtain fused implicit information; and constructing the patent knowledge database according to the explicit information and the fused implicit information.
1. A method for constructing a patent knowledge database, comprising:
obtaining patent resource data,
analyzing the patent resource data to obtain explicit information and implicit information,
fusing the implicit information to filter out noise data and obtain fused implicit information, and
constructing the patent knowledge database according to the explicit information and the fused implicit information, wherein:
the explicit information comprises attribute information of each entity,
the implicit information comprises various entity relationships of technical elements,
the patent resource data comprises a patent specification and one or more drawings corresponding to the patent specification,
analyzing the patent resource data to obtain the implicit information comprises processing the patent specification and the one or more drawings to obtain the various entity relationships of the technical elements through a specification-drawing matching analysis method,
processing the patent specification and the one or more drawings through the specification-drawing matching analysis method comprises:
extracting basic feature information of the one or more drawings through a feature extraction convolutional network in a deep neural network model,
processing the basic feature information through a specification detection convolutional network in the deep neural network model to detect reference numbers of the technical elements and position information of the reference numbers from the one or more drawings,
extracting description information corresponding to each of the reference numbers from the patent specification,
determining a directional relationship between the reference numbers according to the position information of the reference numbers, and
matching the reference numbers, the directional relationship between the reference numbers, and the description information corresponding to the reference numbers to obtain the various entity relationships of the technical elements,
fusing the implicit information comprises:
generating corresponding feature vectors according to the various entity relationships of the technical elements comprised in the implicit information,
using a linear classifier to judge the corresponding feature vectors of the various entity relationships of the technical elements according to the corresponding feature vectors,
calculating a probability that the various entity relationships of the technical elements are real relationships, and
filtering the various entity relationships of the technical elements whose probability are less than a set threshold, and
a loss function of the deep neural network model is:
L
(
x
,
c
,
1
,
g
)
=
1
N
(
L
conf
(
x
,
c
)
+
a
L
l
o
c
(
x
,
l
,
g
)
)
wherein, x represents a matching matrix of the reference numbers, c represents a confidence of a detection, 1 represents a predicted reference number position, g represents a real reference number position, N represents a number of default bounding boxes that matches real bounding boxes, L conf (x,c) represents a loss of the confidence, L loc (x, l, g) represents a loss of the position information, and a represents a balance factor.
2. The method for constructing the patent knowledge database according to claim 1 , wherein analyzing the patent resource data to obtain the explicit information comprises extracting the explicit information from structural information from the patent specification through a named entity recognition algorithm.
3. The method for constructing the patent knowledge database according to claim 1 , wherein processing the patent specification and the one or more drawings through the specification-drawing matching analysis method further comprises:
matching the reference numbers, the position information of the reference numbers, and the description information corresponding to the reference numbers.
4. The method for constructing the patent knowledge database according to claim 1 , wherein extracting the description information corresponding to each of the reference numbers from the patent specification comprises:
extracting named entities from the patent specification through a supervised learning algorithm based on a bi-LSTM (Long Short Term Memory) network model, and
associating the named entities and the reference numbers to obtain the description information corresponding to each of the reference numbers.
5. The method for constructing the patent knowledge database according to claim 1 , wherein matching the reference numbers, the directional relationship between the reference numbers, and the description information corresponding to the reference numbers comprises:
matching the description information corresponding to the reference numbers to the reference numbers in the one or more drawings to obtain a directional relationship between various description information, and
generating a relationship graph according to the directional relationship among the various description information and storing the relationship graph.
6. The method for constructing the patent knowledge database according to claim 1 , wherein:
the implicit information further comprises an entity relationship of technical terms, and
analyzing the patent resource data to obtain the implicit information comprises processing the patent specification to obtain the entity relationship of the technical terms by a reasoning algorithm based on specification analysis.
7. The method for constructing the patent knowledge database according to claim 6 , wherein processing the patent specification by the reasoning algorithm based on the specification analysis comprises:
processing the patent specification through at least one of a term frequency statistical algorithm, a co-word analysis algorithm, or a term embedding analysis algorithm.
8. A device for constructing a patent knowledge database, comprising:
a resource acquisition module,
a knowledge extraction module,
a fusion processing module, and
a construction module, wherein:
the resource acquisition module is used to acquire patent resource data,
the knowledge extraction module is used to analyze the patent resource data to obtain explicit information and implicit information,
the explicit information comprises attribute information of each entity,
the implicit information comprises various entity relationships of technical elements,
the fusion processing module is used to fuse the implicit information to filter out noise data and obtain fused implicit information,
the construction module is used to construct the patent knowledge database to obtain the patent resource data according to the explicit information and the fused implicit information,
the patent resource data comprises a patent specification and one or more drawings corresponding to the patent specification,
analyzing the patent resource data to obtain the implicit information comprises processing the patent specification and the one or more drawings to obtain the various entity relationships of the technical elements through a specification-drawing matching analysis method,
processing the patent specification and the one or more drawings through the specification-drawing matching analysis method comprises:
extracting basic feature information of the one or more drawings through a feature extraction convolutional network in a deep neural network model,
processing the basic feature information through a specification detection convolutional network in the deep neural network model to detect reference numbers of the technical elements and position information of the reference numbers from the one or more drawings,
extracting description information corresponding to each of the reference numbers from the patent specification,
determining a directional relationship between the reference numbers according to the position information of the reference numbers, and
matching the reference numbers, the directional relationship between the reference numbers, and the description information corresponding to the reference numbers to obtain the various entity relationships of the technical elements,
fusing the implicit information comprises:
generating corresponding feature vectors according to the various entity relationships of the technical elements comprised in the implicit information,
using a linear classifier to judge the corresponding feature vectors of the various entity relationships of the technical elements according to the feature vectors,
calculating a probability that the various entity relationships of the technical elements are real relationships, and
filtering the various entity relationships of the technical elements whose probability are less than a set threshold, and
a loss function of the deep neural network model is:
L
(
x
,
c
,
1
,
g
)
=
1
N
(
L
conf
(
x
,
c
)
+
a
L
l
o
c
(
x
,
l
,
g
)
)
wherein, x represents a matching matrix of the reference numbers, c represents a confidence of a detection, 1 represents a predicted reference number position, g represents a real reference number position, N represents a number of default bounding boxes that matches real bounding boxes, L conf (x,c) represents a loss of the confidence, L loc (x, l, g) represents a loss of the position information, and a represents a balance factor.
9. A computing device, comprising:
a memory storing a computer program, and
a processor, wherein:
the processor executes the computer program to implement a method for constructing a patent knowledge database, comprising:
obtaining patent resource data,
analyzing the patent resource data to obtain explicit information and implicit information,
fusing the implicit information to filter out noise data and obtain fused implicit information, and
constructing the patent knowledge database according to the explicit information and the fused implicit information, wherein:
the explicit information comprises attribute information of each entity,
the implicit information comprises various entity relationships of technical elements,
the patent resource data comprises a patent specification and one or more drawings corresponding to the patent specification,
analyzing the patent resource data to obtain the implicit information comprises processing the patent specification and the one or more drawings to obtain the various entity relationships of the technical elements through a specification-drawing matching analysis method,
processing the patent specification and the one or more drawings through the specification-drawing matching analysis method comprises:
extracting basic feature information of the one or more drawings through a feature extraction convolutional network in a deep neural network model,
processing the basic feature information through a specification detection convolutional network in the deep neural network model to detect reference numbers of the technical elements and position information of the reference numbers from the one or more drawings,
extracting description information corresponding to each of the reference numbers from the patent specification,
determining a directional relationship between the reference numbers according to the position information of the reference numbers, and
matching the reference numbers, the directional relationship between the reference numbers, and the description information corresponding to the reference numbers to obtain the various entity relationships of the technical elements,
fusing the implicit information comprises:
generating corresponding feature vectors according to the various entity relationships of the technical elements comprised in the implicit information,
using a linear classifier to judge the corresponding feature vectors of the various entity relationships of the technical elements according to the corresponding feature vectors,
calculating a probability that the various entity relationships of the technical elements are real relationships, and
filtering the various entity relationships of the technical elements whose probability are less than a set threshold, and
a loss function of the deep neural network model is:
L
(
x
,
c
,
1
,
g
)
=
1
N
(
L
conf
(
x
,
c
)
+
a
L
l
o
c
(
x
,
l
,
g
)
)
wherein, x represents a matching matrix of the reference numbers, C represents a confidence of a detection, 1 represents a predicted reference number position, g represents a real reference number position, N represents a number of default bounding boxes that matches real bounding boxes, L conf (x,c) represents a loss of the confidence, L loc (x, l, g) represents a loss of the position information, and a represents a balance factor.
10. A non-transitory computer-readable storage medium, wherein:
a computer program is stored on the non-transitory computer-readable storage medium, and when the computer program is executed by a processor, the computer program implements a method for constructing a patent knowledge database, comprising:
obtaining patent resource data,
analyzing the patent resource data to obtain explicit information and implicit information,
fusing the implicit information to filter out noise data and obtain fused implicit information, and
constructing the patent knowledge database according to the explicit information and the fused implicit information, wherein:
the explicit information comprises attribute information of each entity,
the implicit information comprises various entity relationships of technical elements,
the patent resource data comprises a patent specification and one or more drawings corresponding to the patent specification,
analyzing the patent resource data to obtain the implicit information comprises processing the patent specification and the one or more drawings to obtain the various entity relationships of the technical elements through a specification-drawing matching analysis method,
processing the patent specification and the one or more drawings through the specification-drawing matching analysis method comprises:
extracting basic feature information of the one or more drawings through a feature extraction convolutional network in a deep neural network model,
processing the basic feature information through a specification detection convolutional network in the deep neural network model to detect reference numbers of the technical elements and position information of the reference numbers from the one or more drawings,
extracting description information corresponding to each of the reference numbers from the patent specification,
determining a directional relationship between the reference numbers according to the position information of the reference numbers, and
matching the reference numbers, the directional relationship between the reference numbers, and the description information corresponding to the reference numbers to obtain the various entity relationships of the technical elements,
fusing the implicit information comprises:
generating corresponding feature vectors according to the various entity relationships of the technical elements comprised in the implicit information,
using a linear classifier to judge the corresponding feature vectors of the various entity relationships of the technical elements according to the corresponding feature vectors,
calculating a probability that the various entity relationships of the technical elements are real relationships, and
filtering the various entity relationships of the technical elements whose probability are less than a set threshold, and
a loss function of the deep neural network model is:
L
(
x
,
c
,
1
,
g
)
=
1
N
(
L
conf
(
x
,
c
)
+
a
L
l
o
c
(
x
,
l
,
g
)
)
wherein, x represents a matching matrix of the reference numbers, c represents a confidence of a detection, 1 represents a predicted reference number position, g represents a real reference number position, N represents a number of default bounding boxes that matches real bounding boxes, L conf (x,c) represents a loss of the confidence, L loc (x, l, g) represents a loss of the position information, and a represents a balance factor.
11. The method for constructing the patent knowledge database according to claim 1 , wherein:
the deep neural network model inherits a VGG-16 network structure of a convolutional neural network, retains a network from conv1_1 to conv5_3 of the VGG-16 network structure, and converts a last two layers of fully connected network into convolutional networks, namely conv6 and conv7, and four convolutional networks, conv8 to conv11, are added to the convolutional network conv7 to define the feature extraction convolutional network; and
the specification detection convolutional network has a 6-layer convolutional network structure, an input of the 6-layer convolutional network structure is basic feature information of an output of the feature extraction convolutional network, and an output of the specification detection convolutional network undergoes a non-maximum compression to generate final reference numbers and position information of the final reference numbers.
12. The method for constructing the patent knowledge database according to claim 1 , wherein:
each of the corresponding feature vectors is a two-dimensional digital feature vector,
a first dimension of the two-dimensional digital feature vector represents a square root of a number of the patent resource data configured to obtain the various entity relationships of the technical elements, and
a second dimension of the two-dimensional digital feature vector represents an average value of scores of the various entity relationships of the technical elements by a scoring method.
13. The method for constructing the patent knowledge database according to claim 4 , wherein:
the patent specification and corresponding reference symbol of the patent specification is two sequences, and
past features and future features of the two sequences are simultaneously modeled by a forward network and a backward network of LSTM (Long Short Term Memory),
a sliding window having a length of n is disposed respectively in the two sequences,
the forward network and the backward network are used to describe characteristics of the sliding window, and
a CRF (conditional random field) method is used to mark words in the sliding window.