CN111160564B - A Chinese Knowledge Graph Representation Learning Method Based on Feature Tensor - Google Patents

Abstract

The present invention provides a Chinese knowledge map representation learning method based on feature tensor, the method includes: data preparation, establishment of data structure, construction of entity feature vector matrix, definition of relational vector and distance formula of marked triplets, and obtaining training set 1. Train the knowledge graph representation learning model, update the model parameters, iteratively train, and use the model to predict the relationship between unlabeled triples, and iteratively train again until no new unlabeled triples can be learned. The present invention proposes to use Chinese pinyin, word information, word information, and description information to form feature tensors and convert them into feature vectors to replace the method of randomly initializing entity vectors in traditional knowledge representation learning, making full use of Chinese characteristics. In addition, the training corpus is supplemented by a two-layer iterative method, so that the relationship matrix can be continuously revised, and the accuracy and convergence speed of the knowledge graph representation learning model can be improved.

Description

A Chinese knowledge graph representation learning method based on feature tensor

Technical Field

The present invention relates to the field of knowledge graphs, and in particular to a Chinese knowledge graph representation learning method based on feature tensors.

Background Art

Knowledge graphs describe the complex relationships between concepts and entities in the objective world in a structured form, providing a better way to organize, manage and understand the massive amount of information on the Internet. Knowledge graph technology usually includes three aspects of research: knowledge representation, knowledge graph construction and knowledge graph application. Among them, knowledge representation is the basis for knowledge graph construction and application, reflecting human cognition of the objective world and being able to express the semantics presented by the objective world from different levels and granularities. First, we need to understand how humans themselves represent knowledge and use them to solve problems, and then formalize it into an expression form that computers can reason and calculate, establish a knowledge-based system, and provide intelligent knowledge services. At the same time, knowledge representation also needs to combine the computer's ability to represent, process and calculate symbols. The key issues that need to be solved in knowledge representation are: 1) what kind of knowledge representation form can accurately reflect the knowledge of the objective world; 2) what kind of knowledge representation can have semantic representation capabilities; 3) how knowledge representation can support efficient knowledge reasoning and calculation, so that knowledge representation has the ability to reason about new knowledge. Current knowledge representation methods can be divided into knowledge representation based on symbolic logic, open knowledge representation methods of Internet resources, and representation learning based on knowledge graphs.

1) Knowledge representation based on symbolic logic: mainly includes logical representation, production representation and framework representation. Although knowledge representation technology based on symbolic logic can well describe logical reasoning, the ability of machines to generate rules in reasoning is very weak, the acquisition of reasoning rules requires a lot of manpower, and the quality of data is relatively high. In the current era of large-scale data, knowledge representation based on symbolic logic can no longer solve the problem of knowledge representation well.

2) Knowledge representation of Web content: Tim Berners-Lee proposed the concept of the semantic web, in which all web content should have a definite meaning and can be easily understood, acquired and integrated by computers. Web content knowledge representation includes semi-structured tag-based markup language XML, RDF-based Web resource semantic metadata description framework, and description logic-based OWL ontology description language, as well as triple-based knowledge graph knowledge representation methods that are currently widely used in the industry. A triple can be represented as <h, r, t>, indicating that there is a relationship r between the head entity h and the tail entity t. These technologies enable us to publish semantic information that can be understood and processed by machines on the Web. However, the content of the Web is in the hundreds of trillions, which is a huge challenge for knowledge storage and knowledge representation learning.

3) Representation learning: The goal of representation learning is to represent the semantic information of the research object as a dense low-dimensional vector through machine learning or deep learning. Implicit vectorization is performed on knowledge units of different granularities to support the rapid calculation of knowledge in a big data environment. Representation learning mainly includes tensor reconstruction and potential energy function methods. Tensor reconstruction integrates the information of the entire knowledge base, but in a big data environment, the tensor dimension is very high and the reconstruction calculation is large; the potential energy function method believes that the relationship is a translation operation from the head entity to the tail entity. The TransE model proposed by Bordes et al. is a representative of the translation model, but it lacks explicit semantic information, especially for Chinese, which has phonetic, structural, and word information. The low-dimensional Chinese vector obtained through machine learning or deep learning is only a computer parameter fitting and lacks interpretability.

In summary, knowledge representation based on symbolic logic and open knowledge representation methods based on Internet resources give knowledge an explicit semantic definition, but there is a data sparsity problem, making it difficult to realize large-scale knowledge graph applications; knowledge representation based on deep learning can map knowledge units (entities, relationships, and rules) to a low-dimensional continuous real number space representation, but lacks an explicit semantic definition.

In addition, there are many studies on learning knowledge graph representation abroad, but they are limited to English knowledge graphs. Due to language differences, English words only have simple string information and phrase information. When learning knowledge representation, only random initialization vectors are needed. Chinese contains rich semantic information. Existing research methods cannot achieve good results on Chinese knowledge graphs. At present, China is still mainly at the stage of how to build knowledge graphs, and lacks research on learning Chinese knowledge graph representation.

Summary of the invention

In response to the above problems, the present invention proposes a Chinese knowledge graph representation learning method based on feature tensor. Compared with random initialization vectors, the present invention introduces four features of Chinese pinyin, characters, words, and description information as Chinese display semantic information to constitute the feature tensor, which makes the learning process of Chinese knowledge graph representation explainable. It is combined with deep learning to map the learned knowledge representation to a low-dimensional continuous real number space, which facilitates the learning of Chinese knowledge and the relationships between them.

The present invention proposes a Chinese knowledge graph representation learning method based on feature tensor, which includes the following steps:

Step 1) Data preparation

The data from an open Chinese link dataset zhishi.me is used to form triple data, wherein the triple data consists of a large number of triples, and the triple is in the form of <h, r, t>, where h represents the head entity, t represents the tail entity, and r represents the relationship between the head entity h and the tail entity t;

Step 2) Create data structure

The triple data are divided into labeled triples and unlabeled triples, and data structures of a dictionary, an entity dictionary, a relationship dictionary, an entity pinyin matrix, a character embedding matrix, a word embedding matrix and a description matrix are constructed;

Step 3) Construct entity feature vector matrix

For each entity in the tag triple, firstly, the entity's feature tensor is constructed by the entity's pinyin vector, character vector, word vector and description vector; then the feature tensors of all entities in the tag triple are converted into the entity's feature vector, and the entity feature vector matrix is constructed in the order of the entity dictionary;

Step 4) Take a labeled triple T _l = <h, r, t>, and obtain the feature vectors h _ft and t _ft of the head entity h and the tail entity t through the entity feature vector matrix. In order to indicate that there is a relationship r between entity h and entity t, that is, h+r=t, the relationship vector of the labeled triple T _l = <h, r, t> can be expressed as:

r = t _ft -h _ft

In order to calculate the distance between entity h and entity t, the relationship between entities is represented by vector transformation, and the distance formula of the triple <h, r, t> is defined using Euclidean distance:

The subscript "2" indicates the 2-norm, i.e. the Euclidean norm, and the superscript "2" indicates square.

Step 5) Use the labeled triples as training sets, and initialize entity vectors, i.e., entity feature vector matrices, initialize relationship vectors, and construct relationship vector matrices. The order is consistent with the relationship dictionary. The relationship calculation is obtained by the formula r = t _ft -h _ft . If multiple entity pairs have the same relationship, the relationship vector is the average of the differences between the vectors of multiple entity pairs. All relationship vectors are normalized after initialization to improve accuracy and enhance convergence.

所述知识图谱表示学习模型的损失函数定义为：Step 6) Randomly select a positive triplet <h, r, t> in the training set, select <h′, r, t> and <h, r, t′> from the negative triplet, and form a tuple pair with <h, r, t> to form a training batch T _batch = [(<h, r, t>, <h′, r, t>), (<h, r, t>, <h, r, t′>)], use S _p = {<h, r, t>} to represent the positive triplet, S _f = {<h′, r, t>|h′∈E, <h, r, t′>|t′∈E} to represent the negative triplet, use T _batch as the input of the knowledge graph representation learning model, and perform knowledge graph representation learning training on T _batch , where E represents the entity set, combined with the formula

The loss function of the knowledge graph representation learning model is defined as:

Where γ is a distance parameter greater than 0, which is a hyperparameter. [x] ₊ represents a positive function, that is, when x>0, [x] ₊ =x; when x≤0, [x] ₊ =0. This training method is called margin-based ranking criterion, and its purpose is to separate positive triplets and negative triplets as much as possible and find the support vector with the maximum distance.

Step 7) Use stochastic gradient descent (SGD) to update the parameters of the knowledge graph representation learning model. The gradient update only needs to calculate the distance d(h+r, t) and d(h′+r, t′). Assume there are a total of |E| entities and |R| relationships, each entity vector is m-dimensional, and the relationship vector is n-dimensional, then a total of (|E|*m+|R|*n) parameters need to be updated;

Step 8) Repeat step 6)-step 7) to perform iterative training. After the iterative training is completed, the knowledge graph is used to represent the learning model parameters to predict the relationship between unlabeled triples. The prediction step is: randomly select a triple <h, r, t> _unlabel from the unlabeled triples, and use the knowledge graph to represent the learning model parameters to predict the relationship r′ between h and t. If r′=r, the prediction is correct; at the same time, the correctly predicted triples are used as positive triples, and the head entities or tail entities of these positive triples are randomly replaced as negative triples. Then, these new positive triples and negative triples are combined with the original labeled triples to form new labeled triples;

Step 9) Repeat steps 4)-8) for iterative training using new labeled triples until no new unlabeled triples can be learned, indicating that the knowledge graph representation learning model can no longer learn more Chinese knowledge features from the current training data. The entity vectors and relationship vectors output by the knowledge graph representation learning model are the best Chinese knowledge graph representation corresponding to the Chinese link dataset zhishi.me, and the training is completed.

In view of the problem that the existing knowledge representation learning method cannot combine Chinese word information, the present invention proposes to use Chinese pinyin, character information, word information, and description information to form a feature tensor, and convert it into a feature vector to replace the method of randomly initializing entity vectors in traditional knowledge representation learning, making full use of the characteristics of Chinese. In addition, the present invention adopts a double-layer iteration method to supplement the training corpus, so that the relationship matrix can be continuously corrected, improving the accuracy and convergence speed of the knowledge graph representation learning model.

BRIEF DESCRIPTION OF THE DRAWINGS

FIG. 1 is a framework diagram of the method processing process of the present invention

FIG. 2 is a flow chart of the method processing process of the present invention

Figure 3 is the description matrix LSTM encoding entity description vector

Figure 4 is a schematic diagram of converting a tensor into a vector

DETAILED DESCRIPTION

The technical solutions in the embodiments of the present invention will be described clearly and completely below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments.

As shown in FIG2 , a Chinese knowledge graph representation learning method based on feature tensor proposed in the present invention includes the following steps:

Step 1) Data preparation

The triple data used in the present invention comes from an open Chinese link data set zhishi.me, which consists of a large number of triples in the form of <h, r, t>, where h represents the head entity, t represents the tail entity, and r represents the relationship between the head entity h and the tail entity t.

Step 2) Create data structure

As shown in FIG1 , the triple data are divided into labeled triples and unlabeled triples, and data structures such as dictionaries, entity dictionaries, relationship dictionaries, entity pinyin matrices, character embedding matrices, word embedding matrices, and description matrices are constructed.

Marking triples: Randomly extract triples from the data set zhishi.me to obtain a triple data set, take all triples in the triple data set as positive triples, remove the head entity or tail entity of each positive triple, and randomly select an entity different from itself in the entity dictionary to replace it, forming a negative triple. Only one entity in the triple is replaced each time, so that there is a comparison. Mark the above triples, mark the positive triples as 1, and the negative triples as 0.

Unlabeled triples: any unlabeled triples in the dataset zhishi.me.

Dictionary: All the characters appearing in the dataset zhishi.me, including all head entities, tail entities and relations. The dictionary format is "character: serial number", where the serial number is a number and increases from zero.

Entity dictionary: The entity set in the dataset zhishi.me, including the dictionary consisting of all head entities and tail entities. The dictionary format is "entity name: serial number", where the serial number is a number and increases from zero.

Relation dictionary: a dictionary consisting of the relationship sets in the dataset zhishi.me. The dictionary format is "relation name: sequence number". The sequence number is a number, starting from zero and increasing.

Entity Pinyin Matrix: To solve the problem of different meanings of polyphones, call the Baidu Translate API to get the entity Pinyin and construct an entity Pinyin matrix. The number of rows is consistent with the number of entities in the entity dictionary, and each row is an entity Pinyin vector obtained using one-hot encoding.

Word embedding matrix: The number of rows is the same as the number of words in the dictionary, and each row is a word vector obtained using word2vec.

Word embedding matrix: The number of rows is consistent with the number of entities in the entity dictionary, and each row is a word vector obtained using word2vec.

Description matrix: The number of rows is consistent with the number of entities in the entity dictionary. The Baidu Encyclopedia API is called to obtain entity description information, and the description information is input into the bidirectional long short-term memory network (Bi-directional Long Short-Term Memory, BiLSTM) to encode the entity description vector, as shown in Figure 3. This vector introduces the description information of the entity and can solve the problem of Chinese synonyms.

Step 3) Construct entity feature vector matrix

For each entity in a tag triple, the entity's feature tensor is first constructed by the entity's pinyin vector, character vector, word vector, and description vector, which is used as the predefined entity's feature tensor in subsequent steps. The construction process is as follows: the tag triple is represented as T _l , E is used to represent the entity set in the knowledge graph, and R is used to represent the relationship set. Take any entity e∈E in a tag triple, search the entity's pinyin matrix to obtain the entity's pinyin vector _ep ; let the entity name be c ₁ c ₂ ...c _m , where c _m represents the mth character that makes up the entity name, and search the character embedding matrix based on the character to obtain the entity's character vector e _c =c ₁ c ₂ ...c _m ; search the word embedding matrix to obtain the word vector e _w ; search the description matrix to obtain the entity's description vector _ed , and the entity's feature tensor FeatureTens or is represented as

The feature tensor of the entity is converted into the feature vector of the entity, and the entity feature vector matrix is constructed. As shown in Figure 4, the different dimensions of the feature tensor of the entity are connected, and the connection method is vector splicing. For example _, given vectors A = [x ₁ , x ₂ , x ₃ , ..., x _m ] and B = [y ₁ , y _{2, y 3, ..., yn], the connection is obtained to obtain vector C = [x 1, x 2 , x 3 ,} ..., x _m , y ₁ , y ₂ , y ₃ , ..., _yn ], and dropout is used to randomly lose the vector to prevent overfitting of knowledge representation learning. In this way, the pinyin vector _ep , character vector _ec , word vector _ew , and description vector _ed of the same entity e are connected to obtain the entity's feature vector e _ft = [ _ep ; e _c ; e _w ; _ed ].

Convert the feature tensors of all entities in the tag triplet into feature vectors of the entity, and build the entity feature vector matrix in the order of the entity dictionary.

Step 4) Take a labeled triple T _l = <h, r, t>, and obtain the feature vectors h _ft and t _ft of the head entity h and the tail entity t through the entity feature vector matrix. In order to indicate that there is a relationship r between entity h and entity t, that is, h+r=t, the relationship vector of the labeled triple T _l = <h, r, t> can be expressed as:

r＝t _ft -h _ft (1)

In order to calculate the distance between entity h and entity t, the relationship between entities is represented by vector transformation, and the distance formula of the triple <h, r, t> is defined using Euclidean distance:

The subscript "2" in formula (2) represents the 2-norm, i.e., the Euclidean norm, and the superscript "2" represents square.

Step 5) Use the labeled triples as training sets, initialize entity vectors, i.e., entity feature vector matrices, initialize relationship vectors, and construct relationship vector matrices. The order is consistent with the relationship dictionary, and the relationship calculation is obtained by formula (1). If there are multiple entity pairs with the same relationship, the relationship vector is the average of the differences between the vectors of multiple entity pairs. After all vectors are initialized, they should be normalized to increase accuracy and enhance convergence.

Step 6) Randomly select a positive triplet <h, r, t> in the training set, select <h′, r, t> and <h, r, t′> from the negative triplet, and form a tuple pair with <h, r, t> to form a training batch T _batch = [(<h, r, t>, <h′, r, t′>), (<h, r, t>, <h, r, t′>)]. Use _Sp = {<h, r, t>} to represent the positive triplet, and _Sf = {<h′, r, t>|h′∈E, <h, r, t′>|t′∈E} to represent the negative triplet. Take T _batch as the input of the knowledge graph representation learning model, perform knowledge graph representation learning training on T _batch , and combine formula (2), the loss function of the knowledge graph representation learning model is defined as:

Where γ is a distance parameter greater than 0, which is a hyperparameter, and [x] ₊ represents a positive function, that is, when x>0, [x] ₊ = x; when x≤0, [x] ₊ = 0. This training method is called margin-based ranking criterion, and its purpose is to separate positive triplets and negative triplets as much as possible and find the support vector with the maximum distance.

Step 7) Use stochastic gradient descent (SGD) to update the parameters of the knowledge graph representation learning model. Gradient update only needs to calculate the distance d(h+r, t) and d(h′+r, t′). Assume there are a total of |E| entities and |R| relationships, each entity vector is m-dimensional, and the relationship vector is n-dimensional, then a total of (|E|*m+|R|*n) parameters need to be updated.

Step 8) Repeat steps 6)-7) for iterative training. After the iterative training is completed, the knowledge graph representation learning model is used to predict the relationship between unlabeled triples. The prediction step is to randomly select a triple <h, r, t> _unlabel from the unlabeled triples, and use the knowledge graph representation learning model to predict the relationship r′ between h and t. If r′＝r, the prediction is correct. At the same time, the correctly predicted triples are used as positive triples, and the head entities or tail entities of these positive triples are randomly replaced as negative triples. Then, these new positive triples and negative triples are combined with the original labeled triples to form new labeled triples.

Step 9) Repeat steps 4)-8) for iterative training using new labeled triples until no new unlabeled triples can be learned, indicating that the knowledge graph representation learning model can no longer learn more Chinese knowledge features from the current training data. The entity vectors and relationship vectors output by the knowledge graph representation learning model are the best Chinese knowledge graph representation corresponding to the data set zhishi.me, and the training is completed.

Chinese knowledge graph representation learning methods usually use link prediction as an evaluation task. The evaluation indicators include mean rank (MR), mean reciprocal rank (MRR), the percentage of correct entities in the top ten results (Hits@10), the percentage of correct entities in the top three results (Hits@3), and the percentage of correct entities in the first-ranked results (Hits@1). The smaller the MR value, the better, and the larger the MRR, Hits@10, Hits@3, and Hits@1 values, the better. Randomly remove some entities or relations from the triples in the zhishi.me dataset, and link prediction is to predict the randomly removed entities or relations in the triples.

In an embodiment of the present invention, representation learning is performed on the open source Chinese dataset zhishi.me and link prediction task evaluation is performed, and the test results of two knowledge graph representation learning methods, the TransE model and the TransR model, are compared. The results are shown in Table 1:

Table 1 Test results

Evaluation indicators MR MRR Hits@10 Hits@3 Hits@1 TransE 713 0.458 0.812 0.723 0.556 TranR 687 0.519 0.839 0.768 0.646 The present invention 611 0.843 0.875 0.801 0.692

The experimental results show that the test results of the present invention are better than those of the TransE model and the TransR model, reaching a usable level.

Although the above is a description of the illustrative specific embodiments of the present invention to facilitate the understanding of the present invention by those skilled in the art, it should be clear that the present invention is not limited to the scope of the specific embodiments. Where equivalent substitution or equivalent substitution is adopted, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims (4)

1. The Chinese knowledge graph representation learning method based on the feature tensor is characterized by comprising the following steps of:

step 1) data preparation

Forming data from an open Chinese linked data set zhishi. Me into triplet data, wherein the triplet data is formed by a plurality of triples, the triples are shown as < h, r, t >, h represents a head entity, t represents a tail entity, and r represents a relation between the head entity h and the tail entity t;

step 2) establishing a data structure

Dividing the triplet data into marked triples and unmarked triples, and constructing a dictionary, an entity dictionary, a relation dictionary, an entity pinyin matrix, a word embedding matrix and a data structure of a description matrix, wherein,

marking the triplet: randomly extracting triplet data from the Chinese linked data set zhishi. Me to obtain a triplet data set, taking all triples in the triplet data set as positive triples, removing head entities or tail entities of each positive triplet, randomly selecting an entity different from the entity dictionary to replace the entity dictionary to form negative triples, and only replacing one entity in the triples each time, so that the triples are marked with contrast, and marking the positive triples as 1 and the negative triples as 0;

unlabeled triples: any unlabeled triples in the chinese linked dataset zhishi;

dictionary: all words appearing in the Chinese linked dataset zhishi. Me, including all head entities, tail entities and dictionaries of relationships, are in the form of "words: a serial number, wherein the serial number is a number, and the serial number is increased from zero;

entity dictionary: the entity set in the Chinese link data set zhishi. Me is represented by E, and comprises a dictionary formed by all head entities and tail entities, wherein the dictionary form is 'entity name': a serial number, wherein the serial number is a number, and the serial number is increased from zero;

relationship dictionary: the dictionary is formed by relation sets in the Chinese link data set zhishi. Me, and the dictionary form is' relation name: a serial number, wherein the serial number is a number, and the serial number is increased from zero;

entity pinyin matrix: in order to solve the problem of different meanings of polyphones, calling hundred-degree translation API to obtain entity pinyin, and constructing an entity pinyin matrix, wherein the number of rows of the entity pinyin matrix is consistent with the number of entities in the entity dictionary, and each action of the entity pinyin matrix uses an entity pinyin vector obtained by a one-hot coding mode;

word embedding matrix: the number of lines of the word embedding matrix is consistent with the number of words in the dictionary, and word2vec is used for each line of the word embedding matrix to obtain a word vector;

word embedding matrix: the number of lines of the word embedding matrix is consistent with the number of entities in the entity dictionary, and word vectors obtained by word2vec are used for each action of the word embedding matrix;

describing a matrix: the number of lines of the description matrix is consistent with the number of entities in the entity dictionary, an hundred-degree encyclopedia API is called to acquire entity description information, the entity description information is input into a two-way long-short-term memory network for coding to acquire an entity description vector, and the entity description vector introduces the entity description information, so that the problem of Chinese synonyms can be solved;

step 3) constructing entity feature vector matrix

For the entity in each marked triplet, firstly, an entity pinyin vector, a word vector and an entity description vector form a feature tensor of the entity; converting feature tensors of all entities in the marked triplet into feature vectors of the entities, and constructing an entity feature vector matrix according to the sequence of the entity dictionary;

step 4) taking a marked triplet T _l = < h, r, t >, and obtaining the eigenvectors h of the head entity h and the tail entity t through the entity eigenvector matrix _ft And t _ft To indicate that there is a relation r between entity h and entity T, i.e. h+r=t, the triplet T is marked _l The relationship vector for = < h, r, t > can be expressed as:

r＝t _ft -h _ft

in order to calculate the distance between the entity h and the entity t, the relation between the entities is represented by vector conversion, and the distance formula of the Euclidean distance definition triplet < h, r, t > is as follows:

wherein the subscript "2" denotes the 2 norm, i.e. euclidean norm, and the superscript "2" denotes squaring;

step 5) taking all marked triples as training sets, initializing entity vectors, namely entity feature vector matrixes, initializing relation vectors, constructing a relation vector matrix, wherein the sequence of the relation vector matrix is consistent with that of the relation dictionary, and calculating the relation by the formula r=t _ft -h _ft If the plurality of entity pairs have the same relation, the relation vector is obtained by averaging the vector difference values of the plurality of entity pairs, and normalization is carried out after all the relation vectors are initialized, so that the precision is improved, and convergence is enhanced;

step 6) randomly selecting a positive triplet < h, r, t > in the training set, selecting < h ', r, t > and < h, r, t' in the negative triplet, and forming a tuple pair with < h, r, t > to form a training batch

T _batch ＝[(<h，r，t>，<h′，r，t>)，(<h，r，t>，<h，r，t′>)]，

By S _p ＝{<h，r，t>Positive triples, S _f ＝{<h′，r，t>|h′∈E，<h，r，t′>T' E represents a negative triplet, T _batch As input of knowledge graph representation learning model, for T _batch Training knowledge graph representation learning and combining formula

The knowledge graph represents the loss function definition of the learning model as follows:

wherein gamma is a spacing distance parameter greater than 0, is a super parameter, [ x ]] ₊ Representing a positive function, i.e. when x > 0, [ x ]] ₊ =x; when x is less than or equal to 0, [ x ]] ₊ =0; this training method is called margin-based ranking criterion, which aims to separate the positive triplet and the negative triplet as far as possible and find the support vector of the maximum distance;

step 7) updating parameters of the knowledge graph representation learning model by adopting random gradient descent, wherein gradient updating only needs to calculate distances d (h+r, t) and d (h '+r, t'), and a total of |E| entities and |R| relations are set, the length of each entity vector is m-dimension, the length of each relation vector is n-dimension, and the total of |E|m+ |R| n parameters need to be updated;

and 8) repeating the steps 6) -7) for iterative training, and carrying out relation prediction on unlabeled triples by using a knowledge graph representation learning model after the iterative training is completed, wherein the prediction steps are as follows: taking one of the triples from the unlabeled triplet<h，r，t＞ _unlabel Using knowledge graph to represent learning model to predict relation r 'between h and t, if r' =r, then predictingCorrect; meanwhile, taking the predicted correct triples as positive triples, randomly replacing head entities or tail entities of the positive triples as negative triples, and then combining the new positive triples and the new negative triples into the original marked triples to form new marked triples;

and 9) repeating the steps 4) -8) by adopting a new marked triplet until a new unmarked triplet cannot be learned, wherein the description knowledge graph indicates that the learning model cannot learn more Chinese knowledge features from the current training set, and the entity vector and the relation vector output by the knowledge graph indicates that the learning model is the best Chinese knowledge graph representation corresponding to the Chinese link data set zhishi.

2. The learning method of the feature tensor-based chinese knowledge graph representation according to claim 1, wherein the process of constructing the feature tensor of the entity in the step 3) is as follows: the tag triplet is denoted as T _l E is used for representing an entity set in the knowledge graph, R is used for representing a relation set, an entity E E in one marked triplet is selected, and the entity pinyin matrix is searched for obtaining a pinyin vector E of the entity E _p The method comprises the steps of carrying out a first treatment on the surface of the Let entity name be c ₁ c ₂ ...c _m Wherein c _m The mth word representing the name of the entity is searched for the word embedding matrix according to the word to obtain the word vector e of the entity e _c ＝c ₁ c ₂ ...c _m The method comprises the steps of carrying out a first treatment on the surface of the Searching the word embedding matrix to obtain a word vector e of the entity e _w The method comprises the steps of carrying out a first treatment on the surface of the Searching the description matrix to obtain a description vector e of the entity e _d The feature tensor FeatureTensor of entity e is expressed as

3. The learning method of the feature tensor-based chinese knowledge graph representation according to claim 2, wherein the step 3) of converting the feature tensor of the entity into the feature vector of the entity comprises: entity is put into practiceThe different dimensions of the feature tensor of (2) are connected in a vector splicing way, the vectors are randomly lost by adopting dropout, knowledge representation is prevented from learning and fitting, and the pinyin vectors e of the same entity e are obtained by the method _p Word vector e _c Word vector e _w Description vector e _d Connecting to obtain a feature vector e of the entity e _ft ＝[e _p ；e _c ；e _w ；e _d ]。

4. A feature tensor-based chinese knowledge graph representation learning method according to any one of claims 1-3, characterized in that the feature tensor-based chinese knowledge graph representation learning method employs link prediction as an evaluation task, the evaluation indexes comprising rank average MR, average reciprocal rank MRR, percentage of top ten results of correct entity hit @10, percentage of top three results of correct entity hit @3, percentage of top three results of correct entity hit @1, wherein the smaller the value of MR, the better the values of MRR, hit @10, hit @3, hit @ 1; and randomly removing part of entities or relations of the triples in the Chinese linked data set zhishi.

Images