CN108052683A - A kind of knowledge mapping based on cosine measurement rule represents learning method - Google Patents

Abstract

The invention discloses a learning method for knowledge graph representation based on cosine metric rules. Firstly, entities and relationships in the knowledge graph are randomly embedded into two vector spaces; secondly, by using the statistical rules of candidate entities, the triplet sets and Candidate entity vector set; again use the cosine similarity to construct the scoring function of the target vector and the candidate entity, and evaluate the candidate entity; finally, use the loss function to uniformly train all the candidate entity vectors and target vectors related to each other, and use stochastic gradient descent The algorithm minimizes a loss function. When the optimization goal is achieved, the best representation of each entity vector and relationship vector in the knowledge graph can be obtained, so as to better represent the connection between the entity and the relationship, and can be well applied to large-scale knowledge graph complementation. All of them.

Description

A Knowledge Graph Representation Learning Method Based on Cosine Metric Rule

technical field

The invention relates to the technical field of knowledge graphs, in particular to a knowledge graph representation learning method based on cosine metric rules.

Background technique

With the advent of the era of big data, the dividend of data has made artificial intelligence technology develop at an unprecedented speed. Great progress has been made in related fields such as knowledge engineering represented by knowledge graph and machine learning represented by representation learning. On the one hand, as the dividends of big data are exhausted by representation learning, the effect of representation learning models tends to become a bottleneck. On the other hand, with the continuous emergence of a large number of knowledge graphs, these treasures of human prior knowledge have not been effectively utilized by representation learning. The fusion of knowledge graph and representation learning has become one of the important ideas to further improve the effect of representation learning model. The symbolism represented by the knowledge map and the connectionism represented by the representation learning are increasingly deviating from the original independent development track and embarking on a new path of synergy.

The knowledge graph is essentially a semantic network, which expresses various entities, concepts and the semantic relationship between them. Compared with traditional knowledge representation forms (such as ontology and traditional semantic network), knowledge graph has the advantages of high entity/concept coverage, diverse semantic relations, friendly structure (usually expressed as RDF format) and high quality, which makes knowledge graph increasingly become a It is the most important knowledge representation method in the era of big data and artificial intelligence.

Triple is a general representation of knowledge graph. The basic form of triple includes (entity 1, relationship, entity 2) and (concept, attribute, attribute value), etc. Entity is the most basic element in knowledge graph. , there are different relationships between different entities. Concepts mainly include collections, object types, and types of things, such as geography, people, etc.; attributes refer to the attribute characteristics and characteristics of objects, such as nationality and date of birth; attribute values refer to the values corresponding to attributes, such as China, 1993-01 -12 etc. Usually (head, relation, tail) (abbreviated as (h, r, t)) is used to represent a triple, where r represents the relationship between the head entity h and the tail entity t. For example, the knowledge that Paris is the capital of France can be represented by the triplet (Paris, is...the capital, France) in the knowledge graph.

Representation learning for knowledge graphs aims to embed entities and relationships in knowledge graphs into low-dimensional vector spaces and represent them as dense low-dimensional real-valued vectors. The key is to reasonably define the vectorized representation of the loss function f _r (h, t) of the fact (triple (h, r, t)) in the knowledge graph and the two entities h, t of the triple. In general, it is desirable to minimize f _r (h,t) when the fact (h,r,t) holds. Considering the facts of the entire knowledge graph, the vector representation of entities and relationships can be learned by minimizing the loss function. Different representation learning can use different principles and methods to define corresponding loss functions. The current translation model represented by the TransE model has received extensive attention for its outstanding performance and simple model parameters. However, existing translation models can effectively deal with simple relations of 1-1, but are still limited for complex relations of 1-N, N-1, NN. This leads to the fact that existing knowledge graph representation learning methods cannot be well applied to large-scale knowledge graphs.

Contents of the invention

What the present invention aims to solve is the problem that the existing knowledge graph representation learning method cannot handle complex relationships of 1-N, N-1, and N-N well, and provides a knowledge graph representation learning method based on the cosine metric rule.

In order to solve the above problems, the present invention is achieved through the following technical solutions:

A knowledge map representation learning method based on the cosine metric rule, comprising the following steps:

Step 1. Use the vector random generation method to embed the entity set and relation set in the knowledge map into the entity vector space and the relation vector space respectively to obtain the entity vector and relation vector;

Step 2, using the statistical rules of candidate entities to obtain a set of candidate entity vectors for randomly selected triples, and randomly generate a set of wrong entity vectors according to the set of candidate entity vectors;

Step 3, using the cosine similarity to construct the scoring function between the target vector and the candidate entity vector, and at the same time standardize the value range of its function value;

Step 4. Use the scoring function to construct a boundary-based loss function for distinguishing the candidate entity vector set from the wrong entity vector set, and then make the candidate entity vector set uniformly constrain the target vector according to the loss function;

Step 5. Use the optimization algorithm to optimize the loss function value, so that the scoring function value of the candidate entity vector set is close to 1, and the scoring function value of the wrong entity vector set is close to 0, so as to learn the best vector representation of entities and relationships. reach the optimization goal.

The specific sub-steps of the above step 2 are as follows:

Step 2.1, randomly select a triplet from the triplet set in the knowledge map;

Step 2.2, find all tail entity vectors that match the head entity vector and relation vector of the selected triple in the triple set, and form the candidate tail entity vector set with the found tail entity vectors; meanwhile, in the three Find all head entity vectors that match the tail entity vector and relation vector of the selected triple in the tuple set, and form the candidate head entity vector set with the found head entity vectors;

Step 2.3, performing random replacement operation on the candidate tail entity vector set to generate the error tail entity vector set; at the same time, performing random replacement operation on the candidate head entity vector set to generate the error head entity vector set.

The scoring function constructed in the above step 3 is as follows:

The scoring function f _t (g _t ,t) of the candidate tail entity vector is:

The scoring function f _h (g _h ,h) of the candidate head entity vector is:

The scoring function f _t ′(g _t ,t′) of the wrong tail entity vector is:

The scoring function f′ _h (g _h ,h′) of the wrong head entity vector is:

Among the above formulas, α is the scoring function value range specification parameter; g _t is the target vector of the tail entity, g _t =h ₀ +r ₀ ; g _h is the target vector of the head entity, g _h =t ₀ -r ₀ ; h ₀ is the head entity vector of the selected triple, t ₀ is the tail entity vector of the selected triple, r ₀ is the relation vector of the selected triple; t is the candidate tail entity vector, h is Candidate head entity vector, t' is the error tail entity vector, h' is the error head entity vector.

The above scoring function value range specification parameter α∈[0,1].

The loss function L constructed in step 4 above is:

In the formula, γ is the set boundary value; f _t ′(g _t ,t′) represents the scoring function of the wrong tail entity vector, f _t (g _t ,t) represents the scoring function of the candidate tail entity vector, f′ _h (g _h , h′) represents the scoring function of the wrong head entity vector, f _h (g _h , h) represents the scoring function of the candidate head entity vector; g _t is the target vector of the tail entity, g _t = h ₀ + r ₀ ; g _h is the target vector of the head entity, g _h =t ₀ -r ₀ ; h ₀ is the head entity vector of the selected triple, t ₀ is the tail entity vector of the selected triple, r ₀ is The relationship vector of the selected triplet; t is the candidate tail entity vector, t _c is the candidate tail entity vector set, t′ is the wrong tail entity vector, t′ _c is the wrong tail entity vector set, h is the candidate head entity vector , h _c is the candidate header entity vector set, h′ is the error header entity vector, h′ _c is the error header entity vector set.

The optimization algorithm described in step 5 above is a stochastic gradient descent algorithm.

Compared with prior art, the present invention has following characteristics:

First, a candidate entity statistics rule is proposed to obtain the candidate entity vector set of correlation relationship;

Second, by introducing the cosine similarity between the two vectors, the cosine value between the target vector and the candidate entity vector is calculated as a measure of the difference between the two individuals. Compared with the Euclidean distance, it is not a simple calculation between the two vectors The cosine similarity pays more attention to the difference in the direction of the two vectors; this solves the shortcomings of the existing models when dealing with complex relationships such as 1-N, N-1, and N-N, and enriches the expressive ability of entities and relationships. Overall improved model performance;

Thirdly, by uniting all candidate entities and target vectors of related relations to form a unified constraint, the interaction between candidate entity vectors and target vectors is improved.

Description of drawings

Fig. 1 is a flow chart of a knowledge map representation learning method based on the cosine metric rule of the present invention.

Figure 2 An example diagram of entity and relation triples in the knowledge graph.

Fig. 3 is an example diagram of the training target of a knowledge map representation learning method based on the cosine metric rule of the present invention, wherein (a) is before training, and (b) is after training.

Detailed ways

In order to make the object, technical solution and advantages of the present invention clearer, the present invention will be further described in detail below in combination with specific examples and with reference to the accompanying drawings.

The present invention discloses a knowledge map representation learning method based on the cosine metric rule. As shown in Figure 1, firstly, the entities and relations in the knowledge map are randomly embedded into two vector spaces; secondly, the statistical rules of the candidate entities are used to count the relevant relations The corresponding triplet set and the candidate entity vector set, and randomly generate the wrong entity vector set according to the candidate entity vector set; again use the cosine similarity to construct the scoring function between the target vector and the candidate entity, and evaluate the candidate entity; finally use the loss function The candidate entity vectors and target vectors of all relevant relations are uniformly trained, and the loss function is minimized by the stochastic gradient descent algorithm. When the optimization goal is achieved, the best representation of each entity vector and relationship vector in the knowledge graph can be obtained, so as to better represent the connection between the entity and the relationship, and can be well applied to large-scale knowledge graph complementation. All of them. Through the cosine similarity and unified training, the present invention well overcomes the problem that the existing models cannot deal with the complex relationship of 1-N, N-1, and N-N well, and has strong feasibility and good practicability.

The present invention considers the triplet structured information in the knowledge map, and uses the typical triplet form of (head, relation, tail) to represent the knowledge. The relation is used to connect the entity head and tail, and describe two Relationships between entities. Figure 2 is an example diagram of a typical knowledge graph triplet structure. Among them, the circle represents the entity node (such as "Tang**", "Yi**", "Ti**", etc.), and the connection between two entities represents the relationship (such as "nationality", "president", " daughter", etc.). In addition, it can be seen that there are multiple relationships between the entity "Tang **" and "United States", and there are multiple entity pairs corresponding to the "daughter" and "nationality" relationships.

A knowledge graph representation learning method based on the cosine metric rule, using the knowledge graph in Figure 2 as the training set, which includes 5 entities, 4 relationships, and 8 triples. The concrete implementation of method comprises the following steps:

Step 1. Use the vector random generation method to embed the entities and relationships in the knowledge map into two vector spaces respectively;

Step 11. Express the entity set {"Tang**", "Ma*", "Ti**", "伊**", "United States"} of the knowledge graph in Figure 2 as {e ₁ , e ₂ , e ₃ , e ₄ , e ₅ } and the relation set {"wife", "daughter", "president", "nationality"} are expressed as {r ₁ ,r ₂ ,r ₃ ,r ₄ }. The knowledge map in Figure 2 includes triple sets {(e ₁ ,r ₁ ,e ₂ ), (e ₁ ,r ₂ ,e ₃ ), (e ₁ ,r ₂ ,e ₄ ), (e ₁ ,r ₃ ,e ₅ ),(e ₂ ,r ₂ ,e ₃ ),(e ₃ ,r ₃ ,e ₅ ),(e ₄ ,r ₃ ,e ₅ ),(e ₅ ,r ₄ ,e ₁ )};

Step 12. Embed them into the entity vector space and relation vector space respectively to obtain entity vectors {e ₁ , e ₂ , e ₃ , e ₄ , e ₅ } and relation vectors {r ₁ , r ₂ , r ₃ , r ₄ }.

Step 2, using the statistical rules of candidate entities to obtain the triplet set of related relations and the vector set of candidate entities, and at the same time randomly generate the set of wrong entity vectors.

The statistical rules for candidate entities are as follows:

In the first step, a triplet is randomly extracted from the knowledge graph (eg, (h ₀ ,r ₀ ,t ₀ ));

Step 21, obtain a triplet by random extraction (eg, (Tang **, daughter, Ti **) namely (e ₁ , r ₂ , e ₃ ));

The second step is to match the triplet sets (h _{0 ,} r _{0 ,} t _c ) and _{( h c ,} r ₀ ,t ₀ ), where is all candidate tail entity sets with g _t = h ₀ +r ₀ as the target vector, including |t _c | different numbers of tail entities, is all candidate head entity sets of g _h =t ₀ -r ₀ target vectors, including |h _c | different numbers of head entities;

Step 22. According to ((e ₁ ,r ₂ ,?), match the triple set {(e ₁ ,r ₂ ,e ₃ ), (e ₁ ,r ₂ ,e ₄ )}, and obtain the candidate tail entity set at the same time {e ₃ ,e ₄ };

Step 23. According to (?, r ₂ , e ₃ ), match the triplet set {(e ₁ , r ₂ , e ₃ ), (e ₂ , r ₂ , e ₃ )}, and obtain the candidate head entity set { e ₁ ,e ₂ };

The third step is to randomly generate the error entity set corresponding to the candidate entity set. In the process of generating the error entity set, the generated error tail entity will be compared with the candidate entity in the candidate entity set, and only the error tail entity that does not exist in the candidate entity set will be included in the error entity set.

Step 24: Randomly generate an error tail entity set {e ₅ , e ₂ } and an error head entity set {e ₄ , e ₅ }.

Step 3. Using the cosine similarity to construct the scoring function between the target vector and the candidate entity vector, and at the same time standardize the value range of the function value, which is generally [0,1].

Step 31. Using the cosine value formula, express the cosine value cos<a, b> of vector a and vector b as:

Step 32. Construct a scoring function based on the above cosine formula:

The scoring function f _t (g _t ,t) of the candidate tail entity according to the cosine similarity between the target vector and the candidate tail entity vector is constructed as follows:

According to the cosine similarity between the target vector and the candidate head entity vector, construct the scoring function f _h (g _h ,h) of the candidate head entity as follows:

Among them, α∈[0,1] is the specification parameter of the scoring function value range, g _t is the target vector of the tail entity, g _t =h ₀ +r ₀ , g _h is the target vector of the head entity, g _h =t ₀ − r ₀ , t is the candidate tail entity vector, h is the candidate head entity vector.

According to the cosine similarity between the target vector and the wrong tail entity vector, the scoring function f _t ′(g _t ,t′) of the wrong tail entity is constructed as follows:

According to the cosine similarity between the target vector and the error head entity vector, the scoring function f′ _h (g _h ,h′) of the error head entity is constructed as follows:

Among them, α∈[0,1] is the specification parameter of the scoring function value range, g _t is the target vector of the tail entity, g _t =h ₀ +r ₀ , g _h is the target vector of the head entity, g _h =t ₀ − r ₀ , t' is the error tail entity vector, h' is the error head entity vector.

Step 4. Use the scoring function to construct a boundary-based loss function for distinguishing the candidate entity vector set from the wrong entity vector set, and then make the candidate entity vector set uniformly constrain the target vector according to the loss function;

Step 41, the boundary-based loss function is constructed as follows:

Among them, [γ+f′-f] ₊ ＝max(0,γ+f′-f); γ is the set boundary value; f _t ′(g _t ,t′) represents the scoring function of the wrong tail entity vector , f _t (g _t ,t) represents the scoring function of the candidate tail entity vector, f′ _h (g _h ,h′) represents the scoring function of the wrong head entity vector, f _h (g _h ,h) represents the candidate head entity vector scoring function; g _t is the target vector of the tail entity, g _t = h ₀ + r ₀ ; g _h is the target vector of the head entity, g _h = t ₀ -r ₀ ; h ₀ is the selected triplet head entity vector, t ₀ is the tail entity vector of the selected triple, r ₀ is the relationship vector of the selected triple; t is the candidate tail entity vector, t _c is the candidate tail entity vector set, and t′ is Error tail entity vector, t′ _c is the error tail entity vector set, h is the candidate head entity vector, h _c is the candidate head entity vector set, h′ is the error head entity vector, h′ _c is the error head entity vector set.

Step 42. Combine the triplet set {(e ₁ , r ₂ , e ₃ ), (e ₂ , r ₂ , e ₃ )} obtained in step 22 and the candidate tail entity set {e ₃ , e ₄ }, respectively Calculate the score by the scoring function in step 32;

Step 421, (e ₁ , r ₂ , e ₃ ) score is

Its corresponding error triplet (e ₁ ,r ₂ ,e ₅ ) score is

Step 422, (e ₁ , r ₂ , e ₄ ) score is

Its corresponding error triplet (e ₁ , r ₂ , e ₂ ) score is

Step 43. Combine the triplet set {(e ₁ , r ₂ , e ₃ ), (e ₂ , r ₂ , e ₃ )} obtained in step 22 with the candidate head entity set {e ₁ , e ₂ }, respectively Calculate the score by the scoring function in step 32;

Step 431, (e ₁ , r ₂ , e ₃ ) score is

Its corresponding error triplet (e ₄ , r ₂ , e ₃ ) score is

Step 432, (e ₂ , r ₂ , e ₃ ) score is

The corresponding error triplet (e ₅ ,r ₂ ,e ₂ ) score is

Step 44. Substituting the scores calculated in steps 42 and 43 into the loss function L to obtain the loss function value. All candidate entity vectors and target vectors of related relations are uniformly constrained by a boundary-based loss function.

Step 5, use the optimization algorithm to optimize the loss function value, so that the scoring function value of the candidate entity vector set is close to 1, and the scoring function value of the wrong entity vector set is close to 0, represented by the best vector of the learned entity and relationship, reach the optimization goal.

Step 51, the optimization algorithm will use the stochastic gradient descent algorithm to minimize the loss function, and obtain the best scores of all candidate entity vectors, so as to learn the best vector representation of entities and relationships, and achieve the optimization goal.

The translation principle adopted by the knowledge map representation learning method of the present invention is shown in Fig. 3, and its basic idea is: first construct the training target vector g=h+r (or g=tr) according to the triplet (h, r, t), At the same time, obtain the set of candidate entity vectors, such as {e ₁ , e ₂ , e ₃ , e ₄ } in Figure 3; secondly, use the cosine similarity between the two vectors to calculate the score; finally, use the stochastic gradient descent algorithm to score according to the loss function to change the candidate entity vector, while changing the target vector according to the overall change direction of the preferred entity. The present invention uses the cosine similarity between two vectors to better calculate the difference between the target vector and the candidate entity vector. Compared with the Euclidean distance, it is not a simple calculation of the distance between two vectors, and the cosine similarity pays more attention to The difference in direction of the two vectors. This solves the shortcomings of existing models when dealing with complex relationships such as 1-N, N-1, and NN, enriches the expressive ability of entities and relationships, and improves the performance of the model as a whole.

The present invention uses the cosine similarity between two vectors to better calculate the similarity between the target vector and the candidate entity vector. Using the embedding-based model of entity vectors and relationship vectors, using the statistical rules of candidate entities, the triple set of related relations and the set of candidate entity vectors are obtained; and the cosine similarity between the target vector and candidate entity vectors is introduced to enhance the model pair 1- The ability to express complex relationships such as N, N-1, and N-N, and construct a unique scoring function for target vectors and candidate entity vectors. Finally, a new loss function is constructed, and the loss function is optimized through the random gradient descent algorithm. When the optimal optimization goal is reached, each optimal entity vector and relationship vector in the knowledge map can be obtained, so as to better integrate entities and relationships. Represent and save the connection between entities and relationships, so that it can be well applied to large-scale knowledge map completion.

It should be noted that although the above-mentioned embodiments of the present invention are illustrative, they are not intended to limit the present invention, so the present invention is not limited to the above specific implementation manners. Without departing from the principles of the present invention, all other implementations obtained by those skilled in the art under the inspiration of the present invention are deemed to be within the protection of the present invention.

Claims (6)

1. A knowledge map representation learning method based on the cosine metric rule, is characterized in that, comprises steps as follows: Step 1. Use the vector random generation method to embed the entity set and relation set in the knowledge map into the entity vector space and the relation vector space respectively to obtain the entity vector and relation vector; Step 2, using the statistical rules of candidate entities to obtain a set of candidate entity vectors for randomly selected triples, and randomly generate a set of wrong entity vectors according to the set of candidate entity vectors; Step 3, using the cosine similarity to construct the scoring function between the target vector and the candidate entity vector, and at the same time standardize the value range of its function value; Step 4. Use the scoring function to construct a boundary-based loss function for distinguishing the candidate entity vector set from the wrong entity vector set, and then make the candidate entity vector set uniformly constrain the target vector according to the loss function; Step 5. Use the optimization algorithm to optimize the loss function value, so that the scoring function value of the candidate entity vector set is close to 1, and the scoring function value of the wrong entity vector set is close to 0, so as to learn the best vector representation of entities and relationships. reach the optimization goal. 2. a kind of knowledge map representation learning method based on cosine metric rule according to claim 1, it is characterized in that, the specific sub-steps of step 2 are as follows: Step 2.1, randomly select a triplet from the triplet set in the knowledge graph; Step 2.2, find all tail entity vectors that match the head entity vector and relation vector of the selected triple in the triple set, and form the candidate tail entity vector set with the found tail entity vectors; meanwhile, in the three Find all head entity vectors that match the tail entity vector and relation vector of the selected triple in the tuple set, and form the candidate head entity vector set with the found head entity vectors; Step 2.3, performing random replacement operation on the candidate tail entity vector set to generate the error tail entity vector set; at the same time, performing random replacement operation on the candidate head entity vector set to generate the error head entity vector set. 3. A kind of knowledge map representation learning method based on cosine metric rule according to claim 1 or 2, it is characterized in that, the scoring function constructed in step 3 is as follows: The scoring function f _t (g _t ,t) of the candidate tail entity vector is: <mrow><msub><mi>f</mi><mi>t</mi></msub><mrow><mo>(</mo><msub><mi>g</mi><mi>t</mi></msub><mo>,</mo><mi>t</mi><mo>)</mo></mrow><mo>=</mo><mi>&amp;alpha;</mi><mo>+</mo><mrow><mo>(</mo><mn>1</mn><mo>-</mo><mi>&amp;alpha;</mi><mo>)</mo></mrow><mfrac><mrow><msub><mi>g</mi><mi>t</mi></msub><mo>&amp;CenterDot;</mo><mi>t</mi></mrow><mrow><msqrt><msub><mi>g</mi><mi>t</mi></msub></msqrt><mo>*</mo><msqrt><mi>t</mi></msqrt></mrow></mfrac><mo>,</mo></mrow> The scoring function f _h (g _h ,h) of the candidate head entity vector is: <mrow><msub><mi>f</mi><mi>h</mi></msub><mrow><mo>(</mo><msub><mi>g</mi><mi>h</mi></msub><mo>,</mo><mi>h</mi><mo>)</mo></mrow><mo>=</mo><mi>&amp;alpha;</mi><mo>+</mo><mrow><mo>(</mo><mn>1</mn><mo>-</mo><mi>&amp;alpha;</mi><mo>)</mo></mrow><mfrac><mrow><msub><mi>g</mi><mi>h</mi></msub><mo>&amp;CenterDot;</mo><mi>h</mi></mrow><mrow><msqrt><msub><mi>g</mi><mi>h</mi></msub></msqrt><mo>*</mo><msqrt><mi>h</mi></msqrt></mrow></mfrac><mo>,</mo></mrow> The scoring function f _t ′(g _t ,t′) of the wrong tail entity vector is: <mrow><msubsup><mi>f</mi><mi>t</mi><mo>&amp;prime;</mo></msubsup><mrow><mo>(</mo><msub><mi>g</mi><mi>t</mi></msub><mo>,</mo><msup><mi>t</mi><mo>&amp;prime;</mo></msup><mo>)</mo></mrow><mo>=</mo><mi>&amp;alpha;</mi><mo>+</mo><mrow><mo>(</mo><mn>1</mn><mo>-</mo><mi>&amp;alpha;</mi><mo>)</mo></mrow><mfrac><mrow><msub><mi>g</mi><mi>t</mi></msub><mo>&amp;CenterDot;</mo><msup><mi>t</mi><mo>&amp;prime;</mo></msup></mrow><mrow><msqrt><msub><mi>g</mi><mi>t</mi></msub></msqrt><mo>*</mo><msqrt><msup><mi>t</mi><mo>&amp;prime;</mo></msup></msqrt></mrow></mfrac><mo>,</mo></mrow> The scoring function f _h ′(g _h ,h′) of the error head entity vector is: <mrow><msubsup><mi>f</mi><mi>h</mi><mo>&amp;prime;</mo></msubsup><mrow><mo>(</mo><msub><mi>g</mi><mi>h</mi></msub><mo>,</mo><msup><mi>h</mi><mo>&amp;prime;</mo></msup><mo>)</mo></mrow><mo>=</mo><mi>&amp;alpha;</mi><mo>+</mo><mrow><mo>(</mo><mn>1</mn><mo>-</mo><mi>&amp;alpha;</mi><mo>)</mo></mrow><mfrac><mrow><msub><mi>g</mi><mi>h</mi></msub><mo>&amp;CenterDot;</mo><msup><mi>h</mi><mo>&amp;prime;</mo></msup></mrow><mrow><msqrt><msub><mi>g</mi><mi>h</mi></msub></msqrt><mo>*</mo><msqrt><msup><mi>h</mi><mo>&amp;prime;</mo></msup></msqrt></mrow></mfrac><mo>,</mo></mrow> Among the above formulas, α is the scoring function value range specification parameter; g _t is the target vector of the tail entity, g _t =h ₀ +r ₀ ; g _h is the target vector of the head entity, g _h =t ₀ -r ₀ ; h ₀ is the head entity vector of the selected triple, t ₀ is the tail entity vector of the selected triple, r ₀ is the relation vector of the selected triple; t is the candidate tail entity vector, h is Candidate head entity vector, t' is the error tail entity vector, h' is the error head entity vector. 4. A knowledge map representation learning method based on the cosine metric rule according to claim 3, characterized in that the scoring function value range specification parameter α∈[0,1]. 5. A kind of knowledge map representation learning method based on cosine metric rule according to claim 1, it is characterized in that, the loss function L constructed in step 4 is: <mrow><mi>L</mi><mo>=</mo><munder><mo>&amp;Sigma;</mo><mrow><mi>t</mi><mo>&amp;Element;</mo><msub><mi>t</mi><mi>c</mi></msub><mo>,</mo><msup><mi>t</mi><mo>&amp;prime;</mo></msup><mo>&amp;Element;</mo><msubsup><mi>t</mi><mi>c</mi><mo>&amp;prime;</mo></msubsup></mrow></munder><msub><mrow><mo>&amp;lsqb;</mo><mi>&amp;gamma;</mi><mo>+</mo><msubsup><mi>f</mi><mi>t</mi><mo>&amp;prime;</mo></msubsup><mrow><mo>(</mo><msub><mi>g</mi><mi>t</mi></msub><mo>,</mo><msup><mi>t</mi><mo>&amp;prime;</mo></msup><mo>)</mo></mrow><mo>-</mo><msub><mi>f</mi><mi>t</mi></msub><mrow><mo>(</mo><msub><mi>g</mi><mi>t</mi></msub><mo>,</mo><mi>t</mi><mo>)</mo></mrow><mo>&amp;rsqb;</mo></mrow><mo>+</mo></msub><mo>+</mo><munder><mo>&amp;Sigma;</mo><mrow><mi>h</mi><mo>&amp;Element;</mo><msub><mi>h</mi><mi>c</mi></msub><mo>,</mo><msup><mi>h</mi><mo>&amp;prime;</mo></msup><mo>&amp;Element;</mo><msubsup><mi>h</mi><mi>c</mi><mo>&amp;prime;</mo></msubsup></mrow></munder><msub><mrow><mo>&amp;lsqb;</mo><mi>&amp;gamma;</mi><mo>+</mo><msubsup><mi>f</mi><mi>h</mi><mo>&amp;prime;</mo></msubsup><mrow><mo>(</mo><msub><mi>g</mi><mi>h</mi></msub><mo>,</mo><msup><mi>h</mi><mo>&amp;prime;</mo></msup><mo>)</mo></mrow><mo>-</mo><msub><mi>f</mi><mi>h</mi></msub><mrow><mo>(</mo><msub><mi>g</mi><mi>h</mi></msub><mo>,</mo><mi>h</mi><mo>)</mo></mrow><mo>&amp;rsqb;</mo></mrow><mo>+</mo></msub></mrow> In the formula, γ is the set boundary value; f _t ′(g _t ,t′) represents the scoring function of the wrong tail entity vector, f _t (g _t ,t) represents the scoring function of the candidate tail entity vector, f _h ′ (g _h , h′) represents the scoring function of the wrong head entity vector, f _h (g _h , h) represents the scoring function of the candidate head entity vector; g _t is the target vector of the tail entity, g _t = h ₀ +r ₀ ; g _h is the target vector of the head entity, g _h =t ₀ -r ₀ ; h ₀ is the head entity vector of the selected triple, t ₀ is the tail entity vector of the selected triple, r ₀ is The relationship vector of the selected triplet; t is the candidate tail entity vector, t _c is the candidate tail entity vector set, t′ is the wrong tail entity vector, t′ _c is the wrong tail entity vector set, h is the candidate head entity vector , h _c is the candidate header entity vector set, h′ is the error header entity vector, h′ _c is the error header entity vector set. 6. A knowledge map representation learning method based on the cosine metric rule according to claim 1, wherein the optimization algorithm in step 5 is a stochastic gradient descent algorithm.