CN105630901A - Knowledge graph representation learning method - Google Patents

Abstract

The invention discloses a knowledge graph representation learning method. The method comprises the following steps: defining the correlation between entity vectors and relation vectors in a relation triple (head, relation, tail) by utilizing a translation-based model between the entity vector and the relation vector; defining the correlation between entity vectors and feature vectors in a feature triple (entity, attribute, value) by utilizing a neural network classification model; and correlating the entity vectors, the relation vectors and the feature vectors with one another through an evaluation function, and minimizing the evaluation function to learn the entity vectors, the relation vectors and the feature vectors so as to achieve the optimization aim. By adopting the method disclosed in the invention, the relation among the entity, the relation and the feature can be accurately represented.

Description

A kind of knowledge mapping represents learning method

Technical field

The present invention relates to natural language processing and knowledge mapping field, represent learning method particularly to a kind of knowledge mapping.

Background technology

Along with society develops rapidly, we come into the information explosion epoch, and the entity that magnanimity is new all can be had every day to produce with information. The Internet is as current information acquisition platform the most easily, and user is day by day urgent with the demand concluded to effective information screening, and how obtaining valuable information from mass data becomes a difficult problem. Knowledge mapping arises at the historic moment in this.

Knowledge mapping the proper noun such as all personages, place name, title, team's name and things will be expressed as entity in the world, internal relation between entity is expressed as relation, it is intended to be expressed as between entity using the mass knowledge in data base utilizing relation as the ternary relation group of bridge. Such as, Beijing is this knowledge of capital of China, then utilizes tlv triple relation (Beijing is ... capital, China) to be indicated in knowledge mapping. The different semantemes of one key word can be set up different entities by knowledge mapping, eliminate the interference of language polysemy, deeper wider array of internal relation between target entity and other related entities can be represented simultaneously, be widely used in multiple fields such as data mining, information retrieval, question answering systems. Owing to existing entity is huge, and entity and relation every day all change with increase, it is difficult to manual maintenance with excavate new information, be study hotspot important now to expression and the auto-complete of knowledge mapping.

Knowledge mapping physical quantities is huge, and network structure is openness seriously. And in the research of knowledge mapping, achieve significant progress in recent years, represent that study applies in knowledge mapping, by in all entities and relationship map to low-dimensional vector row space, the openness and efficiency produced during knowledge mapping study before solving. But, current existing knowledge mapping represents that the characteristic of the relation between entity and entity is lumped together by learning method when study, uses same model to be modeled, it is impossible to accurate presentation-entity, connecting each other between relation and characteristic.

Summary of the invention

In view of this, the goal of the invention of the present invention is: solve prior art exists cannot Precise Representation entity, the problem of contact between relation and characteristic, to improve the quality that knowledge mapping represents.

For reaching above-mentioned purpose, technical scheme is specifically achieved in that

The invention provides a kind of knowledge mapping and represent learning method, the method includes: utilize the model based on translation between entity vector and relation vector, in definition relation tlv triple (head, relation, tail) entity vector and relation vector between interrelated; Utilize neural network classification model, in defined property tlv triple (entity, attribute, value) entity vector and eigen vector between interrelated; By evaluation function, entity vector, relation vector and eigen vector are associated, and minimize evaluation function, with learning object vector, relation vector and eigen vector, reach optimization aim.

As seen from the above technical solutions, relation and characteristic are distinguished and are treated by the present invention, so the form that knowledge mapping adopts relation tlv triple and characteristic tlv triple represents knowledge. Therefore, the present invention adopts the model between entity vector and relation vector based on translation, and what represent between the entity vector sum relation vector in relation tlv triple is interrelated; Adopting neural network classification model, that comes between the entity vector sum eigen vector in characterization tlv triple is interrelated; Then pass through evaluation function entity vector, relation vector and eigen vector to be associated, and optimize this evaluation function, when reaching optimization aim, each entity vector, relation vector and the eigen vector that just can simultaneously learn in knowledge mapping, thus accurate presentation-entity, connecting each other between relation and characteristic.

Accompanying drawing explanation

Fig. 1 is the exemplary plot of inclusion relation tlv triple and characteristic tlv triple in knowledge mapping.

Fig. 2 is the schematic flow sheet that knowledge mapping of the present invention represents learning method.

According to prior art knowledge mapping, Fig. 3 a represents that the triple table that learning method obtains advises the exemplary plot of knowledge.

According to knowledge mapping of the present invention, Fig. 3 b represents that the triple table that learning method obtains advises the exemplary plot of knowledge.

Detailed description of the invention

For making the purpose of the present invention, technical scheme and advantage clearly understand, developing simultaneously embodiment referring to accompanying drawing, the present invention is described in more detail.

In prior art, and do not differentiate between relation and characteristic, characteristic is also served as the one of relation, knowledge mapping mainly adopts the form of (entity 1, relation, entity 2) tlv triple to represent knowledge, namely relation tlv triple (head, relation, tail) is adopted to represent. What therefore represent between the entity vector sum relation vector in relation tlv triple in prior art only with a kind of model is interrelated, knowledge mapping represents that the characteristic of the relation between entity and entity be cannot be distinguished by out by learning method when study, it is impossible to accurate presentation-entity, connecting each other between relation and characteristic.

Relation and characteristic are distinguished and are treated by the present invention, so the form that knowledge mapping adopts relation tlv triple and characteristic tlv triple represents knowledge. Relation tlv triple represents with (head, relation, tail), and relation is used for connecting two entities, portrays the association between two entities. Characteristic tlv triple represents with (entity, attribute, value), and (a v) is used for portraying the intrinsic characteristic of correspondent entity to each characteristic-value. In knowledge mapping, relation tlv triple interior joint presentation-entity, even limit represents relation; Connecting limit characterization in characteristic tlv triple, even one end node presentation-entity on limit, even the other end node on limit represents the characteristic value of this entity. Fig. 1 is the exemplary plot of inclusion relation tlv triple and characteristic tlv triple in knowledge mapping. Wherein, the node " Clinton " that circle represents and " Hillary " are entity, Lian Bianwei " wife " relation between them. It addition, it will be seen that two entities are each with one's own characteristic, such as " occupation ", " sex ", " birthplace " etc., characteristic value for entity " Clinton " occupation is US President, and the characteristic value for entity " Hillary " occupation is U.S. Secretary of State.

Embodiment one

The invention discloses a kind of knowledge mapping and represent learning method, its schematic flow sheet is as in figure 2 it is shown, the method includes:

Step 21, utilize the model based on translation between entity vector and relation vector, interrelated between entity vector and relation vector in definition relation tlv triple (head, relation, tail).

Wherein, utilizing the model based on translation between entity vector and relation vector, in definition relation tlv triple, the method for being mutually related between entity vector and relation vector includes:

S211, definition relation tlv triple probability are $p\left(h\right|r,t,X)=\frac{\exp \left(g\right(h,r,t\left)\right)}{\underset{\stackrel{\‾}{h}}{\Σ}\exp \left(g\right(\stackrel{\‾}{h},r,t\left)\right)};$

Represent any entity in knowledge mapping;It is the normalization factor of relation tlv triple probability function, refers to all entity h in traversal knowledge mapping so that normalization factor is 1. $p\left(h\right|r,t,X)=\frac{\exp \left(g\right(h,r,t\left)\right)}{\underset{\stackrel{\‾}{h}}{\Σ}\exp \left(g\right(\stackrel{\‾}{h},r,t\left)\right)}$ It it is softmax function.

S212, utilize entity vector and relation vector between based on translation model, definition measurement relation r and entity to (h, the function g connected each other between t).

Between entity vector and relation vector, the model based on translation can have multiple, for instance, TransE and TransR etc., if what adopt is the energy function of TransE, then g can be defined as:

G (h, r, t)=-| | h+r-t | |_L1/L2+b₁

If what adopt is the energy function of TransR, then g can be defined as:

G (h, r, t)=-| | hM_r+r-tM_r||_L1/L2+b₁

Wherein, L1 is L1 normal form, and L2 is L2 normal form, M_rFor projection matrix relevant to relation in TransR model, b₁It is an offset constant, for making the average of g function return value be maintained at about 0.

It should be noted that relation tlv triple probability tables is shown as p (h | r, t, X) by the embodiment of the present invention, it is also possible to replace with p (t | r, h, X) or p (r | h, t, X). X is the vector representation of r, h, t.

Step 22, utilize neural network classification model, interrelated between entity vector and eigen vector in defined property tlv triple (entity, attribute, value).

Wherein, utilizing neural network classification model, in defined property tlv triple, the method for being mutually related between entity vector and eigen vector includes:

S221, define the first characteristic tlv triple probability and be $p\left(v\right|e,a,X)=\frac{\exp \left(k\right(e,a,v\left)\right)}{\underset{\stackrel{\‾}{e}}{\Σ}\exp \left(k\right(\stackrel{\‾}{e},a,v\left)\right)};$

Represent any entity in knowledge mapping;It is the normalization factor of the first characteristic tlv triple probability function, refers to all entity e in traversal knowledge mapping so that normalization factor is 1. $p\left(v\right|e,a,X)=\frac{\exp \left(k\right(e,a,v\left)\right)}{\underset{\stackrel{\‾}{e}}{\Σ}\exp \left(k\right(\stackrel{\‾}{e},a,v\left)\right)}$ It it is softmax function.

S222, utilize neural network classification model, definition weigh characteristic-value (a, v) and the function k connected each other between entity e.

When neural network classification model is monolayer neural networks model, k (e, a, v)=-| | f (eW_a+b_a)-V_av||_L1/L2+b₂; W_aAnd b_aFor the model parameter in monolayer neural networks model; | | f (eW_a+b_a)-V_av||_L1/L2Represent representing entity e in the vectorial subspace being projected to corresponding characteristic a by the neural network model of a monolayer, then calculate the similarity represented between vector of the vector sum correspondence characteristic value v after projection; b₂It is an offset constant, for making the average of k function return value be maintained at about 0.

Step 23, by evaluation function, entity vector, relation vector and eigen vector are associated, and minimize evaluation function, with learning object vector, relation vector and eigen vector, reach optimization aim.

Specifically include:

S231, definition evaluation function are O (X)=log (P (S, Y | X))+�� C (X);

S232, minimizing described evaluation function, study obtains each entity vector, relation vector and eigen vector in knowledge mapping. The method minimizing evaluation function can have multiple, it is possible to the method adopting stochastic gradient descent, etc.

S represents the set of all relation tlv triple in knowledge mapping, and Y represents the set of all characteristic tlv triple in knowledge mapping, and P (S, Y | X) represent the product of all relation tlv triple probability and all characteristic tlv triple probability in knowledge mapping; Characteristic tlv triple probability is the first characteristic tlv triple probability;

�� is hyper parameter, for controlling the weight of penalty term; C (X) is penalty, is used for preventing parameter learning over-fitting, and penalty C (X) definition is as follows:

$C\left(X\right)=\underset{e\∈E}{\Σ}{\[\left|\right|e\left|\right|-1\]}_{+}+\underset{r\∈R}{\Σ}{\[\left|\right|r\left|\right|-1\]}_{+}+\underset{e\∈E}{\Σ}\underset{i}{\Σ}{\[\left|\right|{\mathrm{eW}}_i+b_i\left|\right|-1\]}_{+}+\underset{i}{\Σ}{\[\left|\right|V_i\left|\right|-1\]}_{+},$ Wherein, [x]₊=max (0, x) represent that an input is for x, return value is the function of number bigger between 0 and x.

It should be noted that the process minimizing evaluation function is exactly the process reaching optimization aim. If the g in relation tlv triple probability function, what adopt is TransE model, then minimize in the process of evaluation function, by constantly adjusting h, r and the t vector of relation and tail (head), make each (h+r) in P (S | X) equal with t as far as possible, i.e. h+r=t. If the k in the first characteristic tlv triple probability function, what adopt is monolayer neural networks model, then minimize in the process of evaluation function, by constantly adjusting e, a and v (entity, the vector of attribute and value) so that in P (Y | X), the value v probability of each entity e correspondence characteristic a is 100%.

Thus, study obtains each entity vector, relation vector and eigen vector in knowledge mapping. According to prior art knowledge mapping, Fig. 3 a represents that the triple table that learning method obtains advises the exemplary plot of knowledge. According to knowledge mapping of the present invention, Fig. 3 b represents that the triple table that learning method obtains advises the exemplary plot of knowledge. In Fig. 3 a, not differentiating between relation and the characteristic of entity, still using the characteristic one as relation, knowledge mapping adopts the form of (entity 1, relation, entity 2) tlv triple to represent knowledge. And in Fig. 3 b, distinguish relation and the characteristic of entity, knowledge mapping adopts the form of relation tlv triple and characteristic tlv triple to represent knowledge. It can be seen that e6, e7, e8 and e9 are characteristic value from Fig. 3 b, e6, e7 belong to the value of a kind of characteristic A1, and e8 and e9 belongs to the value of another kind of characteristic A2, and specifically, in characteristic tlv triple, the value of entity e1 correspondence characteristic A1 is e6; The value of entity e2 correspondence characteristic A1 is e6; The value of entity e3 correspondence characteristic A1 is e7; The value of entity e3 correspondence characteristic A2 is e8; The value of entity e5 correspondence characteristic A2 is e8; The value of entity e4 correspondence characteristic A2 is e9. In simul relation tlv triple, the relation of entity e1 and entity e3 is r1; The relation of entity e1 and entity e2 is r5; The relation of entity e2 and entity e4 is r4; The relation of entity e3 and entity e4 is r3; The relation of entity e3 and entity e5 is r2; The relation of entity e4 and entity e5 is r4. It can thus be seen that the prior art of Fig. 3 a is compared with the present invention of Fig. 3 b, the knowledge mapping of the present invention represents learning method, it is possible to accurately represent connecting each other between entity, relation and characteristic.

Embodiment two

Treat owing to the knowledge mapping of the present invention represents that relation and characteristic are distinguished by learning method, it is possible to further consider connecting each other between characteristic.

The knowledge mapping of the embodiment of the present invention two represents that learning method comprises the following steps:

Step 31, utilize the model based on translation between entity vector and relation vector, interrelated between entity vector and relation vector in definition relation tlv triple (head, relation, tail).

Wherein, utilizing the model based on translation between entity vector and relation vector, in definition relation tlv triple, the method for being mutually related between entity vector and relation vector includes:

S311, definition relation tlv triple probability are $p\left(h\right|r,t,X)=\frac{\exp \left(g\right(h,r,t\left)\right)}{\underset{\stackrel{\‾}{h}}{\Σ}\exp \left(g\right(\stackrel{\‾}{h},r,t\left)\right)};$ Represent any entity in knowledge mapping;

S312, utilize entity vector and relation vector between based on translation model, definition measurement relation r and entity to (h, the function g connected each other between t).

Between entity vector and relation vector, the model based on translation can have multiple, for instance, TransE and TransR etc., if what adopt is the energy function of TransE, then g can be defined as:

G (h, r, t)=-| | h+r-t | |_L1/L2+b₁

If what adopt is the energy function of TransR, then g can be defined as:

G (h, r, t)=-| | hM_r+r-tM_r||_L1/L2+b₁

Wherein, L1 is L1 normal form, and L2 is L2 normal form, M_rFor projection matrix relevant to relation in TransR model, b₁It is an offset constant, for making the average of g function return value be maintained at about 0.

It should be noted that relation tlv triple probability tables is shown as p (h | r, t, X) by the embodiment of the present invention, it is also possible to replace with p (t | r, h, X) or p (r | h, t, X). X is the vector representation of r, h, t.

Step 32, utilize neural network classification model, interrelated between entity vector and eigen vector in defined property tlv triple (entity, attribute, value).

Wherein, utilizing neural network classification model, in defined property tlv triple, the method for being mutually related between entity vector and eigen vector includes:

S321, define the first characteristic tlv triple probability and be $p\left(v\right|e,a,X)=\frac{\exp \left(k\right(e,a,v\left)\right)}{\underset{\stackrel{\‾}{e}}{\Σ}\exp \left(k\right(\stackrel{\‾}{e},a,v\left)\right)};$ Represent any entity in knowledge mapping.

S322, utilize neural network classification model, definition weigh characteristic-value (a, v) and the function k connected each other between entity e.

When neural network classification model is monolayer neural networks model, k (e, a, v)=-| | f (eW_a+b_a)-V_av||_L1/L2+b₂; W_aAnd b_aFor the model parameter in monolayer neural networks model; | | f (eW_a+b_a)-V_av||_L1/L2Represent representing entity e in the vectorial subspace being projected to corresponding characteristic a by the neural network model of a monolayer, then calculate the similarity represented between vector of the vector sum correspondence characteristic value v after projection; b₂It is an offset constant, for making the average of k function return value be maintained at about 0.

Step 323,

Define the second characteristic tlv triple Probability p ((e, a, v) | X) �� p (v | e, a, X) p (v | e, a, Y (e)); Wherein, $p\left(v\right|e,a,Y\left(e\right))=\frac{\exp \left(z\right(e,a,v,Y\left(e\right)\left)\right)}{\underset{\stackrel{\‾}{v}\∈V_a}{\Σ}\exp \left(z\right(e,a,\stackrel{\‾}{v},Y\left(e\right)\left)\right)},$ Y (e) for entity e in knowledge mapping except characteristic-value (a, all known characteristics outside v); $p\left(v\right|e,a,Y\left(e\right))=\frac{\exp \left(z\right(e,a,v,Y\left(e\right)\left)\right)}{\underset{\stackrel{\‾}{v}\∈V_a}{\Σ}\exp \left(z\right(e,a,\stackrel{\‾}{v},Y\left(e\right)\left)\right)}$ It it is softmax function.

Step 324, definition weigh characteristic-value (a, v) and the function z connected each other between other characteristic-values. Assume characteristic-value (the combination positive correlation of every other characteristic-value in a, conditional probability v) and knowledge mapping, it is defined as: $z(e,a,v,Y(e))\∝\underset{(e,\stackrel{\‾}{a},\stackrel{\‾}{v})\∈Y\left(e\right)}{\Σ}P\left(\right(a,v\left)\right|\left(\stackrel{\‾}{a},\stackrel{\‾}{v}\right))(A_a\·A_{\stackrel{\‾}{a}});$ Wherein,For A_aWithInner product, for weighing degree of contact between two characteristics;Characteristic is had for known entitiesTime (a, conditional probability v).

Step 33, by evaluation function, entity vector, relation vector and eigen vector are associated, and minimize evaluation function, with learning object vector, relation vector and eigen vector, reach optimization aim.

Specifically include:

S331, definition evaluation function are O (X)=log (P (S, Y | X))+�� C (X);

S232, minimizing described evaluation function, study obtains each entity vector, relation vector and eigen vector in knowledge mapping.

S represents the set of all relation tlv triple in knowledge mapping, and Y represents the set of all characteristic tlv triple in knowledge mapping, and P (S, Y | X) represent the product of all relation tlv triple probability and all characteristic tlv triple probability in knowledge mapping; Characteristic tlv triple probability is the second characteristic tlv triple probability;

�� is hyper parameter, for controlling the weight of penalty term; C (X) is penalty, is used for preventing parameter learning over-fitting, and penalty C (X) definition is as follows:

$C\left(X\right)=\underset{e\∈E}{\Σ}{\[\left|\right|e\left|\right|-1\]}_{+}+\underset{r\∈R}{\Σ}{\[\left|\right|r\left|\right|-1\]}_{+}+\underset{e\∈E}{\Σ}\underset{i}{\Σ}{\[\left|\right|{\mathrm{eW}}_i+b_i\left|\right|-1\]}_{+}+\underset{i}{\Σ}{\[\left|\right|V_i\left|\right|-1\]}_{+},$ Wherein, [x]₊=max (0, x) represent that an input is for x, return value is the function of number bigger between 0 and x.

Further, calculating softmax function owing to there being many places to need in former evaluation function, wherein the amount of calculation of normalization item is very big, greatly reduces the speed of algorithm, the embodiment of the present invention is before minimizing evaluation function, it is preferred to use softmax function is similar to by negative sampling algorithm.

When including relation tlv triple Probability p (h | r, t, X) and the first characteristic tlv triple probability when evaluation function, softmax function therein is converted into formula calculated below by negative sampling by us:

$p\left(h\right|r,t,X)=\underset{(h,r,t)\∈S}{\Π}\[\mathrm{\σ}\left(g\right(h,r,t\left)\right)\underset{i=1}{\overset{C}{\Π}}{E}_{(h_i,r,t)~P\left(S^{-}\right)}\mathrm{\σ}\left(g\right(h_i,r,t\left)\right)\];$

$p\left(v\right|e,a,X)=\underset{(h,r,t)\∈Y}{\Π}\[\mathrm{\σ}\left(k\right(e,a,v\left)\right)\underset{i=1}{\overset{C}{\Π}}{E}_{(e,a,v_i)~P\left(Y^{-}\right)}\mathrm{\σ}\left(k\right(e,a,v_i\left)\right)\].$

When evaluation function include p (h | r, t, X) and the second characteristic tlv triple Probability p ((e, a, v) | X) �� p (v | e, a, X) p (v | e, a, Y (e)) time, softmax function therein is converted into formula calculated below by negative sampling by us:

$p\left(h\right|r,t,X)=\underset{(h,r,t)\∈S}{\Π}\[\mathrm{\σ}\left(g\right(h,r,t\left)\right)\underset{i=1}{\overset{C}{\Π}}{E}_{(h_i,r,t)~P\left(S^{-}\right)}\mathrm{\σ}\left(g\right(h_i,r,t\left)\right)\];$

$p\left(v\right|e,a,X)=\underset{(h,r,t)\∈Y}{\Π}\[\mathrm{\σ}\left(k\right(e,a,v\left)\right)\underset{i=1}{\overset{C}{\Π}}{E}_{(e,a,v_i)~P\left(Y^{-}\right)}\mathrm{\σ}\left(k\right(e,a,v_i\left)\right)\];$

$p\left(v\right|e,a,Y\left(e\right))=\underset{(h,r,t)\∈Y}{\Π}\[\mathrm{\σ}\left(z\right(e,a,v,Y\left(e\right)\left)\underset{i=1}{\overset{C}{\Π}}{E}_{(e,a,v_i)~P\left(Y^{-}\right)}\mathrm{\σ}\right(z\left(e,a,v_i,Y\left(e\right)\right)\].$

Obviously, above-mentioned formula $p\left(h\right|r,t,X)=\underset{(h,r,t)\∈S}{\Π}\[\mathrm{\σ}\left(g\right(h,r,t\left)\right)\underset{i=1}{\overset{C}{\Π}}{E}_{(h_i,r,t)~P\left(S^{-}\right)}\mathrm{\σ}\left(g\right(h_i,r,t\left)\right)\],$ Can also be replaced by:

$p\left(r\right|h,t,X)=\underset{(h,r,t)\∈S}{\Π}\[\mathrm{\σ}\left(g\right(h,r,t\left)\right)\underset{i=1}{\overset{C}{\Π}}{E}_{(h_i,r,t)~P\left(S^{-}\right)}\mathrm{\σ}\left(g\right(h,r_i,t\left)\right)\];$ Or,

$p\left(h\right|r,t,X)=\underset{(h,r,t)\∈S}{\Π}\[\mathrm{\σ}\left(g\right(h,r,t\left)\right)\underset{i=1}{\overset{C}{\Π}}{E}_{(h_i,r,t)~P\left(S^{-}\right)}\mathrm{\σ}\left(g\right(h_i,r,t\left)\right)\].$

��=1/ (1+exp (-x)) is sogmoid function, S^-For the negative example set of relation tlv triple, P (S^-) bear the probability function of all elements, Y in example set for relation tlv triple^-For the negative example set of characteristic tlv triple, P (Y^-) bear the probability function of all elements in example set for characteristic tlv triple.

Relation tlv triple bears example set S^-Producing method as follows: we are by each tlv triple (head in positive example set S, relation, tail) random with arbitrarily replacing head entity head for his entity or random with arbitrarily replacing tail entity tail for his entity or random with arbitrarily replacing head entity relation for his relation, so we will produce a new tlv triple not appeared in positive example set, we are treated as a negative example, and the combination of all this negative examples just constitutes S^-��

Same, characteristic tlv triple bears example set Y^-Producing method as follows: we are by each tlv triple (entity in positive example set Y, attribute, it is value) random that with arbitrarily replacing existing characteristic value value for his possible characteristic value, so we will produce a new tlv triple not appeared in positive example set, we are treated as a negative example, and the combination of all this negative examples just constitutes Y^-��

To sum up, the invention has the beneficial effects as follows: compared with prior art, the while that the present invention proposing, in learning knowledge collection of illustrative plates, the knowledge mapping of entity, relation and personality presentation represents learning method, what solve to exist in prior art cannot Precise Representation entity, the problem of contact between relation and characteristic, to improve the quality that knowledge graph represents, there is good practicality.

The foregoing is only presently preferred embodiments of the present invention, be not intended to limit protection scope of the present invention. All any amendment of making, equivalent replace and improvement etc. within the spirit and principles in the present invention, should be included within protection scope of the present invention.

Claims (7)

1. knowledge mapping represents a learning method, and the method includes:

Utilize the model based on translation between entity vector and relation vector, interrelated between entity vector and relation vector in definition relation tlv triple (head, relation, tail);

Utilize neural network classification model, in defined property tlv triple (entity, attribute, value) entity vector and eigen vector between interrelated;

By evaluation function, entity vector, relation vector and eigen vector are associated, and minimize evaluation function, with learning object vector, relation vector and eigen vector, reach optimization aim.

2. the method for claim 1, it is characterised in that

Utilizing the model based on translation between entity vector and relation vector, in definition relation tlv triple, the method for being mutually related between entity vector and relation vector includes:

Definition relation tlv triple probability is Represent any entity in knowledge mapping;

Utilizing the model based on translation between entity vector and relation vector, definition measurement relation r and entity are to (h, the function g connected each other between t).

3. method as claimed in claim 2, it is characterised in that

Utilizing neural network classification model, in defined property tlv triple, the method for being mutually related between entity vector and eigen vector includes:

Defining the first characteristic tlv triple probability is Represent any entity in knowledge mapping;

Utilize neural network classification model, definition weigh characteristic-value (a, v) and the function k connected each other between entity e.

4. method as claimed in claim 3, it is characterised in that when neural network classification model is monolayer neural networks model, k (e, a, v)=-| | f (eW_a+b_a)-V_av||_L1/L2+b₂;

W_aAnd b_aFor the model parameter in monolayer neural networks model; | | f (eW_a+b_a)-V_av||_L1/L2Represent representing entity e in the vectorial subspace being projected to corresponding characteristic a by the neural network model of a monolayer, then calculate the similarity represented between vector of the vector sum correspondence characteristic value v after projection; b₂It is an offset constant, for making the average of k function return value be maintained at about 0.

5. method as claimed in claim 4, it is characterised in that after defined function k, the method farther includes:

Define the second characteristic tlv triple Probability p ((e, a, v) | X) �� p (v | e, a, X) p (v | e, a, Y (e)); Wherein,Y (e) for entity e in knowledge mapping except characteristic-value (a, all known characteristics outside v);

Definition weigh characteristic-value (a, v) and the function z connected each other between other characteristic-values; Assume characteristic-value (the combination positive correlation of every other characteristic-value in a, conditional probability v) and knowledge mapping, it is defined as:Wherein,For A_aWithInner product, for weighing degree of contact between two characteristics;Characteristic is had for known entitiesTime (a, conditional probability v).

6. the method described in claim 5, it is characterised in that

Entity vector, relation vector and eigen vector being associated by evaluation function, and minimize evaluation function, with each entity vector, relation vector and eigen vector in learning knowledge collection of illustrative plates, the method reaching optimization aim includes:

Definition evaluation function is O (X)=log (P (S, Y | X))+�� C (X);

Minimizing described evaluation function, study obtains each entity vector, relation vector and eigen vector in knowledge mapping;

Wherein, S represents the set of all relation tlv triple in knowledge mapping, and Y represents the set of all characteristic tlv triple in knowledge mapping, and P (S, Y | X) represent the product of all relation tlv triple probability and all characteristic tlv triple probability in knowledge mapping; Characteristic tlv triple probability is the first characteristic tlv triple probability or the second characteristic tlv triple probability;

�� is hyper parameter, for controlling the weight of penalty term; C (X) is penalty, is used for preventing parameter learning over-fitting, and penalty C (X) definition is as follows:

Wherein, [x]₊=max (0, x), it is a function returning number bigger between 0 and x.

7. method as claimed in claim 6, it is characterised in that before minimizing, adopts negative sampling algorithm, the softmax function in evaluation function is similar to, to accelerate pace of learning;

Softmax function

Softmax function

Softmax function

��=1/ (1+exp (-x)) is sogmoid function, S^-For the negative example set of relation tlv triple, P (S^-) bear the probability function of all elements, Y in example set for relation tlv triple^-For the negative example set of characteristic tlv triple, P (Y^-) bear the probability function of all elements in example set for characteristic tlv triple.