CN109960810A

CN109960810A - A kind of entity alignment schemes and device

Info

Publication number: CN109960810A
Application number: CN201910243854.5A
Authority: CN
Inventors: 何芙珍; 李直旭; 陈志刚
Original assignee: Hkust Technology (suzhou) Technology Co Ltd
Current assignee: Hkust Technology (suzhou) Technology Co Ltd
Priority date: 2019-03-28
Filing date: 2019-03-28
Publication date: 2019-07-02
Anticipated expiration: 2039-03-28
Also published as: CN109960810B

Abstract

This application discloses a kind of entity alignment schemes and devices, this method iteratively executes two screening steps and obtains first instance to set, and, two steps are as follows: according to objective attribute target attribute to filtering out target entity pair in two knowledge mappings, and new objective attribute target attribute pair filtered out in two knowledge mappings according to target entity.Due to first instance to each target entity in set to being obtained by the attribute information between different entities, and each entity attributes information more really can comprehensively represent the entity, thus, when carrying out entity alignment using entity attributes information, the accuracy of entity alignment result can be improved.

Description

A kind of entity alignment schemes and device

Technical field

This application involves knowledge mapping technical field more particularly to a kind of entity alignment schemes and devices.

Background technique

With the continuous development and breakthrough of artificial intelligence, knowledge mapping (Knowledge Graph, abbreviation KG) is used as future The technology foundation stone for realizing strong artificial intelligence, has received widespread attention.At this stage when constructing knowledge mapping, it will usually from major hundred Triple is collected in the semi-structured text of section website, or is extracted from various non-structured texts using information extraction technique Triple, for constructing knowledge mapping, wherein there are two class triples, one kind be relationship triple, it is another kind of be attribute ternary Group.

Due to major encyclopaedia and the non-structured text in various sources, for entity itself and relationship and attribute There are semantic gaps in expression, therefore, for a unified, consistent, succinct form that different knowledge mappings permeates, are Using between the application program of different knowledge mappings interaction provide semantic interoperability, scholars propose semantic intergration this Research topic.And entity alignment is an important prerequisite and technological means for semantic intergration, its target is by two isomeries Those of same target in the real world entity is directed toward in knowledge mapping to find out.

But since different knowledge mappings is for the different expression of relational structure between entity various aspects information and entity Larger, so entity alignment work is challenging, and existing entity alignment techniques are mainly based upon between entity Relationship carry out entity alignment, but in this way obtained by entity alignment as a result, being also unable to reach higher accuracy.

Summary of the invention

The main purpose of the embodiment of the present application is to provide a kind of entity alignment schemes and device, in knowing two isomeries When knowing map progress entity alignment, the accuracy of entity alignment result can be improved.

The embodiment of the present application provides a kind of entity alignment schemes, comprising:

In two knowledge mappings, known each objective attribute target attribute pair is determined, refer to attribute pair as each；

Each target entity pair is filtered out in two knowledge mappings with reference to attribute according to each；

According to the target entity pair filtered out, each new objective attribute target attribute pair is filtered out in two knowledge mappings, as It is each to refer to attribute pair, and continue to execute and described filter out each target in two knowledge mappings with reference to attribute according to each The step of entity pair forms first instance to set until it can not filter out target entity pair；

Wherein, the objective attribute target attribute is identical to two included attributes and the two attributes are belonging respectively to two knowledge Map；The target entity is identical to two included entities and the two entities are belonging respectively to two knowledge mappings.

Optionally, described to filter out each target entity pair in two knowledge mappings with reference to attribute according to each, it wraps It includes:

According to it is each with reference to attribute to the attribute value of two respectively included attributes, filtered out in two knowledge mappings Each target entity pair.

Optionally, it is described according to each with reference to attribute to the attribute value of two respectively included attributes, in two knowledge Each target entity pair is filtered out in map, comprising:

Determine that each initial solid pair, the initial solid refer to attribute to at least one in two knowledge mappings To and have it is each with reference to attribute to a corresponding attribute value similarity, the attribute value similarity is corresponding to refer to attribute pair Similarity between the attribute value of two included attributes；

According at least one attribute value similarity of the initial solid pair, determine the initial solid to whether belonging to institute State target entity pair.

Optionally, described at least one attribute value similarity according to the initial solid pair, determines the initial solid To whether belonging to the target entity pair, comprising:

Calculate the average value of at least one attribute value similarity of the initial solid pair；

If the average value being calculated is greater than the first preset threshold, determine the initial solid to for the target entity It is right.

Optionally, the target entity pair that the basis filters out filters out each new target in two knowledge mappings Attribute pair, comprising:

For each target entity pair filtered out, each to be selected attribute of the target entity under is subjected to group two-by-two It closes, obtains the attribute pair to be selected under every kind of combination, the attribute to be selected is not belonging to two included attributes fixed each A objective attribute target attribute to and be belonging respectively to two entities of the target entity centering；

The attribute to be selected is calculated to the similarity between the attribute value of two included attributes；

If the similarity being calculated is greater than the second preset threshold, determine the attribute to be selected to for new objective attribute target attribute It is right.

Optionally, the method also includes:

The entity alignment model obtained using preparatory training, filters out each target entity pair in two knowledge mappings, Second instance is formed to set, the entity alignment model is used to screen entity pair based on entity relationship.

Optionally, the entity alignment model is trained using model training data, the model training Data include the target entity pair high to the correctness filtered out in set from the first instance.

Optionally, after the formation second instance is to set, further includes:

Merge the first instance to set and the second instance to set, forms third entity to set；

From the third entity to the low target entity pair of accuracy is rejected in set, as the 4th entity to set.

Optionally, the target entity pair low to rejecting accuracy in set from the third entity, comprising:

For belonging to target entity pair of the first instance to set with the second instance to set simultaneously, according to this The first similarity and the second similarity of target entity pair, determine the final similarity of the target entity pair；

Wherein, first similarity is the target entity that obtains when forming the first instance to set to being wrapped The similarity between two entities included, second similarity are the mesh obtained when forming the second instance to set Entity is marked to the similarity between two included entities；

If the final similarity of the target entity pair is less than third predetermined threshold value, by the target entity to from the third Entity in set to rejecting.

Optionally, the final similarity of the determination target entity pair, comprising:

Based on first similarity and the respective confidence level of the second similarity, the final of the target entity pair is determined Similarity.

Optionally, the confidence level is to be arrived using the regression model constructed in advance in the middle school's acquistion of model learning data, The model learning data include the target entity pair high to the correctness filtered out in set from the first instance.

The embodiment of the present application also provides a kind of entity alignment means, comprising:

With reference to attribute to acquiring unit, in two knowledge mappings, determining known each objective attribute target attribute pair, as It is each to refer to attribute pair；

Target entity is to screening unit, for filtering out each mesh in two knowledge mappings with reference to attribute according to each Mark entity pair；

With reference to attribute to screening unit, for being filtered out in two knowledge mappings according to the target entity pair filtered out Each new objective attribute target attribute pair refers to attribute pair as each；

Target entity refers to attribute to Cycle Screening unit, for calling the reference attribute pair filtered out, and according to each Each target entity pair is filtered out in two knowledge mappings, until it can not filter out target entity pair, forms first Entity is to set；

Optionally, the target entity is specifically used for screening unit:

Optionally, the target entity is to screening unit, comprising:

Initial solid determines subelement, for determining each initial solid pair, the initial reality in two knowledge mappings Body to at least one with reference to attribute to and have it is each with reference to attribute to a corresponding attribute value similarity, the attribute Value similarity is the corresponding attribute that refers to the similarity between the attribute value of two included attributes；

Target entity sentences determining subelement at least one attribute value similarity according to the initial solid pair Whether the fixed initial solid is to belonging to the target entity pair.

Optionally, the target entity is to determining subelement, comprising:

Similarity mean value calculation module, for calculate the initial solid pair at least one attribute value similarity it is flat Mean value；

Target entity is to determining module, if average value for being calculated is greater than the first preset threshold, determine described in Initial solid is to for the target entity pair.

Optionally, the reference attribute is to screening unit, comprising:

Attribute to be selected is to acquisition subelement, for each target entity pair for filtering out, by the target entity under Each attribute to be selected carry out combination of two, obtain the attribute pair to be selected under every kind of combination, the attribute to be selected is to included Two attributes be not belonging to fixed each objective attribute target attribute to and be belonging respectively to two entities of the target entity centering；

Attributes similarity computation subunit, for calculate the attribute to be selected to the attribute values of two included attributes it Between similarity；

Objective attribute target attribute, if the similarity for being calculated is greater than the second preset threshold, determines institute to subelement is determined Attribute to be selected is stated to for new objective attribute target attribute pair.

Based on the above-mentioned technical proposal, the application has the advantages that

Entity alignment schemes provided by the present application iteratively execute two screening steps and obtain first instance to collection It closes, moreover, two steps are as follows: filter out target entity pair in two knowledge mappings according to objective attribute target attribute, and according to mesh Mark entity filters out new objective attribute target attribute pair in two knowledge mappings.Since first instance is real to each target in set Body is obtained by the attribute information between different entities to being, and each entity attributes information can more it is true comprehensively Ground represents the entity, thus, when carrying out entity alignment using entity attributes information, it can be improved the accurate of entity alignment result Property.In addition, since objective attribute target attribute according to target entity to can obtain to screening from knowledge mapping, so that semantic identical but table Two attributes different up to mode can make up objective attribute target attribute pair, overcome the asking for can not being aligned due to attribute expression way multiplicity Topic, to further increase the accuracy of entity alignment result.Further, since objective attribute target attribute to and target entity to be repeatedly It generates during generation, without using including the training data for the entity pair being largely aligned in advance, overcomes because of training data matter The low problem of the accuracy of low caused entity alignment result is measured, to improve the accuracy of entity alignment result.

Detailed description of the invention

In order to illustrate the technical solutions in the embodiments of the present application or in the prior art more clearly, to embodiment or will show below There is attached drawing needed in technical description to be briefly described, it should be apparent that, the accompanying drawings in the following description is the application Some embodiments for those of ordinary skill in the art without creative efforts, can also basis These attached drawings obtain other attached drawings.

Fig. 1 is the structural schematic diagram of different knowledge mappings provided by the embodiments of the present application；

Fig. 2 is the flow chart for the entity alignment schemes that the application embodiment of the method one provides；

Fig. 3 is the flow example figure provided by the embodiments of the present application for iteratively screening target entity pair；

Fig. 4 is the flow chart for the entity alignment schemes that the application embodiment of the method two provides；

Fig. 5 is the flow chart for the entity alignment schemes that the application embodiment of the method three provides；

Fig. 6 is the schematic diagram of the first set amalgamation result provided by the embodiments of the present application；

Fig. 7 is the schematic diagram of second provided by the embodiments of the present application set amalgamation result；

Fig. 8 is a kind of process signal of the specific embodiment for the entity alignment schemes that the application embodiment of the method three provides Figure；

Fig. 9 is a kind of flow chart of the specific embodiment for the entity alignment schemes that the application embodiment of the method three provides；

Figure 10 is the structural schematic diagram for the entity alignment means that the application Installation practice one provides.

Specific embodiment

Knowledge mapping can be used for describing attribute possessed by relationship and each entity between different entities.Wherein, Entity refers to the things that present in objective world and can be distinguished one another, and entity can be people, is also possible to object material object, It can also be abstract concept.For example, " Qiao Busi " and " San Francisco " is all entity.

Relationship between different entities refers to certain association being present between two entities, for example, according to " Qiao Busi goes out It is born in San Francisco " it is found that being there are associated between entity " Qiao Busi " and entity " San Francisco ", which is specifically: " Qiao Bu This " birthplace be " San Francisco ".

Attribute possessed by entity refers to certain characteristics possessed by entity itself, moreover, each attribute is related to attribute-name And attribute value.For example, according to " date of birth of Qiao Busi is 24 days 2 months nineteen fifty-five " it is found that possessed by entity " Qiao Busi " Attribute is to be born in nineteen fifty-five 24 days 2 months, wherein " date of birth " is attribute-name, and " 24 days 2 months nineteen fifty-five " is attribute value.

In addition, when constructing knowledge mapping, it will usually triple is collected from the semi-structured text of major encyclopaedia website, Or triple is extracted from various non-structured texts using information extraction technique.Wherein, triple can be using unification Representation: (subject subject, predicate predicate, object object)；Moreover, in some cases, triple can be with Using certain special representation, specifically: it can be with for describing between different entities certain associated relationship triple Using the representation of (entity entity, relationship relation, entity entity), for example, " Qiao Busi is born in San Francisco " It can be indicated with relationship triple (Qiao Busi, birthplace, San Francisco)；For describing the category of certain attribute of each entity Property triple can use (entity entity, attribute attribute, value value) representation, for example, " the body of Qiao Busi Height is that 188cm " can be indicated with attribute triple (Qiao Busi, height, 188cm).

However, due to the non-structured text of major encyclopaedia website and various sources, for entity itself and relationship There are semantic gaps in the expression of attribute, therefore, a unified, consistent, succinct in order to which different knowledge mappings permeates Form, the interaction between application program to use different knowledge mappings provides semantic interoperability, and scholars propose language Justice integrates this research topic.And entity alignment be semantic intergration an important prerequisite and technological means, its target be by Those of same target in the real world entity is directed toward in the knowledge mapping of two isomeries to find out.However, due to different Knowledge mapping is larger for the different expression of relational structure between entity various aspects information and entity, so entity is aligned work It is challenging.

For the ease of explanation and understanding, the entity statement difference in different knowledge mappings is said below in conjunction with Fig. 1 It is bright, wherein Fig. 1 is the structural schematic diagram of different knowledge mappings provided by the embodiments of the present application.

As an example, Fig. 1 includes the first knowledge mapping KG₁With the second knowledge mapping KG₂.Wherein, the first knowledge mapping KG₁ Including first instanceSecond instanceWith third entitySecond instanceWith third entityBetween have the first relationshipFirst instanceAnd second instanceBetween have the second relationshipFor example, second instanceThe attribute information having and pass Be that information specifically includes: " birth-time " is " 1955-2-24 ", and " name " is " Steve Jobs ", and " birth-place " is " San Francisco, California, USA ", " height " is " 188cm ".Second knowledge mapping KG₂Including the 4th entity5th entityWith the 6th entity5th entityWith the 6th entityBetween have third relationship4th entityWith the 5th entityBetween have the 4th relationshipFor example, the 5th entityThe attribute information and relation information having are specific It include: " birthdate " is " 1955.02.24 ", " name " is " Steve Jobs ", and " birth-place " is " San Francisco ", " nick name " are " Apple godfather ", and " height " is " 188centi-meter ".

As can be seen from FIG. 1, second instanceWith the 5th entityIt is directed to " Qiao Busi " this people, still, for calculating For machine, judge " birth-time " and " birthdate " be same attribute, " San Francisco, California, USA " and " San Francisco " are same entity and " 188cm " and " 188centi-meter " are that equal attribute value is equal Be it is very difficult, so result in entity alignment work be challenging.

In the prior art, what the thought that entity alignment schemes are normally based on term vector insertion (embedding) carried out, It will be illustrated by taking two kinds of common entity alignment schemes as an example below.

The first entity alignment schemes be by the entity in knowledge mapping and the relationship map between different entities to In quantity space, the similarity between different entities is obtained by the distance between calculating vector, that is, independent of any text This information gets deep structure information of the entity on entire knowledge mapping, and carries out entity alignment based on this.

Second of entity alignment schemes is improved based on the first entity alignment schemes, moreover, this method Improvements are as follows: according to the semantic description and attribute of relationship, each entity between different entities, obtain each entity Vector indicates, so overcomes and leads because the first entity alignment schemes only considers between different entities relationship to a certain extent The low defect of the entity alignment result accuracy of cause.

But it has been investigated that, above two entity alignment schemes have the disadvantage that

The defect of the first entity alignment schemes is: since this method needs the entity being largely aligned in advance to as training Data, but the acquisition of high quality training data be it is very difficult, thus, the quality of training data used in this method compared with Low, the accuracy so as to cause the entity alignment result of this method is low.In addition, due between the entity in knowledge mapping relationship have Sparsity, that is to say, that each entity in knowledge mapping and other entities only have it is seldom contact, even without connection (example Such as, the stand-alone entity in knowledge mapping), therefore, entity alignment only is carried out using entity relationship, will lead to entity alignment result Accuracy is low.

The defect of second of entity alignment schemes is: since this method is the improved method of the first entity alignment schemes, It still has the shortcomings that the first entity alignment schemes.In addition, in the building process of entity vector, in order to get around attribute-name and Attribute value has been simplified to attribute Value Types (for example, date type, numeric type etc.), caused by the diversity that attribute value table reaches Noise in attribute value is larger, from without effectively utilize attribute information, and then cause entity alignment result accuracy still It is lower.

And attribute information is effectively utilized when carrying out entity alignment in the embodiment of the present application, and then improves entity pair The accuracy of neat result.

To keep the purposes, technical schemes and advantages of the embodiment of the present application clearer, below in conjunction with the embodiment of the present application In attached drawing, the technical scheme in the embodiment of the application is clearly and completely described, it is clear that described embodiment is Some embodiments of the present application, instead of all the embodiments.Based on the embodiment in the application, those of ordinary skill in the art Every other embodiment obtained without making creative work, shall fall in the protection scope of this application.

Embodiment of the method one

Referring to fig. 2, which is the flow chart for the entity alignment schemes that the application embodiment of the method one provides.

Entity alignment schemes provided by the embodiments of the present application, comprising:

S201: in two knowledge mappings, determining known each objective attribute target attribute pair, refers to attribute pair as each.

S202: each target entity pair is filtered out in two knowledge mappings with reference to attribute according to each.

S203: whether the number for the target entity pair that judgement currently filters out is 0, if so, executing S206；If it is not, then Execute S204.

S204: according to the target entity pair filtered out, each new objective attribute target attribute pair is filtered out in two knowledge mappings.

S205: each new objective attribute target attribute pair that will be filtered out refers to attribute pair as each, and returns and execute S202.

S206: all target entities pair that will have been filtered out form first instance to set.

The above are the specific execution steps for the entity alignment schemes that the application embodiment of the method one provides, in order to facilitate understanding With the entity alignment schemes for explaining that the application embodiment of the method one provides, the specific implementation of S201 to S206 will be successively introduced below Mode.

The specific embodiment of S201 is introduced first.

In S201, each objective attribute target attribute is to including two attributes, wherein the two attributes are identical, and the two attributes It is belonging respectively to two knowledge mappings.As an example, first object attribute is to including the first attribute and the second attribute, wherein first Attribute and the second attribute are identical, moreover, the first attribute belongs to first knowledge mapping, the second attribute belongs to second knowledge graph Spectrum.

In addition, the first attribute and the identical attribute-name that can refer to the first attribute of the second attribute and attribute value are respectively with The attribute-name and attribute value of two attributes are identical, and the attribute-name and attribute value that may also mean that the first attribute are respectively with The attribute-name of two attributes and the semanteme of attribute value are identical.

For the ease of explanation and understanding objective attribute target attribute pair, it is illustrated below in conjunction with Fig. 1 and two example.

First example specifically: as shown in Figure 1, when the first attribute is second instanceAn attribute, and the first attribute Attribute it is entitled " name ", the attribute value of the first attribute is " Steve Jobs "；Moreover, the second attribute is the 5th entityOne A attribute, and the attribute of the second attribute is entitled " name ", when the attribute value of the second attribute is " Steve Jobs ", due to the first category Property attribute-name and attribute value it is identical with the attribute-name of the second attribute and attribute value, thus, the first attribute and the Two attributes are identical；Moreover, because second instanceBelong to the first knowledge mapping KG₁, and the 5th entityBelong to the second knowledge graph Compose KG₂, thus, belong to second instanceThe first attribute belong to the first knowledge mapping KG₁, belong to the 5th entity? Two attributes belong to the second knowledge mapping KG₂, so that the first attribute and the second attribute are belonging respectively to two different knowledge mappings.It can See, since the attribute-name and attribute value of the first attribute are identical with the attribute-name of the second attribute and attribute value respectively, and And first attribute from the second attribute be belonging respectively to two different knowledge mappings, thus, the first attribute and the second attribute composition one A objective attribute target attribute pair.

The above are the first exemplary related contents, in this example, are distinguished with the attribute-name of the first attribute and attribute value To objective attribute target attribute to being introduced for identical with the attribute-name of the second attribute and attribute value.

Second example specifically: as shown in Figure 1, when the first attribute is second instanceAn attribute, and the first attribute Attribute it is entitled " birth-time ", the attribute value of the first attribute is " 1955-2-24 "；Moreover, the second attribute is the 5th entityAn attribute, and the attribute of the second attribute is entitled " birthdate ", and the attribute value of the second attribute is " 1955.02.24 " When, since attribute-name " birth-time " and " birthdate " they are to indicate the date of birth, attribute value " 1955-2-24 " and " 1955.02.24 " is to indicate nineteen fifty-five 24 days 2 months, thus, the attribute-name and attribute value of the first attribute are and the second attribute Attribute-name and attribute value semanteme it is identical so that the first attribute and the second attribute are identical；Moreover, because second instanceBelong to In the first knowledge mapping KG₁, and the 5th entityBelong to the second knowledge mapping KG₂, thus, belong to second instanceFirst Attribute belongs to the first knowledge mapping KG₁, belong to the 5th entityThe second attribute belong to the second knowledge mapping KG₂, so that first Attribute and the second attribute are belonging respectively to two different knowledge mappings.As it can be seen that due to the attribute-name and attribute value of the first attribute It is identical as the semanteme of the attribute-name of the second attribute and attribute value respectively, and the first attribute and the second attribute are belonging respectively to difference Two knowledge mappings, thus, the first attribute and the second attribute form an objective attribute target attribute pair.

The above are the second exemplary related contents, in this example, are distinguished with the attribute-name of the first attribute and attribute value To objective attribute target attribute to being introduced for identical as the semanteme of the attribute-name of the second attribute and attribute value.

Additionally, it is known that each objective attribute target attribute to can be preset attribute pair, be also possible to advance with default The attribute pair that algorithm obtains.Wherein, preset algorithm can be any algorithm that can determine objective attribute target attribute pair, and the application is implemented Example is not specifically limited in this embodiment.

Since known each objective attribute target attribute is to that can use different acquisition modes, thus, S201 can be adopted accordingly With numerous embodiments, below will in one embodiment for be illustrated.

As an implementation, S201 is specifically as follows: utilizing preset algorithm, in two knowledge mappings, determines The each objective attribute target attribute pair known, so that each objective attribute target attribute is referred to attribute pair to as each.

The above are the specific embodiments of S201, in this embodiment, can use preset algorithm from two knowledge graphs Each objective attribute target attribute pair is determined in spectrum, and determining each objective attribute target attribute is referred into attribute pair to as each.

The specific embodiment of S202 is explained below.

In S202, each target entity is to including two entities, wherein the two entities are identical (that is, the two entities The content of reference is identical), and the two entities are belonging respectively to two knowledge mappings.

As an example, as shown in Figure 1, second instanceWith the 5th entityThat all refer to is " Qiao Busi " this people, and And second instanceBelong to the first knowledge mapping KG₁, the 5th entityBelong to the second knowledge mapping KG₂, thus, second instance With the 5th entityTarget entity pair can be formed.

Moreover, the embodiment of the present application can be determined according to entity attributes information first instance and second instance whether phase Together.

Since each attribute information may include attribute-name and attribute value, thus can be according to the category of each attribute of entity Property name and attribute value determine target entity pair.At this point, in order to further increase the accuracy of target entity pair, the application is provided A kind of embodiment of S202, in this embodiment, S202 is specifically as follows: being wrapped with reference to attribute to respective according to each The attribute value of two attributes included filters out each target entity pair in two knowledge mappings.

For the ease of the embodiment of the S202 of the above-mentioned offer of explanation and understanding, it is illustrated below in conjunction with Fig. 1.

As an example it is supposed that by S201 obtain it is each with reference to attribute to include first with reference to attribute to and second reference Attribute pair.Specifically, the first reference attribute is to including the first attribute and the second attribute, wherein the first attribute is second instance An attribute, the attribute of the first attribute entitled " birth-time " and attribute value are " 1955-2-24 ", and the second attribute is the Five entitiesAn attribute, the attribute of the second attribute entitled " birthdate " and attribute value are " 1955.02.24 "；Second With reference to attribute to including third attribute and the 4th attribute, wherein third attribute is second instanceAnother attribute, third category Property attribute entitled " height " and attribute value be " 188cm ", the 4th attribute is the 5th entityAnother attribute, the 4th The attribute of attribute entitled " height " and attribute value are " 188centi-meter ".

It is assumed immediately when above-mentioned, then S202 is specifically as follows: according to the category of the first attribute of the first reference attribute centering Property value " 1955-2-24 " and the second attribute attribute value " 1955.02.24 " and second with reference to attribute centering third attribute Attribute value " 188cm " and the 4th attribute attribute value " 188centi-meter ", in the first knowledge mapping KG₁With the second knowledge Map KG₂In filter out including second instanceWith the 5th entityTarget entity pair.

It should be noted that be above by according to two with reference to attributes to being illustrated for filtering out target entity pair , still, in this application, at least a pair of to can be with reference to attribute, the target entity being screened out is to being also possible at least A pair, moreover, according at least one with reference to attribute to the process and above embodiment for filtering out at least one target entity pair Identical, for the sake of brevity, details are not described herein.

A kind of embodiment of S202 based on above-mentioned introduction, in order to further increase the accuracy of target entity pair, into And the accuracy of entity alignment schemes is improved, present invention also provides the another embodiments of S202, in this embodiment, S202 can specifically include step S2021-S2022:

S2021: each initial solid pair is determined in two knowledge mappings.

Initial solid to at least one with reference to attribute to and have it is each with reference to attribute to a corresponding attribute value phase Like degree, which is the corresponding attribute that refers to the similarity between the attribute value of two included attributes.

It in order to facilitate understanding with explanation initial solid pair, is illustrated below in conjunction with Fig. 3, wherein Fig. 3 is that the application is real The flow example figure for iteratively screening target entity pair of example offer is provided.

As shown in Figure 3, it is assumed that the first knowledge mapping KG₁Including first instanceSecond instanceWith third entityThe Two knowledge mapping KG₂Including the 4th entity5th entityWith the 7th entityMoreover,Indicate the first knowledge mapping KG₁ In ith attribute,Indicate the second knowledge mapping KG₂In j-th of attribute, i and j are positive integer；Moreover, the first reference Attribute is to including attributeAnd attributeAnd first instanceAttributeWith the 4th entityAttributeAttribute value it is similar Degree is 0.78；Second reference attribute is to including attributeAnd attributeAnd first instanceAttributeWith the 4th entity's AttributeAttribute value similarity be 0.90；Third is with reference to attribute to including attributeAnd attributeAnd first instanceCategory PropertyWith the 4th entityAttributeAttribute value similarity be 0.85 and second instanceAttributeIt is real with the 5th BodyAttributeAttribute value similarity be 0.80；4th reference attribute is to including attributeAnd attributeAnd second instanceAttributeWith the 5th entityAttributeAttribute value similarity be 0.95；5th reference attribute is to including attributeWith AttributeAnd second instanceAttributeWith the 7th entityAttributeAttribute value similarity be 0.95.

It assumes immediately, then knows in the first knowledge mapping KG when above-mentioned₁With the second knowledge mapping KG₂In, first instance With the 4th entityIncluding first with reference to attribute to, second with reference to attribute to and third with reference to attribute to this three pairs refer to attribute It is right；Second instanceWith the 5th entityIncluding third with reference to attribute to and the 4th with reference to attribute to these two pair refer to attribute pair； Second instanceWith the 7th entityAttribute pair is referred to this pair with reference to attribute including the 5th.At this point, S2021 specifically can be with Are as follows: in the first knowledge mapping KG₁With the second knowledge mapping KG₂Middle determining first instanceWith the 4th entitySecond instanceWith 5th entityAnd second instanceWith the 7th entityRespectively initial solid pair.

The above are the specific embodiments of S2021, in this embodiment, can be known according to reference attribute pair from two Know and determines each initial solid pair in map.

S2022: according at least one attribute value similarity of initial solid pair, determine the initial solid to whether belonging to mesh Mark entity pair.

Attribute value similarity refers to the corresponding attribute that refers to the similarity between the attribute value of two included attributes.Example Such as, in Fig. 3, attribute value similarity 0.78 indicates first instanceAttributeAttribute value and the 4th entityAttribute Attribute value between similarity.

As an example, then S2022 is specifically as follows: according to initial when initial solid is to including N number of attribute value similarity M attribute value similarity of entity centering determines the initial solid to whether belonging to target entity pair；Wherein, M is positive integer, And M≤N.

As an implementation, in order to further increase the accuracy of target entity pair, and then entity alignment side is improved The accuracy of method, S2022 can specifically include S2022a-S2022b:

S2022a: the average value of at least one attribute value similarity of initial solid pair is calculated, as the initial solid pair Attributes similarity.

Attributes similarity refers to correspondent entity to the similarity between two included entity attributes.For example, such as Fig. 3 It is shown, when the first initial solid is to including first instanceWith the 4th entityWhen, then the attribute of the first initial solid pair is similar Degree can indicate first instanceAttribute and the 4th entityAttribute between similarity.

As an example, then S2022a is specifically as follows when initial solid is to including N number of attribute value similarity: calculating just The average value of M attribute value similarity of beginning entity pair, and using the average value as the attributes similarity of the initial solid pair；Its In, M is positive integer, and M≤N.

For the ease of explanation and understanding, it is illustrated below in conjunction with Fig. 3 and by taking M=N as an example.

As an example, when what S2021 was provided assuming immediately to what Fig. 3 was carried out, then known to the first initial solid to including First instanceWith the 4th entityAnd first initial solid to including 0.78,0.90 and 0.85 totally three attribute value similarities； Second initial solid is to including second instanceWith the 5th entityAnd second initial solid to including 0.80 and 0.95 totally two Attribute value similarity；Third initial solid is to including second instanceWith the 7th entityAnd third initial solid is to including 0.3 Totally one attribute value similarity.At this point, S2022a is specifically as follows: 0.78,0.90 and 0.85 average value 0.843 is calculated, and By the attributes similarity of the average value 0.843 as the first initial solid pair；0.80 and 0.95 average value 0.875 is calculated, and By the attributes similarity of the average value 0.875 as the second initial solid pair；0.3 average value 0.3 is calculated, and by the average value 0.3 attributes similarity as third initial solid pair.

It should be noted that be to be illustrated by taking M=N as an example to S2022a above, still, in this application, In S2022a, M can not only be equal to N, can also be any positive integer less than N, moreover, when M takes different value, S2022a's Implementation procedure is identical with above-mentioned example, and for the sake of brevity, details are not described herein.

The above are the specific embodiments of S2022a, in this embodiment, can calculate at least the one of initial solid pair The average value of a attribute value similarity, and using the average value as the attributes similarity of the initial solid pair.

S2022b: if the attributes similarity for the initial solid pair being calculated is greater than the first preset threshold, determine that this is first Beginning entity is to for target entity pair.

First preset threshold can be preset, for example, the first preset threshold can be set previously according to application scenarios.

As an example it is supposed that then S2022b is specifically as follows: judgement is initial when to preset the first preset threshold be 0.7 Whether the attributes similarity of entity pair is greater than the first preset threshold, if so, determining the initial solid to being target entity pair；If It is no, it is determined that the initial solid is not to being target entity pair.

For the ease of explanation and understanding S2022b, it is illustrated below in conjunction with Fig. 3.

As an example, when the S2021 hypothesis carried out to Fig. 3 provided is set up, and the attribute phase of the first initial solid pair It is 0.843 like degree, the attributes similarity of the second initial solid pair is 0.875, and the attributes similarity of third initial solid pair is 0.3, and preset the first preset threshold be 0.7 when, then S2022b is specifically as follows: since 0.843 and 0.875 are all larger than 0.7, thus, it is default that the attributes similarity of the attributes similarity of the first initial solid pair and the second initial solid pair is all larger than first Threshold value, at this point it is possible to determine the first initial solid to and the second initial solid to being target entity pair.

The above are the specific embodiments of S2022b, in this embodiment, can be calculated by judgement initial Whether whether the attributes similarity of entity pair is greater than the first preset threshold, determine the initial solid to being target entity pair.

The above are the specific embodiments of S202, in this embodiment, can be according to each each to determination with reference to attribute The attributes similarity of a entity pair, and according to the attributes similarity of each entity pair, it is filtered out from two knowledge mappings each Target entity pair.In this way, since each entity attributes more really comprehensively can represent the entity, thus, according to entity pair The accuracy of target entity pair that filters out of attributes similarity it is higher.

The specific embodiment of S203 is described below.

In S203, the target entity that currently filters out is to referring in the currently screening period by executing step S202 sieve The target entity pair selected.

Since the application can iteratively execute step S202 to S205, the screening to target entity pair is realized Journey, thus, the screening process of target entity pair provided by the present application may include at least one screening period.In addition, each Screening the period in can by judge currently screen the period in whether can filter out target entity pair, come determine whether after It is continuous to carry out next screening period, moreover, the determination process is specifically as follows: if can be filtered out in the currently screening period At least twin target entity pair then continues to execute next screening period；If mesh can not be filtered out in the currently screening period Entity pair is marked, then the screening process of target end entity pair, and will be obtained in all screening periods before the currently screening period The target entity pair obtained forms first instance to set.

The above are the specific embodiments of S203, in this embodiment, the target that can be currently filtered out by judgement Whether the number of entity pair is 0, it is determined whether continues to execute next screening period.

The specific embodiment of S204 is described below.

It is identical due to belonging to the reference content of two entities of same target entity pair in S204, and every A entity includes multiple attributes, therefore, it is possible to determine two categories with identical semanteme for being respectively belonging to two entities Property is also identical.In this way, each new mesh can be filtered out in two knowledge mappings according to the target entity pair filtered out Mark attribute pair.

As an implementation, S204 can specifically include step S2041-S2043:

S2041: for each target entity pair filtered out, each to be selected attribute of the target entity under is carried out two Two combinations, obtain the attribute pair to be selected under every kind of combination.

Attribute to be selected to two included attributes be not belonging to fixed each objective attribute target attribute to and be belonging respectively to the mesh Mark two entities of entity centering.

For the ease of explanation and understanding attribute pair to be selected, it is illustrated below in conjunction with Fig. 3.

When the hypothesis carried out to Fig. 3 that S2021 is provided is set up, and first instanceWith the 4th entityFor first object Entity clock synchronization, then can be by first instanceAttributeAnd attributeAnd the 4th entityAttributeAnd attributeMake For the attribute to be selected of first object entity pair, and by attributeAnd attributeRespectively with attributeAnd attributeCarry out group two-by-two It closes, obtains four groups of attributes pair to be selected: including attributeAnd attributeThe first attribute to be selected to including attributeAnd attribute's Second attribute to be selected to including attributeAnd attributeThird attribute to be selected to and including attributeAnd attribute? Four attributes pair to be selected.

In addition, working as second instanceWith the 5th entityIt, then can be by second instance for the second target entity clock synchronization's AttributeAnd attributeAnd the 5th entityAttributeAnd attributeAs the attribute to be selected of the second target entity pair, and By attributeAnd attributeRespectively with attributeAnd attributeCombination of two is carried out, obtains four groups of attributes pair to be selected: including attributeAnd attributeThe 5th attribute to be selected to including attributeAnd attributeThe 6th attribute to be selected to including attributeAnd category PropertyThe 7th attribute to be selected to and including attributeAnd attributeThe 8th attribute pair to be selected.

It should be noted that it is above-mentioned be with according to first object entity to and the second target entity to obtaining attribute pair to be selected Process for the specific embodiment of S2041 is explained and illustrated, still, in this application, do not limit S2041 only Above embodiment can be used, can also other embodiments, for the sake of brevity, details are not described herein.

S2042: attribute to be selected is calculated to the similarity between the attribute value of two included attributes, as the category to be selected The attribute value similarity of property pair.

As an example, when the first attribute to be selected is to including attributeAnd attributeWhen, then S2042 is specifically as follows: calculating The attribute of first attribute centering to be selectedAttribute value and attributeAttribute value between similarity, as the first attribute to be selected Pair attribute value similarity.

S2043: if the attribute value similarity for the attribute pair to be selected being calculated is greater than the second preset threshold, determining should be to Select attribute to for new objective attribute target attribute pair.

Second preset threshold can be preset, for example, the second preset threshold can be set previously according to practical application scene It is fixed.

For the ease of explanation and understanding S2043, it is illustrated below in conjunction with Fig. 3.

As an example, when the S2021 hypothesis carried out to Fig. 3 provided is set up, and the first attribute to be selected is to including attributeAnd attributeAnd first attribute pair to be selected attribute value similarity be 0.8；Second attribute to be selected is to including attributeAnd attributeAnd second attribute pair to be selected attribute value similarity be 0.2；Third attribute to be selected is to including attributeAnd attributeAnd the The attribute value similarity of three attributes pair to be selected is 0.1；4th attribute to be selected is to including attributeAnd attributeAnd the 4th category to be selected Property pair attribute value similarity be 0.15；5th attribute to be selected is to including attributeAnd attributeAnd the 5th attribute pair to be selected Attribute value similarity is 0.2；6th attribute to be selected is to including attributeAnd attributeAnd the 6th attribute pair to be selected attribute value Similarity is 0.3；7th attribute to be selected is to including attributeAnd attributeAnd the 7th attribute pair to be selected attribute value similarity It is 0.85；8th attribute to be selected is to including attributeAnd attributeAnd the 8th the attribute value similarity of attribute pair to be selected be 0.13；And second preset threshold be 0.6 when, then due to there was only the attribute value similarity 0.8 and the 7th of the first attribute pair to be selected The attribute value similarity 0.85 of attribute pair to be selected is all larger than the second preset threshold 0.6, then can determine: including attributeAnd attributeThe first attribute to be selected to include attributeAnd attributeThe 7th attribute to be selected to being new objective attribute target attribute pair.

The above are the specific embodiments of S204, in this embodiment, can be according to the target entity filtered out to obtaining Take at least a pair of attribute pair to be selected, and according to the attribute value similarity of each attribute pair to be selected, determine the attribute to be selected to whether For new objective attribute target attribute pair, new objective attribute target attribute pair can be so extracted from existing target entity pair, in order to should New objective attribute target attribute to as reference attribute to and continue next screening period.

The specific embodiment of S205 is described below.

As an implementation, when the total quantity of new objective attribute target attribute pair is N number of, then S205 is specifically as follows: by N A new objective attribute target attribute refers to attribute pair to as N number of, to be screened according to N number of attribute that refers to in two knowledge mappings Each new target entity pair out.

For the ease of explanation and understanding S205, it is illustrated below in conjunction with Fig. 3.

As an example, when the S2021 hypothesis carried out to Fig. 3 provided is set up, and including attributeAnd attribute? One attribute to be selected to include attributeAnd attributeThe 7th attribute to be selected to being new objective attribute target attribute clock synchronization, then S205 It is specifically as follows: will will include attributeAnd attributeThe first attribute to be selected to as the 6th refer to attribute pair, and will include belong to PropertyAnd attributeThe 7th attribute to be selected to as the 7th refer to attribute pair, so as to it is subsequent can in S202 according to the 6th ginseng Examine attribute to and the 7th refer to attribute pair, in the first knowledge mapping KG₁With the second knowledge mapping KG₂In filter out including third reality BodyWith the 7th entityThird target entity pair.

The above are the specific embodiments of S205, in this embodiment, each new target category that can will be filtered out Property pair, as it is each refer to attribute pair, and return execute S202, so as to based on this with reference to attribute to next screening period progress The screening of target entity pair.

The specific embodiment of S206 is described below

In S206, all target entities for having filtered out are to referring to that the target screened within all screening periods is real Body pair.

For the ease of explanation and understanding S206, it is illustrated below in conjunction with Fig. 3.

As an example, when the S2021 hypothesis carried out to Fig. 3 provided is set up, and screened within first screening period Having gone out includes first instanceWith the 4th entityFirst object entity pair, and including second instanceWith the 5th entity The second target entity pair；Having filtered out within second screening period includes second instanceWith the 7th entityThird mesh Mark entity clock synchronization, then S206 is specifically as follows: by first object entity to, the second target entity to and third target entity into Row set obtains first instance to set.

The above are the specific embodiments of S206, in this embodiment, all target entities that can will have been filtered out To gathering, first instance is obtained to set.

The above are the specific embodiments of embodiment of the method one, in this embodiment, iteratively execute two sieves Step is selected to obtain first instance to set, moreover, two steps are as follows: screen according to objective attribute target attribute in two knowledge mappings Target entity pair out, and new objective attribute target attribute pair is filtered out in two knowledge mappings according to target entity.In fact due to first Body obtains each target entity in set to being by the attribute information between different entities, and each entity Attribute information more really comprehensively can represent the entity, thus, it, can when carrying out entity alignment using entity attributes information Improve the accuracy of entity alignment result.In addition, since objective attribute target attribute according to target entity to from knowledge mapping to can sieve Choosing obtains, so that semantic two identical but different expression way attributes can make up objective attribute target attribute pair, overcomes because of attribute list Up to mode multiplicity the problem of can not be aligned, thus the further accuracy for improving entity alignment result.Further, since target Attribute to and target entity to generating in an iterative process, without using including the instruction for the entities pair being largely aligned in advance Practice data, overcome because training data quality it is low caused by entity alignment result the low problem of accuracy, to improve reality The accuracy of body alignment result.

The entity alignment schemes that above method embodiment one provides, it is real by obtaining at least twin target using entity attribute Body pair, since entity attribute more really comprehensively can represent the entity, thus, the accuracy of target entity pair is improved, from And improve the accuracy of entity alignment result.

In addition, in order to further increase the accuracy of target entity pair, to further increase the standard of entity alignment result True property can also obtain target entity pair using entity attribute and entity relationship simultaneously, thus, present invention also provides another kinds Entity alignment schemes, are explained and illustrated below in conjunction with attached drawing.

Embodiment of the method two

Embodiment of the method is second is that the improvement carried out on the basis of embodiment of the method one, for the sake of brevity, method are implemented Part identical with content in embodiment of the method one in example two, details are not described herein.

Referring to fig. 4, which is the flow chart for the entity alignment schemes that the application embodiment of the method two provides.

Entity alignment schemes provided by the embodiments of the present application, including S401-S407:

It should be noted that S401 to S406 is identical as the S201 to S206 in embodiment of the method one, for the sake of brevity, Details are not described herein.

S407: it is real to filter out each target in two knowledge mappings for the entity alignment model obtained using preparatory training Body pair forms second instance to set.

It should be noted that the application do not limit S407 execute sequence, can be executed before S401-S406 or it It executes or synchronizes afterwards and execute.

In S407, entity alignment model is used to screen target entity pair based on entity relationship, moreover, entity is aligned mould Type can be any model using entity relationship progress entity to screening.

As an example, entity alignment model can be any model based on term vector insertion (embedding), and And in the model, each pair of entity can map in vector space corresponding relationship triple (h, r, t) and obtainsTo measure the similarity between different entities using distance between the vector between different entities, it is, real The relationship similarity of body pair.Wherein, h indicates the head entity of relationship triple；R indicates the head entity and relationship three of relationship triple Relationship between the tail entity of tuple；The tail entity of t expression relationship triple；The head entity of expression relationship triple is corresponding Vector；The corresponding vector of expression relationship；The corresponding vector of tail entity of expression relationship triple.

Entity alignment model, which can be, advances with what model training data were trained, which can To include at least twin target entity pair, moreover, model training data source is extensive.Wherein, model training data can be by Training dataset of at least a pair of target entity manually marked to composition；Model training data are also possible to by least a pair of of benefit The target entity obtained with default dimensioning algorithm is to the training data of composition, and default dimensioning algorithm can be preset.

For the ease of explanation and understanding entity alignment model, below by a kind of embodiment to train entity alignment model For be illustrated.

As an implementation, in order to improve the efficiency and accuracy of entity alignment schemes, above-mentioned default mark is calculated Method can be any entity alignment schemes of the offer of embodiment of the method one (it is, any reality of step S401 to S406 Apply mode), thus, the model training data of entity alignment model may include correct to what is filtered out in set from first instance The high target entity pair of property.

Wherein, first instance to set is generated by S406 step；Moreover, from first instance to filtering out in set The process of the high target entity pair of correctness is specifically as follows: firstly, for first instance to each target entity in set It is right, the average value of the attribute value similarity of at least one same alike result of the target entity pair is determined, using the average value as mesh Mark the alignment correctness of entity pair；Secondly, the alignment correctness of each target entity pair and default correctness threshold value are compared Compared with acquisition alignment correctness is higher than the target entity pair of the default correctness threshold value, as model training data.

The above are the specific embodiments of S407, in this embodiment, can use by from first instance in set The high target entity of the correctness filtered out is trained, and utilize the model training data of composition to entity alignment model The entity alignment model that training obtains filters out each target entity pair in two knowledge mappings, forms second instance to collection It closes.

The above are the specific embodiments for the entity alignment schemes that embodiment of the method two provides, in this embodiment, when After obtaining first instance to set according to entity attribute, the entity alignment model based on entity relationship can also be utilized, Each target entity pair is filtered out in two knowledge mappings, forms second instance to set.It so ensure that this finally obtained A little target entities are to being both to obtain based on entity attribute and based on entity relationship, so that it is real to improve the target finally obtained The accuracy of body pair, and then improve the entity alignment accuracy of result and comprehensive.

The entity alignment schemes that above method embodiment one and embodiment of the method two provide, can according to entity attribute and/ Or entity relationship obtains target entity pair.

In addition, in order to further increase the accuracy of entity alignment result, can also attribute to target entity pair it is similar Degree and relationship similarity carry out overall merit, to obtain final target entity to set, thus, present invention also provides Another entity alignment schemes, is explained and illustrated below in conjunction with attached drawing.

Embodiment of the method three

Embodiment of the method is third is that the improvement carried out on the basis of embodiment of the method two, for the sake of brevity, method are implemented Part identical with content in embodiment of the method two in example three, details are not described herein.

Referring to Fig. 5, which is the flow chart for the entity alignment schemes that the application embodiment of the method three provides.

Entity alignment schemes provided by the embodiments of the present application, including S501-S509:

It should be noted that S501 to S507 is identical as the S401 to S407 in embodiment of the method two, for the sake of brevity, Details are not described herein.

S508: merge first instance to set and second instance to set, form third entity to set.

S509: from third entity to the low target entity pair of accuracy is rejected in set, as the 4th entity to set.

The above are the specific execution steps for the entity alignment schemes that the application embodiment of the method three provides, in order to facilitate understanding With the entity alignment schemes for explaining that the application embodiment of the method three provides, the specific implementation of S508 and S509 will be successively introduced below Mode.

The specific embodiment of S508 is introduced first.

S508 can use three kinds of embodiments, successively be introduced below in conjunction with attached drawing.

As the first embodiment, as shown in fig. 6, S508 is specifically as follows: by first instance to included in set Target entity pair and second instance to target entity included in set to gathering, obtain third entity to collection It closes.

The above are the first embodiments of S508, in this embodiment, can be by first instance to set and the Two entities, to gathering, can form third entity to set to all target entities in set.

In addition, third entity can be made in set, and there is no phases in order to further increase the efficiency of entity alignment schemes Same target entity pair, thus, present invention also provides second of embodiment of S508 and the third embodiments, below will Successively introduce.

As second of embodiment, as shown in fig. 7, S508 is specifically as follows: obtaining first instance to set and second Entity is to union of sets collection, and using the union as third entity to set.

As the third embodiment, S508 is specifically as follows: firstly, by first instance to target included in set Entity pair and second instance, to gathering, obtain initial third entity to set to target entity included in set (as shown in Figure 6)；Then, from initial third entity to duplicate target entity pair is deleted in set, third entity is obtained to collection It closes (as shown in Figure 7).

The above are three kinds of embodiments of S508, in this embodiment, can merge first instance to set and second Entity forms third entity to set to set.

The specific embodiment of S509 is described below.

In S509, the accuracy of target entity pair can judge according to multiple indexs, for example, target entity pair Accuracy can be judged according only to attributes similarity, can also be judged according only to relationship similarity, can also basis The integrated value of attributes similarity and relationship similarity is judged.

It, below will be to judge target according to the integrated value of attributes similarity and relationship similarity for the ease of explanation and understanding It is illustrated for the accuracy of entity pair.

As an implementation, in order to improve the accuracys of entity alignment schemes, S509 can specifically include S5091- S5092:

S5091: for belonging to first instance simultaneously to set and second instance to the target entity pair of set, according to the mesh The first similarity and the second similarity for marking entity pair, determine the final similarity of the target entity pair.

Wherein, while belonging to first instance to set and second instance to the target entity pair of set, be first instance pair Set and second instance are to the entity pair in intersection of sets collection (intersection as shown in Figure 7).

First similarity is to form first instance to the target entity obtained when gathering to two included entities Between similarity (it is, attributes similarity of entity pair), the second similarity be formed second instance to set when obtain The target entity arrived is to the similarity (it is, relationship similarity of entity pair) between two included entities.

As an implementation, S5091 is specifically as follows: for and meanwhile belong to first instance to set and second instance To the target entity pair of set, according to the first similarity and the second similarity of the target entity pair, and it is based on the first similarity With the respective confidence level of the second similarity, the final similarity of the target entity pair is determined.

Wherein, confidence level can be preset, for example, the confidence level of the confidence level of the first similarity and the second similarity It can be set previously according to application scenarios.

In addition, in order to further increase the accuracy of confidence level, so that the accuracy of final similarity is further increased, into And the accuracy of entity alignment schemes is improved, confidence level can use the regression model constructed in advance in model learning data middle school What acquistion was arrived, and the model learning data include the target entity high to the correctness filtered out in set from first instance It is right.Wherein, it is referred to from the specific screening process of the first instance target entity pair high to the correctness filtered out in set " the target entity high to the correctness filtered out in set from first instance provided in the specific embodiment of step S407 Pair process ", for the sake of brevity, details are not described herein.

Based on the related content of above-mentioned confidence level, as an implementation, S5091 can specifically include S50911- S50913:

S50911: using the regression model constructed in advance the middle school's acquistion of model learning data to formula (1) in set Reliability parameter.

Wherein, model learning data include the target entity pair high to the correctness filtered out in set from first instance, Moreover, it is also possible to by model learning data withForm is indicated.

In formula,Indicate i-th of entity in the first knowledge mapping,Indicate j-th of reality in the second knowledge mapping Body, moreover,WithTwin target entity pair can be constituted；Expression includesWithTarget entity pair The first similarity；Expression includesWithTarget entity pair the second similarity； Expression includesWithTarget entity pair final similarity；λ indicates confidence level parameter.

S50912: according to confidence level parameter lambda, the confidence level of the first similarity and the confidence level of the second similarity are obtained.

As an implementation, S50912 is specifically as follows: using 1- λ as the confidence level of the first similarity, and λ being made For the confidence level of the second similarity.

S50913: for belonging to first instance simultaneously to set and second instance to the target entity pair of set, according to this The first similarity and the second similarity of target entity pair, and it is based on the first similarity and the respective confidence level of the second similarity, Determine the final similarity of the target entity pair.

As an implementation, when the confidence level of the first similarity is 1- λ, and the confidence level of the second similarity is λ, Then S50913 is specifically as follows: for and meanwhile belong to first instance to set and second instance target entity pair to set, root According to the first similarity of the target entity pair, the second similarity, the confidence level of the confidence level of the first similarity and the second similarity, Using formula (2), the final similarity of the target entity pair is obtained.

Final similarity=first similarity × (1- λ)+the second similarity × λ (2)

In formula, 1- λ indicates the confidence level of the first similarity；λ indicates the confidence level of the second similarity.

S5092: if the final similarity of the target entity pair is less than third predetermined threshold value, by the target entity to from the Three entities will reject the entity after operation to set, as the 4th entity to set to rejecting in set.

Third predetermined threshold value can be preset, for example, third predetermined threshold value can be determined previously according to application scenarios.

The above are the specific embodiments of S509, in this embodiment, can be quasi- to rejecting in set from third entity The low target entity pair of exactness, as the 4th entity to set.

In addition, in order to understand and explain entity alignment schemes provided by the embodiments of the present application, below in conjunction with Fig. 8 to entity A kind of specific embodiment of alignment schemes is illustrated.

The concrete meaning of each symbol in Fig. 8: KG is introduced first₁Indicate the first knowledge mapping；KG₂Indicate the second knowledge graph Spectrum；Indicate i-th of entity in the first knowledge mapping；Indicate the attribute-name of the ith attribute in the first knowledge mapping； It indicates in the first knowledge mappingCorresponding attribute value；Indicate t-th of relationship in the first knowledge mapping；Indicate second J-th of entity in knowledge mapping；Indicate the attribute-name of j-th of attribute in the second knowledge mapping；Indicate the second knowledge In mapCorresponding attribute value；Indicate s-th of relationship in the second knowledge mapping；Indicate first instance to set In m-th of target entity pair attributes similarity (it is, first similarity)；Indicate second instance in set N-th of target entity pair relationship similarity (it is, second similarity).

Then, a kind of specific embodiment party of the entity alignment schemes of the offer of embodiment of the method three is provided in conjunction with Fig. 8 and Fig. 9 Formula, in this embodiment, entity alignment schemes are specifically as follows:

S901: from the first knowledge mapping KG₁With the second knowledge mapping KG₂In, obtain attribute triple and relationship ternary Group.

Wherein, attribute triple is for indicating each entity attributes information；Relationship triple is for indicating different entities Between relation information.

S902: using default normalizing algorithm, the attribute value in attribute triple is standardized.

Wherein, normalizing algorithm is preset to be used to be converted into using the attribute value of different expression ways using same expression side The attribute value of formula；Moreover, default normalizing algorithm can use any standardized algorithm, for example, default normalizing algorithm can be adopted With the canonical matching algorithm based on artificial constituting criterion.

It is indicated as an example, " date of birth " corresponding attribute value can use " 1955-02-24 ", it can also benefit It is indicated with " 02/24/1955 ", " 24th Feb.1955 " can also be utilized to be indicated, at this point, S902 is specifically as follows: Using default normalizing algorithm, " 1955-02-24 ", " 02/24/1955 " and " 24th Feb.1955 " is standardized respectively, Obtain " 1955/02/24 ".

S903: according to attribute triple, step 501 is executed to 506 corresponding specific embodiments, obtains first instance pair Set.

Wherein, step 501 can be realized the process of " interactive mode " in Fig. 8 to 506 corresponding specific embodiments.

S904: by the first instance target entity high to the correctness filtered out in set to as model training data, Entity alignment model is trained.

S905: it is corresponding specific to execute step S507 for the entity alignment model obtained according to relationship triple and training Embodiment obtains second instance to set.

S906: it is corresponding specific to be executed to set to set and second instance according to first instance by step S508 to S509 Embodiment obtains the 4th entity to set.

The above are a kind of embodiments of entity alignment schemes, in this embodiment, can will be in attribute triple Attribute value standardizes, and can be avoided attribute value noise problem caused by the expression way multiplicity because of attribute value, improves real The accuracy and consistency of body attribute description to improve the accuracy of target entity pair, and then improve entity alignment As a result accuracy.

The above are the specific embodiments of the entity alignment schemes of the offer of embodiment of the method three in this embodiment can To merge the first instance of acquisition to set to set and second instance, third entity is obtained to set, and from third Entity is to the low target entity pair of accuracy is rejected in set, as the 4th entity to set.It is accurate due to target entity pair Degree can be obtained by the first similarity and the second similarity of the overall merit target entity pair, moreover, the first similarity is The attributes similarity of the target entity pair, the second similarity is the relationship similarity of the target entity pair, so that the target entity Pair accuracy can be obtained by the attributes similarity and relationship similarity of the overall merit target entity pair, thus, improve The accuracy of entity alignment result and comprehensive.

Based on any entity alignment schemes that above method embodiment one to embodiment of the method three provides, the application is also mentioned A kind of entity alignment means have been supplied, have been explained and illustrated below in conjunction with attached drawing.

Installation practice one

Referring to Figure 10, which is the structural schematic diagram for the entity alignment means that the application Installation practice one provides.

Entity alignment means 1000 provided by the embodiments of the present application, comprising:

With reference to attribute to acquiring unit 1001, in two knowledge mappings, determining known each objective attribute target attribute pair, Attribute pair is referred to as each；

Target entity is to screening unit 1002, for according to each each to filtering out in two knowledge mappings with reference to attribute A target entity pair；

With reference to attribute to screening unit 1003, for being sieved in two knowledge mappings according to the target entity pair filtered out Each new objective attribute target attribute pair is selected, refers to attribute pair as each,

Target entity is to Cycle Screening unit 1004, for calling the reference attribute pair filtered out, and according to each reference Attribute filters out each target entity pair in two knowledge mappings, until it can not filter out target entity pair, is formed First instance is to set；

As an implementation, in order to further increase the accuracy that entity is aligned result, the target entity is to sieve Menu member 1002, is specifically used for:

As an implementation, in order to further increase the accuracy that entity is aligned result, the target entity is to sieve Menu member 1002, comprising:

As an implementation, in order to further increase the accuracy that entity is aligned result, the target entity is to true Stator unit, comprising:

As an implementation, in order to further increase the accuracy that entity is aligned result, the reference attribute is to sieve Menu member 1003, comprising:

As an implementation, in order to further increase entity alignment result accuracy, the entity alignment means 1000 further include:

Second instance is to set generation unit, for training obtained entity alignment model using preparatory, in two knowledge Each target entity pair is filtered out in map, forms second instance to set, the entity alignment model is used to close based on entity System's screening entity pair.

As an implementation, in order to further increase entity alignment result accuracy, the entity alignment model It is to be trained using model training data, the model training data include from the first instance to sieving in set The high target entity pair of the correctness selected.

Third entity is to set generation unit, for merging the first instance after forming second instance to set Third entity is formed to set to set with the second instance to set；

4th entity is to set generation unit, for real to the low target of accuracy is rejected in set from the third entity Body pair, as the 4th entity to set.

As an implementation, in order to further increase the accuracy that entity is aligned result, the 4th entity is to collection Close generation unit, comprising:

Final similarity determines subelement, for for belonging to the first instance simultaneously to set and the second instance The target entity pair is determined according to the first similarity and the second similarity of the target entity pair to the target entity pair of set Final similarity；

Target entity is to subelement is rejected, if the final similarity for the target entity pair is less than third predetermined threshold value, Then by the target entity to from the third entity to being rejected in set.

As an implementation, in order to further increase the accuracy that entity is aligned result, the final similarity is true Stator unit is specifically used for:

As an implementation, in order to further increase the accuracy that entity is aligned result, the confidence level is to utilize The regression model constructed in advance is arrived in the middle school's acquistion of model learning data, and the model learning data include real from described first The body target entity pair high to the correctness filtered out in set.

Further, the embodiment of the present application also provides a kind of entity alignment apparatus, comprising: processor, memory, system Bus；

The processor and the memory are connected by the system bus；

The memory includes instruction, described instruction for storing one or more programs, one or more of programs The processor is set to execute any one implementation of above-mentioned entity alignment schemes when being executed by the processor.

Further, described computer-readable to deposit the embodiment of the present application also provides a kind of computer readable storage medium Instruction is stored in storage media, when described instruction is run on the terminal device, so that the terminal device executes above-mentioned entity Any one implementation of alignment schemes.

Further, the embodiment of the present application also provides a kind of computer program product, the computer program product exists When being run on terminal device, so that the terminal device executes any one implementation of above-mentioned entity alignment schemes.

As seen through the above description of the embodiments, those skilled in the art can be understood that above-mentioned implementation All or part of the steps in example method can be realized by means of software and necessary general hardware platform.Based on such Understand, substantially the part that contributes to existing technology can be in the form of software products in other words for the technical solution of the application It embodies, which can store in storage medium, such as ROM/RAM, magnetic disk, CD, including several Instruction is used so that a computer equipment (can be the network communications such as personal computer, server, or Media Gateway Equipment, etc.) execute method described in certain parts of each embodiment of the application or embodiment.

It should be noted that each embodiment in this specification is described in a progressive manner, each embodiment emphasis is said Bright is the difference from other embodiments, and the same or similar parts in each embodiment may refer to each other.For reality For applying device disclosed in example, since it is corresponded to the methods disclosed in the examples, so being described relatively simple, related place Referring to method part illustration.

It should also be noted that, herein, relational terms such as first and second and the like are used merely to one Entity or operation are distinguished with another entity or operation, without necessarily requiring or implying between these entities or operation There are any actual relationship or orders.Moreover, the terms "include", "comprise" or its any other variant are intended to contain Lid non-exclusive inclusion, so that the process, method, article or equipment including a series of elements is not only wanted including those Element, but also including other elements that are not explicitly listed, or further include for this process, method, article or equipment Intrinsic element.In the absence of more restrictions, the element limited by sentence "including a ...", it is not excluded that There is also other identical elements in process, method, article or equipment including the element.

The foregoing description of the disclosed embodiments makes professional and technical personnel in the field can be realized or use the application. Various modifications to these embodiments will be readily apparent to those skilled in the art, as defined herein General Principle can be realized in other embodiments without departing from the spirit or scope of the application.Therefore, the application It is not intended to be limited to the embodiments shown herein, and is to fit to and the principles and novel features disclosed herein phase one The widest scope of cause.

Claims

1. a kind of entity alignment schemes characterized by comprising

According to the target entity pair filtered out, each new objective attribute target attribute pair is filtered out in two knowledge mappings, as each With reference to attribute pair, and continues to execute and described filter out each target entity in two knowledge mappings with reference to attribute according to each Pair step, until it can not filter out target entity pair, formed first instance to set；

Wherein, the objective attribute target attribute is identical to two included attributes and the two attributes are belonging respectively to two knowledge mappings； The target entity is identical to two included entities and the two entities are belonging respectively to two knowledge mappings.

2. the method according to claim 1, wherein it is described according to it is each with reference to attribute in two knowledge mappings In filter out each target entity pair, comprising:

3. according to the method described in claim 2, it is characterized in that, it is described according to each with reference to attribute to respectively included two The attribute value of a attribute filters out each target entity pair in two knowledge mappings, comprising:

Determine each initial solid pair in two knowledge mappings, the initial solid to have at least one with reference to attribute to, And have it is each with reference to attribute to a corresponding attribute value similarity, the attribute value similarity refers to attribute to wrapping for correspondence Similarity between the attribute value of two attributes included；

According at least one attribute value similarity of the initial solid pair, determine the initial solid to whether belonging to the mesh Mark entity pair.

4. according to the method described in claim 3, it is characterized in that, described at least one attribute according to the initial solid pair It is worth similarity, determines the initial solid to whether belonging to the target entity pair, comprising:

If the average value being calculated is greater than the first preset threshold, determine the initial solid to for the target entity pair.

5. the method according to claim 1, wherein the target entity pair that the basis filters out, knows at two Know in map and filter out each new objective attribute target attribute pair, comprising:

For each target entity pair filtered out, each to be selected attribute of the target entity under is subjected to combination of two, is obtained Attribute pair to be selected under every kind of combination, the attribute to be selected are not belonging to fixed each target to two included attributes Attribute to and be belonging respectively to two entities of the target entity centering；

If the similarity being calculated is greater than the second preset threshold, determine the attribute to be selected to for new objective attribute target attribute pair.

6. method according to any one of claims 1 to 5, which is characterized in that the method also includes:

The entity alignment model obtained using preparatory training, filters out each target entity pair in two knowledge mappings, is formed To set, the entity alignment model is used to screen entity pair based on entity relationship second instance.

7. according to the method described in claim 6, it is characterized in that, the entity alignment model be using model training data into Row training obtains, and the model training data include the target high to the correctness filtered out in set from the first instance Entity pair.

8. according to the method described in claim 6, it is characterized in that, the formation second instance to set after, further includes:

9. according to the method described in claim 8, it is characterized in that, it is described from the third entity to rejecting accuracy in set Low target entity pair, comprising:

For belonging to target entity pair of the first instance to set with the second instance to set simultaneously, according to the target The first similarity and the second similarity of entity pair, determine the final similarity of the target entity pair；

Wherein, first similarity is to form the first instance to the target entity obtained when gathering to included Similarity between two entities, second similarity are the target realities obtained when forming the second instance to set Body is to the similarity between two included entities；

If the final similarity of the target entity pair is less than third predetermined threshold value, by the target entity to from the third entity To being rejected in set.

10. according to the method described in claim 9, it is characterized in that, the final similarity of the determination target entity pair, packet It includes:

Based on first similarity and the respective confidence level of the second similarity, the final similar of the target entity pair is determined Degree.

11. according to the method described in claim 10, it is characterized in that, the confidence level is to utilize the regression model constructed in advance It is arrived in the middle school's acquistion of model learning data, the model learning data include from the first instance to filtering out in set The high target entity pair of correctness.

12. a kind of entity alignment means characterized by comprising

With reference to attribute to acquiring unit, in two knowledge mappings, determining known each objective attribute target attribute pair, as each With reference to attribute pair；

Target entity is to screening unit, for according to each real to each target is filtered out in two knowledge mappings with reference to attribute Body pair；

With reference to attribute to screening unit, for being filtered out in two knowledge mappings each according to the target entity pair filtered out New objective attribute target attribute pair refers to attribute pair as each；

Target entity is to Cycle Screening unit, for calling the reference attribute pair filtered out, and according to it is each with reference to attribute to Each target entity pair is filtered out in two knowledge mappings, until it can not filter out target entity pair, forms first instance To set；

13. device according to claim 12, which is characterized in that the target entity is specifically used for screening unit:

14. device according to claim 13, which is characterized in that the target entity is to screening unit, comprising:

Initial solid determines subelement, for determining each initial solid pair, the initial solid pair in two knowledge mappings With at least one with reference to attribute to and have it is each with reference to attribute to a corresponding attribute value similarity, the attribute value phase It is the corresponding attribute that refers to the similarity between the attribute value of two included attributes like degree；

Target entity, at least one attribute value similarity according to the initial solid pair, determines institute to subelement is determined Initial solid is stated to whether belonging to the target entity pair.

15. device according to claim 14, which is characterized in that the target entity is to determining subelement, comprising:

Similarity mean value calculation module, at least one attribute value similarity for calculating the initial solid pair are averaged Value；

Target entity is to determining module, if the average value for being calculated is greater than the first preset threshold, determines described initial Entity is to for the target entity pair.

16. device according to claim 12, which is characterized in that the reference attribute is to screening unit, comprising:

Attribute to be selected is each under by the target entity for each target entity pair for filtering out to acquisition subelement A attribute to be selected carries out combination of two, obtains the attribute pair to be selected under every kind of combination, the attribute to be selected is to included two Attribute be not belonging to fixed each objective attribute target attribute to and be belonging respectively to two entities of the target entity centering；

Attributes similarity computation subunit, for calculating the attribute to be selected between the attribute value of two included attributes Similarity；

Objective attribute target attribute is to subelement is determined, if similarity for being calculated is greater than the second preset threshold, determine it is described to Select attribute to for new objective attribute target attribute pair.