CN107145523B - Alignment method for large heterogeneous knowledge bases based on iterative matching - Google Patents

Abstract

The invention discloses a large-scale heterogeneous knowledge base alignment method based on iterative matching. The specific implementation is as follows: 1) Screen the data in the original knowledge base, unify the data format, and obtain the relationship and initial knowledge base in the knowledge base on this basis. Match entity pairs; 2) use the relationship in the knowledge base to partition the preprocessed knowledge base, and simplify the blocks; 3) use the matched entity pairs to match blocks to obtain matching block pairs; 4) match block pairs Select candidate entity pairs, and combine the similarity measurement method and threshold to confirm candidate entity pairs; 5) Repeat the above steps until no new candidate entity pairs can be found, and all matching entity pairs are obtained. The invention combines the idea of iterative matching to align heterogeneous knowledge bases, and has broad application prospects in the fields of knowledge base alignment, data fusion, automatic question answering and the like.

Description

Alignment method for large heterogeneous knowledge bases based on iterative matching

technical field

The invention relates to the field of knowledge base alignment, in particular to a large-scale heterogeneous knowledge base alignment method based on iterative matching.

Background technique

With the advent of Web 3.0, structured knowledge bases are appearing more and more frequently on the Internet. These knowledge bases are widely used in various semantic applications, such as automatic question answering, search services, and social services. However, the limited information of a single knowledge base limits the functionality of these applications. In this context, knowledge base alignment has huge room for development. Knowledge Base Alignment (Knowledge Base Alignment) usually refers to the entity alignment of the knowledge base, that is, to automatically discover two entities representing the same thing in reality and connect them.

Due to the ever-increasing size of knowledge bases, knowledge base alignment methods usually divide the alignment process into two steps: discovering candidate entity pairs and confirming candidate entity pairs. Discovering candidate entity pairs usually uses a small number of attributes to quickly screen out several candidate entities for each entity, and confirming candidate entity pairs comprehensively compares two entities, and uses similarity and threshold to determine whether the two entities match. This approach greatly improves the overall efficiency of the method by avoiding exact pairwise comparisons of entities. At present, the bottleneck of the knowledge base alignment method is that the discovered candidate entity pairs are often missed, which further leads to the undiscovered matching entity pairs.

In order to improve the quality of candidate entity pairs, researchers propose the idea of using iterative matching, that is, a small number of matching entity pairs are found in each round, and used as the basis for the next round of candidate entity pairs. However, traditional knowledge base alignment methods usually focus on the alignment of isomorphic knowledge bases, that is, there are more alignable relations between two knowledge bases. The basic assumption is: if a pair of entity pairs match and they have an aligned relationship, then their "compatible neighbors" have a higher probability of matching, so "compatible neighbors" are used as candidate entity pairs. However, due to the lack of alignable relationships between knowledge bases, traditional methods will miss some candidate entity pairs. To solve this problem, researchers propose to use a class-based knowledge base alignment method. This method classifies instances with the same characteristics into the same class and excludes candidate entities that are not related to the content of the class to confirm candidate entity pairs. However, since this method only obtains candidate entity pairs through classical partitioning techniques at the initial stage of the model, when there are fewer attributes aligned between the two knowledge bases, this method will also miss more candidate entity pairs.

Contents of the invention

In view of the above, the present invention proposes a large-scale heterogeneous knowledge base alignment method based on iterative matching. This method combines the idea of iterative matching to align the knowledge base, uses the iterative framework to traverse the relationship to partition the knowledge base, and expands the search space for candidate entity pairs; at the same time, uses the idea of divide and conquer to select and confirm candidate entity pairs, so that each entity Only a few candidate entities need to be comprehensively compared, increasing the efficiency of the method.

A method for aligning large heterogeneous knowledge bases based on iterative matching, including:

Data preprocessing stage: filter the data in any two original knowledge bases KB ₁ and KB ₂ , unify the data format and eliminate meaningless characters, and obtain the relation set R corresponding to the processed knowledge base KB′ _{1 1.} The relationship set R ₂ corresponding to the processed knowledge base KB′ ₂ is compared to obtain the initial matching entity pair set

Knowledge base alignment stage: use the relationship in relation set R ₁ and relation set R ₂ to partition knowledge base KB′ ₁ and knowledge base KB′ ₂ , and simplify each block to obtain a simplified block set B′ ₁ and B′ ₂ ; then, use the initial matching entity pair set Match the blocks in the reduced block sets B′ ₁ and B′ ₂ to obtain matching block pairs. Finally, select candidate entity pairs from the matching block pairs, and combine the similarity measurement method and the threshold δ _e to confirm the candidate entity pairs .

The concrete steps of the described data preprocessing stage are:

(1-1) Input any two original knowledge bases KB ₁ and KB ₂ , and remove the information irrelevant to the alignment task in the knowledge bases KB ₁ and KB ₂ ;

(1-2) unify the data format to the literal quantity L ₁ in the knowledge base KB ₁ and the literal quantity L ₂ in the knowledge base KB ₂ , date, numeral, full name are represented as unified format;

(1-3) Remove the stop word characters, symbol characters, and language label characters in the literal quantity L ₁ in the knowledge base KB ₁ and the literal quantity L ₂ in the knowledge base KB ₂ , and obtain the processed knowledge base KB′ ₁ and KB '₂;

(1-4) statistically obtain the relation set R ₁ corresponding to the knowledge base KB′ ₁ and the relation set R ₂ corresponding to the knowledge base KB′ ₂ ;

(1-5) Compare all entities in knowledge base KB′ ₁ and knowledge base KB′ ₂ , and obtain the initial matching entity pair set

The knowledge base is defined as a six-tuple (E, L, R, P, F _R , F _P ), where E, L, R, and P respectively represent a collection of entities, literals, relationships, and attributes; Represents the triplet set of entity-relation-entity, and represents the relationship fact that the object is an entity; Represents the set of triples of entity-attribute-literal, and represents the fact that the object is a literal attribute; there are _meaningless information in both FR and _FP , for example: some knowledge bases contain triples for extracting The original text corpus, this information will affect the efficiency of the algorithm. In addition, some triples containing the "sameAs" relationship should also be removed.

The concrete process of described step (1-4) is:

For the knowledge base KB′ ₁ , traverse all the triples (entity-relationship-entity) in the triple set F _R1 belonging to the knowledge base, and obtain the relation set R ₁ through statistics; for the knowledge base KB′ ₂ , traverse all the triples belonging to the All the triples (entity-relationship-entity) in the triplet set F _R2 of the knowledge base are counted to obtain the relational set R ₂ , the relational set R ₁ and the relational set R ₂ for subsequent knowledge base partitioning operations.

In step (1-5), the initial matching entity pair set The acquisition process is:

First, extract all the entities in the knowledge base KB′ ₁ to form the entity set E ₁ , extract all the entities in the knowledge base KB’ ₂ to form the entity set E ₂ ; and use any entity in the entity set E ₁ to form the entity set E ₂ The Cartesian product of any entity in is used as an entity pair to form an entity pair set;

Then, filter and obtain the entity pairs whose strings of the two entity name attributes in the entity pair set represent identical entity pairs, and obtain the pre-initial matching entity pair set;

Finally, filter the entity pairs with one-to-one matching relationship in the pre-initial matching entity pair set as the initial matching entity pair set

The specific steps of the knowledge base alignment phase are as follows:

(2-1) Input knowledge base KB′ ₁ , knowledge base KB′ ₂ , relation set R ₁ , relation set R ₂ , initial matching entity pair set Set the block similarity threshold δ _b , the entity similarity threshold δ _e , the threshold of the number of entities in the block δ ₁ , and the threshold of the ratio of matched entities in the block δ ₂ , the matching entity pair set M _e is initialized as the initial matching entity pair set

(2-2) Randomly select any relation in relation set R ₁ or relation set R ₂ , use this relation to divide entities in knowledge base KB′ ₁ and knowledge base KB′ ₂ into several blocks, and obtain Block set B ₁ corresponding to KB′ ₁ , block set B ₂ corresponding to knowledge base KB′ ₂ ;

(2-3) Remove block sets B ₁ and block sets B ₂ that are prone to generate high computational load or blocks that are difficult to generate matching entity pairs, and obtain a simplified block set B' ₁ and a simplified block set B'₂;

( _2-4 ) Use all matching entity pairs in the matching entity pair set Me to measure the similarity between any block in the reduced block set _B'1 and any block in the reduced block set _B'2 , Select two blocks whose similarity value is greater than the block similarity threshold δ _b to match, and obtain a matching block pair set;

(2-5) For any matching block pair belonging to the matching block pair set, any unmatched entity in one block of the matching block pair and any unmatched entity in the other block of the matching block pair The Cartesian product of any unmatched entity is used as a candidate entity pair to form a candidate entity pair set;

(2-6) Judging whether no new candidate entity pair has been found, if not, jump to step (2-7), if so, end the iteration, and output the matching entity pair set M _e ;

(2-7) Calculate the similarity between two entities in each candidate entity pair in the candidate entity pair set, and add the candidate entity pair whose similarity value is greater than the entity similarity threshold δ _e to the matching entity pair set M _e , The remaining candidate entity pairs are discarded;

(2-8) Determine whether the number of iterations is less than the iteration threshold, if not, skip to step (2-2); if yes, end the iteration, and output the matching entity pair set M _e .

In step (2-2), the specific process of dividing the entity in the knowledge base into several blocks by using the relationship is;

First, for the triplet set F _R1 in the knowledge base KB′ ₁ , get n types of object entities in the triplet set F _R1 through statistics;

Then, for each object entity, put together all the subject entities corresponding to it in the triple set _FR1 to obtain 1 block, and n types of object entities to obtain n blocks to form a block set B ₁ ;

Use the same method to get the block set B ₂ .

In step (2-3), the blocks that are prone to high computational load or difficult to generate matching entity pairs include: blocks with the number of entities exceeding the threshold δ1, blocks with the ratio of matched entities less _than the threshold _δ2 , and entities blocks that are already matched.

In step (2-4), the method for obtaining the similarity between blocks is:

Each block is regarded as a collection of entities, and the matched entity pairs are regarded as the same elements between the two collections, and the similarity between blocks is measured by using the similarity of the collection, similarity sim _block (b _k , b _l ) The calculation formula is:

Among them, b _k and b _l represent two blocks, |b _k ∩ b _l | represents the number of matching entity pairs in the two blocks, and |b _k ∪ b _l | represents the total number of entities in the two blocks.

In step (2-7), the formula for obtaining the similarity between entities is:

sim(e _i ,e _j )=αsim _string (e _i ,e _j )+(1-α)sim _block (b _k ,b _l )

ste _i ∈ b _k , e _j ∈ b _l

Among them, b _k and b _l represent the blocks where entities e _i and e _j are located respectively, and sim _string (e _i , e _j ) and sim _block (b _k , b _l ) represent the string similarity and block Block similarity, α is the weight of string similarity, and the value range is [0,1].

Preferably, similarity functions based on Levenshtein distance, Jaro-Winker distance, q-gram and I-SUB are used, and these similarity measurement functions are combined in a linear weighted manner to obtain string similarity.

The invention combines the idea of iterative matching to align heterogeneous knowledge bases, uses the iterative framework to traverse the relationship to partition the knowledge base, and expands the search space for candidate entity pairs; at the same time, adopts the idea of divide and conquer to select and confirm candidate entity pairs, so that each Entities only need to be comprehensively compared with a few candidate entities, which improves the efficiency of the method. Compared with existing methods, its advantages are:

(1) View knowledge base alignment as an iterative process. In different iterations, each relation is traversed to partition the knowledge base, and the matching block pairs are used to select candidate entity pairs, so that the alignment method does not depend on the alignable relations and attributes between the knowledge bases.

(2) In each round of iteration, only a small number of matching entity pairs are found, and these matching entity pairs are used for the selection of candidate entity pairs. Since the process of selecting candidate entity pairs uses more information about matching entity pairs, the improvement of The quality of the candidate entity pairs.

Description of drawings

Fig. 1 is the flow diagram of the method for aligning large-scale heterogeneous knowledge bases based on iterative matching in the present invention;

Fig. 2 is a flow chart of the data preprocessing stage in the large-scale heterogeneous knowledge base alignment method based on iterative matching in the present invention;

Fig. 3 is a flowchart of the knowledge base alignment stage in the iterative matching-based large-scale heterogeneous knowledge base alignment method of the present invention.

Detailed ways

In order to describe the present invention more specifically, the technical solutions of the present invention will be described in detail below in conjunction with the accompanying drawings and specific embodiments.

As shown in FIG. 1 , the method for aligning large-scale heterogeneous knowledge bases based on iterative matching in the present invention is divided into two parts: data preprocessing and knowledge base alignment. Data preprocessing part: filter the data in the original knowledge base KB, unify the data format, and obtain the relationship in the knowledge base and the initial matching entity pair; knowledge base alignment part: first use the relationship in the knowledge base to pair the preprocessed The knowledge base is partitioned, and the blocks are simplified, and then the matched entity pairs are used to match the blocks to obtain the matched block pairs, and then the candidate entity pairs are selected from the matched block pairs, and the candidate entity pairs are confirmed by combining the similarity measurement method and the threshold value, Finally, the above steps are repeated until no new candidate entity pairs can be found, and all matching entity pairs can be obtained.

Figure 2 shows the flow chart of the data preprocessing stage; according to Figure 2, this stage is divided into the following steps:

S1-1, input any two original knowledge bases KB ₁ and KB ₂ , and remove information irrelevant to the alignment task in the knowledge bases KB ₁ and KB ₂ .

The knowledge base is defined as a six-tuple (E, L, R, P, F _R , F _P ), where E, L, R, and P respectively represent a collection of entities, literals, relationships, and attributes; Represents the triplet set of entity-relation-entity, and represents the relationship fact that the object is an entity; Represents the set of triples of entity-attribute-literal, and represents the fact that the object is a literal attribute; there are _meaningless information in both FR and _FP , for example: some knowledge bases contain triples for extracting The original text corpus, this information will affect the efficiency of the algorithm. In addition, some triples containing "same As" relations should also be removed.

_S1-2 , unify the data format for the literal quantity L1 in the knowledge base _KB1 and the literal quantity L2 in the knowledge base _KB2 , and express the date, number, and name in _a unified format.

Literal values such as names, dates, and numbers may be expressed in different ways in different knowledge bases, for example: "2016-01-01" and "01.01.2016". Unifying these information is beneficial for subsequent comparisons. In addition, the method unifies literals into lowercase.

S1-3, remove meaningless characters such as stop word characters, symbol characters, and language tags in the literal amount L ₁ in the knowledge base KB ₁ and the literal amount L ₂ in the knowledge base KB ₂ , and obtain the processed knowledge base KB′ ₁ and KB′ ₂ .

There may be some meaningless characters in the attribute description of the entity in the knowledge base, such as stop words such as "the", "a" and "an", symbols such as "#", "!" and "*", and " @en" and other language tags. These characters affect the similarity measure of entity pairs, so these characters are removed.

S1-4. Obtain statistically the relation set R ₁ corresponding to the knowledge base KB′ ₁ and the relation set R ₂ corresponding to the knowledge base KB′ ₂ .

In this step, for the knowledge base KB′ ₁ , traverse all triples (entity-relationship-entity) in the triple set F _R1 belonging to the knowledge base, and obtain the relation set R ₁ through statistics; for the knowledge base KB′ ₂ , traverse all triples (entity-relationship-entity) in the triplet set F _R2 belonging to the knowledge base, and obtain the relation set R ₂ , relation set R ₁ and relation set R ₂ for subsequent knowledge base partitioning operate.

S1-5, compare all the entities in the knowledge base KB′ ₁ and the knowledge base KB′ ₂ , and obtain the initial matching entity pair set

In this step, the initial matching entity pair set The acquisition process is:

First, extract all the entities in the knowledge base KB′ ₁ to form the entity set E ₁ , extract all the entities in the knowledge base KB’ ₂ to form the entity set E ₂ ; and use any entity in the entity set E ₁ to form the entity set E ₂ The Cartesian product of any entity in is used as an entity pair to form an entity pair set;

Then, filter and obtain the entity pairs whose strings of the two entity name attributes in the entity pair set represent identical entity pairs, and obtain the pre-initial matching entity pair set;

Finally, filter the entity pairs with one-to-one matching relationship in the pre-initial matching entity pair set as the initial matching entity pair set

Figure 3 shows the flowchart of the knowledge base alignment phase; according to Figure 3, this phase is divided into the following steps:

S2-1, input knowledge base KB′ ₁ , knowledge base KB′ ₂ , relation set R ₁ , relation set R ₂ , initial matching entity pair set Set the block similarity threshold δ _b to 0.2, the entity similarity threshold δ _e to 0.65, the threshold of the number of entities in the block δ ₁ to 50, and the threshold of the ratio of matched entities in the block δ ₂ to 0.3, the matching entity pair set M _e initialized to the initial set of matching entity pairs

S2-2. Randomly select any relation in relation set R ₁ or relation set R ₂ , use this relation to divide entities in knowledge base KB′ ₁ and knowledge base KB′ ₂ into several blocks, and obtain The block set B ₁ corresponding to ′ ₁ and the block set B ₂ corresponding to the knowledge base KB′ ₂ .

In this step, the specific process of dividing the entities in the knowledge base into several blocks by using the relationship is as follows;

First, for the triplet set F _R1 in the knowledge base KR′ ₁ , get n types of object entities in the triplet set F _R1 through statistics;

Then, for each object entity, put together all the subject entities corresponding to it in the triple set _FR1 to obtain 1 block, and n types of object entities to obtain n blocks to form block set B ₁ .

Use the same method to get the block set B ₂ , namely:

First, for the triplet set F _R2 in the knowledge base KB′ ₂ , get n types of object entities in the triplet set F _R2 through statistics;

Then, for each object entity, put together all the subject entities corresponding to it in the triple set _FR2 to obtain 1 block, and n types of object entities to obtain n blocks to form block set B ₂ .

S2-3, remove the blocks in the block set B ₁ and the block set B ₂ that are prone to high computational load or difficult to generate matching entity pairs, and obtain the simplified block set B' ₁ and the simplified block set B' ₂ .

In this step, the blocks that tend to generate high computational load or are difficult to generate matching entity pairs include: blocks with the number of entities exceeding the threshold δ1, blocks with the ratio of matched entities less _than the threshold _δ2 , and blocks with entities that have already been matched.

S2-4, using all matching entity pairs in the matching entity pair set M _e to measure the similarity between any block in the reduced block set B′ ₁ and any block in the reduced block set B′ ₂ , select Two blocks whose similarity value is greater than the block similarity threshold δ _b are matched to obtain a matching block pair set.

Each block is regarded as a collection of entities, and the matched entity pairs are regarded as the same elements between the two collections, and the similarity between blocks is measured by using the similarity of the collection, similarity sim _block (b _k , b _l ) The calculation formula is:

Among them, b _k and b _l represent two blocks, |b _k ∩ b _l | represents the number of matching entity pairs in the two blocks, and |b _k ∪ b _l | represents the total number of entities in the two blocks.

S2-5, for any matching block pair belonging to the matching block pair set, any unmatched entity in one block of the matching block pair and any unmatched entity in the other block of the matching block pair The Cartesian product of an unmatched entity is used as a candidate entity pair to form a candidate entity pair set.

S2-6, judging whether no new candidate entity pair is found, if not, skip to S2-7, if so, end the iteration, and output the matching entity pair set M _e .

S2-7. Calculate the similarity between two entities in each candidate entity pair in the candidate entity pair set, and add the candidate entity pair whose similarity value is greater than the entity similarity threshold δ _e to the matching entity pair set M _e , and the remaining The following candidate entity pairs are discarded.

In this step, the similarity between entities is measured in two ways: string similarity and block similarity, and these two similarities are combined with a certain weight. The formula is as follows:

sim(e _i ,e _j )=αsim _string (e _i ,e _j )+(1-α)sim _block (b _k ,b _l )

ste _i ∈ b _k , e _j ∈ b _l

Among them, sim _string (e _i , e _j ) and sim _block (b _k , b _l ) represent the string similarity and block similarity between entities respectively, and b _k and b _l represent the locations of entities e _i and e _j respectively block, α is the weight of string similarity, with a value of 0.6. For the attribute pairs shared by entities e _i and e _j (for example: name), the string similarity measures the similarity of these attribute values. The method uses a variety of similarity measurement functions, such as: based on Levenshtein distance, based on Jaro-Winker distance, based on q-gram and based on I-SUB, and combined these similarity measurement functions by linear weighting. Block similarity represents the similarity of entities through the similarity between the blocks where entities are located. After obtaining the similarity between entities, combine the threshold δ _e to judge whether the pair of entities match, and add the newly discovered matching entity pair to all matching entity pairs.

S2-8, judging whether the number of iterations is less than the iteration threshold, if not, skip to S2-2; if so, end the iteration, and output the matching entity pair set M _e .

The above-mentioned specific embodiments have described the technical solutions and beneficial effects of the present invention in detail. It should be understood that the above-mentioned are only the most preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, supplements and equivalent replacements made within the scope shall be included in the protection scope of the present invention.

Claims (10)

1. A large-scale heterogeneous knowledge base alignment method based on iterative matching, including: Data preprocessing stage: filter the data in any two original knowledge bases KB ₁ and KB ₂ , unify the data format and eliminate meaningless characters, and obtain the relation set R corresponding to the processed knowledge base KB′ _{1 1.} The relationship set R ₂ corresponding to the processed knowledge base KB′ ₂ is compared to obtain the initial matching entity pair set Knowledge base alignment stage: use the relationship in relation set R ₁ and relation set R ₂ to partition knowledge base KB′ ₁ and knowledge base KB′ ₂ , and simplify each block to obtain a simplified block set B′ ₁ and B′ ₂ ; then, use the initial matching entity pair set Match the blocks in the reduced block sets B' ₁ and B' ₂ to obtain matching block pairs. Finally, select candidate entity pairs from the matching block pairs, and combine similarity measurement methods and thresholds to confirm candidate entity pairs. 2. the large-scale heterogeneous knowledge base alignment method based on iterative matching as claimed in claim 1, is characterized in that, the specific steps of described data preprocessing stage are: (1-1) Input any two original knowledge bases KB ₁ and KB ₂ , and remove the information irrelevant to the alignment task in the knowledge bases KB ₁ and KB ₂ ; (1-2) unify the data format to the literal quantity L ₁ in the knowledge base KB ₁ and the literal quantity L ₂ in the knowledge base KB ₂ , date, numeral, full name are represented as unified format; (1-3) Remove the stop word characters, symbol characters, and language label characters in the literal quantity L ₁ in the knowledge base KB ₁ and the literal quantity L ₂ in the knowledge base KB ₂ , and obtain the processed knowledge base KB′ ₁ and KB '₂; (1-4) statistically obtain the relation set R ₁ corresponding to the knowledge base KB′ ₁ and the relation set R ₂ corresponding to the knowledge base KB′ ₂ ; (1-5) Compare all entities in knowledge base KB′ ₁ and knowledge base KB′ ₂ , and obtain the initial matching entity pair set 3. the large-scale heterogeneous knowledge base alignment method based on iterative matching as claimed in claim 2, is characterized in that, the specific process of described step (1-4) is: For the knowledge base KB′ ₁ , traverse all the entity-relationship-entity triples in the triple set F _R1 belonging to the knowledge base, and obtain the relation set R ₁ through statistics; for the knowledge base KB′ ₂ , traverse all the triples belonging to the knowledge All the entity-relationship-entity triplets in the triplet set F _R2 of the library are counted to obtain the relational set R ₂ . 4. The large-scale heterogeneous knowledge base alignment method based on iterative matching as claimed in claim 2, characterized in that, in step (1-5), the initial matching entity pair set The acquisition process is: First, extract all the entities in the knowledge base KB′ ₁ to form the entity set E ₁ , extract all the entities in the knowledge base KB’ ₂ to form the entity set E ₂ ; and use any entity in the entity set E ₁ to form the entity set E ₂ The Cartesian product of any entity in is used as an entity pair to form an entity pair set; Then, filter and obtain the entity pairs whose strings of the two entity name attributes in the entity pair set represent identical entity pairs, and obtain the pre-initial matching entity pair set; Finally, filter the entity pairs with one-to-one matching relationship in the pre-initial matching entity pair set as the initial matching entity pair set 5. The large-scale heterogeneous knowledge base alignment method based on iterative matching as claimed in claim 1, wherein the specific steps of the knowledge base alignment stage are: (2-1) Input knowledge base KB′ ₁ , knowledge base KB′ ₂ , relation set R ₁ , relation set R ₂ , initial matching entity pair set Set the block similarity threshold δ _b , the entity similarity threshold δ _e , the threshold of the number of entities in the block δ ₁ , and the threshold of the ratio of matched entities in the block δ ₂ , the matching entity pair set M _e is initialized as the initial matching entity pair set (2-2) Randomly select any relation in relation set R ₁ or relation set R ₂ , use this relation to divide entities in knowledge base KB′ ₁ and knowledge base KB′ ₂ into several blocks, and obtain Block set B ₁ corresponding to KB′ ₁ , block set B ₂ corresponding to knowledge base KB′ ₂ ; (2-3) Remove block sets B ₁ and block sets B ₂ that are prone to generate high computational load or blocks that are difficult to generate matching entity pairs, and obtain a simplified block set B' ₁ and a simplified block set B'₂; ( _2-4 ) Use all matching entity pairs in the matching entity pair set Me to measure the similarity between any block in the reduced block set _B'1 and any block in the reduced block set _B'2 , Select two blocks whose similarity value is greater than the block similarity threshold δ _b to match, and obtain a matching block pair set; (2-5) For any matching block pair belonging to the matching block pair set, any unmatched entity in one block of the matching block pair and any unmatched entity in the other block of the matching block pair The Cartesian product of any unmatched entity is used as a candidate entity pair to form a candidate entity pair set; (2-6) Judging whether no new candidate entity pair has been found, if not, jump to step (2-7), if so, end the iteration, and output the matching entity pair set M _e ; (2-7) Calculate the similarity between two entities in each candidate entity pair in the candidate entity pair set, and add the candidate entity pair whose similarity value is greater than the entity similarity threshold δ _e to the matching entity pair set M _e , The remaining candidate entity pairs are discarded; (2-8) Determine whether the number of iterations is less than the iteration threshold, if not, skip to step (2-2); if yes, end the iteration, and output the matching entity pair set M _e . 6. the large-scale heterogeneous knowledge base alignment method based on iterative matching as claimed in claim 5, is characterized in that, in step (2-2), described utilization relation divides the entity in the knowledge base into several blocks The specific process is; First, for the triplet set F _R1 in the knowledge base KB′ ₁ , get n types of object entities in the triplet set F _R1 through statistics; Then, for each object entity, put together all the subject entities corresponding to it in the triple set _FR1 to obtain 1 block, and n types of object entities to obtain n blocks to form a block set B ₁ ; Use the same method to get the block set B ₂ . 7. The large-scale heterogeneous knowledge base alignment method based on iterative matching as claimed in claim 5, characterized in that, in step (2-3), the blocks that are prone to generate high computational load or are difficult to generate matching entity pairs Including: blocks with the number of entities exceeding the threshold δ ₁ , blocks with the ratio of matched entities less than the threshold δ ₂ , and blocks with entities that have already been matched. 8. the large-scale heterogeneous knowledge base alignment method based on iterative matching as claimed in claim 5, is characterized in that, in step (2-4), the acquisition method of the similarity between blocks is: Each block is regarded as a collection of entities, and the matched entity pairs are regarded as the same elements between the two collections, and the similarity between blocks is measured by using the similarity of the collection, similarity sim _block (b _k , b _l ) The calculation formula is: Among them, b _k and b _l represent two blocks, |b _k ∩ b _l | represents the number of matching entity pairs in the two blocks, and |b _k ∪ b _l | represents the total number of entities in the two blocks. 9. the large-scale heterogeneous knowledge base alignment method based on iterative matching as claimed in claim 8, is characterized in that, in step (2-7), the acquisition formula of the similarity between entities is: sim(e _i ,e _j )=αsim _string (e _i ,e _j )+(1-α)sim _block (b _k ,b _l ) ste _i ∈ b _k , e _j ∈ b _l Among them, b _k and b _l represent the blocks where entities e _i and e _j are located respectively, and sim _string (e _i , e _j ) and sim _block (b _k , b _l ) represent the string similarity and block Block similarity, α is the weight of string similarity, and the value range is [0,1]. 10. the large-scale heterogeneous knowledge base alignment method based on iterative matching as claimed in claim 9, is characterized in that, adopts the similarity function based on Levenshtein distance, based on Jaro-Winker distance, based on q-gram and based on I-SUB, And the string similarity is obtained by combining these similarity measurement functions in a linear weighted manner.