A knowledge-fused low-resource language entity extraction system

Abstract

The application provides a low-resource language entity extraction system fusing knowledge, comprising a target language knowledge base, a related language knowledge base, a processor and a memory storing a computer program, and when the processor executes the computer program, the following steps are implemented: obtaining a target entity in a target sentence of unlabeled corpus, performing entity linking on the target entity with the target language knowledge base and the related language knowledge base, obtaining a first loss function by taking a masked entity as a target, obtaining a second loss function by taking a context of an entity as a target, obtaining a third loss function by taking an entity type as a target, and obtaining an XLM-RK model by minimizing a weighted sum function of the three loss functions; and thus the entity recognition effect of the target language is improved according to the cross-language entity alignment capability obtained by the XLM-RK model in the pre-training stage.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of natural language processing, in particular to a low-resource language entity extraction system fusing knowledge. BACKGROUND

[0002] The named entity recognition task refers to identifying entities with specific meanings in text, mainly including names, place names and organization names, which is widely used in information extraction, public opinion analysis, knowledge graph construction and question answering system scenarios.

[0003] In recent years, cross-lingual pre-training language models such as mBERT and XLM-RoBERTa (XLM-R) have significantly improved the performance of named entity recognition tasks in multiple languages. Cross-lingual pre-training models implicitly learn universal semantic and syntactic knowledge through self-supervised learning on large-scale multilingual unlabelled corpus. The model transfers the knowledge learned from the open domain to the downstream task to improve the performance of low-resource tasks.

[0004] Although the mBERT model has cross-lingual capability, it does not introduce language-related labels during pre-training, but only transfers the context information of multiple languages by sharing the dictionary between languages to learn cross-lingual representations. However, for knowledge-oriented tasks such as named entity recognition, the cross-lingual knowledge alignment and transfer capability of existing models is relatively limited. In actual experiments, for low-resource languages such as Japanese and Dutch that lack corpus or labeled data, using labeled data of high-resource languages to expand the training corpus of low-resource languages still has a large gap in entity recognition performance compared to high-resource languages. SUMMARY

[0005] To solve the above technical problems, the technical scheme adopted by the present application is as follows: a low-resource language entity extraction system fusing knowledge, the system comprising a target language knowledge base, an associated language knowledge base, a processor and a memory storing a computer program, the associated language being a language of the same language family as the target language or a language having a degree of coincidence with the target language, and the associated language being a high-resource language, when the processor executes the computer program, the following steps are implemented: S100, obtaining a target entity in an unlabelled corpus target sentence.

[0006] S200, performing entity linking on the target entity with the target language knowledge base and the associated language knowledge base to obtain a first entity corresponding to the target entity, wherein the first entity is an entity corresponding to the target entity in the target language knowledge base and the associated language knowledge base.

[0007] S300, based on the first entity knowledge short sentence list Z={Z1, Z2, …, Z i , …, Z m} and the first intermediate sentence X0 to construct a first training corpus X = {X1, X2, …, X i , …, X m}, wherein the i-th first training corpus X i is obtained by concatenating X0 after Z i ; Z i is a sentence constructed according to a first preset syntax rule from the i-th triple corresponding to the first entity, i has a value range of 1 to m, m is the number of triples of the target language and the associated language corresponding to the first entity, and the first intermediate sentence X0 is a sentence in which the entities in the target sentence are masked according to a first mask rule.

[0008] S400, inputting the first training corpus X into the XLM-R model to obtain a first loss function L1 by taking the masked entity as a target.

[0009] S500, constructing a second training corpus P = {P1, P2, …, P j , …, Y n} and a second intermediate sentence X' to construct a second training corpus P = {P1, P2, …, P j , …, P n}, wherein n is the number of triples in the target language knowledge base that match the target entity and correspond to triples that satisfy the isA relationship, Y j is a sentence constructed according to a second preset syntax rule from the j-th triple in the target language knowledge base that matches the target entity and corresponds to triples that satisfy the isA relationship, j has a value range of 1 to n, and the second intermediate sentence X' is a sentence in which non-entities in the target sentence are masked according to a second mask rule.

[0010] S600, inputting the second training corpus P into the XLM-R model to obtain a second loss function L2 by taking the context of the entity as a target.

[0011] S700, concatenating the third entity knowledge short sentence A0 and the target sentence to construct a third training corpus B, wherein A0 is a sentence in which the tail entity corresponding to the isA relationship is masked according to a third mask rule from a second entity knowledge short sentence in Y that satisfies a preset filtering rule.

[0012] S800, inputting the third training corpus B into the XLM-R model to obtain a third loss function L3 by taking the entity type as a target according to a preset candidate entity type set.

[0013] S900, marking the XLM-R model that satisfies the preset loss condition as an XLM-RK model by minimizing the loss function L = aL1 + bL2 + gL3, wherein a, b, and g are all preset weight parameters.

[0014] The present application has at least the following beneficial effects: the present application obtains the target entity in the target sentence, links the target entity to obtain the first entity, predicts the mask entity based on the first entity knowledge short sentence and the target sentence of the entity mask to obtain the first loss function, predicts the context of the entity based on the second entity knowledge short sentence and the second intermediate short sentence of the non-entity mask to obtain the second loss function, predicts the entity type based on the third entity short sentence and the target sentence to obtain the third loss function, and obtains the XLM-RK model by minimizing the weighted sum function of the three loss functions, improves the cross-language knowledge transfer and knowledge alignment capability of the obtained XLM-RK model, fuses the public multilingual knowledge base, better realizes cross-language knowledge alignment and language understanding, and improves the performance on the low-resource entity extraction task. BRIEF DESCRIPTION OF DRAWINGS

[0015] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings needed in the embodiment description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0016] Figure 1 The flowchart of the low-resource language entity extraction system provided by the embodiment of the present application when executing the computer program is provided.

[0017] Figure 2 The schematic diagram of predicting the mask entity provided by the embodiment of the present application is provided.

[0018] Figure 3 The schematic diagram of predicting the context of the entity provided by the embodiment of the present application is provided.

[0019] Figure 4 The schematic diagram of predicting the entity type provided by the embodiment of the present application is provided. DETAILED DESCRIPTION

[0020] The technical solutions in the embodiments of the present application will be described clearly and completely with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some embodiments of the present application, not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0021] The embodiment of the present application provides a low-resource language entity extraction system fusing knowledge, which comprises a target language knowledge base, an associated language knowledge base, a processor and a memory storing a computer program, the associated language is a language of the same language family as the target language or a language with a coincidence degree with the target language, and the associated language is a high-resource language. It is exemplarily explained that when the low-resource language is Japanese, the associated language is Chinese.

[0022] Specifically, in the present application, the target language knowledge base and the associated language knowledge base can be obtained through Wikidata (Wikipedia data). Wikidata extracts information from websites such as Wikipedia, and after cleaning, forms high-quality structured triple data, which constitutes the target language knowledge base and the associated language knowledge base.

[0023] When the processor executes the computer program, the following steps are implemented, as shown in the figure. Figure 1

[0024] S100, obtaining a target entity in a target sentence in unannotated corpus.

[0025] Specifically, the target entity contained in the target sentence can be one or multiple, and when the target entity is multiple, the subsequent steps are executed for each target entity.

[0026] S200, performing entity linking on the target entity and the target language knowledge base and the associated language knowledge base to obtain a first entity corresponding to the target entity, wherein the first entity is an entity corresponding to the target entity in the target language knowledge base and the associated language knowledge base.

[0027] Specifically, any method for performing entity linking on the target entity in the prior art belongs to the protection scope of the present application, and will not be described here.

[0028] Specifically, S200 specifically comprises:

[0029] S201, performing entity linking on the target entity and the target language knowledge base to obtain a first entity corresponding to the target entity in the target language knowledge base.

[0030] S202, performing entity linking on the target entity and the associated language knowledge base to obtain a first entity corresponding to the target entity in the associated language knowledge base.

[0031] ​Further, S202 specifically includes linking the target entity and the associated language knowledge base as entity: querying the API provided by Wikidata (Wiki Data) to obtain the name of the target entity in the corresponding associated language, and linking the target entity to the associated language knowledge base based on the name of the target entity in the corresponding associated language, to obtain the first entity corresponding to the target entity in the associated language knowledge base.

[0032] In another embodiment of the application, in S202, obtaining the first entity corresponding to the target entity in the associated language knowledge base further includes: using the cross-language link to obtain the first entity corresponding to the associated language knowledge base through the first entity corresponding to the target language knowledge base.

[0033] S300, based on the first entity knowledge short sentence list Z={Z1, Z2, …, Z i , …, Z m} and the first intermediate sentence X0, a first training corpus set X={X1, X2, …, X i , …, X m} is constructed, wherein the i-th first training corpus X i is obtained by concatenating X0 after Z i ; Z i is the i-th triple corresponding to the first entity constructed according to the first preset grammar rule, i is in the range of 1 to m, m is the number of triples corresponding to the target language and the associated language of the first entity, and the first intermediate sentence X0 is the sentence obtained by performing entity masking on the target sentence according to the first masking rule.

[0034] Specifically, the first entity knowledge short sentence is constructed based on the first entity according to the first preset grammar rule. In the prior art, a triple (head entity, relation, tail entity) is simply concatenated to form a short sentence, i.e. "head entity relation tail entity", which lacks syntactic coherence and has the problem of missing morphemes. If the triple short sentence is concatenated with normal text as training corpus, it is equivalent to mixing two different syntactic structures together, which will cause the model to learn incorrect syntactic knowledge. The present application constructs a first entity knowledge short sentence with correct syntax by the first preset grammar rule, i.e. the method of completing the missing morphemes. Exemplary explanation: by simply concatenating the triple (Hayao Miyazaki, nationality, Japan), a short sentence "Hayao Miyazaki nationality Japan" is formed, which has the problem of missing morphemes. By completing the missing morphemes according to the first preset grammar rule, a first entity knowledge short sentence with correct syntax "Hayao Miyazaki's nationality is Japan" is constructed.

[0035] Further, when there is an entity description of the target entity in the target language knowledge base or the associated language knowledge base, the first sentence of the entity description is added to the first entity knowledge short sentence list.

[0036] For S400, input the first training corpus X into the XLM-R model to obtain the first loss function L1 with the goal of predicting masked entities.

[0037] Specifically, input the first training corpus X into the XLM-R model, adopt the pre-training task of the Masked Language Model, and obtain the first loss function L1 with the goal of predicting masked entities.

[0038] In an example of the present invention, as Figure 2 shown, mask the Japanese entity "Hayao Miyazaki", and use the delimiter of the XLM-R model to splice the first entity knowledge short sentence Z i "Hayao Miyazaki's nationality is Japan" and the first intermediate sentence X0 " <mask> <mask> <mask>The first training corpus X is obtained from the supervised work "Laputa Castle in the Sky" i "Studio Ghibli director Hayao Miyazaki's nationality is Japanese <mask> <mask> <mask>Supervise the work "Castle in the Sky", and train the XLM-R model to predict masked entities.

[0039] In summary, when the first entity knowledge short sentence comes from the target language knowledge base, the present invention injects the knowledge and context information associated with the target entity into the language representation of the first entity. When the first entity knowledge short sentence comes from the associated language knowledge base, on the one hand, the present invention injects cross-language entity knowledge, and on the other hand, enhances the cross-language entity alignment ability, improving the effectiveness of using the labeled corpus in the associated language for data augmentation in the downstream NER task. By integrating the explicit knowledge in the target language knowledge base and the associated language knowledge base into the pre-training process, more context information is introduced for the first entity, allowing different neighbors of the first entity to be focused on in different contexts, while learning diverse implicit knowledge behind the first entity.

[0040] S500, based on the second entity knowledge short sentence list Y = {Y1, Y2,..., Y j ,..., Y n} and the second intermediate sentence X', construct the second training corpus P = {P1, P2,..., P j ,..., P n}, where n is the number of triples in the target language knowledge base that match the target entity and whose corresponding triples conform to the isA relationship. Y j is the sentence constructed according to the second preset grammar rule for the j-th triple in the target language knowledge base that matches the target entity and whose corresponding triple conforms to the isA relationship. The value range of j is from 1 to n. The second intermediate sentence X' is the sentence obtained by performing non-entity masking on the target sentence according to the second masking rule.

[0041] Among them, the IsA relationship is the membership relationship between an entity and a concept. For example, given an entity "Hayao Miyazaki", the concepts found in the knowledge base that satisfy the isA relationship with the entity include "person's name", "Japanese film director", etc., and "Japanese film director" implies the information that Hayao Miyazaki is a person.

[0042] Specifically, the second masking rule is to randomly mask 15% of the non-entity characters in the target sentence.

[0043] Specifically, in S500, the second entity knowledge short sentence is obtained through the following steps:

[0044] S501, obtain the initial triple list of the first entity corresponding to the target language knowledge base.

[0045] S503, screen the initial triple list, and obtain the triples that satisfy the isA relationship in the initial triple list as the target triples.

[0046] S503, constructing the target triple according to a second preset syntax rule to obtain a second entity knowledge short sentence.

[0047] S600, inputting the second training corpus set P into the XLM-R model to obtain a second loss function L2 by taking the context of the entity as a target.

[0048] In an example of the present application, as shown in Figure 3 the entity "Miyazaki Hayao" is linked to the triple <Miyazaki Hayao <isa><name>, based on the second entity knowledge snippet Y j "miyazaki jun wa jinmei" and the second intermediate sentence X' "miyazaki jun <mask>The Sky of the Governor's Mansion <mask>"Raputa", obtaining a second training corpus P j "Shinji is a person's name Miyazaki Shinji <mask>The Sky of the Governor's Mansion <mask>"Raphtor" and train the XLM-R model to predict the masked context of the entity.

[0049] Based on S500-S600, the present application masks the non-entity characters in the target sentence, predicts the context of the target entity in the case of the corresponding first entity and entity concept in the known target language knowledge base, and can alleviate the problem of "shortcut" in the prediction process through S500-S600, that is, the model does not fully schedule the network structure and understanding ability, but only remembers the entities "seen" in the pre-training stage, and performs poorly on "unseen" entities. The model learns the context and position of the entity in the sentence, thereby helping to generalize more effectively in the prediction process.

[0050] S700, based on the third entity knowledge sentence A0 and the target sentence, a third training corpus B is constructed, wherein A0 is a second entity knowledge sentence in the second entity knowledge sentence list Y that satisfies the preset filtering rule. The tail entity corresponding to the isA relationship is masked according to the third masking rule.

[0051] Specifically, the preset filtering rule is that the corresponding tail entity type of the triple corresponding to the isA relationship of the triple matching the target entity in the target language knowledge base and the corresponding triple is a preset candidate entity type.

[0052] Further, the preset candidate entity type set includes person, place, and organization. For example, the target language is Japanese, the first entity is Miyazaki Hayao, and the second entity knowledge sentence list Y includes "Miyazaki Hayao is a person" and "Miyazaki Hayao is a Japanese film director". According to the preset filtering rule, "person" satisfies the preset candidate entity type set, so "Miyazaki Hayao is a person" is masked according to the third masking rule, and "Miyazaki Hayao is a person" is masked. "Miyazaki Hayao is a person" is used as the third entity knowledge sentence A0. Because the entity type can only be one of the preset candidate entity types, only one third entity knowledge sentence A0 is obtained.

[0053] Specifically, those skilled in the art know that any one of the prior art can achieve the preset filtering rule that the tail entity type of the triple corresponding to the isA relationship of the corresponding first entity in the target language knowledge base is the preset candidate entity type, which belongs to the protection scope of the present application, and will not be repeated here.

[0054] S800, input the third training corpus B into the XLM-R model, and acquire the third loss function L3 according to the preset candidate entity type set with the target entity type.

[0055] Specifically, the third loss function is wherein C r is the true probability that the third entity is of the r-th preset candidate entity type, is the predicted probability that the third entity is of the r-th preset candidate entity type, and the third entity is the corresponding entity in the third entity knowledge short sentence A0.

[0056] In an example of the present invention, as Figure 4 shown, the third entity knowledge short sentence A0 "Hayao Miyazaki <mask>" and the concatenation target sentence "Castle in the Sky Laputa" directed by Miyazaki, a third training sentence "Miyazaki is the director of Castle in the Sky" is obtained. <mask>"The work of director Miyazaki Shun, Sky Castle Rapport", is input into the XLM-R model to predict the type of Miyazaki Shun.

[0057] Therefore, this invention only introduces the triplet knowledge of the isA relation, i.e., the type of the entity, in steps S500-S800, thus avoiding the introduction of excessive knowledge and causing a shift in the model's focus. The prediction of the first entity and its corresponding type in the target language knowledge base supplements the supervision signal, encouraging the XLM-R model to fully utilize its potential and identify the contextual knowledge of this type of entity. On the other hand, this task has certain similarities to downstream named entity recognition tasks. In this way, the model acquires specific knowledge of the named entity recognition task during pre-training, thereby achieving better performance in downstream named entity recognition tasks.

[0058] S900, by minimizing the loss function L=αL1+βL2+γL3, the XLM-R model that meets the preset loss condition is marked as the XLM-RK model, where α, β and γ are preset weight parameters.

[0059] Specifically, the present invention further includes performing the following steps before S100:

[0060] S90, Obtain the list of unlabeled statements in the target language F = {F1, F2, ..., F...} x F q }, F x It is the xth unlabeled statement in the target language, where x ranges from 1 to q, and q is the number of unlabeled statements, where q > the preset number threshold Q0.

[0061] Specifically, the preset quantity threshold Q0 can be determined according to actual needs to ensure the quantity of unannotated sentences in the target language.

[0062] S93, when F x When it contains entities, F x Mark the target statement and obtain the target statement list GF = {GF1, GF2, ..., GF...} y , ..., GF p }, for target statements GF1 to GF p Execute S100 respectively, where GF y It is the y-th target statement, where y ranges from 1 to p, and p is the number of target statements.

[0063] In a specific implementation example, S900 also includes:

[0064] S910, Obtain the target statement GF y The corresponding objective loss function L is achieved by minimizing the objective statement GF. y A corresponding target loss function L is used to train the XLM-R model.

[0065] S930, evaluate the trained XLM-R model, and when the evaluation result meets the preset loss condition, mark the trained XLM-R model as an XLM-RK model.

[0066] In summary, the present application obtains the target entity in the target sentence, links the target entity to obtain the first entity, predicts the mask entity based on the first entity knowledge short sentence and the target sentence with entity mask to obtain the first loss function, predicts the context of the entity based on the second entity knowledge short sentence and the second intermediate short sentence with non-entity mask to obtain the second loss function, predicts the entity type based on the third entity short sentence and the target sentence to obtain the third loss function, obtains the XLM-RK model by minimizing the weighted sum function of the three loss functions, and the obtained XLM-RK model improves the cross-language knowledge transfer and knowledge alignment capability, integrates the public multilingual knowledge base, better realizes cross-language knowledge alignment and language understanding, and improves the performance on the low-resource entity extraction task.

[0067] Further, after S900, it further includes:

[0068] S901, fine-tuning the XLM-RK model using multilingual entity recognition annotated corpus to obtain the fine-tuned XLM-RK model.

[0069] S903, using the annotated corpus of the target language and the associated language to fine-tune the fine-tuned XLM-RK model to obtain the final XLM-RK model.

[0070] Specifically, those skilled in the art know that fine-tuning the XLM-RK model means constructing the model structure required by the task based on the XLM-RK model, defining the loss function, updating the specified parameters of the model by minimizing the loss function, and obtaining the final model.

[0071] Further, the specified parameters of the model are determined according to actual needs.

[0072] Based on S901-S903, the XLM-RK model is fine-tuned, the knowledge related to the task is obtained in S901 using large-scale multilingual annotated corpus, the task focus is readjusted to the target language named entity recognition in S903, the data noise problem introduced in S901 by the language with large difference from the target language is alleviated, and at the same time, the training data is expanded based on the high-quality annotated corpus of the associated language, the cross-language entity alignment capability obtained by the model in the pre-training stage is utilized, and the target language entity recognition effect is improved.

[0073] While certain specific embodiments of the application have been described in detail herein for the purposes of exemplification, numerous other variations and modifications will be apparent to persons skilled in the art. Alterations and modifications of detail can be made by those skilled in the art, having the benefit of the above description, without departing from the spirit and scope of the application. It is intended that all such alterations and modifications be included within the scope of the application whose limits are to be determined only by the appended claims.< / mask> < / mask> < / mask> < / mask> < / mask> < / mask> < / isa> < / mask> < / mask> < / mask> < / mask> < / mask> < / mask>

Claims

1. A low-resource language entity extraction system integrating knowledge, characterized in that, The system includes a target language knowledge base, a related language knowledge base, a processor, and a memory storing computer programs. The related languages ​​are languages ​​from the same language family as the target language or languages ​​that overlap with the target language, and the related languages ​​are high-resource languages. When the processor executes the computer program, it performs the following steps: S100, retrieve the target entity from the target sentence in the unlabeled corpus; S200, link the target entity with the target language knowledge base and the associated language knowledge base to obtain the first entity corresponding to the target entity, wherein the first entity is the entity corresponding to the target entity in the target language knowledge base and the associated language knowledge base; S300, Based on the first entity knowledge short sentence list Z = {Z1, Z2, ..., Z...} i , ..., Z m The first training corpus X = {X1, X2, ..., X0} is constructed using the first intermediate statement X0. i , ..., X m }, where the i-th first training corpus X i By splicing X0 onto Z i Then we get Z. i The statement is constructed according to the first preset syntax rule for the i-th triple corresponding to the first entity. The value of i is from 1 to m, where m refers to the number of target language and related language triples corresponding to the first entity. The first intermediate statement X0 is the statement after entity masking of the target statement according to the first masking rule. S400, input the first training corpus X into the XLM-R model and obtain the first loss function L1 with the goal of predicting masked entities; S500, Based on the second entity knowledge short sentence list Y = {Y1, Y2, ..., Y...} j , ..., Y n The second training corpus P = {P1, P2, ..., P'} is constructed from the second intermediate statement X'. j , ..., P n }, where n is the number of triplets in the target language knowledge base that match the target entity and whose corresponding triplets satisfy the isA relation, and Y j It is the statement constructed according to the second preset syntax rule of the j-th triplet in the target language knowledge base that matches the target entity and whose corresponding triplet conforms to the isA relation. The value of j ranges from 1 to n. The second intermediate statement X′ is the statement after the target statement is non-entity masked according to the second masking rule. S600, input the second training corpus P into the XLM-R model, and obtain the second loss function L2 with the goal of predicting the context of entities; S700, based on the concatenation of the third entity knowledge short sentence A0 and the target sentence, a third training corpus B is constructed. A0 is a second entity knowledge short sentence in Y that satisfies the preset filtering rules and the statement after the tail entity corresponding to the isA relation is masked according to the third masking rules. S800, input the third training corpus B into the XLM-R model, and obtain the third loss function L3 based on the preset candidate entity type set with the goal of predicting entity type; S900, by minimizing the loss function L=αL1+βL2+γL3, the XLM-R model that meets the preset loss condition is marked as the XLM-RK model, where α, β and γ are preset weight parameters.

2. The system according to claim 1, characterized in that, Prior to S100, it also includes: S90, Obtain the list of unlabeled statements in the target language F = {F1, F2, ..., F...} x F q }, F x It is the xth unlabeled statement in the target language, where x ranges from 1 to q, q is the number of unlabeled statements, and q > the preset number threshold Q0; S93, when F x When it contains entities, F x Mark the target statement and obtain the target statement list GF = {GF1, GF2, ..., GF...} y , ..., GF p }, for target statements GF1 to GF p Execute S100 respectively, where GF y It is the y-th target statement, where y ranges from 1 to p, and p is the number of target statements.

3. The system according to claim 1, characterized in that, Obtain the target language knowledge base and related language knowledge base through Wikidata.

4. The system according to claim 1, characterized in that, The third loss function is Among them, C r Is the third entity the true probability that it is the r-th preset candidate entity type? The third entity is the predicted probability of the r-th preset candidate entity type, and the third entity is the entity corresponding to the third entity knowledge phrase A0.

5. The system according to claim 1, characterized in that, Following the S900 are: S901, the XLM-RK model is fine-tuned using multilingual entity recognition labeled corpus to obtain the fine-tuned XLM-RK model; S903 uses annotated corpora in the target language and related languages ​​to perform a second fine-tuning on the fine-tuned XLM-RK model to obtain the final XLM-RK model.

6. The system according to claim 1, characterized in that, S200 specifically includes: S201, Connect the target entity to the target language knowledge base to obtain the first entity corresponding to the target entity in the target language knowledge base; S202, link the target entity with the associated language knowledge base to obtain the first entity corresponding to the target entity in the associated language knowledge base.

7. The system according to claim 1, characterized in that, Specifically, S202 includes linking the target entity and the related language knowledge base as follows: query the name in the related language corresponding to the target entity through the API provided by Wikidata, and link the name entity in the related language corresponding to the target entity to the related language knowledge base to obtain the first entity corresponding to the target entity in the related language knowledge base.

8. The system according to claim 1, characterized in that, The preset candidate entity type set includes people, places, and organizations.

9. The system according to claim 1, characterized in that, If an entity description of the target entity exists in the target language knowledge base or the related language knowledge base, add the first sentence of the entity description to the first entity knowledge short sentence list.

10. The system according to claim 1, characterized in that, In S500, the second entity knowledge phrase is obtained through the following steps: S501, Obtain the initial triplet list of the first entity corresponding to the target language knowledge base; S503, filter the initial triple list and obtain the triples in the initial triple list that satisfy the isA relationship as the target triples; S503, construct the target triple according to the second preset grammar rules to obtain the second entity knowledge short sentence.

Images