A cross-language entity linking method, system, device and terminal

Abstract

The application belongs to the technical field of cross-language entity linking, and discloses a cross-language entity linking method, system, medium, equipment and terminal.The cross-language entity linking method comprises the following steps: searching for candidate entities and constructing an entity name index; constructing an entity linking model based on different clues; performing vector embedding and fusion based on attribute-based entity linking; and performing multi-clue entity linking based on co-occurring entities.The application uses three kinds of clues to realize entity linking, the clues are entity attributes, co-occurring entities and context descriptions, and neural network structures such as LSTM, CNN and GCN are combined to perform word embedding representation and network modeling.Compared with traditional entity linking technology, the application has stronger semantic representation ability and the linking method has cross-language capability.Through analyzing different forms of information, the application realizes a multi-clue cross-language entity linking algorithm with three different structures, and completes accurate linking of news text cross-language entities such as persons and institutions and a given knowledge base.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of knowledge graph, and particularly relates to a cross-language entity linking method, system, medium, device and terminal. BACKGROUND

[0002] At present, in the big data era, with the exponential growth of network data, it also poses a challenge to quickly and accurately obtain target information. The knowledge graph has been initially formed by using information extraction technology to obtain structured data from text data, but there are still problems in its real and substantial use, because natural language has multiple characteristics of complexity, polysemy and fuzziness. Specifically, the same entity can be expressed by different texts (polysemy), and the same text can express multiple different entities (monosemy). By performing entity linking, that is, linking the entity reference in the text with the entity in the knowledge base, the text data can be converted into text with entity annotation, thereby helping people and computers understand the specific meaning of the text. Its potential applications include information extraction, information retrieval and knowledge base filling.

[0003] At the same time, the structured or unstructured data obtained from the network is often composed of multiple languages, and how to integrate multi-language data and structure and uniformly describe and store is a big problem in the industry at present.

[0004] Cross-language entity linking can link knowledge bases of different languages, and the linked knowledge base allows users to simultaneously search and view the same knowledge expressed in different languages, which has practical significance.

[0005] Therefore, in view of the problem that the prior art cannot perform cross-language linking in the scenario where multiple language entities exist in the knowledge graph, it is urgent to design a cross-language entity linking method and system.

[0006] Through the above analysis, the problems and defects of the prior art are that the prior art cannot perform cross-language linking in the scenario where multiple language entities exist in the knowledge graph. SUMMARY

[0007] In view of the problems existing in the prior art, the present application provides a cross-language entity linking method, system, medium, device and terminal, and particularly relates to a cross-language entity linking method, system, medium, device and terminal based on multiple clues.

[0008] The application is implemented as a cross-language entity linking method, which comprises searching for candidate entities and constructing an entity name index; and implementing cross-language entity linking by using a constructed entity linking model; the entity linking model comprises three entity linking calculation models: an attribute-based, a co-occurring entity-based and a context-based entity linking model; according to different clues of input information, a corresponding model is selected for calculation to obtain an entity linking result.

[0009] Further, the construction method of the entity linking model comprises:

[0010] constructing an entity linking model based on different clues;

[0011] vector embedding and fusion based on attribute-based entity linking;

[0012] co-occurring entity-based multi-clue entity linking.

[0013] Further, the cross-language entity linking method further comprises:

[0014] for a given language, candidate entities in a knowledge graph are found; according to entity names and entity descriptions in Chinese and English, each different description text is mapped to a target entity name, and an entity name index is constructed.

[0015] after the index is constructed, a referential item is subjected to word embedding, similarity with entities in the graph is calculated, after candidate items with similarity greater than a specified threshold are screened, a vector fusion method is used to map entities in other languages into the same vector space.

[0016] constructing an entity linking model based on different clues: attribute-based, co-occurring entity-based and context-based entity linking, and selecting different entity linking models through a candidate entity set and provided clues to obtain a cross-language linking result of the referential item in different languages.

[0017] attribute-based entity linking: after vector embedding of the entity referential item and attribute description, fusion is performed through a graph convolution model, the sum average of respective attribute vectors is calculated as a similarity coefficient before the entity, and thus the linked entity is obtained.

[0018] co-occurring entity-based entity linking: attention features of the proportion of co-occurring entities in the center entity are calculated by using a similarity coefficient, a vectorization feature of the co-occurring relationship is combined to calculate the similarity of the target entity and the entity in the graph, and thus the linked entity is obtained.

[0019] context-based entity linking: the entity referential and the context belonging to the entity coreference chain are combined together to form a sentence sequence, a neural network model is used to model the sequence, similarity calculation is performed with the entity in the graph, and thus the linked entity is obtained.

[0020] Further, the cross-language entity linking method comprises the following steps:

[0021] Step one, obtain entity name-entity description text from existing knowledge base, map entity description text in different languages to its entity name in the knowledge base according to the existing knowledge base, and construct entity name index using multilingual knowledge base; constructing index can effectively reduce the retrieval time of candidate entities.

[0022] Step two, obtain the target text to be linked, extract the entity referring item and entity context description to be linked from the text, obtain the attributes possessed by the entity, and the co-occurring entities appearing in the text.

[0023] Step three, calculate the similarity between the entity referring item and the entity in the knowledge base through the entity name index, compare the similarity with the preset candidate entity threshold, and determine whether there is a similarity score less than the candidate entity threshold;

[0024] If yes, give up the entity for the referring item; if no, add the entity to the candidate entity set of the referring item, and go to step three; the first step of screening candidate entities can reduce the calculation amount of subsequent models.

[0025] Step four, use vector fusion method to map word vectors in other languages to English word vectors, and map vectors in different languages to the same vector space; map all entities to the same vector space to realize cross-language entity comparison.

[0026] Step five, select different entity linking models through the candidate entity set and the extracted clues, merge the linking results of each model, and obtain the final cross-language linking result of the referring item in different languages; selecting different linking models according to the clues can more comprehensively utilize the information of the entity for entity linking calculation, so that the linking result is more accurate and reasonable.

[0027] Further, in step one, for the entity name index of cross-language entity linking, use English entity name as the basic entity name index, and construct entity name index from the respective target language knowledge base; wherein, the CBOW (Continuous Bag-of-Words) model in the classic word2vec is used to generate word embeddings in each language.

[0028] After step one, the description text-title needs to be used as positive samples, and other samples in a batch are used as negative samples to calculate the similarity between the referring item and the entity; using positive and negative samples to obtain more accurate entity similarity, and fine-tuning two pre-training models through positive and negative samples, one for embedding the referring item and the other for embedding the candidate entity; on the basis of training completion, store all knowledge base entities in the form of vectors in the database.

[0029] Further, in step four, a plurality of different vector fusion methods are compared, including:

[0030] CCA (Canonical Correlation Analysis): Perform SVD (Singular Value Decomposition) on texts in different languages, and then apply canonical correlation analysis CCA to the vector pairs of aligned words in parallel corpus to learn vectors; for cross-language entity linking, use the title mapping constructed from the cross-language links in Wikipedia to build embeddings.

[0031] MultiCCA: On the basis of CCA, use linear operators to project pre-trained monolingual embeddings in each language except the mother tongue into the vector space of pre-trained mother tongue word embeddings.

[0032] LS (Weighted Least Squares): Other language embeddings are directly projected onto the mother tongue, and the mapping is constructed by multiple regression.

[0033] Further, in step five, attribute-based linking, co-occurrence entity-based linking, and context-based linking.

[0034] The attribute vector is calculated using the following distance method:

[0035] O AE =-∑ (a,c)∈H w a,c ·logp(c|a);

[0036] Where w represents the type weight, and if the attributes have the same type, the weight value will be higher, so that the attributes with relevance are embedded in similar spaces.

[0037] The proportion of co-occurrence entities in the center entity is calculated using the similarity coefficient, and the calculation method is:

[0038] e ij =a([Wh i ][Wh j ]),j∈N i ;

[0039] Where W is a parameter matrix, and h is the vector of each entity after vector fusion.

[0040] Vector embedding of the referent is performed using multi-granularity text:

[0041]

[0042] The context of a mention (entity mention) is modeled by using an NTN (Neural Tensor Network) function to form a fine-grained context representation of the mention.

[0043] Another object of the present application is to provide a cross-language entity linking system applying the cross-language entity linking method, which comprises:

[0044] A candidate entity searching module is configured to find candidate entities in a knowledge graph for a given language mention;

[0045] An entity name index construction module is configured to map each different description text to its target entity name according to the entity name and entity description in Chinese and English, and construct an entity name index;

[0046] A vector fusion module is configured to perform word embedding on the mention after the index construction, calculate the similarity between the mention and entities in the graph, filter candidate items with a similarity greater than a specified threshold, and map entities in other languages to the same vector space by using a vector fusion method;

[0047] An entity linking model construction module is configured to perform attribute-based, co-occurring entity-based and context-based entity linking, and select different entity linking models based on the candidate entity set and the provided clues to obtain cross-language linking results of the mention in different languages;

[0048] A linked entity acquisition module is configured to perform attribute-based entity linking, perform vector embedding on the entity mention and attribute description, fuse them through a graph convolution model, calculate the sum average of the respective attribute vectors as the similarity coefficient between the entities, and thus obtain the linked entity; and / or

[0049] A co-occurring entity-based entity linking is configured to calculate the attention feature of the proportion of the co-occurring entity in the center entity by using a similarity coefficient, calculate the similarity between the target entity and the entity in the graph by combining the vectorization feature of the co-occurring relationship, and obtain the linked entity; and / or

[0050] A context-based entity linking is configured to combine the entity mention and the context belonging to the entity coreference chain together to form a sentence sequence, model the sequence by using a neural network model, calculate the similarity with the entity in the graph, and obtain the linked entity.

[0051] Another object of the present application is to provide a computer device comprising a memory and a processor, wherein the memory stores a computer program, and the computer program is executed by the processor to make the processor perform the following steps:

[0052] Search candidate entities, build entity name index; build entity linking model based on different clues; vector embedding and fusion based on attribute-based entity linking; multi-clue entity linking based on co-occurring entities.

[0053] Another object of the present application is to provide a computer-readable storage medium storing a computer program which, when executed by a processor, causes the processor to perform the following steps:

[0054] Search candidate entities, build entity name index; build entity linking model based on different clues; vector embedding and fusion based on attribute-based entity linking; multi-clue entity linking based on co-occurring entities.

[0055] Another object of the present application is to provide an information data processing terminal for implementing the cross-language entity linking system.

[0056] In combination with the above technical solutions and the technical problems solved, the advantages and positive effects of the technical solution to be protected by the present application are analyzed from the following aspects:

[0057] First, in view of the technical problems existing in the prior art and the difficulty in solving the problems, the technical solution to be protected by the present application and the results and data in the research and development process are closely combined to analyze in detail and profoundly how the technical solution solves the technical problems and brings some creative technical effects after solving the problems. The specific description is as follows:

[0058] The present application uses three clues to realize entity linking, the clues are entity attributes, co-occurring entities and context description, and neural network structures such as LSTM, CNN and GCN are combined to realize word embedding representation and network modeling. Compared with traditional entity linking technology, the present application mainly has the following two technical features: (1) stronger semantic representation ability; (2) cross-language linking method.

[0059] Second, taking the technical solution as a whole or from the perspective of the product, the technical solution to be protected by the present application has the following technical effects and advantages, which are described in detail as follows:

[0060] The cross-language entity linking solves the entity linking problem of different language entities in a multi-language knowledge graph. The present application realizes three different structure multi-clue cross-language entity linking algorithms by analyzing different forms of information, and completes the precise linking of news text cross-language entities such as persons and institutions and a given knowledge base.

[0061] Third, as the creative auxiliary evidence of the claims of the present application, it is also embodied in the following important aspects:

[0062] The technical scheme of the present application solves the technical problem that people have been eager to solve but have failed to solve successfully:

[0063] The prior art in the industry cannot perform cross-language linking in the scenario of multiple languages of entities existing in a knowledge graph at the same time, and the present application solves this problem. The present application constructs a cross-language vector index, uses three different structures of multi-thread cross-language entity linking algorithms, and completes the accurate linking of cross-language news text entities such as persons and institutions and a given cross-language knowledge base. BRIEF DESCRIPTION OF DRAWINGS

[0064] In order to more clearly illustrate the technical scheme of the embodiments of the present application, the drawings needed to be used in the embodiments of the present application will be briefly introduced as follows. Obviously, the drawings described below are only some embodiments of the present application, and other drawings can also be obtained from these drawings by those skilled in the art without creative labor.

[0065] Figure 1 is a cross-language entity linking method flowchart provided by the embodiments of the present application;

[0066] Figure 2 is a cross-language entity linking system structure block diagram provided by the embodiments of the present application;

[0067] Figure 3 is an attribute-based entity linking search engine page diagram provided by the application embodiments of the present application;

[0068] Figure 4 is a co-occurring entity-based entity linking search engine page diagram provided by the application embodiments of the present application;

[0069] Figure 5 is a context-based entity linking search engine page diagram provided by the application embodiments of the present application; in the diagram: 1, candidate entity search module; 2, entity name index construction module; 3, vector fusion module; 4, entity linking model construction module; 5, linked entity acquisition module. DETAILED DESCRIPTION

[0070] In order to make the purpose, technical scheme and advantages of the present application more clear and apparent, the present application will be further described in detail below in combination with embodiments. It should be understood that the specific embodiments described here are only used to explain the present application and not to limit the present application.

[0071] In view of the problems existing in the prior art, the present application provides a cross-language entity linking method, system, medium, device and terminal, which will be described in detail below in combination with the drawings.

[0072] As Figure 1As shown, the cross-language entity linking method provided by the embodiment of the application comprises the following steps:

[0073] S101, obtaining entity name-entity description text from an existing knowledge base, mapping entity description texts in different languages to entity names in the knowledge base according to the existing knowledge base, and constructing an entity name index by using a multilingual knowledge base; constructing the index can effectively reduce the retrieval time of candidate entities.

[0074] S102, obtaining a target text to be linked, extracting an entity referring term and an entity context description to be linked from the text, obtaining attributes possessed by the entity, and co-occurring entities appearing in the text.

[0075] S103, calculating the similarity between the entity referring term and the entity in the knowledge base through the entity name index, comparing the similarity with a preset candidate entity threshold, and determining whether there is a similarity score less than the candidate entity threshold; if yes, the entity is abandoned for the referring term; if no, the entity is added to a candidate entity set of the referring term, and S103 is turned to; the first step of screening the candidate entity can reduce the calculation amount of the subsequent model.

[0076] S104, mapping word vectors of other languages into English word vectors by using a vector fusion method, and mapping vectors in different languages into the same vector space; all entities are mapped into the same vector space to realize cross-language entity comparison.

[0077] S105, selecting different entity linking models through the candidate entity set and the extracted clues, merging the linking results of the models, and obtaining the final cross-language linking result of the referring term in different languages; different linking models are selected according to the clues, the information of the entity can be more comprehensively used for entity linking calculation, so that the linking result is more accurate and reasonable.

[0078] As shown, Figure 2 The cross-language entity linking system provided by the embodiment of the application comprises:

[0079] The candidate entity search module 1 finds candidate entities in the knowledge graph for a given language referring term;

[0080] The entity name index construction module 2 is used for mapping each different description text to its target entity name according to the entity name and the entity description in Chinese and English, and constructing an entity name index;

[0081] The vector fusion module 3 is used for performing word embedding on the referring term after the index is constructed, calculating the similarity between the referring term and the entity in the graph, screening candidate items with a similarity greater than a specified threshold, and mapping entities in other languages into the same vector space by using a vector fusion method;

[0082] An entity linking model construction module 4 is configured to construct an attribute-based entity linking model, a co-occurrence entity-based entity linking model, and a context-based entity linking model, and select different entity linking models according to the candidate entity set and the provided clues, so as to obtain cross-language linking results of the referring term in different languages.

[0083] A linked entity acquisition module 5 is configured to perform attribute-based entity linking, fuse the vector embedding of the entity referring term and the attribute description through a graph convolution model, calculate the sum average of the respective attribute vectors as a similarity coefficient between the entities, and thus obtain the linked entity; and / or

[0084] The co-occurrence entity-based entity linking model is configured to calculate the attention feature of the proportion of the co-occurrence entity in the central entity by using a similarity coefficient, combine the vectorized features of the co-occurrence relationship to calculate the similarity between the target entity and the entities in the graph, and thus obtain the linked entity; and / or

[0085] The context-based entity linking model is configured to combine the entity referring term and the context belonging to the entity co-reference chain to form a sentence sequence, model the sequence by using a neural network model, and calculate the similarity with the entities in the graph, so as to obtain the linked entity.

[0086] Cross-language entity linking can exacerbate the ambiguity of language, and involves linking the referring term in a non-native language document to an entity on a native language knowledge graph. In order to compare cross-language text clues, the present application needs to calculate the similarity between cross-language text segments. The present application adopts a model combining convolution and tensor network to train the fine-grained similarity and difference between the query and the candidate document from multiple angles. In addition, by using effective use of multilingual embedding, this training system can be applied to other languages in zero-shot learning. The main process is: candidate entity search, vector embedding and fusion, and multi-clue entity linking.

[0087] For a given language referring term, the first task is to find the candidate entity in the knowledge graph. According to the entity name and the entity description of Chinese and English, the present application maps each different description text to the target entity name, and constructs an entity name index.

[0088] After the index is constructed, the referring term is subjected to word embedding, and the similarity with the entities in the graph is calculated to screen the candidate items with a similarity greater than a specified threshold. Then, the vector fusion method is used to map the entities of other languages into the same vector space.

[0089] An entity linking model based on different clues is constructed: an attribute-based entity linking model, a co-occurrence entity-based entity linking model, and a context-based entity linking model, and different entity linking models are selected according to the candidate entity set and the provided clues, so as to obtain cross-language linking results of the referring term in different languages.

[0090] The attribute-based entity linking fuses the vector embedding of the entity referring item and the attribute description through a graph convolution model, calculates the sum average of each attribute vector as the similarity coefficient between entities, and thus obtains the linked entity.

[0091] The co-occurring entity-based entity linking calculates the attention feature of the proportion of the co-occurring entity in the center entity by using a similarity coefficient, calculates the similarity between the target entity and the entity in the graph by combining the vectorization feature of the co-occurring relationship, and obtains the linked entity.

[0092] The context-based entity linking combines the entity reference and the context belonging to the entity co-reference chain to form a sentence sequence, models the sequence by using a neural network model, calculates the similarity with the entity in the graph, and obtains the linked entity.

[0093] The cross-language entity linking method based on a knowledge graph provided by the embodiment comprises the following steps:

[0094] Step one, obtaining entity name-entity description text from an existing knowledge base, mapping entity description texts in different languages to the entity names in the knowledge base according to the existing knowledge base, and constructing an entity name index by using a multilingual knowledge base; the index construction can effectively reduce the retrieval time of candidate entities.

[0095] Step two, obtaining a target text to be linked, extracting the entity referring item and the entity context description to be linked from the text, obtaining the attributes possessed by the entity, and the co-occurring entity appearing in the text.

[0096] Step three, calculating the similarity between the entity referring item and the entity in the knowledge base through the entity name index, comparing the similarity with a preset candidate entity threshold, and determining whether there is a similarity score less than the candidate entity threshold;

[0097] If yes, the entity is abandoned for the referring item; if no, the entity is added to the candidate entity set of the referring item, and step three is turned to; the one-step selection of the candidate entity can reduce the calculation amount of the subsequent model.

[0098] Step four, mapping the word vectors of other languages into English word vectors by using a vector fusion method, mapping the vectors in different languages into the same vector space, and mapping all entities into the same vector space to realize cross-language entity comparison.

[0099] Step five, selecting different entity linking models by using the candidate entity set and the extracted clues, merging the linking results of each model, and obtaining the final cross-language linking result of the referring item in different languages; according to the clues, different linking models are selected, the information of the entity is more comprehensively used for entity linking calculation, the linking result is more accurate and reasonable.

[0100] The S1 further comprises: for cross-language entity linking entity name indexing, in addition to using English Wikipedia index (built from English snapshot), entity name index is built from respective target language Wikipedia.

[0101] Word embeddings for each language are generated using CBOW model in classic word2vec. All knowledge base entities are stored in database as vector representations.

[0102] The S4 uses a variety of different vector fusion methods for comparison, including:

[0103] Canonical Correlation Analysis (CCA): first, SVD is performed on texts in different languages, and then canonical correlation analysis CCA is applied to vector pairs of aligned words in parallel corpus, so as to learn vectors. For cross-language entity linking, the present application uses embeddings built from title mappings obtained from cross-language links in Wikipedia.

[0104] MultiCCA: on the basis of CCA, and using linear operators, pre-trained monolingual embeddings in each language (except mother tongue) are projected to the vector space of pre-trained mother tongue word embeddings.

[0105] Weighted Least Squares (LS): other language embeddings are directly projected onto the mother tongue, and the mapping is built through multivariate regression.

[0106] The S5 further comprises: attribute-based linking, co-occurring entity-based linking and context-based linking.

[0107] The attribute vectors are calculated by using the following distance method:

[0108]

[0109] Where H represents the attribute set, a and c represent different attributes, and w represents the type weight. If the attributes have the same type, the weight value will be higher, so that the attributes with correlation are embedded in similar spaces.

[0110] The proportion of co-occurring entities in the central entity is calculated by using the similarity coefficient, and the calculation method is:

[0111] e ij =a([Wh i ][Wh j ]),j∈N i

[0112] Where W is a parameter matrix, and h is the vector of each entity after vector fusion.

[0113] A multi-granularity text is used to vector embed a mention:

[0114]

[0115] An NTN function is used to model a context of the mention to form a fine-grained context representation of the mention, wherein l represents 4 words on the left side of the mention, and r represents 4 words on the right side of the mention.

[0116] An application embodiment of the present application is a retrieval engine based on multi-thread entity linking.

[0117] Figure 3 It is an attribute-based entity linking retrieval engine page, such as Figure 3 In the present embodiment, candidate entities are first retrieved in a knowledge graph through entity mentions, and then an entity mention item and an attribute description input by a user are converted into attribute vectors, candidate entities and their neighbor information are fused through a graph convolution model, and then a similarity coefficient between the entities is calculated through an average sum of the respective attribute vectors, several entities most similar to the user input are calculated, and the entities are ranked from high to low according to the similarity, and are used as a result of linked entities; if there is a cross-language entity, the cross-language entity is displayed in the form of a multi-language label, such as "CN / EN".

[0118] Figure 4 It is an entity linking retrieval engine page based on co-occurring entities, such as Figure 4 In the present embodiment, candidate entities are first retrieved in a knowledge graph through entity mentions, then graph set neural networks are used to fuse co-occurring entity information of co-occurring entities in combination with a center entity, an internal dependency degree between co-occurring relationships and an importance degree of neighbor nodes to the center entity are calculated, candidate entities are modeled in the same way, and finally, similarity between an input entity of a user and the candidate entities is calculated in combination with co-occurring relationship vectorization features and attention features of the center importance, candidate entities exceeding a threshold are ranked from high to low according to the similarity, and are used as a result of linked entities; if there is a cross-language entity, the cross-language entity is displayed in the form of a multi-language label.

[0119] Figure 5 It is an entity linking retrieval engine page based on a context, such as Figure 5In the embodiment, first, candidate entities are retrieved in the knowledge graph through entity mentions, then the entity mention items and context descriptions are combined to form a sentence sequence containing entity mentions, then a sentence embedding is generated using a coupon neural network, then a context window size of 4 of the mention is taken, the context is modeled using an NTN model and bidirectional features of the text are extracted using a BiLSTM, finally, the candidate entities are also modeled in the same way, in the top NTN model, the vector representation of the user input mention and the vector representation of the candidate entity are spliced to perform feature fusion, after overall normalization, a similarity is obtained, the candidate entities exceeding a threshold are sorted in descending order of similarity, and the result of the linked entity is obtained; if there is a cross-language entity, the result is displayed in the form of a multilingual label.

[0120] It should be noted that the embodiments of the present application can be realized by hardware, software or a combination of software and hardware. The hardware part can be realized by special logic; the software part can be stored in a memory and executed by a suitable instruction execution system, such as a microprocessor or a specially designed hardware. Those skilled in the art can understand that the above-mentioned devices and methods can be realized by computer executable instructions and / or included in processor control code, such as provided on a carrier medium, such as a magnetic disk, CD or DVD-ROM, a programmable memory, such as a read-only memory (firmware), or a data carrier, such as an optical or electronic signal carrier. The devices of the present application and their modules can be realized by hardware circuits, such as very large scale integrated circuits or gate arrays, semiconductors, such as logic chips, transistors, etc., or programmable hardware devices, such as field programmable gate arrays, programmable logic devices, etc., can also be realized by software executed by various types of processors, and can also be realized by a combination of the above-mentioned hardware circuits and software, such as firmware.

[0121] The embodiments of the present application realize three cross-language entity linking algorithms based on different clues, and experiments are performed in a 4000 English news text test set and a Chinese encyclopedia knowledge graph, and the experimental results are as follows:

[0122] Accuracy Recall Attribute-based entity linking 80.38％ 82.50％ Co-occurring entity-based entity linking 80.73％ 89.30％ Context-based entity linking 82.35％ 91.47％

[0123] The above is only a specific embodiment of the present application, but the protection scope of the present application is not limited thereto, any modification, equivalent replacement and improvement made by those skilled in the art within the technical range disclosed by the present application, within the spirit and principles of the present application, should be covered within the protection scope of the present application.

Claims

1. A cross-language entity linking method, characterized in that, The cross-language entity linking method includes: searching for candidate entities and constructing an entity name index; and using the constructed entity linking model to implement cross-language entity linking. The entity linking model includes three entity linking calculation models: attribute-based, co-existing entity-based, and context-based entity linking models. The appropriate model is selected for calculation based on different clues in the input information to obtain the entity linking results. The cross-language entity linking method also includes: For a given language reference, find candidate entities in the knowledge graph; based on the entity names and descriptions in Chinese and English, map each different description text to its target entity name and construct an entity name index; After the index is built, the referential items are word-embedded and their similarity to entities in the graph is calculated. After filtering candidates with similarity greater than a specified threshold, the entities of other languages ​​are mapped to the same vector space using the vector fusion method. Construct entity linking models based on different cues: attribute-based, co-existing entity-based, and context-based entity linking. By selecting different entity linking models through candidate entity sets and providing cues, cross-language linking results of referents in different languages ​​are obtained. Attribute-based entity linking involves embedding entity references and attribute descriptions into vectors, fusing them using a graph convolution model, and calculating the summation average of the attribute vectors as the similarity coefficient between entities, thereby identifying the linked entities. Based on entity links of co-occurrence entities, attention features are used to calculate the proportion of co-occurrence entities in the central entity using similarity coefficients. The similarity between the target entity and entities in the graph is calculated by combining the vectorized features of co-occurrence relationships, thus obtaining the linked entities. Context-based entity linking combines entity references with the context belonging to the entity's core association chain to form a sentence sequence. A neural network model is used to model the sequence and calculate the similarity with entities in the graph to obtain the linked entities. The cross-language entity linking method includes the following steps: Step 1: Obtain entity name-entity description text from the existing knowledge base. Based on the existing knowledge base, map the entity description text in different languages ​​to their entity names in the knowledge base, and build an entity name index using a multilingual knowledge base. Step 2: Obtain the target text to be linked, extract the entity reference and entity context description from the text, obtain the attributes of the entity, and the co-existing entities appearing in the text; Step 3: Calculate the similarity between the entity reference and the entity in the knowledge base using the entity name index, compare the similarity with the preset candidate entity threshold, and determine whether there is a similarity score that is less than the candidate entity threshold. If yes, then discard the entity for that reference; if no, then add the entity to the candidate entity set of the reference and proceed to step three; screening candidate entities in advance reduces the computational load of subsequent models. Step four: Use vector fusion to map word vectors from other languages ​​to English word vectors, mapping vectors from different languages ​​to the same vector space; map all entities to the same vector space to achieve cross-language entity comparison; Step 5: Select different entity linking models using the candidate entity set and extracted clues, merge the linking results of each model, and obtain the final cross-language linking results of the referent in different languages.

2. The cross-language entity linking method as described in claim 1, characterized in that, The method for constructing the entity linking model includes: Construct an entity linking model based on different clues; Vector embedding and fusion based on attribute-based entity links; Multi-thread entity linking based on shared realities.

3. The cross-language entity linking method as described in claim 1, characterized in that, In step one, for cross-language entity link entity name indexes, the English Wikipedia index is used, and entity name indexes are constructed from the respective target language Wikipedia; wherein, the English Wikipedia index is constructed from English snapshots; Word embeddings for each language are generated using the CBOW model from the classic word2vec; all entities in the knowledge base are represented by vectors and stored in the database.

4. The cross-language entity linking method as described in claim 1, characterized in that, In step four, several different vector fusion methods are compared, including: Canonical correlation analysis (CCA): After performing SVD on texts in different languages, CCA is applied to the vector pairs of aligned words in parallel corpora to learn the vectors; for cross-language entity links, embeddings are constructed using title mappings obtained from cross-language links in Wikipedia. MultiCCA: Based on canonical correlation analysis (CCA), linear operators are used to project the pre-trained monolingual embeddings in each language other than the mother tongue onto the vector space of the pre-trained mother tongue word embeddings. Least Squares (LS): Embeddings of other languages ​​are directly projected onto the native language, and the mapping is constructed through multivariate regression.

5. The cross-language entity linking method as described in claim 1, characterized in that, In step five, there are attribute-based links, co-existing entity-based links, and context-based links; Calculate the attribute vector using the following distance method: O AE =-∑ (a,c)∈H w a,c ·logp(c|a); Where H represents the set of attributes, a and c represent different attributes, and w represents the type weight. If the attributes have the same type, the weight value will be higher, thus embedding related attributes into a similar space. The proportion of the co-existing entity in the central entity is calculated using the similarity coefficient. The calculation method is as follows: e ij =a([Wh i ][Wh j ]),j∈N i ; Where W is the parameter matrix and h is the vector obtained by vector fusion of each entity; Vector embedding of referential terms using multi-granular text: The NTN function is used to model the context of a mention, forming a fine-grained context representation of the mention, where l represents the four words to the left of the mention and r represents the four words to the right of the mention.

6. A cross-language entity linking system implementing the cross-language entity linking method according to any one of claims 1 to 5, the cross-language entity linking system comprising: The candidate entity search module finds candidate entities in the knowledge graph for a given language reference. The entity name index building module is used to map each different description text to its target entity name based on the entity name and entity description in Chinese and English, and build the entity name index. The vector fusion module, after constructing the index, is used to embed the referents into words, calculate the similarity with entities in the graph, filter candidates with similarity greater than a specified threshold, and then use the vector fusion method to map entities from other languages ​​to the same vector space. The entity linking model building module is used for entity linking based on attributes, co-existing entities, and context. It selects different entity linking models through candidate entity sets and provided clues to obtain cross-language linking results of referents in different languages. The link entity acquisition module is used for attribute-based entity linking. It embeds entity designations and attribute descriptions into vectors, then fuses them using a graph convolution model. The average of the summed attribute vectors is used as the similarity coefficient between entities, thus identifying the linked entities; and / or Based on entity links of co-occurring entities, attention features are calculated using similarity coefficients to determine the proportion of co-occurring entities within the central entity. Combined with vectorized features of co-occurrence relationships, the similarity between the target entity and entities in the graph is calculated to obtain the linked entities; and / or Context-based entity linking combines entity references with contexts belonging to the entity's core association chain to form a sentence sequence. A neural network model is used to model the sequence and calculate its similarity with entities in the graph to obtain the linked entities.

7. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the cross-language entity linking method as described in any one of claims 1 to 5.

8. An information data processing terminal, characterized in that, The information data processing terminal is used to implement the cross-language entity linking system as described in claim 6.

Images