Entity linking method and apparatus

Abstract

This disclosure provides an entity linking method and apparatus. The method includes: using a vectorization method that matches the target language of the text to be processed, vectorizing the entity semantic descriptions of the text to be processed and candidate entities respectively, to obtain a text vector matrix and a candidate entity vector matrix; fusing the text vector matrix and the candidate entity vector matrix to obtain a first vector matrix; determining association information associated with candidate entities from a pre-established knowledge graph, and determining association feature vectors based on the association relationship between the association information and the text to be processed; the knowledge graph is constructed based on a multi-source heterogeneous entity library; fusing the first vector matrix and the association feature vectors to obtain a second vector matrix corresponding to the candidate entities; and determining the linked entities corresponding to the text to be processed based on the second vector matrix corresponding to at least one candidate entity. This method can be applied to complex text scenarios, improving the accuracy of entity linking results in complex text scenarios.

Description

Technical Field

[0001] This disclosure relates to the field of computer technology, and more particularly to a method and apparatus for linking entities. Background Technology

[0002] Entity linking is fundamental to natural language processing and knowledge graphs. It primarily associates entity references in text with entities in an entity database, thereby establishing a link between textual objects and knowledge objects, aiding in the understanding of content resources and user needs. Applications of entity linking include: question answering, information flow recommendation, text semantic understanding, webpage recommendation, and semantic search.

[0003] In practical applications, the text scenarios that require entity linking are diverse and complex, and the entity linking methods provided in related technologies are difficult to obtain accurate entity linking results for different text scenarios. Summary of the Invention

[0004] This disclosure provides a method and apparatus for linking entities.

[0005] In a first aspect, embodiments of this disclosure provide an entity linking method, the method comprising: using a vectorization method matching the target language of the text to be processed, vectorizing the entity semantic descriptions of the text to be processed and candidate entities respectively, to obtain a text vector matrix and a candidate entity vector matrix; fusing the text vector matrix and the candidate entity vector matrix to obtain a first vector matrix; determining association information associated with the candidate entities from a pre-established knowledge graph, and determining association feature vectors based on the association relationship between the association information and the text to be processed; the knowledge graph being constructed based on a multi-source heterogeneous entity library; fusing the first vector matrix and the association feature vectors to obtain a second vector matrix corresponding to the candidate entities; and determining linking entities corresponding to the text to be processed based on the second vector matrix corresponding to at least one candidate entity.

[0006] Secondly, embodiments of this disclosure provide an entity linking device, comprising: a first vectorization unit, configured to vectorize the entity semantic descriptions of the text to be processed and candidate entities respectively using a vectorization method matching the target language of the text to be processed, to obtain a text vector matrix and a candidate entity vector matrix; a first fusion unit, configured to fuse the text vector matrix and the candidate entity vector matrix to obtain a first vector matrix; a second vectorization unit, configured to determine association information associated with the candidate entities from a pre-established knowledge graph, and determine association feature vectors based on the association relationship between the association information and the text to be processed; the knowledge graph is constructed based on a multi-source heterogeneous entity library; a second fusion unit, configured to fuse the first vector matrix and the association feature vectors to obtain a second vector matrix corresponding to the candidate entities; and a determination unit, configured to determine the linking entity corresponding to the text to be processed based on the second vector matrix corresponding to at least one candidate entity.

[0007] Thirdly, embodiments of this disclosure provide an electronic device, including: a processor and a memory; the memory stores computer execution instructions; the processor executes the computer execution instructions stored in the memory, causing the at least one processor to perform the first aspect and various possible entity linking methods described above.

[0008] Fourthly, embodiments of this disclosure provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the first aspect and various possible entity linking methods described above.

[0009] Fifthly, embodiments of this disclosure provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect above and various possible entity linking methods of the first aspect.

[0010] The entity linking method and device provided in this embodiment vectorize the entity semantic descriptions of the text to be processed and the candidate entities respectively using a vectorization method that matches the target language of the text to be processed, obtaining a text vector matrix and a candidate entity vector matrix; the text vector matrix and the candidate entity vector matrix are fused to obtain a first vector matrix; association information associated with the candidate entities is determined from a pre-established knowledge graph, and association feature vectors are determined based on the association relationship between the association information and the text to be processed; the knowledge graph is constructed based on a multi-source heterogeneous entity library; the first vector matrix and the association feature vectors are fused to obtain a second vector matrix corresponding to the candidate entities; based on the second vector matrix corresponding to at least one candidate entity, the linking entity corresponding to the text to be processed is determined, realizing an entity linking method applicable to multi-text scenarios. In addition, when applied to complex text scenarios, this entity linking method can improve the accuracy of determining linking entities for complex texts. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this disclosure or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this disclosure. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic diagram illustrating an application scenario of entity linking methods in related technologies;

[0013] Figure 2 Flowchart of the entity linking method provided in the embodiments of this disclosure Figure 1 ;

[0014] Figure 3A A schematic diagram of a structured semantic association network;

[0015] Figure 3B This is a schematic diagram of an entity-word association knowledge graph;

[0016] Figure 4 Flowchart of the entity linking method provided in the embodiments of this disclosure Figure 2 ;

[0017] Figure 5 A schematic diagram of the model structure for implementing the entity linking method provided in this disclosure;

[0018] Figure 6 A structural block diagram of a physical linking device provided in an embodiment of this disclosure;

[0019] Figure 7This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of this disclosure. Detailed Implementation

[0020] To make the objectives, technical solutions, and advantages of the embodiments of this disclosure clearer, the technical solutions of the embodiments of this disclosure will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this disclosure, and not all embodiments. Based on the embodiments of this disclosure, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this disclosure.

[0021] Entity links are involved in scenarios such as knowledge-based question answering, information flow recommendation, text semantic understanding, webpage recommendation, and semantic search. Taking knowledge-based question answering as an example, the server receives a question sent by a user through a terminal device, searches for matching link entities in a specified knowledge base using the question text, and then uses the knowledge corresponding to the aforementioned link entities in the specified knowledge base as the answer matching the question, sending the answer to the user's terminal device.

[0022] In related technologies, the aforementioned knowledge base is a knowledge base built for a specific application scenario. This application scenario could be, for example, an abbreviation scenario or a short text scenario.

[0023] Figure 1 The diagram illustrates a question-and-answer application scenario.

[0024] like Figure 1 As shown, the user inputs the question "TPU" through terminal device 101. The server then determines the link entity that matches the abbreviation in the specified abbreviation knowledge base. Based on the link entity, the server determines the knowledge in the abbreviation knowledge base that matches "TPU" and feeds this knowledge back to terminal device 101 as the answer.

[0025] Similarly, related technologies also include entity linking methods for short Chinese texts and entity linking methods for short English texts.

[0026] In real-world applications, text scenarios requiring entity linking are extremely complex. These include texts composed of multiple languages, texts containing abbreviations and other words, texts containing internet slang and technical terms, short texts, or texts that simultaneously include multiple languages, abbreviations, internet slang, and technical terms.

[0027] Current entity linking methods are based on single text scenarios. For example, for abbreviations, entity linking methods are created based on abbreviations, and for various languages, separate entity linking methods are created for each language. Currently, there is no effective entity linking method designed for complex text scenarios. For complex text scenarios, the accuracy of the linked entities determined using the above entity linking methods is not high. Complex text scenarios can include, for example, texts containing multiple languages, mixed entity words / abbreviations, and a mix of internet slang and technical terms.

[0028] To address the aforementioned issues, this disclosure provides an entity linking method and apparatus. It uses a vectorization method that matches the target language corresponding to the text to be processed to vectorize the text and the entity semantic descriptions of candidate entities. The vectorization results are then fused to obtain a first vector matrix. A multi-source heterogeneous knowledge graph is pre-constructed, and the knowledge graph corresponding to the candidate entities is matched with the text to be processed to determine the associated feature vectors. Finally, based on the first vector matrix and the associated feature vectors, it is determined whether a candidate entity is a linked entity. This enables effective entity linking in complex text scenarios.

[0029] Please refer to Figure 2 , Figure 2 This is a flowchart illustrating the entity linking method provided in the embodiments of this disclosure. Figure 1 This method can be applied to terminal devices or server-side applications.

[0030] S201: Using a vectorization method that matches the target language of the text to be processed, the entity sense descriptions of the text to be processed and the candidate entities are vectorized respectively to obtain the text vector matrix and the candidate entity vector matrix.

[0031] Here, "text" can refer to a body of information composed of certain language and script.

[0032] The text to be processed here can be long or short. It can be a passage extracted from an article or a message extracted from an instant messaging message. The text in the text to be processed can include one language or multiple languages.

[0033] In some applications, the text to be processed may be a segment extracted from a longer article. The target language of this text can be determined based on the language used in the context of the text. For example, if the text preceding the text is in Chinese, the target language of the text to be processed can be determined as Chinese.

[0034] Text vectorization refers to representing text as vectors or vector matrices that can express the semantics of the text. Text vectorization methods can include bag-of-words (BOB) methods and word embedding methods. BOB methods include one-hot encoding, term frequency-inverse document frequency (TF-IDF), and n-grams. Word embedding methods include word2vec (word to vector), global vectors for word representation (GloVe), and Bidirectional Encoder Representation from Transformers (BERT).

[0035] Text vectorization methods for different languages ​​can involve different processing procedures. Taking Chinese and English texts as examples, English text has delimiters between different words, so we can directly use this feature to obtain the different words in the English text and then vectorize them. However, Chinese text does not have delimiters between characters / words. Therefore, for Chinese text, it is necessary to first segment the text into multiple words, then remove numbers and stop words from the resulting segments, obtaining a string of words separated by spaces, and then vectorize it.

[0036] Therefore, when vectorizing text, the language of the text can be determined first. Then, a vectorization method matching that language is used to vectorize the text to be processed and the entity sense descriptions of candidate entities.

[0037] The text vector matrix and candidate entity vector matrix can be either K×1 or K×L. A K×1 matrix can contain one K-dimensional vector. A K×L matrix can contain L K-dimensional vectors. K is an integer greater than or equal to 1. L is an integer greater than 1.

[0038] Vectorization methods that match the target language of the text to be processed are used to vectorize the entity sense descriptions of the text to be processed and the candidate entities. Different vectorization methods can be adopted for different languages, thereby improving the accuracy of the vectorization results of the entity sense descriptions of the text to be processed and the candidate entities.

[0039] In some application scenarios, step S201 above may include:

[0040] First, the target language of the text to be processed is determined based on multiple characters in the text to be processed.

[0041] Secondly, using a vectorization method that matches the target language, the semantic descriptions of the text to be processed and the candidate entities are vectorized respectively.

[0042] In these application scenarios, the target language of the text to be processed can be determined based on the text itself.

[0043] Specifically, the proportion of each language in the text to be processed can be calculated, and the target language of the text to be processed can be determined based on the proportion.

[0044] For example, if Chinese characters account for 51% of the text to be processed, then Chinese can be identified as the target language of the text to be processed.

[0045] Vectorization methods matching multiple languages ​​can be pre-stored. After determining the target language of the text to be processed, a vectorization method matching the target language can be selected from the previously saved vectorization methods for each language.

[0046] In some application scenarios, if the entity sense description of a candidate entity includes non-target language text, the non-target language text in the entity sense description can be translated into the target language text. Then, a vectorization method matching the target language is used to vectorize the text to be processed and the entity sense description of the candidate entity. In these application scenarios, converting the entity sense description into the target language first can improve the accuracy of the vectorization results of the entity sense description text.

[0047] S202: Merge the text vector matrix with the candidate entity vector matrix to obtain the first vector matrix.

[0048] Various fusion methods can be used to fuse the text vector matrix and the candidate entity vector matrix to obtain a first vector matrix. Illustratively, the above fusion includes concatenating the text vector matrix and the candidate entity vector matrix, or adding the text vector matrix and the candidate entity vector matrix according to their positions, etc.

[0049] S203: Determine the association information with candidate entities from the pre-established knowledge graph, and determine the association feature vector based on the association relationship between the association information and the text to be processed; the knowledge graph is constructed based on a multi-source heterogeneous entity library.

[0050] Knowledge graphs are primarily used to describe various entities and concepts that exist in the real world, as well as the relationships between them.

[0051] The multi-source heterogeneous entity libraries here include various encyclopedic entity libraries, abbreviation libraries, and domain-specific entity libraries corresponding to various fields. In this way, the constructed knowledge graph can establish relationships between entities in different entity libraries, thereby expanding the application scope of this entity linking method. Here, "entity library" refers to a knowledge base containing multiple entities.

[0052] In some application scenarios, the aforementioned knowledge graph includes a structured semantic association network between multiple entities mined from multiple heterogeneous entity databases. This structured semantic association network between multiple entities constitutes a structured knowledge graph. The constituent units of this structured semantic association network are "entity-relationship-entity" triples, as well as entities and their related attribute-value pairs. Entities are interconnected through relations, forming a network-like knowledge structure.

[0053] For illustrative purposes, please refer to the following: Figure 3A This illustrates a schematic structured semantic association network. Actor A appeared in movie B. Actor A is an entity word, and movie B is also an entity word. The triples between actor A and movie B are as follows: Figure 3A As shown, the relationship between entity "Actor A" and entity "Movie B" is "acting". Each entity can also correspond to its own entity semantic description. For example, the entity semantic description for Movie B may include: movie title, release date, movie type, number of episodes, etc. Actor A's wife is C1. The relationship between entity "Actor A" and entity "C1" is husband and wife. Actor A's son is C2. The relationship between entity "Actor A" and entity "C2" is father and son.

[0054] In these application scenarios, candidate entities can be searched within the aforementioned knowledge graph. Then, the associated information corresponding to the candidate entities is determined. This associated information can be the semantic network within the knowledge graph, including the candidate entities. This semantic network is then matched against the text to be processed. The associated feature vector is determined based on the matching results. For example, if there are 10 entities in the semantic network, and 4 of them successfully match the text to be processed, the associated feature vector can be considered to be 0.4. In these application scenarios, because the associated information corresponding to candidate entities can be determined from multi-source heterogeneous knowledge graphs, and then the associated feature vector used to analyze linked entities is determined based on the relationship between the associated information and the text to be processed, the entity linking method can have a high disambiguation capability for entities from different entity databases. This can solve the mixed disambiguation problem in scenarios where entity words, abbreviations, internet slang, and professional terminology are mixed.

[0055] In other application scenarios, the aforementioned knowledge graph can include an entity-word association knowledge graph obtained from entity / word mining of multiple heterogeneous entity databases. This entity-word association knowledge graph is built based on the following steps: extracting concept words from entity semantic descriptions in multiple heterogeneous entity databases, identifying multiple concept word pairs, determining the association relationships between different concept words based on these pairs, and then establishing an entity-word association knowledge graph between entity words and other concept words within the concept words. In these application scenarios, it can provide the association relationships between contextual words and candidate entities mentioned in short texts, helping to solve the entity linking problem in short texts.

[0056] Figure 3B This is an exemplary illustration of an entity-word association knowledge graph. The entity-word association knowledge graph includes entity D, and keywords D11, D12, D21, D22, D31, etc., which are associated with entity D based on concept word pairs.

[0057] In these application scenarios, the association information corresponding to candidate entities can be retrieved from the entity-word association knowledge graph. This association information can be the related words corresponding to the candidate entities in the entity-word association knowledge graph. Then, each related word in the entity-word association knowledge graph corresponding to the candidate entity is matched with the text to be processed. The association feature vector is determined based on the matching results. For example, if the entity-word association knowledge graph has 5 keywords, and one of these 5 keywords successfully matches a keyword in the text to be processed, then the association feature vector can be 0.2.

[0058] Step S204: Fuse the first vector matrix with the associated feature vector to obtain the second vector matrix corresponding to the candidate entity.

[0059] Specifically, each element in the first vector matrix can be added to the corresponding element in the associated feature vector to obtain the second vector matrix corresponding to the candidate entity.

[0060] Alternatively, the associated feature vector can be concatenated with each vector in the first vector matrix to obtain the second vector matrix.

[0061] Step S205: Based on the second vector matrix corresponding to at least one candidate entity, determine the link entity corresponding to the text to be processed.

[0062] For each candidate entity, the second vector matrix of the candidate entity can be analyzed and processed according to preset rules to obtain the probability that the candidate entity is a link entity of the text to be processed.

[0063] Schematic, the aforementioned preset rule may include: assigning a weight to each element in the second vector matrix, then calculating the sum of the products of each element and its corresponding weight, and using the sum as the probability that the candidate entity is a link entity of the text to be processed.

[0064] After obtaining the probabilities corresponding to at least one candidate entity, the candidate entity with the highest probability value can be used as the link entity of the text to be processed. Alternatively, all candidate entities with probability values ​​greater than a preset probability threshold can be used as link entities of the text to be processed.

[0065] In this embodiment, a vectorization method matching the target language of the text to be processed is used to vectorize the entity sense descriptions of the text to be processed and the candidate entities, respectively, to obtain a text vector matrix and a candidate entity vector matrix. The text vector matrix and the candidate entity vector matrix are fused to obtain a first vector matrix. Association information associated with the candidate entities is determined from a pre-established knowledge graph. Based on the association relationship between the association information and the text to be processed, association feature vectors are determined. The knowledge graph is constructed based on a multi-source heterogeneous entity library. The first vector matrix and the association feature vectors are fused to obtain a second vector matrix corresponding to the candidate entities. Based on the second vector matrix corresponding to at least one candidate entity, the linked entities corresponding to the text to be processed are determined. Therefore, this entity linking method can be applied to different text scenarios. Furthermore, for texts to be processed, including complex scenarios, this entity linking method can obtain relatively accurate linked entities.

[0066] Please continue to refer to this. Figure 4 , Figure 4 Flowchart of the entity linking method provided in the embodiments of this disclosure Figure 2 .like Figure 4 As shown, the entity linking method provided in this embodiment includes the following steps:

[0067] S401: Input the text to be processed and the entity sense descriptions of the candidate entities into the pre-trained vectorization model. The vectorization model uses a vectorization method that matches the target language of the text to be processed to vectorize the entity sense descriptions of the text to be processed and the candidate entities respectively, to obtain the text vector matrix and the candidate entity vector matrix.

[0068] The vectorization model described above is used to vectorize the input text. This vectorization model can include the bag-of-words model, neural network models, etc.

[0069] The vectorization model here can vectorize the text to be processed and vectorize the entity sense descriptions of candidate entities.

[0070] Schematic illustration: The vectorization model described above may include a word embedding layer. The vectorization model can use the word embedding layer to vectorize the text to be processed, obtaining a text vector matrix of the text. Furthermore, the vectorization model can also use the word embedding layer to vectorize the entity semantic descriptions of candidate entities, obtaining a candidate entity vector matrix.

[0071] In some application scenarios, if the text to be processed is a segment extracted from a longer text, the vectorization model can determine the target language of the text to be processed based on the target language of other segments of the longer text determined in the past.

[0072] In other scenarios, the vectorization model described above can first determine the proportion of each language in the text to be processed, and then determine the target language of the text to be processed based on the proportion of each language.

[0073] In one implementation, the aforementioned vectorization model includes a preprocessing module and a vectorization module; wherein,

[0074] The aforementioned preprocessing module is used to determine the target language corresponding to the input text to be processed;

[0075] The vectorization module is used to call a vectorization network that matches the target language to perform vectorization processing on the text to be processed and the entity sense descriptions of candidate entities.

[0076] In these implementations, the preprocessing module can be connected to the vectorization module. The preprocessing module can pass language information, including the target language, to the vectorization module.

[0077] The preprocessing module described above can determine the language of the text to be processed in various ways. For example, it can determine the target language of the text to be processed by calculating the proportion of each language in the input text.

[0078] The aforementioned vectorization module can call the vectorization network corresponding to the target language to vectorize the entity sense descriptions of the text to be processed and the candidate entities, based on the target language.

[0079] In practice, for each language, the corresponding vectorization network can be trained in advance using training samples of that language, so that the vectorization network of that language can vectorize the input text using that language.

[0080] In some embodiments, the preprocessing module described above is further configured to translate the candidate entity semantic descriptions into the corresponding symbols in the target language.

[0081] In these embodiments, the entity sense description may include text expressed in a language different from the target language. The preprocessing module may translate the text expressed in other languages ​​in the entity sense description into the text corresponding to the target language, and then vectorize the entity sense description.

[0082] As an illustrative example, if the target language is Chinese and the entity semantic descriptions of the candidate entities are in English, the English entity semantic descriptions can be translated into Chinese entity semantic descriptions. Then, the text to be processed and the translated entity semantic descriptions of the candidate entities are input into the vectorization module. The vectorization module then vectorizes the text to be processed and the entity semantic descriptions.

[0083] As another illustrative example, if the target language is Chinese, the entity semantic descriptions of the candidate entities are derived from semantic descriptions in multiple languages. The Chinese entity semantic description can be extracted from these multiple language descriptions. Then, the text to be processed and the Chinese entity semantic description are input into the vectorization module.

[0084] In some embodiments, the vectorization model described above is a multilingual vector model. This multilingual vectorization model can vectorize the text to be processed and entity semantic descriptions.

[0085] In these embodiments, the multilingual vector model described above can be a vectorized model pre-trained using multilingual training text. This vectorized model can vectorize input text in different languages.

[0086] In these embodiments, the structure of the vectorization model can be simplified by vectorizing the text to be processed and the entity sense descriptions through a general multilingual vectorization model.

[0087] S402: The text vector matrix and the candidate entity vector matrix are fused based on the cross-attention mechanism to obtain the first vector matrix.

[0088] In this embodiment, the dimensions of the text vector matrix and the candidate entity vector matrix can be the same.

[0089] The aforementioned cross-attention mechanism can include various cross-attention algorithms in related technologies.

[0090] Through the aforementioned cross-attention mechanism, the text vector matrix and the candidate entity vector matrix can be fused. Furthermore, since the first vector matrix is ​​obtained based on the cross-attention mechanism, it can reflect the dependency relationship between the text vector matrix and the candidate entity vector matrix.

[0091] Schematic, the text vector matrix can be represented as [M1, M2, M3]; the entity vector matrix can be represented as [N1, N2, N3]. Where M1, M2, M3, N1, N2, and N3 are all multi-dimensional vectors. The process of determining the first vector matrix based on the cross-attention mechanism can be characterized by the following formula:

[0092] in

[0093] i and j are integers greater than or equal to 1 and less than or equal to 3. N j Let a be the j-th vector among the candidate entity vectors. ij The cosine of the i-th vector in the text vector matrix and the j-th vector in the candidate entity vector matrix.

[0094] ′

[0095] Similarity. M i Let be the i-th vector in the first vector matrix.

[0096] Each vector in the first vector matrix represents the importance of each vector in the candidate entity vector matrix to the corresponding vector in the text vector matrix.

[0097] S403: Determine the association information with candidate entities from the pre-established knowledge graph, and determine the association feature vector based on the association relationship between the association information and the text to be processed; the knowledge graph is based on a multi-source heterogeneous entity library.

[0098] For the specific implementation of step S403 above, please refer to Figure 2 The specific implementation of step S203 in the illustrated embodiment will not be elaborated here.

[0099] S404: Fuse the first vector matrix with the associated feature vector to obtain the second vector matrix corresponding to the candidate entity.

[0100] Specifically, each element in the first vector matrix can be added to the corresponding element in the associated feature vector to obtain the second vector matrix corresponding to the candidate entity.

[0101] Alternatively, the associated feature vector can be concatenated with each vector in the first vector matrix to obtain the second vector matrix.

[0102] Step S405: Input at least one of the above-mentioned second vector matrices into the classifier, and have the classifier output the confidence level of each candidate entity associated with the text to be processed.

[0103] Step S406: Based on the confidence level of each candidate entity, determine the link entities of the text to be processed.

[0104] In some application scenarios, the above classifier can be various binary classifiers.

[0105] In these application scenarios, the classifier outputs a result indicating whether the candidate entity is a link entity that matches the text to be processed. This output includes "1" and "0". A "1" indicates a confidence level of "1" for the candidate entity being a link entity that matches the text to be processed, while a "0" indicates a confidence level of "0" for the candidate entity not being a link entity that matches the text to be processed.

[0106] In some application scenarios, the classifier described above can be a multi-classifier, and its input can be a second vector matrix corresponding to at least one candidate entity. The classifier includes at least one output, each output corresponding to a candidate entity. Each output of the classifier can be the probability (confidence score) that the candidate entity corresponding to that output is a link entity of the text to be processed. In these application scenarios, the values ​​of the classifier's outputs can be used to determine which candidate entity(s) are the link entities matching the text to be processed. For example, the candidate entity with the largest output value can be used as the link entity corresponding to the text to be processed. Alternatively, at least one candidate entity with an output value greater than a preset probability threshold can be used as the link entity corresponding to the text to be processed.

[0107] In this embodiment, the text to be processed and entity sense descriptions are input into a pre-trained vectorization model. The vectorization model uses a vectorization network that matches the target language of the text to be processed to vectorize the text and entity sense descriptions respectively, obtaining a text vector matrix and a candidate entity vector matrix. The text vector matrix and the candidate entity vector matrix are fused based on a cross-attention mechanism to obtain a first vector matrix. Association information associated with candidate entities is determined from a pre-established knowledge graph. Based on the association relationship between the association information and the text to be processed, association feature vectors are determined. The knowledge graph is constructed based on a multi-source heterogeneous entity database. The second vector matrix of at least one candidate entity is input into a classifier, which outputs the confidence score of each candidate entity's association with the text to be processed. Based on the confidence scores corresponding to each candidate entity, the linking entities of the text to be processed are determined, realizing the determination of linking entities in complex text scenarios by the model. This improves the efficiency of determining the linking entities matching the text to be processed. Furthermore, it expands the application scenarios of the model used to determine entity links.

[0108] exist Figure 2 and Figure 4 In some embodiments of the entity linking method provided in the illustrated examples, the entity linking method further includes constructing a multi-source heterogeneous knowledge graph based on the following steps:

[0109] First, a mapping relationship between entities in different entity databases is established based on a multi-source heterogeneous knowledge association method.

[0110] For example, entity alignment methods from multiple entity libraries can be used to find identical entities across multiple entity libraries. Entity features can be extracted from different entity libraries based on entity-relationship-entity triples and entity attributes. Each entity feature in each entity library is then vectorized. One entity library can be used as a baseline entity library. For each entity in this library, the distance between its vector and vectors in other entity libraries is calculated. The correlation between this entity and other entities is determined based on these distances. Based on these correlations, entities associated with this entity are identified in other entity libraries. For example, entities identical to this entity in other entity libraries.

[0111] Secondly, based on the semantic relationships within each entity database, the association mapping relationship is expanded to obtain a structured multi-source heterogeneous knowledge graph.

[0112] For each entity in each entity library, a semantic association network corresponding to that entity in that entity library can be constructed. This semantic association network can be supplemented into the aforementioned association mapping relationship. For example, semantic associations between entities can be captured from a general entity library. Semantic associations between entity words and abbreviations can be captured from an abbreviation library; and semantic associations between terminology entries and entities, and between specialized terminology entries and abbreviations, can be captured from a domain entity library.

[0113] In this way, the semantic associations corresponding to each entity word in the association mapping relationship can be extended to the above association mapping relationship, thereby constructing a structured knowledge graph.

[0114] In these embodiments, by establishing a multi-source heterogeneous structured knowledge graph, the structured semantic relationships between entities in a multi-source heterogeneous entity library are demonstrated. This provides support for disambiguation in complex text scenarios where entity words, abbreviations, internet slang, and technical terms are mixed. Therefore, it helps apply the entity linking method using this knowledge graph to different text scenarios. Furthermore, this multi-source heterogeneous structured knowledge graph helps improve the accuracy of identifying linked entities in texts including complex scenarios.

[0115] exist Figure 2 and Figure 4 In some embodiments of the entity linking method provided in the illustrated examples, the entity linking method further includes constructing an entity-word association knowledge graph based on the following steps:

[0116] First, concept words are extracted from the entity semantic descriptions corresponding to multiple entities in different entity databases to obtain multiple concept words, including entity words, nouns and technical terms.

[0117] Secondly, multiple pairs of concept words are identified from a variety of concept words.

[0118] Next, determine the number of times each concept pair co-occurs in the description of each entity's meaning.

[0119] Next, based on the number of co-occurrences, multiple target concept word pairs are determined for constructing an entity-word association knowledge graph.

[0120] Then, for each entity word, based on multiple target concept word pairs, multiple other concept words associated with the entity word are determined, and these multiple other concept words are used as associated words of the entity word.

[0121] Finally, the entity words are associated with each related word to obtain the entity-word association knowledge graph.

[0122] In these embodiments, for each of the multiple entity libraries, concept words are extracted from the entity semantic descriptions of the entities in that entity library, resulting in multiple concept words. These concept words include entity words, nouns, and technical terms. Nouns and technical terms can include abbreviations. In this way, multiple concept words extracted from different entity libraries can be obtained.

[0123] Then, concept word pairs are determined from the multiple concept words. Each concept word pair consists of two concept words. For example, for any given concept word, that concept word is paired with each of the other multiple concept words.

[0124] When determining the frequency of each concept word pair, specifically, for each concept word pair, it can be searched in different entity sense descriptions. If the concept word pair is found in only one entity sense description, then the frequency of the concept word pair is 1. If the concept word pair appears multiple times in the entity sense description, it can also be considered that the frequency of the concept word is 1. If the concept word pair is found in 3 entity sense descriptions, then the frequency of the concept word pair is 3.

[0125] As an illustrative example, the entity "TPU" is described as follows: "A TPU is a dedicated chip customized for machine learning, designed specifically for the deep learning framework TensorFlow." The concept words extracted from this entity description are: "TPU, TensorFlow, machine learning, chip, deep learning." Here, "TPU (abbreviation)," "TensorFlow," and "machine learning" are entity words. "Chip" and "deep learning framework" are technical terms. Any two of these concept words can be paired to form multiple concept word pairs. In this entity description, each concept word pair co-occurs once. Here, "TPU" is an abbreviation for "Tensor Processing Unit."

[0126] A co-occurrence threshold can be preset, and concept word pairs with a co-occurrence frequency greater than the threshold can be identified as target concept word pairs for constructing an entity-word association knowledge graph. In this way, multiple target concept word pairs can be obtained.

[0127] The co-occurrence threshold mentioned above can be set according to the specific application scenario, and there is no restriction here.

[0128] After obtaining multiple target concept word pairs, the relationships between the concept words are mined from the multiple target concept word pairs, and then the related words of the entity words are found.

[0129] Schematic, the concept word pair consisting of entity word a and noun b is the target concept word pair, and the concept word pair consisting of noun b and technical term p is also the target concept word pair. Noun b and technical term c can be identified as related words to entity word a.

[0130] When building an entity-word association knowledge graph by associating entity words with their related terms, one can use the entity word as the center and connect it with each related term to create the entity-word association knowledge graph. Please refer to [reference needed]. Figure 3B The entity-word association knowledge graph shown.

[0131] In these embodiments, a multi-source heterogeneous entity-word association knowledge graph is established, thereby constructing a knowledge base between multi-source heterogeneous entities and associated words. During subsequent entity linking, the association information of candidate entity associations can be searched within this entity-word association knowledge graph. This provides a knowledge base for determining linked entities for the text to be processed. Since the aforementioned entity-word association knowledge graph is constructed from multiple concept words extracted from a multi-source heterogeneous entity library, it can be applied to entity linking in different domains, supporting the application of this entity linking method to different application scenarios. Furthermore, this entity-word association knowledge graph helps determine linked entities for texts with complex structures.

[0132] In other application scenarios, the aforementioned knowledge graph can include an entity-word association knowledge graph obtained from entity / word mining of multiple heterogeneous entity databases. This entity-word association knowledge graph is built based on the following steps: extracting concept words from entity semantic descriptions in multiple heterogeneous entity databases, identifying multiple concept word pairs, determining the association relationships between different concept words based on these pairs, and then establishing an entity-word association knowledge graph between entity words and other concept words within the concept words. In these application scenarios, it can provide the association relationships between contextual words and candidate entities mentioned in short texts, helping to solve the entity linking problem in short texts.

[0133] Please refer to Figure 5 This illustrates a schematic diagram of the model structure for implementing the entity linking method provided in this disclosure. For example... Figure 5 As shown, the model includes a vectorization model 51, a cross-attention layer 52, an associated feature extraction module 53, a fusion layer 54, and a classifier 55.

[0134] In some embodiments, the vectorization model 51 described above may include a preprocessing module 511 and a vectorization module 512. The preprocessing module 511 can determine the target language corresponding to the input text to be processed based on the input text to be processed. Then, the target language information and the text to be processed are sent to the vectorization module 512. The vectorization module 512 can retrieve a vectorization network that matches the language information based on the language information to vectorize the text to be processed and the semantic descriptions of candidate entities.

[0135] In some application scenarios, the semantic descriptions of the candidate entities may include characters / words from non-target languages. In such cases, these non-target language characters / words can be translated into their corresponding characters / words in the target language. Then, a vectorization network matching the target language can be used to vectorize the semantic descriptions of the candidate entities expressed in the target language.

[0136] In some other embodiments, the vectorization model 51 described above includes a multilingual vectorization model 513. This multilingual vectorization model can be pre-trained using training samples in multiple languages. Each language's training samples may include multiple training texts in that language and the corresponding vectors for those training texts.

[0137] The vectorization model can vectorize the input text to be processed and the entity sense descriptions of candidate entities, outputting a text vector matrix and a candidate entity vector matrix. These text vector matrices and candidate entity vector matrices can be passed to a cross-attention layer 52. This cross-attention layer can be a hidden layer. It can perform cross-representation on the two input vector matrices. The cross-attention layer outputs the first vector after cross-representation of the text vector matrix and the candidate entity vector matrix.

[0138] Simultaneously, for candidate entities, the association feature extraction module 53 can search for related association information in a pre-obtained knowledge graph. This knowledge graph includes multi-source heterogeneous structured knowledge graphs and / or multi-source heterogeneous entity-word association knowledge graphs. Based on the association relationship between the association information and the text to be processed, an association feature vector is determined. Both the first vector and the association feature vector are input to the fusion layer 54. The fusion layer 54 can fuse the first vector and the association feature vector to obtain a second feature vector. The second feature vector is input to the classifier. The classifier analyzes and processes the second feature vector to obtain the confidence score of the candidate entity as a linked entity in the text to be processed. The linked entities in the text to be processed can be determined based on the confidence scores of multiple candidate entities.

[0139] Corresponding to the entity linking method in the above embodiment, Figure 6 This is a structural block diagram of a physical linking device provided for embodiments of this disclosure. For ease of explanation, only the parts relevant to embodiments of this disclosure are shown. (Refer to...) Figure 6 The device 60 includes: a first vectorization unit 601, a first fusion unit 602, a second vectorization unit 603, a second fusion unit 604, and a determination unit 605.

[0140] The first vectorization unit 601 is used to vectorize the entity sense descriptions of the text to be processed and the candidate entities respectively using a vectorization method that matches the target language of the text to be processed, so as to obtain a text vector matrix and a candidate entity vector matrix.

[0141] The first fusion unit 602 is used to fuse the text vector matrix and the candidate entity vector matrix to obtain the first vector matrix;

[0142] The second vectorization unit 603 is used to determine the association information associated with the candidate entities from the pre-established knowledge graph, and to determine the association feature vector based on the association relationship between the association information and the text to be processed; the knowledge graph is constructed based on a multi-source heterogeneous entity library;

[0143] The second fusion unit 604 is used to fuse the first vector matrix and the associated feature vector to obtain the second vector matrix corresponding to the candidate entity;

[0144] The determining unit 605 is used to determine the link entity corresponding to the text to be processed based on the second vector matrix corresponding to at least one candidate entity.

[0145] In some embodiments, the first vectorization unit 601 is further configured to: determine the target language corresponding to the text to be processed based on the characters included in the text to be processed; and vectorize the semantic descriptions of the text to be processed and the candidate entities respectively using a vectorization method that matches the target language.

[0146] In some embodiments, the first vectorization unit 601 is further configured to: input the text to be processed and the entity sense description into a pre-trained vectorization model, and the vectorization model uses a vectorization method that matches the target language of the text to be processed to vectorize the text to be processed and the entity sense description respectively.

[0147] In some embodiments, the language model includes a preprocessing module and a vectorization module; wherein...

[0148] The preprocessing module is used to determine the target language corresponding to the input text to be processed;

[0149] The vectorization module is used to call a vectorization network that matches the target language. The vectorization network then vectorizes the text to be processed and the entity sense descriptions of the candidate entities.

[0150] In some embodiments, the preprocessing module is further configured to translate the entity semantic descriptions of candidate entities into the target language.

[0151] In some embodiments, the first fusion unit 602 is further configured to: fuse the text vector matrix and the candidate entity vector matrix based on a cross-attention mechanism to obtain a first vector matrix.

[0152] In some embodiments, the determining unit 605 is further configured to: input at least one second vector matrix into a classifier, and have the classifier output the confidence level of each candidate entity associated with the text to be processed; and determine the linked entity based on the confidence level corresponding to each candidate entity.

[0153] In some embodiments, the determining unit 605 is further configured to: determine the candidate entity with the highest confidence level as the link entity; or, determine the candidate entity with a confidence level greater than a preset confidence threshold as the link entity.

[0154] In some embodiments, the device further includes a knowledge graph construction unit (not shown in the figure), which is used to construct a structured multi-source heterogeneous knowledge graph based on the following steps:

[0155] Establish association mapping relationships between entity entries in different entity databases based on multi-source heterogeneous knowledge association methods;

[0156] Based on the semantic relationships within each entity database, the association mapping relationships are expanded to obtain a structured multi-source heterogeneous knowledge graph.

[0157] In some embodiments, the knowledge graph construction unit is further used to construct a multi-source heterogeneous entity-word association knowledge graph based on the following steps:

[0158] Concept words are extracted from the entity semantic descriptions corresponding to multiple entities in different entity databases to obtain multiple concept words, including entity words, nouns and technical terms;

[0159] For any given concept word, form a concept word pair with multiple other concept words;

[0160] Determine the number of times each pair of concept terms co-occurs in the descriptions of each entity's meaning;

[0161] Multiple target concept word pairs were determined based on the co-occurrence frequency to construct the entity-word association knowledge graph;

[0162] For each entity word, based on multiple target concept word pairs, determine multiple other concept words associated with the entity word, and use these multiple other concept words as associated words of the entity word;

[0163] By associating entity words with each related word, a multi-source heterogeneous multi-source entity-word association knowledge graph is obtained.

[0164] To implement the above embodiments, this disclosure also provides an electronic device.

[0165] refer to Figure 7 The diagram illustrates a structural schematic of an electronic device 700 suitable for implementing embodiments of the present disclosure. The electronic device 700 can be a terminal device or a server. The terminal device can include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, personal digital assistants (PDAs), portable Android devices (PADs), portable media players (PMPs), and in-vehicle terminals (e.g., in-vehicle navigation terminals), as well as fixed terminals such as digital TVs and desktop computers. Figure 7 The electronic device shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments disclosed herein.

[0166] like Figure 7As shown, the electronic device 700 may include a processing unit (e.g., a central processing unit, a graphics processor, etc.) 701, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 702 or a program loaded from a storage device 708 into a random access memory (RAM) 703. The RAM 703 also stores various programs and data required for the operation of the electronic device 700. The processing unit 701, ROM 702, and RAM 703 are interconnected via a bus 704. An input / output (I / O) interface 705 is also connected to the bus 704.

[0167] Typically, the following devices can be connected to I / O interface 705: input devices 706 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 707 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 708 including, for example, magnetic tapes, hard disks, etc.; and communication devices 709. Communication device 709 allows electronic device 700 to communicate wirelessly or wiredly with other devices to exchange data. Although Figure 7 An electronic device 700 with various devices is shown; however, it should be understood that it is not required to implement or possess all of the devices shown. More or fewer devices may be implemented or possessed alternatively.

[0168] In particular, according to embodiments of this disclosure, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this disclosure include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication device 709, or installed from storage device 708, or installed from ROM 702. When the computer program (computer execution instructions) is executed by processing device 701, the functions defined above in the methods of embodiments of this disclosure are performed.

[0169] It should be noted that the computer-readable medium described in this disclosure can be a computer-readable signal medium or a computer-readable storage medium, or any combination thereof. A computer-readable storage medium can be, for example,—but not limited to—an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, apparatus, or device, or any combination thereof. More specific examples of a computer-readable storage medium may include, but are not limited to: an electrical connection having one or more wires, a portable computer disk, a hard disk, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage device, magnetic storage device, or any suitable combination thereof. In this disclosure, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in connection with an instruction execution system, apparatus, or device. In this disclosure, a computer-readable signal medium can include a data signal propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such propagated data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. A computer-readable signal medium can be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to: wires, optical fibers, RF (radio frequency), etc., or any suitable combination thereof.

[0170] The aforementioned computer-readable medium may be included in the aforementioned electronic device; or it may exist independently and not assembled into the electronic device.

[0171] The aforementioned computer-readable medium carries one or more programs (computer execution instructions) that, when executed by the electronic device, cause the electronic device to perform the method shown in the above embodiments.

[0172] Computer program code for performing the operations of this disclosure can be written in one or more programming languages ​​or a combination thereof, including object-oriented programming languages ​​such as Java, Smalltalk, and C++, and conventional procedural programming languages ​​such as the "C" language or similar programming languages. The program code can be executed entirely on the user's computer, partially on the user's computer, as a standalone software package, partially on the user's computer and partially on a remote computer, or entirely on a remote computer or server. In cases involving remote computers, the remote computer can be connected to the user's computer via any type of network—including a Local Area Network (LAN) or a Wide Area Network (WAN)—or can be connected to an external computer (e.g., via the Internet using an Internet service provider).

[0173] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0174] The units described in the embodiments of this disclosure can be implemented in software or in hardware. The names of the units are not necessarily limiting in certain circumstances; for example, the first fusion unit can also be described as "a unit that fuses a text vector matrix with a candidate entity vector matrix to obtain a first vector matrix".

[0175] The functions described above in this document can be performed, at least in part, by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: Field Programmable Gate Arrays (FPGAs), Application-Specific Integrated Circuits (ASICs), Application Standard Products (ASSPs), System-on-Chip (SoCs), Complex Programmable Logic Devices (CPLDs), and so on.

[0176] In the context of this disclosure, a machine-readable medium can be a tangible medium that may contain or store a program for use by or in conjunction with an instruction execution system, apparatus, or device. A machine-readable medium can be a machine-readable signal medium or a machine-readable storage medium. A machine-readable medium can be, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of machine-readable storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fiber, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.

[0177] The above description is merely a preferred embodiment of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described concept. For example, technical solutions formed by substituting the above features with (but not limited to) technical features disclosed in this disclosure that have similar functions.

[0178] Furthermore, while the operations are described in a specific order, this should not be construed as requiring these operations to be performed in the specific order shown or in a sequential order. In certain environments, multitasking and parallel processing may be advantageous. Similarly, while several specific implementation details are included in the above discussion, these should not be construed as limiting the scope of this disclosure. Certain features described in the context of individual embodiments may also be implemented in combination in a single embodiment. Conversely, various features described in the context of a single embodiment may also be implemented individually or in any suitable sub-combination in multiple embodiments.

[0179] Although the subject matter has been described using language specific to structural features and / or methodological logic, it should be understood that the subject matter defined in the appended claims is not necessarily limited to the specific features or actions described above. Rather, the specific features and actions described above are merely illustrative examples of implementing the claims.

Claims

1. A method for linking entities, comprising: Using a vectorization method that matches the target language of the text to be processed, the entity sense descriptions of the text to be processed and the candidate entities are vectorized respectively to obtain the text vector matrix and the candidate entity vector matrix. The text vector matrix and the candidate entity vector matrix are fused to obtain a first vector matrix; The association information associated with the candidate entity is determined from a pre-established knowledge graph, and the association information is matched with the text to be processed. The association feature vector is determined based on the matching result. The knowledge graph is constructed based on a multi-source heterogeneous entity library. The first vector matrix and the associated feature vector are fused to obtain the second vector matrix corresponding to the candidate entity; Based on the second vector matrix corresponding to at least one candidate entity, determine the link entity corresponding to the text to be processed.

2. The method of claim 1, wherein, The vectorization method, which uses matching with the target language of the text to be processed, vectorizes the text to be processed and the entity sense descriptions of the candidate entities, respectively, including: The target language corresponding to the text to be processed is determined based on the characters included in the text to be processed; Using a vectorization method that matches the target language, the semantic descriptions of the text to be processed and the candidate entities are vectorized respectively.

3. The method of claim 1, wherein, The vectorization method, which uses matching with the target language of the text to be processed, vectorizes the text to be processed and the entity sense descriptions of the candidate entities, respectively, including: The text to be processed and the entity sense description are input into a pre-trained vectorization model, which then uses a vectorization method that matches the target language of the text to be processed to vectorize the text to be processed and the entity sense description respectively.

4. The method of claim 3, wherein, The vectorization model includes a preprocessing module and a vectorization module; wherein... The preprocessing module is used to determine the target language corresponding to the input text to be processed; The vectorization module is used to call a vectorization network that matches the target language, and the vectorization network vectorizes the entity sense descriptions of the text to be processed and the candidate entities.

5. The method of claim 4, wherein, The preprocessing module is also used to translate the entity semantic descriptions of candidate entities into the target language.

6. The method of claim 1, wherein, The process of fusing the text vector matrix and the candidate entity vector matrix to obtain the first vector matrix includes: The text vector matrix and the candidate entity vector matrix are fused based on the cross-attention mechanism to obtain the first vector matrix.

7. The method according to claim 1, characterized in that, The link entity corresponding to the text to be processed is determined based on the second vector matrix corresponding to at least one candidate entity. At least one of the second vector matrices is input into a classifier, which outputs the confidence level of each candidate entity associated with the text to be processed. The linked entity is determined based on the confidence level corresponding to each candidate entity.

8. The method according to claim 7, characterized in that, The step of determining the confidence scores of the linked entity based on the confidence scores corresponding to each candidate entity, according to the output of the classifier, includes: The candidate entity with the highest confidence level is determined as the linked entity; or Candidate entities with a confidence level greater than a preset confidence threshold are identified as the linked entities.

9. The method according to any one of claims 1-8, characterized in that, The method also includes constructing a structured multi-source heterogeneous knowledge graph based on the following steps: Establish association mapping relationships between entity entries in different entity databases based on multi-source heterogeneous knowledge association methods; Based on the semantic associations within each entity database, the association mapping relationship is expanded to obtain a structured multi-source heterogeneous knowledge graph.

10. The method according to any one of claims 1-8, characterized in that, The method also includes constructing a multi-source heterogeneous entity-word association knowledge graph based on the following steps: Concept words are extracted from the entity semantic descriptions corresponding to multiple entities in different entity databases to obtain multiple concept words, which include entity words, nouns and technical terms; For any given concept word, form a concept word pair with multiple other concept words; Determine the number of times each pair of concept terms co-occurs in the descriptions of each entity's meaning; Multiple target concept word pairs were determined based on the co-occurrence frequency to construct the entity-word association knowledge graph; For each entity word, based on multiple target concept word pairs, multiple other concept words associated with the entity word are determined, and these multiple other concept words are used as associated words of the entity word; By associating the entity words with each of the associated words, a multi-source heterogeneous multi-source entity-word association knowledge graph is obtained.

11. A physical linking device, comprising: The first vectorization unit is used to vectorize the text to be processed and the entity sense descriptions of the candidate entities using a vectorization method that matches the target language of the text to be processed, so as to obtain a text vector matrix and a candidate entity vector matrix. The first fusion unit is used to fuse the text vector matrix and the candidate entity vector matrix to obtain a first vector matrix; The second vectorization unit is used to determine the association information associated with the candidate entity from the pre-established knowledge graph, match the association information with the text to be processed, and determine the association feature vector based on the matching result; the knowledge graph is constructed based on a multi-source heterogeneous entity library; The second fusion unit is used to fuse the first vector matrix and the associated feature vector to obtain the second vector matrix corresponding to the candidate entity; The determining unit is used to determine the link entity corresponding to the text to be processed based on the second vector matrix corresponding to at least one candidate entity.

12. An electronic device, characterized in that, include: Processor and memory; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the entity linking method as described in any one of claims 1 to 10.

13. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, implement the entity linking method as described in any one of claims 1 to 10.

14. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the entity linking method as described in any one of claims 1 to 10.

Images