Name disambiguation method based on multi-relation deep retrieval text matching

Abstract

The invention provides a name disambiguation method based on multi-relation deep retrieval text matching, and relates to the technical field of name disambiguation. Cross-modal data related to enterprises and persons are obtained, data alignment and data fusion are conducted on the cross-modal data through an entity alignment algorithm, and a person-enterprise structured data set is formed; a multi-relation deep retrieval model is established based on a pre-training language model to generate a semantic vector, semantic similarity is calculated according to the semantic vector, a character embedding vector is generated according to co-occurrence frequency and time-space correlation characteristics calculated by a human-enterprise structured data set, and a disambiguation recognition model is established based on an adversarial neural network. Performing name disambiguation according to the character embedding vector, establishing a structure updating model through a graph attention network, updating a graph structure in real time, calculating semantic similarity and the confidence coefficient of the co-occurrence frequency according to the Bayesian theorem, and updating the weight of the semantic similarity and the weight of the co-occurrence frequency according to the confidence coefficient.

Description

Technical Field

[0001] The present invention relates to the technical field of name disambiguation, and in particular to a name disambiguation method based on multi-relation deep retrieval text matching. Background Art

[0002] In the past few decades, name disambiguation aims to eliminate the ambiguity of names in different environments. The main task is to identify the real person corresponding to the name based on the name and other information. Existing enterprise and industrial and commercial public data often only involve names for personnel, but no identity information of the corresponding personnel. This makes it impossible to accurately identify the relationship between individuals and enterprises, thus hindering the construction of the person-enterprise relationship database, and further poses a challenge to commercial businesses such as supply chain management and risk access screening. Due to the huge number of companies and personnel involved, existing name disambiguation methods are difficult to complete the task of personnel identification efficiently and accurately. Traditional name disambiguation methods are mostly based on the company's own information and company-related information, and rarely consider personnel information, so it is impossible to identify the situation where there are actually two different people with the same name in the same company. Further, there will be problems of misidentification.

[0003] In the prior art, the publication number CN 114611516 A discloses a method, device, storage medium and electronic device for name disambiguation, which generates a first relationship graph with enterprises and natural persons as nodes based on the acquired enterprise relationship data; splits the first relationship graph to obtain several second relationship graphs; trains a preset graph network model based on the several second relationship graphs to obtain a vector representation model; determines whether the natural person to be identified needs to be merged with the node with the same name in the first relationship graph based on the associated nodes and the vector representation model or based on the acquired attribute information of the natural person to be identified and the vector representation model. However, this solution mainly focuses on the relationship between enterprises and natural persons. However, in name disambiguation, enterprise relationship data cannot cover all contextual information and lacks sufficient information for accurate disambiguation. At the same time, the preset graph network model is trained based on the split second relationship graph. This method is a static model that cannot adapt to dynamic changes and real-time updates of data.

[0004] The above information disclosed in this Background section is only for enhancement of understanding of the background of the present disclosure and therefore it may contain information that does not constitute the prior art that is already known to one of ordinary skill in the art. Summary of the invention

[0005] The purpose of the present invention is to provide a method for disambiguating personal names based on multi-relation deep retrieval text matching to solve the problems raised in the above-mentioned background technology.

[0006] To achieve the above object, the present invention provides the following technical solutions: A method for disambiguating names based on multi-relation deep retrieval text matching, the specific steps of which include: Step 1: Obtain cross-modal data related to enterprises and people, form an initial dataset of people and enterprises based on the names of people and enterprises, align the cross-modal data using the entity alignment algorithm, and fuse the aligned data to form a structured dataset of people and enterprises; Step 2: Based on the structured data set of people and enterprises, a multi-relation deep retrieval model is established based on the pre-trained language model to generate semantic vectors, and semantic similarity is calculated based on the semantic vectors; Step 3: Calculate the co-occurrence frequency and spatiotemporal correlation features based on the structured data set of people and enterprises, generate a person embedding vector based on semantic similarity, co-occurrence frequency, and spatiotemporal correlation features, establish a disambiguation recognition model based on the adversarial neural network, and perform name disambiguation based on the person embedding vector; Step 4: Using the real-time updated structured data set of people and enterprises and the person embedding vectors, a structure update model is established through the graph attention network. The graph structure is updated in real time by capturing the high-order relationship changes between each node and the node relationship with the person name as the node; Step 5: By analyzing the semantic similarity and co-occurrence frequency, the confidence of the semantic similarity and co-occurrence frequency is calculated according to the Bayesian theorem, and the weights of the semantic similarity and co-occurrence frequency are updated according to the confidence.

[0007] Furthermore, the cross-modal data is obtained in the following ways: through an enterprise public disclosure system, through social software, and a news aggregation platform; The cross-modal data includes: text data, structured data, image data and geographic information data; The method for establishing the initial data set of human enterprises is as follows: The initial data set of human enterprises is represented as , where each element Represents a person's name and company-related data, including a person's name and multiple related modal information. The mathematical expression is: in, For the initial data set of human enterprises, The number of data related to names of people and companies in the initial data set of people and companies; Among them, each element The expression is: in, For the Data related to the names of people and companies in the initial data set of personal companies. They are personal and company names, text features, structured data, image data, and geographic information data.

[0008] Furthermore, the calculation method for aligning cross-modal data by the entity alignment algorithm is: in, It is the internal data type of the enterprise name data. , , is the cosine similarity; like When the two data types are similar, it is judged that they are the same internal factors of the name of the person or enterprise; in, is the similarity judgment threshold.

[0009] Furthermore, the calculation formula for data fusion of the aligned data is: in, After fusion Data related to the names of individuals and companies in the initial data set, are the weights of text features, structured data, image data, and geographic information data, respectively. .

[0010] Furthermore, the specific steps of generating the semantic vector are: Use [CLS] and [SEP] tags to define text boundaries in the structured data set, separate fields with commas, and use RoBERTa to extract the semantic vector of the [CLS] tag: in, , They are the names of people and companies in the structured data set. Training models for languages, is the sentence-level representation vector, , They are the semantic vectors of person names and company names respectively.

[0011] Furthermore, the calculation formula of the semantic similarity is: in, Semantic similarity between personal and corporate names.

[0012] Furthermore, the calculation formula of the co-occurrence frequency is: in, is the co-occurrence frequency, is the number of times that names of people and companies appear in the structured dataset of people and companies. is the number of times a person's name appears in the structured data set of people and enterprises. The number of times the company name appears in the structured data set of people and enterprises; The spatiotemporal correlation characteristics are: in, is the time-related feature, , They are the start time and end time of the association between the person name and the company name. , They are the start and end time of the association between the company name and the person name; in, is the spatial correlation characteristic, are the latitude and longitude of the person's name, are the latitude and longitude of the company name, is the radius of the Earth, The spherical distance between a person's name and a company's name; The calculation formula for generating the character embedding vector is: in, is the joint eigenvector, is the character embedding vector, is the projection matrix, is the bias term.

[0013] Furthermore, the disambiguation recognition model is established based on the adversarial neural network, and name disambiguation is performed according to the character embedding vector: A disambiguation recognition model is established based on an adversarial neural network. The character embedding vectors of different historical names are input into the recognition model. Whether different names are the same person is used as the result of the recognition model to train the recognition model. The character embedding vectors of the names to be recognized are used as the input of the trained recognition model to recognize and disambiguate the names: The disambiguation recognition model includes a generator and a judge: The generator maps the input to the output of the name disambiguation result through a neural network. The calculation formula is: in, For the name disambiguation result, Embedding vectors for different names, are the parameters of the generator, is the mapping function of the generator; The goal of the discriminator is to determine whether the name disambiguation result is generated by the generator, and based on this judgment it gives a probability value: in, Indicates the output name disambiguation result. is the weight of the discriminator, is the bias of the discriminator.

[0014] Furthermore, the calculation formula for establishing the structure update model through the graph attention network is: in, is the updated feature vector of the node, is the set of neighbor nodes of the node, is the attention weight between the node’s neighbor nodes, is the node’s own characteristics, is the activation function; The calculation method of attention weight between neighbor nodes of a node is: in, is the attention score between nodes, is the learnable parameter vector, is the transpose operation, are node feature vectors respectively; The calculation formula of the attention score between nodes is: in, Indicates the node neighbor nodes.

[0015] Furthermore, the specific steps of calculating the confidence of semantic similarity and co-occurrence frequency according to Bayesian theorem and updating the weights of semantic similarity and co-occurrence frequency according to the confidence are as follows: The formula for calculating the prior probability is: in, In is the prior probability of semantic similarity, In is the prior probability of co-occurrence frequency, is the number of correct disambiguation of semantic similarity and co-occurrence frequency, is the total number of disambiguations for semantic similarity and co-occurrence frequency; The calculation formula for likelihood probability is: in, To disambiguate the correct probability through semantic similarity and co-occurrence frequency, is the semantic similarity and co-occurrence probability of the person name and company name when disambiguating. To disambiguate the historical semantic similarity and co-occurrence probability of personal and corporate names, The variance of semantic similarity and co-occurrence probability of names of people and companies for disambiguation; The posterior probability is: in, is the confidence level, is a normalization constant; The calculation formula for updating the weight is: in, is the update weight for semantic matching and co-occurrence analysis. hour, for represents the probability that semantic similarity disambiguation is correct, when hour, for Represents the probability that the co-occurrence frequency disambiguation is correct.

[0016] Compared with the prior art, the present invention has the following beneficial effects: By acquiring cross-modal data related to enterprises and people, the cross-modal data is aligned and fused through the entity alignment algorithm to form a structured data set of people and enterprises. A semantic vector is generated by establishing a multi-relation deep retrieval model based on a pre-trained language model. The semantic similarity is calculated based on the semantic vector, and the person embedding vector is generated based on the co-occurrence frequency and spatiotemporal correlation features calculated from the structured data set of people and enterprises. A disambiguation recognition model is established based on an adversarial neural network, and name disambiguation is performed based on the person embedding vector. A structural update model is established through a graph attention network to update the graph structure in real time. The confidence of semantic similarity and co-occurrence frequency is calculated according to the Bayesian theorem, and the weights of semantic similarity and co-occurrence frequency are updated according to the confidence. The present invention can significantly improve the accuracy and robustness of name disambiguation by integrating cross-modal data from multiple channels, multi-relational deep retrieval technology based on pre-trained language models, and the structural update capability of graph neural networks. First, the scheme uses entity alignment algorithms and cross-modal data fusion to form a high-quality structured dataset of people and enterprises, providing a more comprehensive information basis for subsequent disambiguation tasks. Secondly, through deep learning and semantic vector generation, this method can capture the deep semantic relationships between characters, and combine the characteristics of spatiotemporal correlation and co-occurrence frequency to further optimize the matching and disambiguation effects of names. In particular, the disambiguation model based on adversarial neural networks makes the system more accurate in processing complex and ambiguous names; The present invention also dynamically updates structured data through the graph attention network, which can capture the changes in the relationship between enterprises and people in real time, ensuring that the disambiguation model can remain efficient and accurate in a constantly updated information environment, and can better adapt to dynamically changing enterprise and person data, solving the limitations of the static model in Solution 1. In general, this solution not only solves the limitations of traditional methods in cross-modal data fusion and multi-dimensional feature mining through the combination of multiple technical means, but also improves the system's adaptability in a dynamic data environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 It is a schematic diagram of the overall method flow of the present invention. DETAILED DESCRIPTION

[0018] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments.

[0019] It should be noted that, unless otherwise defined, the technical terms or scientific terms used in the present invention should be understood by people with ordinary skills in the field to which the present invention belongs. The words "first", "second" and similar words used in the present invention do not indicate any order, quantity or importance, but are only used to distinguish different components. "Include" or "comprise" and similar words mean that the elements or objects appearing before the word include the elements or objects listed after the word and their equivalents, without excluding other elements or objects. "Connect" or "connected" and similar words are not limited to physical or mechanical connections, but may include electrical connections, whether direct or indirect. "Up", "down", "left", "right" and the like are only used to indicate relative positional relationships. When the absolute position of the described object changes, the relative positional relationship may also change accordingly.

[0020] Example

[0021] See also Figure 1 , the present invention provides a technical solution: A method for disambiguating names based on multi-relation deep retrieval text matching, the specific steps of which include: Step 1: Obtain cross-modal data related to enterprises and people, form an initial data set of people and enterprises based on the names of people and enterprises, align the cross-modal data through the entity alignment algorithm, and fuse the aligned data to form a structured data set of people and enterprises.

[0022] Cross-modal data is obtained through various channels such as corporate public disclosure systems, social software, and news aggregation platforms, covering a variety of data forms, including text, structured data, image data, and geographic information data. Compared with existing technologies, the advantage of this method is that it can comprehensively and systematically collect and combine information from multiple platforms, thereby providing more dimensional information for subsequent disambiguation tasks. The data provided by these channels are not only rich in sources, but also can show different contexts and associations in different backgrounds, making the data set more comprehensive and able to accurately describe the relationship between names and companies. This solution can supplement and verify information between different modalities, thereby improving the integrity of the data and the accuracy of disambiguation. In particular, image data and geographic information data can provide additional contextual information for disambiguation tasks, further reducing the probability of mismatching.

[0023] Aligning cross-modal data through entity alignment algorithms ensures that data from different sources and modalities can be correctly matched and unified, which can reduce information redundancy and ensure information consistency in disambiguation tasks. Aligning through the cosine distance calculation method helps to accurately evaluate the similarity between different modalities, especially in the relationship matching between text data and structured data. Cosine distance, as a common similarity measurement method, can effectively measure semantic similarity and has high computational efficiency. Compared with traditional alignment methods based on simple rules or static matching, the cosine distance calculation method is more efficient and accurate when processing cross-modal data, avoiding information loss or mismatching.

[0024] In this embodiment, the cross-modal data is obtained in the following ways: through an enterprise public disclosure system, through social software, and a news aggregation platform; The cross-modal data includes: text data, structured data, image data and geographic information data; The method for establishing the initial data set of human enterprises is as follows: The initial data set of human enterprises is represented as , where each element Represents a person's name and company-related data, including a person's name and multiple related modal information. The mathematical expression is: in, For the initial data set of human enterprises, The number of data related to names of people and companies in the initial data set of people and companies; Among them, each element The expression is: in, For the Data related to the names of people and companies in the initial data set of personal companies. They are person and company names, text features, structured data, image data, and geographic information data; The calculation method for aligning cross-modal data by the entity alignment algorithm is: in, It is the internal data type of the enterprise name data. , , is the cosine similarity; like When the two data types are similar, it is judged that they are the same internal factors of the name of the person or enterprise; in, is the similarity judgment threshold.

[0025] In this embodiment, the calculation formula for performing data fusion on the aligned data is: in, After fusion Data related to the names of individuals and companies in the initial data set, are the weights of text features, structured data, image data, and geographic information data, respectively. , .

[0026] After data alignment, data of different modalities are combined through the data fusion process to generate a structured data set. The structured data set provides a standardized data format, so that subsequent deep retrieval and name disambiguation algorithms can be calculated on a more efficient basis. Compared with raw unstructured data, structured data has stronger operability and consistency, which can improve the performance of subsequent models and reduce the impact of noise. In the existing technology, many systems rely on a single data source or fail to effectively integrate data of different modalities, so the accuracy and scope of application are relatively limited, while structured data sets can more accurately reflect the multi-dimensional relationship between names and companies. After the fusion and structuring of multimodal data, the disambiguation model can perform deep retrieval and semantic analysis on a more efficient and accurate basis. This process not only improves the learning efficiency of the model, but also makes the disambiguation results more accurate, especially when dealing with complex relationships between names and companies, the model can effectively identify potential matches and erroneous information.

[0027] Step 2: Based on the structured data set of people and enterprises, a multi-relation deep retrieval model is established based on the pre-trained language model to generate semantic vectors, and semantic similarity is calculated based on the semantic vectors.

[0028] By using a pre-trained language model to generate semantic vectors and calculate semantic similarity, the semantic information of the text can be captured efficiently. RoBERTa can provide higher-quality representation vectors and efficiently capture deeper semantic information, especially for complex multiple relationships between names and companies. The model can match and disambiguate more accurately at the semantic level. By using the [CLS] tag as a representative of the sentence and using the [SEP] tag to separate text fields, accurate contextual embeddings can be generated to enhance the relationship and meaning understanding between texts.

[0029] Semantic vectors are a form of representation that converts text into high-dimensional vectors through natural language processing models. These vectors are numerical expressions of the semantics of text and can capture rich information about vocabulary, syntax, and context in the text. Semantic vectors not only represent the literal meaning of a single word, but also reflect the semantic meaning in the context. This type of vector is usually high-dimensional. It learns the deep connection between vocabulary and sentences through pre-trained language models and can better express the diversity and complexity of language.

[0030] In the scenario of name disambiguation, the semantic vector can represent the semantic relationship between each candidate name and the company name, address or other contextual information, thereby helping to identify the similarity between the two.

[0031] In this embodiment, the specific steps of generating the semantic vector are: Use [CLS] and [SEP] tags to define text boundaries in the structured data set, separate fields with commas, and use RoBERTa to extract the semantic vector of the [CLS] tag: in, , They are the names of people and companies in the structured data set. Training models for languages, is the sentence-level representation vector, , They are the semantic vectors of person names and company names respectively.

[0032] In this embodiment, the calculation formula of the semantic similarity is: in, Semantic similarity between personal and corporate names.

[0033] By calculating the similarity between the generated semantic vectors, the semantic similarity between different texts can be compared more accurately. Traditional text similarity calculation methods (such as TF-IDF based on word frequency or traditional vector space models) are difficult to capture the deep semantic relationship of the text, while semantic vectors generated by pre-trained models can identify more detailed similarities. For example, some company names may have similar naming habits, and RoBERTa can identify these potential similarities and make more accurate judgments in the disambiguation process.

[0034] Semantic similarity is a measure of the semantic proximity between two texts. In name disambiguation, semantic similarity is used to calculate the degree of match between candidate names and company names, addresses, and other information. For example, by calculating the semantic similarity between different candidate names and a company name, it is possible to determine whether the person is associated with the company or is an employee of the company.

[0035] Step 3: Calculate the co-occurrence frequency and spatiotemporal correlation features based on the structured data set of people and enterprises, generate a person embedding vector based on semantic similarity, co-occurrence frequency, and spatiotemporal correlation features, establish a disambiguation recognition model based on the adversarial neural network, and perform name disambiguation based on the person embedding vector.

[0036] Co-occurrence frequency refers to the frequency with which two entities, such as a person's name and a company's name, co-appear in the same text. In name disambiguation, if a person's name co-appears with certain companies or places in multiple texts, this co-occurrence relationship can be used as evidence of the potential association between the person's name and the company. Names and companies with a high co-occurrence frequency may have a stronger connection in reality.

[0037] Spatiotemporal correlation features refer to analyzing the correlation between names and other entities based on information in the temporal and spatial dimensions. For example, a person's name may cooperate more closely with a company at a specific time (such as a certain year), or the person's name may only appear in data from a specific location (such as a city or region). Spatiotemporal features can help distinguish people with the same name who appear frequently in different time periods or locations by combining time and geographic information.

[0038] In this embodiment, the calculation formula of the co-occurrence frequency is: in, is the co-occurrence frequency, is the number of times that names of people and companies appear in the structured dataset of people and companies. is the number of times a person's name appears in the structured data set of people and enterprises. The number of times the company name appears in the structured data set of people and enterprises; The spatiotemporal correlation characteristics are: in, is the time-related feature, , They are the start time and end time of the association between the person name and the company name. , They are the start and end time of the association between the company name and the person name; in, is the spatial correlation characteristic, are the latitude and longitude of the person's name, are the latitude and longitude of the company name, is the radius of the Earth, The spherical distance between a person's name and a company's name.

[0039] In this embodiment, the calculation formula for generating the character embedding vector is: in, is the joint eigenvector, is the character embedding vector, is the projection matrix, is the bias term.

[0040] The character embedding vector is a vector representation constructed based on semantic similarity, co-occurrence frequency, and spatiotemporal correlation features. This embedding vector combines the multi-dimensional features of each candidate name to form a multi-dimensional vector representation. It not only contains the basic semantic information of the name (such as through the learning of semantic vectors), but also combines co-occurrence information and spatiotemporal relationships, which can fully reflect the performance of the name in different contexts.

[0041] In this embodiment, the disambiguation recognition model is established based on the adversarial neural network, and the name disambiguation is performed according to the character embedding vector: A disambiguation recognition model is established based on an adversarial neural network. The character embedding vectors of different historical names are input into the recognition model. Whether different names are the same person is used as the result of the recognition model to train the recognition model. The character embedding vectors of the names to be recognized are used as the input of the trained recognition model to recognize and disambiguate the names: The disambiguation recognition model includes a generator and a judge: The generator maps the input to the output of the name disambiguation result through a neural network. The calculation formula is: in, For the name disambiguation result, Embedding vectors for different names, are the parameters of the generator, is the mapping function of the generator; The goal of the discriminator is to determine whether the name disambiguation result is generated by the generator, and based on this judgment it gives a probability value: in, Indicates the output name disambiguation result. is the weight of the discriminator, is the bias of the discriminator.

[0042] Adversarial neural network is a neural network architecture that improves the robustness of the model through adversarial training. The basic idea is to optimize the disambiguation effect by building two networks: a generator network and a discriminator network, and training them against each other. In this scheme, the generator network may generate candidate disambiguation results, while the discriminator network evaluates the correctness of these results, and finally forms an optimized disambiguation model.

[0043] Step 4: By obtaining the real-time updated human-enterprise structured dataset and person embedding vector, a structure update model is established through the graph attention network. The graph structure is updated in real time by capturing the changes in high-order relationships between each node and node relationships using person names as nodes.

[0044] The graph attention network is a deep learning model based on graph structure. It uses the attention mechanism to assign different weights to different adjacent nodes in the graph, thereby dynamically adjusting the flow of information according to the strength of the relationship between nodes during the information propagation process. Specifically for the name disambiguation problem, the graph attention network can model the relationship between nodes (names) and adjacent nodes (such as companies, positions, etc.) in the graph, and weight these relationships through the attention mechanism, thereby effectively capturing high-order relationship changes.

[0045] The graph structure is updated dynamically in this step, which means that as new structured data is introduced (such as the relationship data between companies and people), the graph structure will also be updated. This means that as the data changes, the node relationships in the graph will reflect the new correlations in real time, which can effectively respond to changes and new information in the data, further improving the disambiguation effect.

[0046] The introduction of real-time updated structured data sets of people and companies and character embedding vectors means that we can capture the relationship between new names and companies, as well as the interactions between people, in a timely manner, especially in environments where people or companies change frequently. Traditional static data sets often lead to information lags, affecting the effectiveness of the model. By updating the graph structure in real time, we can ensure that the graph structure and character embedding vectors are always up to date, greatly improving the disambiguation effect. Graph neural networks can integrate global information from different nodes, which makes disambiguation not only dependent on a certain node or a single feature, but can comprehensively consider the relationship between a person's name and other related nodes in the graph (such as company, geographic location, time, etc.). This graph-based modeling method is more comprehensive than the traditional context-based single matching method, and can effectively resolve ambiguous phenomena, especially when faced with people with the same name or polysemous names, graph neural networks can effectively distinguish through the relationship between adjacent nodes.

[0047] In this embodiment, the calculation formula for establishing the structure update model through the graph attention network is: in, is the updated feature vector of the node, is the set of neighbor nodes of the node, is the attention weight between the node’s neighbor nodes, is the node’s own characteristics, is the activation function; The calculation method of attention weight between neighbor nodes of a node is: in, is the attention score between nodes, is the learnable parameter vector, is the transpose operation, are node feature vectors respectively; The calculation formula of the attention score between nodes is: in, Indicates the node neighbor nodes.

[0048] Step 5: By analyzing the semantic similarity and co-occurrence frequency, the confidence of the semantic similarity and co-occurrence frequency is calculated according to the Bayesian theorem, and the weights of the semantic similarity and co-occurrence frequency are updated according to the confidence.

[0049] Bayes’ theorem provides a method to calculate the probability of an event based on prior knowledge combined with observed data. In this application, we can use Bayes’ theorem to calculate the confidence of semantic similarity and co-occurrence frequency, and use these confidences to dynamically update weights to optimize the disambiguation process.

[0050] Combining semantic similarity and co-occurrence frequency, and dynamically updating using Bayesian theorem, this not only evaluates the similarity between names from a semantic level, but also verifies whether these names may refer to the same entity through the statistical information of co-occurrence frequency. Such multi-dimensional evaluation makes the disambiguation process more accurate and reliable.

[0051] This solution enables the model to adapt to changes in data by dynamically updating weights. As new information enters the system, the system can recalculate the confidence of each factor through Bayes' theorem and flexibly adjust the weights according to the current context. This mechanism of dynamically adjusting weights helps to cope with complex and changing text scenarios, especially when faced with homonymous or polysemous names, and can more accurately distinguish different characters.

[0052] In this embodiment, the specific steps of calculating the confidence of semantic similarity and co-occurrence frequency according to Bayesian theorem and updating the weights of semantic similarity and co-occurrence frequency according to the confidence are as follows: The formula for calculating the prior probability is: in, In is the prior probability of semantic similarity, In is the prior probability of co-occurrence frequency, is the number of correct disambiguation of semantic similarity and co-occurrence frequency, is the total number of disambiguations for semantic similarity and co-occurrence frequency; The calculation formula for likelihood probability is: in, To disambiguate the correct probability through semantic similarity and co-occurrence frequency, is the semantic similarity and co-occurrence probability of the person name and company name when disambiguating. To disambiguate the historical semantic similarity and co-occurrence probability of personal and corporate names, The variance of semantic similarity and co-occurrence probability of names of people and companies for disambiguation; The posterior probability is: in, is the confidence level, is a normalization constant; The calculation formula for updating the weight is: in, is the update weight for semantic matching and co-occurrence analysis. hour, for represents the probability that semantic similarity disambiguation is correct, when hour, for Represents the probability that the co-occurrence frequency disambiguation is correct.

[0053] The above formulas are all dimensionless and numerical calculations. The formula is a formula for the most recent real situation obtained by collecting a large amount of data and performing software simulation. The preset parameters in the formula are set by technicians in this field according to actual conditions.

[0054] The above embodiments may be implemented in whole or in part by software, hardware, firmware or any other combination thereof. When implemented by software, the above embodiments may be implemented in whole or in part in the form of a computer program product. Those skilled in the art may appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein may be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed by hardware or software methods depends on the specific application and design constraints of the technical solution.

[0055] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, and may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

[0056] The above description is only a specific implementation manner of the present application, but the protection scope of the present application is not limited thereto. Any technician familiar with the technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application.

Claims

1. A method for disambiguating names based on multi-relation deep retrieval text matching, characterized in that: The specific steps include: Step 1: Obtain cross-modal data related to enterprises and people, form an initial dataset of people and enterprises based on the names of people and enterprises, align the cross-modal data using the entity alignment algorithm, and fuse the aligned data to form a structured dataset of people and enterprises; Step 2: Based on the structured data set of people and enterprises, a multi-relation deep retrieval model is established based on the pre-trained language model to generate semantic vectors, and semantic similarity is calculated based on the semantic vectors; Step 3: Calculate the co-occurrence frequency and spatiotemporal correlation features based on the structured data set of people and enterprises, generate a person embedding vector based on semantic similarity, co-occurrence frequency, and spatiotemporal correlation features, establish a disambiguation recognition model based on the adversarial neural network, and perform name disambiguation based on the person embedding vector; Step 4: Using the real-time updated structured data set of people and enterprises and the person embedding vectors, a structure update model is established through the graph attention network. The graph structure is updated in real time by capturing the high-order relationship changes between each node and the node relationship with the person name as the node; Step 5: By analyzing the semantic similarity and co-occurrence frequency, the confidence of the semantic similarity and co-occurrence frequency is calculated according to the Bayesian theorem, and the weights of the semantic similarity and co-occurrence frequency are updated according to the confidence.

2. According to claim 1, a method for disambiguating names based on multi-relation deep search text matching is characterized by: The cross-modal data acquisition methods are: through the enterprise public disclosure system, through social software, and news aggregation platform; The cross-modal data includes: text data, structured data, image data and geographic information data; The method for establishing the initial data set of human enterprises is as follows: The initial data set of human enterprises is represented as , where each element Represents a person's name and company-related data, including a person's name and multiple related modal information. The mathematical expression is: in, For the initial data set of human enterprises, The number of data related to names of people and companies in the initial data set of people and companies; Among them, each element The expression is: in, For the Data related to the names of people and companies in the initial data set of personal companies. They are personal and company names, text features, structured data, image data, and geographic information data.

3. According to claim 2, a method for disambiguating names based on multi-relation deep search text matching is characterized by: The calculation method for aligning cross-modal data by the entity alignment algorithm is: in, It is the internal data type of the enterprise name data. , , is the cosine similarity; like When the two data types are similar, it is judged that they are the same internal factors of the name of the person or enterprise; in, is the similarity judgment threshold.

4. The method for disambiguating names based on multi-relation deep search text matching according to claim 3 is characterized by: The calculation formula for data fusion of the aligned data is: in, After fusion Data related to the names of individuals and companies in the initial data set, are the weights of text features, structured data, image data, and geographic information data, respectively. .

5. The method for disambiguating names based on multi-relation deep search text matching according to claim 1, characterized in that: The specific steps of generating the semantic vector are: Use [CLS] and [SEP] tags to define text boundaries in the structured data set, separate fields with commas, and use RoBERTa to extract the semantic vector of the [CLS] tag: in, , They are the names of people and companies in the structured data set. Training models for languages, is the sentence-level representation vector, , They are the semantic vectors of person names and company names respectively.

6. The method for disambiguating names based on multi-relation deep search text matching according to claim 5, characterized in that: The calculation formula of the semantic similarity is: in, Semantic similarity between personal and corporate names.

7. The method for disambiguating names based on multi-relation deep search text matching according to claim 1, characterized in that: The calculation formula of the co-occurrence frequency is: in, is the co-occurrence frequency, is the number of times that names of people and companies appear in the structured dataset of people and companies. is the number of times a person's name appears in the structured data set of people and enterprises. The number of times the company name appears in the structured data set of people and enterprises; The spatiotemporal correlation characteristics are: in, is the time-related feature, , They are the start time and end time of the association between the person name and the company name. , They are the start and end time of the association between the company name and the person name; in, is the spatial correlation characteristic, are the latitude and longitude of the person's name, are the latitude and longitude of the company name, is the radius of the Earth, The spherical distance between a person's name and a company's name; The calculation formula for generating the character embedding vector is: in, is the joint eigenvector, is the character embedding vector, is the projection matrix, is the bias term.

8. The method for disambiguating names based on multi-relation deep search text matching according to claim 1, characterized in that: The disambiguation recognition model is established based on the adversarial neural network, and the name disambiguation is performed according to the character embedding vector: A disambiguation recognition model is established based on an adversarial neural network. The character embedding vectors of different historical names are input into the recognition model. Whether different names are the same person is used as the result of the recognition model to train the recognition model. The character embedding vectors of the names to be recognized are used as the input of the trained recognition model to recognize and disambiguate the names: The disambiguation recognition model includes a generator and a judge: The generator maps the input to the output of the name disambiguation result through a neural network. The calculation formula is: in, For the name disambiguation result, Embedding vectors for different names, are the parameters of the generator, is the mapping function of the generator; The goal of the discriminator is to determine whether the name disambiguation result is generated by the generator, and based on this judgment it gives a probability value: in, Indicates the output name disambiguation result. is the weight of the discriminator, is the bias of the discriminator.

9. The method for disambiguating names based on multi-relation deep search text matching according to claim 1, characterized in that: The calculation formula for establishing the structure update model through the graph attention network is: in, is the updated feature vector of the node, is the set of neighbor nodes of the node, is the attention weight between the node’s neighbor nodes, is the node’s own characteristics, is the activation function; The calculation method of attention weight between neighbor nodes of a node is: in, is the attention score between nodes, is the learnable parameter vector, is the transpose operation, are node feature vectors respectively; The calculation formula of the attention score between nodes is: in, Indicates the node neighbor nodes.

10. The method for disambiguating names based on multi-relation deep search text matching according to claim 1, characterized in that: The specific steps of calculating the confidence of semantic similarity and co-occurrence frequency according to Bayesian theorem and updating the weights of semantic similarity and co-occurrence frequency according to the confidence are as follows: The formula for calculating the prior probability is: in, In is the prior probability of semantic similarity, In is the prior probability of co-occurrence frequency, is the number of correct disambiguation of semantic similarity and co-occurrence frequency, is the total number of disambiguations for semantic similarity and co-occurrence frequency; The calculation formula for likelihood probability is: in, To disambiguate the correct probability through semantic similarity and co-occurrence frequency, is the semantic similarity and co-occurrence probability of the person name and company name when disambiguating. To disambiguate the historical semantic similarity and co-occurrence probability of personal and corporate names, The variance of semantic similarity and co-occurrence probability of names of people and companies for disambiguation; The posterior probability is: in, is the confidence level, is a normalization constant; The calculation formula for updating the weight is: in, is the update weight for semantic matching and co-occurrence analysis. hour, for represents the probability that semantic similarity disambiguation is correct, when hour, for Represents the probability that the co-occurrence frequency disambiguation is correct.

Images