Entity name retrieval method and device, electronic equipment and storage medium

Abstract

The invention provides an entity name retrieval method and device, electronic equipment and a storage medium, and the method comprises the steps: obtaining a target query word, determining a pronunciation feature and a spelling feature corresponding to the target query word, and splitting the target query word into a plurality of query sub-words; searching matched pronunciation similar names and spelling similar names in a preset name database according to the pronunciation features and the spelling features; retrieving a matched associated name in a preset name database by taking the query sub-word as a keyword, and retrieving a variant name matched with the target query word in a preset variant name database; taking the similar pronunciation names, the similar spelling names, the associated names and the variant names as candidate names, and determining the matching similarity between each candidate name and the target query word; and sorting the candidate names according to the matching similarity, and outputting the sorted candidate names as a name retrieval result. According to the method, the recall coverage rate and the sorting efficiency can be improved, the robustness of the entity alias ambiguity is enhanced, and the accuracy and consistency of search experience are ensured.

Description

Technical Field

[0001] This disclosure relates to the field of information retrieval technology, and more specifically, to an entity name retrieval method, apparatus, electronic device, and storage medium. Background Technology

[0002] In today's digital age, efficient search and matching technologies for personal and corporate names have significant application value across numerous fields. For example, in the legal and compliance field, accurately identifying and analyzing company names and legal representative names is a crucial step in ensuring compliance; in market research, accurately analyzing company and brand names is essential for in-depth analysis of market trends and competitive landscape; and in enterprise information management, quickly and accurately searching and matching personal and corporate names is central to improving work efficiency.

[0003] However, with the explosive growth of information, the structure of entity names has become increasingly complex, resulting in various variants, abbreviations, and multilingual hybrid forms. In current information retrieval and intelligent search systems, entity recognition and accurate ranking play a central role in user experience. Especially in scenarios involving entities with high homonyms or diverse aliases, such as personal names, organization names, and brand names, improving the completeness of entity recall and the accuracy of ranking has become a significant factor limiting the performance of search systems.

[0004] Traditional search engine recall methods suffer from high maintenance costs, delayed updates, and insufficient coverage, making them particularly vulnerable to rapidly evolving aliases such as trending online terms, nicknames, and pinyin abbreviations. Furthermore, due to the high degree of overlap in pinyin for personal names and the prevalence of identical names, the system struggles to accurately identify the user's true search intent, and the high inference costs make it difficult to meet the real-time requirements of large-scale, high-concurrency search scenarios. Summary of the Invention

[0005] This disclosure provides at least one entity name retrieval method, apparatus, electronic device, and storage medium, which can improve recall coverage and ranking efficiency, enhance robustness to entity alias ambiguity, and ensure the accuracy and consistency of the search experience.

[0006] This disclosure provides an entity name retrieval method, including:

[0007] Obtain the target query term, determine the pronunciation and spelling features corresponding to the target query term, and split the target query term into multiple query sub-words;

[0008] In the preset name database, similar names with similar pronunciations and similar names with similar spellings are retrieved based on the pronunciation features and spelling features, respectively.

[0009] The system retrieves matching related names in the preset name database using the query sub-word as the keyword, and retrieves matching variant names in the preset variant name database.

[0010] The similar-sounding names, similar-spelling names, related names, and variant names are used as candidate names to determine the matching similarity between each candidate name and the target query term;

[0011] The candidate names are sorted according to the matching similarity and then output as the name search results.

[0012] In one alternative implementation, the phonetically similar names are retrieved based on the following steps:

[0013] The target query term is converted into a corresponding query term phoneme sequence, and the entity name stored in the preset name database is converted into a corresponding entity name phoneme sequence;

[0014] Determine the phonetic edit distance between the query word phoneme sequence and the entity name phoneme sequence, or encode both the query word phoneme sequence and the entity name phoneme sequence into corresponding phonetic codes, and calculate the encoding similarity between the phonetic codes;

[0015] The entity names that are greater than the corresponding preset pronunciation similarity threshold in the preset name database are selected as the pronunciation similar names.

[0016] In one optional implementation, the spelling-similar names are retrieved based on the following steps:

[0017] The entity names stored in the preset name database are converted into the corresponding entity pinyin, the entity initials abbreviations corresponding to the entity pinyin are extracted, and an entity index table is constructed based on the entity initials abbreviations.

[0018] The target query term is converted into its corresponding pinyin, and the initial letter abbreviation of the pinyin is extracted.

[0019] Based on the entity index table, select spelling similar names from the entity initials that have a spelling similarity greater than a preset spelling similarity threshold with the query initials.

[0020] In one optional implementation, the associated name is retrieved based on the following steps:

[0021] The entity names stored in the preset name database are divided into corresponding entity sub-words;

[0022] An inverted index is built based on the entity sub-words, and candidate sub-words that match the entity sub-words are retrieved from the preset name database based on the inverted index;

[0023] Based on the position of the candidate word in the corresponding entity name, the part of speech of the candidate word, and the importance of the candidate word, a weight score is generated for each candidate word;

[0024] The entity name to which the candidate sub-words with a weight score greater than a preset weight score threshold belong is selected as the associated name.

[0025] In one alternative implementation, the variant name is retrieved based on the following steps:

[0026] For each entity name stored in the preset name database, the corresponding variant forms of the entity name are collected to construct the preset variant name database. The variant forms include at least abbreviations, acronyms, homophone substitutions, and English translations.

[0027] After standardizing the target query term, the variant similarity between it and the variant form in the preset variant name database is determined. A fuzzy matching algorithm is used to match the variants based on the variant similarity, and the variant forms with a variant similarity greater than a preset variant threshold are selected as the variant names.

[0028] In one optional implementation, determining the matching similarity between each candidate name and the target query term specifically includes:

[0029] Determine the string edit distance and Jaro-Winkler similarity between the candidate name and the target query term;

[0030] The pre-trained text embedding model converts the entity names stored in the preset name database into corresponding entity vectors and the target query terms into corresponding query vectors.

[0031] Calculate the vector similarity between the entity vector and the query vector;

[0032] The string edit distance, the Jaro-Winkler similarity, and the vector similarity are determined as the matching similarity.

[0033] In one optional implementation, the candidate names are sorted according to the matching similarity, specifically including:

[0034] The string edit distance, Jaro-Winkler similarity, and vector similarity are weighted and integrated using a Softmax function carrying a preset temperature coefficient to determine the comprehensive matching score corresponding to each candidate name;

[0035] Based on the comprehensive matching score from largest to smallest, the candidate names are sorted to generate a sequence of name retrieval results.

[0036] In one optional implementation, the method further includes:

[0037] The relationships between corporate entities and individual entities are extracted based on publicly available business registration data;

[0038] Establish an enterprise-person knowledge graph with the enterprise entities and person entities as nodes and the association relationships as connecting edges;

[0039] Extract the request entity that appears in the target query term. When the request entity matches the enterprise node or person node in the enterprise-person knowledge graph, extract the neighbor node corresponding to the matching node.

[0040] Configure corresponding relationship weights between the enterprise entity or person entity corresponding to the matching node and the enterprise entity or person entity corresponding to the neighbor node;

[0041] The enterprise entity or person entity corresponding to the matching node and the enterprise entity or person entity corresponding to the neighbor node are weighted and integrated according to the relationship weight, and then added to the candidate name.

[0042] In one optional implementation, the method further includes:

[0043] In the enterprise-person knowledge graph, identity information attribute tags are added to the person entities, and enterprise status attribute tags are added to the enterprise entities;

[0044] Determine the attribute information corresponding to the request entity, and extract the matching node according to the preset attribute priority.

[0045] This disclosure also provides an entity name retrieval device, including:

[0046] The feature extraction module is used to acquire the target query term, determine the pronunciation features and spelling features corresponding to the target query term, and split the target query term into multiple query sub-words;

[0047] The first recall module is used to retrieve, in a preset name database, matching pronunciation-similar names and spelling-similar names based on the pronunciation features and spelling features, respectively.

[0048] The second recall module is used to retrieve matching associated names in the preset name database using the query sub-word as the keyword, and to retrieve matching variant names in the preset variant name database.

[0049] The similarity determination module is used to determine the matching similarity between each of the candidate names and the target query term by taking the pronunciation-similar names, the spelling-similar names, the associated names and the variant names as candidate names.

[0050] The sorting module is used to sort the candidate names according to the matching similarity and output the results as name retrieval results.

[0051] This disclosure also provides an electronic device, including: a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of the above-described entity name retrieval method or any possible implementation of the above-described entity name retrieval method.

[0052] This disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the above-described entity name retrieval method or any possible implementation of the above-described entity name retrieval method.

[0053] This disclosure also provides a computer program product, including a computer program / instruction, which, when executed by a processor, implements the above-described entity name retrieval method, or the steps in any possible implementation of the above-described entity name retrieval method.

[0054] This disclosure provides an entity name retrieval method, apparatus, electronic device, and storage medium. The method involves acquiring a target query term, determining its corresponding pronunciation and spelling features, and splitting the target query term into multiple query sub-words. It then retrieves matching pronunciation-similar names and spelling-similar names from a preset name database based on the pronunciation and spelling features, respectively. Next, it retrieves matching associated names from the preset name database using the query sub-words as keywords, and retrieves matching variant names from a preset variant name database. Finally, it uses the pronunciation-similar names, spelling-similar names, associated names, and variant names as candidate names, determines the matching similarity between each candidate name and the target query term, and sorts the candidate names according to the matching similarity, outputting the search results. This method can improve recall coverage and ranking efficiency, enhance robustness to entity alias ambiguity, and ensure the accuracy and consistency of the search experience.

[0055] To make the above-mentioned objects, features and advantages of this disclosure more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0056] To more clearly illustrate the technical solutions of the embodiments of this disclosure, the accompanying drawings used in the embodiments will be briefly described below. These drawings are incorporated in and constitute a part of this specification. They illustrate embodiments conforming to this disclosure and, together with the specification, serve to explain the technical solutions of this disclosure. It should be understood that the following drawings only show some embodiments of this disclosure and should not be considered as limiting the scope. Those skilled in the art can obtain other related drawings based on these drawings without creative effort.

[0057] Figure 1 A flowchart of an entity name retrieval method provided by an embodiment of this disclosure is shown;

[0058] Figure 2 A flowchart of another entity name retrieval method provided by an embodiment of this disclosure is shown;

[0059] Figure 3 A schematic diagram of an entity name retrieval device provided in an embodiment of this disclosure is shown;

[0060] Figure 4 A schematic diagram of an electronic device provided in an embodiment of the present disclosure is shown. Detailed Implementation

[0061] 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 a part of the embodiments of this disclosure, and not all of them. The components of the embodiments of this disclosure described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of this disclosure provided in the accompanying drawings is not intended to limit the scope of the claimed disclosure, but merely represents selected embodiments of this disclosure. All other embodiments obtained by those skilled in the art based on the embodiments of this disclosure without inventive effort are within the scope of protection of this disclosure.

[0062] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0063] In this document, the term "and / or" merely describes a relationship, indicating that three relationships can exist. For example, A and / or B can represent three cases: A alone, A and B simultaneously, and B alone. Furthermore, the term "at least one" in this document means any combination of at least two of any one or more elements. For example, including at least one of A, B, and C can mean including any one or more elements selected from the set consisting of A, B, and C.

[0064] Research has revealed that in today's digital age, efficient search and matching technology for personal and corporate names has significant application value across numerous fields. However, with the explosive growth of information, the structure of entity names has become increasingly complex, exhibiting various variants, abbreviations, and multilingual hybrid forms. In current information retrieval and intelligent search systems, entity recognition and accurate ranking play a core role in user experience. Especially in scenarios involving highly homonymous or diverse aliases for personal names, organization names, and brand names, improving the completeness of entity recall and the accuracy of ranking has become a crucial factor restricting search system performance. Traditional recall methods suffer from high maintenance costs, delayed updates, and insufficient coverage, particularly struggling to cope with rapidly evolving aliases such as trending online terms, nicknames, and pinyin abbreviations. Furthermore, due to the high overlap in pinyin for personal names and severe homonym phenomena, the system struggles to accurately identify the user's true query intent, and the high inference costs make it difficult to meet the real-time requirements of large-scale, high-concurrency search scenarios.

[0065] Based on the above research, this disclosure provides an entity name retrieval method, apparatus, electronic device, and storage medium. The method involves acquiring a target query term, determining its corresponding pronunciation and spelling features, and splitting the target query term into multiple query sub-words. It then retrieves matching pronunciation-similar names and spelling-similar names from a preset name database based on the pronunciation and spelling features, respectively. Next, it retrieves matching associated names from the preset name database using the query sub-words as keywords, and retrieves matching variant names from a preset variant name database. Finally, it uses the pronunciation-similar names, spelling-similar names, associated names, and variant names as candidate names, determines the matching similarity between each candidate name and the target query term, and sorts the candidate names according to the matching similarity, outputting the result as the name retrieval result. This method can improve recall coverage and ranking efficiency, enhance robustness to entity alias ambiguity, and ensure the accuracy and consistency of the search experience.

[0066] To facilitate understanding of this embodiment, a detailed description of the entity name retrieval method disclosed in this disclosure is provided first. The entity name retrieval method provided in this disclosure generally executes a computer device with a certain computing capability. This computer device may include, for example, a terminal device, a server, or other processing devices. The terminal device may be a user equipment (UE), mobile device, user terminal, terminal, cellular phone, cordless phone, personal digital assistant (PDA), handheld device, computing device, in-vehicle device, wearable device, etc. In some possible implementations, this entity name retrieval method can be implemented by a processor calling computer-readable instructions stored in memory.

[0067] See Figure 1 The diagram shows a flowchart of an entity name retrieval method provided in an embodiment of this disclosure. The method includes steps S101 to S105, wherein:

[0068] S101. Obtain the target query term, determine the pronunciation features and spelling features corresponding to the target query term, and split the target query term into multiple query sub-words.

[0069] In a specific implementation, the target query term input by the user is obtained. First, the original text input of the target query term is obtained. The target query term is a text input used to search for an entity name. Then, pronunciation features of the target query term are extracted. The Chinese Pinyin conversion module can be used to convert a Chinese character string into a complete Pinyin sequence to solve the problem of different spellings with similar pronunciations, which is applicable to situations such as user spelling mistakes and substitution of homophonic and异形 characters. While obtaining the pronunciation features, spelling features of the target query term are also extracted to handle issues such as fuzzy spelling, swapped word positions, and similar glyphs in the user input. To improve the partial matching ability, the target query term is also split into multiple query sub-terms to achieve fuzzy recall at the sub-term level.

[0070] Here, the target query term is segmented into a word sequence in Chinese, and each word in the word sequence is converted into a Pinyin sequence using a Pinyin conversion tool to obtain pronunciation features. For example, "第一银行" is converted to "diyi yin hang". Further, each Pinyin word can be segmented into phonemes, and optionally, the Pinyin form with tones can be extracted for refined pronunciation matching. In addition, to adapt to pronunciation similarity analysis in different scenarios, a phonetic symbol encoding feature can also be constructed. For example, the Soundex or Double Metaphone algorithm is used to encode the Pinyin features to enhance the fault tolerance ability for speech input or spelling mistakes.

[0071] Among them, while obtaining the pronunciation features, the query term is converted into Pinyin, and the first letter of each Pinyin is taken to obtain the abbreviated spelling feature, including the following representation forms: the original string, the complete Pinyin, the initials abbreviation of the Pinyin, and combinations of different forms. The spelling feature can be used in subsequent fuzzy matching calculations.

[0072] Furthermore, to improve the partial matching ability, the sliding window method and the dictionary segmentation method are used for joint splitting processing to split the target query term into multiple query sub-terms. The sliding window method constructs n-gram segments from the original word according to the character sequence. For example, "第一银行" can be split into "第一", "一银", "银行", "第一银", "一银行", etc. The dictionary segmentation method extracts sub-terms with actual meanings based on a preset entity dictionary or a general word library. For example, "第一银行" is split into "第一" and "银行".

[0073] Finally, the obtained spelling features, pronunciation features, and query sub-terms will be used as the basic inputs in the fuzzy matching stage to perform matching calculations on the Pinyin similarity, spelling similarity, and sub-term coverage of candidate entities, thereby enhancing the ability to recognize entity names under irregular inputs, spelling mistakes, or speech recognition noise.

[0074] S102. Respectively retrieve pronunciation-similar names and spelling-similar names that match according to the pronunciation feature and the spelling feature in a preset name database.

[0075] In specific implementation, after obtaining the pronunciation features and spelling features of the target query term, the system further retrieves matching pronunciation-similar names and spelling-similar names from the preset name database based on the pronunciation features and spelling features, so as to achieve fuzzy recall of potential candidate entity names. This is mainly used to recall names that have the same pronunciation as the query term but different characters.

[0076] Here, the default name database is a database that stores a large amount of entity name information, including entity entries such as enterprise names, bank names, brand names, and organization names. Each entity record contains multi-dimensional features such as its standard spelling, pinyin representation, pinyin initial abbreviation, and phonetic encoding.

[0077] In a preferred embodiment of the present invention, in order to improve the retrieval accuracy of similar-sounding names, the system performs pronunciation similarity judgment based on phoneme-level feature comparison, converts the target query word into a corresponding query word phoneme sequence, and converts the entity names stored in the preset name database into corresponding entity name phoneme sequences; determines the phonetic edit distance between the query word phoneme sequence and the entity name phoneme sequence, or encodes both the query word phoneme sequence and the entity name phoneme sequence into corresponding phonetic codes, and calculates the encoding similarity between the phonetic codes; and filters entity names in the preset name database whose phonetic edit distance or encoding similarity is greater than the corresponding preset pronunciation similarity threshold, and determines them as similar-sounding names.

[0078] First, the target query term is converted into a corresponding query term phoneme sequence. Based on phoneme decomposition algorithms, a phoneme dictionary (CMU Pronouncing Dictionary), or a phonetic transcription conversion algorithm (such as the Grameme-to-Phoneme model), each pinyin syllable in the query term is further parsed into the International Phonetic Alphabet or its corresponding acoustic phoneme. Correspondingly, the same phoneme conversion is performed sequentially on all entity names stored in the name database to obtain the entity name phoneme sequence for each entity name, which serves as a structured pronunciation feature vector.

[0079] Next, calculate the pronunciation similarity between the target query term and each entity name in one of the following two ways: Phonetic edit distance method: Treat the phoneme sequence of the query term and the phoneme sequence of the entity name as two string sequences, and calculate the minimum edit distance (Edit Distance) between them, including the minimum number of transformation steps required for operations such as insertion, deletion, and replacement. The smaller this edit distance, the more similar the pronunciations are. It can be further combined with acoustic distance weighting, for example, setting different weights for the replacement costs between different phonemes. Phonetic encoding similarity method: Encoding the phoneme sequence of the query term and the phoneme sequence of the entity name into fixed-length phonetic encoding vectors respectively. The encoding method can use rule-driven speech encoders such as Soundex, Metaphone, Double Metaphone, Pinyin-soundex, or phoneme embeddings based on pre-trained models (such as Phone2Vec). Subsequently, calculate the encoding similarity between the encoding vectors, and the closeness between the two can be measured by methods such as cosine similarity, reciprocal of Euclidean distance, Jaccard similarity, etc.

[0080] Furthermore, after obtaining the pronunciation similarity score in any of the above methods, further screen out the names with similarity greater than the preset threshold in the preset name database, that is, screen out the entity names that meet any of the following conditions: The phonetic edit distance is less than or equal to the preset distance threshold; The phonetic encoding similarity is greater than or equal to the preset similarity threshold. The screened entity names are determined as the set of pronunciation-similar names for the target query term.

[0081] It should be noted that the preset pronunciation similarity threshold can be set according to actual needs and is not specifically limited here.

[0082] Exemplarily, if the target query term input by the user is "zhaoshangyinhang", and there are entity names such as "China Merchants Bank", "Zhaoshang Bank", "Zhaoshang Bank" in the database, the system may convert these names into similar phoneme sequences. After calculating the phoneme-level edit distance or encoding similarity as described above, if their similarity exceeds the set threshold, these names will be recognized as pronunciation-similar names and participate in the subsequent matching and sorting processes.

[0083] In a preferred embodiment of the present invention, in order to improve the recognition efficiency and matching accuracy of spell-similar names, a fast indexing and screening of spell similarity is also performed based on the initial letter features of pinyin. The entity names stored in the preset name database are converted into corresponding entity pinyins, the initial letter abbreviations corresponding to the entity pinyins are extracted, and an entity index table is constructed according to the initial letter abbreviations of the entities; the target query term is converted into the corresponding query pinyin, and the query initial letter abbreviation corresponding to the query pinyin is extracted; according to the entity index table, spell-similar names with a spell similarity greater than the preset spell similarity threshold are screened out among the initial letter abbreviations of the entities.

[0084] Here, this process mainly targets Chinese personal names and enterprise names. By converting the query term into pinyin, names with similar pinyins can be recalled. Chinese characters have a unique pinyin system, and this feature can be used to retrieve names with different writings but similar pinyins. For example, when querying "Zhang Wei", the system can recall similar names such as "Zhang Wei" through pinyin. In addition, pinyin recall can be further extended to the matching of pinyin abbreviations. For example, "LXH" can recall names such as "Li Xiaohong".

[0085] In a specific implementation, first, spell-structured processing is performed on all entity names in the preset name database. Specifically, each entity name is then, the initial letters of each pinyin are extracted from the pinyin sequence to construct the initial letter abbreviation (referred to as pinyin abbreviation) of the entity name. Then, the system uses the pinyin abbreviation as the index key to construct an entity index table containing all entity names. In the entity index table, each pinyin abbreviation corresponds to a set of entity names, and this index table helps to quickly locate a set of names with similar spell structures. Subsequently, the same processing process is performed on the target query term input by the user: first, it is converted into a query pinyin sequence, and its query initial letter abbreviation is extracted.

[0086] Here, after obtaining the query initial letter abbreviation, the system calculates the spell similarity between all entity pinyin abbreviations and the query pinyin abbreviation in the above entity index table in turn. This similarity can be calculated by any of the following methods: Edit distance method: Calculate the minimum edit distance for two pinyin abbreviation strings; Jaccard similarity: Based on the intersection and union ratio of character sets.

[0087] Among them, when the spell similarity between a certain entity pinyin abbreviation and the query pinyin abbreviation is greater than or equal to the preset spell similarity threshold (for example, the threshold is 0.75), the entity name corresponding to the pinyin abbreviation is determined as the spell-similar name of the target query term.

[0088] It should be noted that the preset spell similarity threshold can be set according to actual needs and is not specifically limited here.

[0089] Exemplarily, when the user inputs "zsyh", the system can screen out abbreviations such as "zsyh", "zsyx", "zsyj", etc. from the index table. Since they have a high spelling similarity with the query abbreviation "zsyh", the corresponding entity names such as "China Merchants Bank" and "China Merchants Bank Credit Card" are recognized as spelling-similar names and participate in the subsequent candidate fusion and ranking processes.

[0090] S103. Retrieve the associated names that match the query sub-word as keywords in the preset name database, and retrieve the variant names that match the target query word in the preset variant name database.

[0091] In specific implementation, to further expand the recall range of candidate entity names and improve the retrieval robustness and recognition ability, a query sub-word matching mechanism and a name variant library matching mechanism are also introduced. The target query word has been split into multiple query sub-words at the initial stage. Using these sub-words as keywords, perform keyword-level matching retrieval in the preset name database to obtain candidate associated names that may be semantically or lexically associated.

[0092] Here, traverse the set of query sub-words, perform fuzzy matching for each sub-word in the preset name database. If an entity name contains any one of the query sub-words or has features such as associated pinyin / abbreviation / synonym, etc. with the query sub-word, add this entity name to the candidate associated name set. Factors such as word segmentation co-occurrence weight, sub-word matching position offset, and context integrity can be introduced to score and rank the matching results.

[0093] In addition, to enhance the adaptability to situations such as users' non-standard expressions, aliases, colloquial abbreviations, typing errors, etc., a preset variant name database is also introduced, which stores a large number of standard names of entity names and their corresponding multiple variant expressions, including but not limited to: abbreviations (such as "CMB" → "China Merchants Bank"), English or pinyin expressions (such as "ICBC" → "Industrial and Commercial Bank of China"), synonymous names, regional names, nicknames, misspelled / mistyped forms (such as "Zhongnong Bank" → "Agricultural Bank of China").

[0094] Here, after the target query word is determined, use this target query word as the retrieval key to perform matching retrieval in the preset variant name database. If a matching variant item is retrieved, extract its corresponding standard entity name and add this standard name to the candidate entity name set as the variant recognition result. For example, when the user inputs "CMB Credit Card", it can be recognized in the variant library that "CMB" is the abbreviation of "China Merchants Bank", and "China Merchants Bank Credit Card" is the corresponding standard name.

[0095] In a preferred embodiment of the present invention, in order to achieve accurate and efficient retrieval of entity names associated with query sub-words in the target query term, an associated name filtering method based on an inverted index and a weighted scoring mechanism is adopted. The entity names stored in the preset name database are divided into corresponding entity sub-words; an inverted index is built based on the entity sub-words; candidate sub-words matching the entity sub-words are retrieved from the preset name database based on the inverted index; a weighted score is generated for each candidate sub-word based on its position in the corresponding entity name, its part of speech, and its importance; and entity names to which candidate sub-words with weighted scores greater than a preset weighted score threshold belong are selected as associated names.

[0096] In practice, the system first performs sub-word-level segmentation on each entity name stored in the pre-defined name database. These sub-words can be obtained by using Chinese word segmentation algorithms (such as dictionary-based matching, bidirectional maximum matching, CRF models, etc.) to segment the entity name, resulting in a set of semantically complete entity sub-words. Subsequently, the system constructs an inverted index table based on all entity sub-words. The index structure records the set of entity names corresponding to each sub-word. This inverted index significantly improves the retrieval efficiency of subsequent sub-word matching entity queries.

[0097] Here, after the target query term is split into several query sub-terms, the inverted index structure constructed above is used to quickly retrieve matching candidate sub-terms and their corresponding entity name sets in the inverted index, using each query sub-term as a key. During the matching process, both exact matching and fuzzy matching (such as pinyin similarity, synonym mapping, simplified / traditional conversion, etc.) are supported to improve recall coverage.

[0098] To assess the relevance of each candidate word within its corresponding entity name, a multi-factor weighted scoring mechanism is introduced. Each candidate word is weighted across the following dimensions: position, part-of-speech, and frequency / importance. Candidate words located at the beginning or in the main body of the entity name are assigned higher weight. Based on part-of-speech tagging (e.g., organization name, person name, job title, geographical area name), their semantic core is determined; for example, "bank" and "securities" are often used as primary core words. Combined with comprehensive database statistics (e.g., TF-IDF), the discriminative power of the word is evaluated. Low-frequency and unique words generally have higher importance.

[0099] Optionally, query context matching can also be introduced, taking into account the semantic relevance or vector similarity between query subwords and candidate subwords.

[0100] Finally, a weighted comprehensive score is generated for each candidate sub-word. After all candidate sub-word weight scores are completed, candidate sub-words with weight scores greater than the preset threshold will be screened out, and the entity names to which they belong will be extracted as the set of associated names corresponding to this query.

[0101] In a further preferred embodiment of the present invention, in order to improve the recognition ability of name variants in the user's query intention and enhance the tolerance for abbreviations, acronyms, homophonic expressions or language differences, a variant name retrieval method based on a name variant library and a fuzzy matching algorithm is designed and implemented. For each entity name stored in the preset name database, variant forms corresponding to the entity name are collected to construct a preset variant name database. The variant forms at least include abbreviations, acronyms, homophone substitutions, and English translations; after standardizing the target query term, the variant similarity with the variant forms in the preset variant name database is determined, and a fuzzy matching algorithm is used for matching according to the variant similarity, and variant forms with variant similarity greater than the preset variant threshold are screened out as variant names.

[0102] In specific implementation, first, each standard entity name stored in the preset name database is expanded to construct various variant forms of the entity name to form a preset variant name database. The construction methods of the variant forms include but are not limited to the following: Abbreviation generation: Extract the core word or representative phrase in the entity name. For example, the abbreviations of "China Merchants Bank Credit Card Center" can be "CMB Credit Card", "CMB Card Center", etc.; Acronym generation: Generate acronymic expressions based on the first letters of pinyin or Chinese-English abbreviation rules. For example, "China Merchants Bank" corresponds to "ZSYH", and "Industrial and Commercial Bank of China" corresponds to "GSYH"; Homophone substitution: Replace the Chinese characters in the name with homophonic characters with the same pinyin but different glyphs. For example, "ZTE" may be mis-entered by users as "Zhongxing Communication", "Zhongxing Telecom", etc.; English translation: According to the general translation rules in the industry, generate English expression forms or abbreviations for Chinese entity names. For example, "Bank of China" corresponds to "Bank of China" or "BOC". Each standard entity name and all its variant forms are stored in association in the database, and a corresponding identification field is attached for subsequent matching recognition and traceability.

[0103] Here, in order to improve the alignment accuracy with variant names, the input target query term is preprocessed and standardized, including but not limited to the following operations: removing stop words, special symbols or formatting characters; converting to pinyin representation or first letter expression; unified encoding representation of Chinese and English; normalizing differences in Chinese and English case, simplified and traditional forms, etc. Through the above standardization process, the target query term can be converted into a set of standardized expression forms for comparison.

[0104] Furthermore, a fuzzy matching algorithm (such as edit distance, Jaccard similarity, Cosine similarity, BERT vector cosine similarity, etc.) is used to calculate the similarity between the standardized form of the target query term and each variant in the preset variant name database, and the variant similarity between each variant form and the target query term is obtained.

[0105] Exemplarily, the edit distance between the target query term "China CITIC Bank" and the variant "CITIC Bank" is small; or the query term "GSYH" of "Industrial and Commercial Bank of China" is exactly matched with the abbreviation "GSYH"; or through BERT encoding, the similarity between "BOC" and "Bank of China" is higher than 0.85.

[0106] After that, after completing all the comparisons, candidate names with similarity values higher than the preset variant similarity threshold (such as 0.75 or 85%) are selected from all the variant forms, and the entity names corresponding to these matching variants are output as the variant name set. Multiple strategies can be set for this screening process: preferentially select candidates with a higher variant level (such as official abbreviations, industry abbreviations); retain multiple high-similarity variants for subsequent fusion and sorting; perform secondary filtering or weighting on the screening results in combination with the context or historical click behavior.

[0107] It should be noted that the preset variant similarity threshold can be set according to actual needs and is not specifically limited here.

[0108] In summary, by constructing a systematic entity name variant library and calculating and screening similarities based on the standardized expression and fuzzy matching mechanism, this embodiment can effectively improve the compatibility with various non-standard expressions in user input, significantly enhance the name recall ability and fault tolerance, and is applicable to various natural language processing application scenarios such as search error correction, knowledge graph entity linking, public opinion recognition, and question answering systems.

[0109] S104. Use the pronunciation-similar name, the spelling-similar name, the associated name, and the variant name as candidate names, and determine the matching similarity between each candidate name and the target query term.

[0110] In specific implementation, to achieve accurate recognition and matching of the target query term, after the preliminary screening of the pronunciation-similar name, the spelling-similar name, the associated name, and the variant name, the names from the above multiple sources are uniformly collected into a candidate set, and the similarity between each candidate name and the target query term is further evaluated.

[0111] Here, names with similar pronunciations are obtained through phoneme matching or phonetic edit distance calculation; names with similar spellings are obtained based on the initial letter abbreviations of pinyin and spelling similarity; related names are selected based on sub-word inverse indexing and semantic weight evaluation; variant names are derived from abbreviations, shortened names, English names, and other modified expressions obtained through standardization and fuzzy matching. To avoid duplicate processing of names in the candidate set, unique entity identifiers and their source types are retained.

[0112] Furthermore, the matching similarity value between each candidate name and the target query term is calculated separately to quantify the degree of similarity between names. The calculation of matching similarity can integrate multiple dimensions, including but not limited to: character-level similarity (e.g., edit distance, Levenshtein distance); pinyin or phoneme-level similarity (e.g., pinyin similarity rate, phoneme nesting ratio); semantic encoding similarity (e.g., cosine similarity calculation after encoding using language models such as BERT, RoBERTa, etc.); context relevance score (e.g., weighted score based on historical search click behavior or semantic context); source credibility weighting: assigning different initial weights to candidate names from different sources, such as lower scores for candidates from sources with similar pronunciations, while scores for variant or related names are moderate; if the name originates from commonly used abbreviations by historical users, the score can be appropriately increased.

[0113] Here, combining the above similarity indicators, the final matching similarity is calculated using a linear weighted model, logistic regression, or neural network scoring model. All candidate names are sorted in descending order based on the matching similarity value, and the names with the highest similarity are selected as the final recommended matching results. Alternatively, a threshold is used to determine if a clear matching target exists. If the highest similarity exceeds a preset matching threshold (e.g., 0.85), it is determined to be the standard entity name corresponding to the target query term. If multiple high-similarity candidates exist, the candidate fusion and sorting step is used for fine-tuning. If all candidates are below the threshold, the system is determined to be in an unmatched state, prompting the user to correct errors or supplement input.

[0114] In another preferred embodiment of the present invention, to further improve the accuracy and robustness of candidate name matching, multiple similarity calculation methods are integrated to achieve precise quantification of the matching degree between candidate names and target query terms, determining the string edit distance and Jaro-Winkler similarity between candidate names and target query terms; through a pre-trained text embedding model, entity names stored in a preset name database are converted into corresponding entity vectors, and target query terms are converted into corresponding query vectors; the vector similarity between entity vectors and query vectors is calculated; and the string edit distance, Jaro-Winkler similarity, and vector similarity are determined as the matching similarity.

[0115] In practice, character-level similarity is first calculated for each candidate name and the target query term. For the Levenshtein Distance, the minimum number of editing operations (insertion, deletion, or replacement) required to convert the candidate name into the target query term is calculated; the smaller this value, the closer the two are. Jaro-Winkler similarity considers the relative positions of characters in the string, giving higher weight to prefix matching, making it suitable for short text and name matching scenarios. The edit distance and Jaro-Winkler similarity values ​​obtained through these two algorithms serve as preliminary character-level matching indicators.

[0116] Here, to incorporate contextual semantic features, pre-trained text embedding models such as BERT, RoBERTa, and Sentence-BERT are used to embed candidate entity names and target query terms into dense semantic vectors. Each entity name in the pre-defined name database is converted into a corresponding entity vector, and the target query term is converted into a corresponding query vector. Then, the cosine similarity function is used to calculate the semantic similarity between each candidate entity vector and the target query vector. This similarity reflects the degree of closeness between the two in the semantic space. The closer the value is to 1, the closer the two are semantically.

[0117] Specifically, by designing positive and negative sample pairs, the model learns to map semantically similar sentences to similar vector spaces. During training, input sentence pairs are encoded using BERT with shared weights to obtain two vectors, and the similarity between the vectors is calculated using cosine similarity. For positive sample pairs, the goal is to make their vector distance as close as possible; for negative sample pairs, the goal is to make their distance as far as possible.

[0118] The construction methods for positive and negative samples are as follows: In unsupervised SimCSE, positive samples are the representations of a sentence generated by two different dropouts, and negative samples are other sentences in the same batch; In supervised SimCSE, "positive equivalence sentences" from semantically similar datasets (such as NLI) are often used as positive samples, and semantically unrelated sentences are used as negative samples.

[0119] Furthermore, after training, the entire sentence can be represented using BERT's [CLS] vector or the token vector after average pooling, which can then be used for subsequent tasks such as retrieval, clustering, and question answering, achieving semantic vector alignment. The Loss function is designed as follows:

[0120] o = softmax(w t (u,v,|uv|))

[0121] Where u and v represent the embeddings of two sentences, that is, the sentence vector representations after processing by the BERT model, and w t This represents a trainable weight matrix with dimension R. 3n×k , where n is the dimension of the sentence embedding and k is the number of category labels.

[0122] S105. The candidate names are sorted according to the matching similarity and then output as name retrieval results.

[0123] In practice, during the fine-ranking stage, although various similarity calculation methods were used to evaluate and rank the recall results, each method has its own limitations. For example, Levenshtein distance mainly focuses on character edit distance while ignoring semantic information; Jaro-Winkler similarity emphasizes prefix matching but may not be sensitive enough to some complex semantic variations; and while vector similarity can capture semantic associations, it may be limited by the embedding model in some cases.

[0124] In another preferred embodiment of the present invention, in order to further improve the accuracy of entity name recognition and user search experience, all candidate names are fused and sorted based on matching similarity. For string edit distance, Jaro-Winkler similarity and vector similarity, a Softmax function with a preset temperature coefficient is used for weighted integration to determine the comprehensive matching score corresponding to each candidate name. The candidate names are sorted from largest to smallest according to the comprehensive matching score to generate a name search result sequence.

[0125] Specifically, for each candidate name, three types of matching similarity indicators have been obtained between it and the target query term: string edit similarity is calculated by normalized edit distance, Jaro-Winkler similarity reflects the local matching degree of character order, and vector semantic similarity is based on the global semantic matching result of the embedded vector.

[0126] Here, to avoid the influence of manually set static weights on the fusion result and to improve the model's adaptability to different feature types, a Softmax function carrying a temperature coefficient is introduced to normalize and integrate the above three similarities, as shown in the following formula:

[0127]

[0128] Where Z represents each matching sublist, and T is a temperature coefficient that controls the sensitivity of the similarity weight distribution. A smaller T value indicates a more concentrated weight distribution, while a larger T value indicates a more uniform distribution. It is typically set within the range of 0.1 to 1 and requires parameter tuning in practical applications. i Let z represent the similarity of the i-th item. jrepresents the similarity of the j-th item, and finalscore represents the comprehensive matching score, which combines the similarity of character matching and semantic matching, and can effectively measure the degree of matching between candidate names and target query terms.

[0129] Furthermore, all candidate names are sorted in descending order of their comprehensive matching scores, with higher comprehensive scores given priority, to obtain the final name retrieval result sequence. The result sequence can be directly used to display to users, provide entity alignment candidates to downstream models, or be used for tasks such as entity linking and named entity disambiguation in knowledge graphs.

[0130] This disclosure provides an entity name retrieval method that involves: acquiring a target query term; determining the pronunciation and spelling features corresponding to the target query term; splitting the target query term into multiple query sub-words; retrieving similarly pronounced and spelled names in a preset name database based on the pronunciation and spelling features; retrieving related names in the preset name database using the query sub-words as keywords; and retrieving variant names in a preset variant name database that match the target query term; using the similarly pronounced, spelled, related, and variant names as candidate names; determining the matching similarity between each candidate name and the target query term; and sorting the candidate names according to the matching similarity and outputting the results as the name retrieval results. This method can improve recall coverage and ranking efficiency, enhance robustness to entity alias ambiguity, and ensure the accuracy and consistency of the search experience.

[0131] See Figure 2 The diagram shows a flowchart of another entity name retrieval method provided in this embodiment of the present disclosure. The method includes steps S201 to S205, wherein:

[0132] S201. Extract the relationship between corporate entities and person entities based on publicly available business registration data.

[0133] S202. Establish an enterprise-person knowledge graph with the enterprise entity and the person entity as nodes and the relationship as connecting edges.

[0134] S203. Extract the request entity appearing in the target query term. When the request entity matches an enterprise node or person node in the enterprise-person knowledge graph, extract the neighbor node corresponding to the matching node.

[0135] S204. Configure corresponding relationship weights between the enterprise entity or person entity corresponding to the matching node and the enterprise entity or person entity corresponding to the neighbor node.

[0136] S205. The enterprise entity or person entity corresponding to the matching node and the enterprise entity or person entity corresponding to the neighbor node are weighted and integrated according to the relationship weight, and then added to the candidate name.

[0137] In a further embodiment of the present invention, to expand the coverage of candidate names and enhance the semantic robustness of name matching, a corporate-person knowledge graph constructed based on business registration data is introduced to achieve candidate name completion based on relationships. First, structured relationships between corporate entities and person entities are extracted from publicly available data sources such as business registration data, annual report information, and company executive filing information. These relationships include, but are not limited to, the following: executive appointment relationships (e.g., "legal representative," "director," "manager"), investment and holding relationships (e.g., "actual controller," "shareholder," "investor"), historical appointment / departure relationships, and person relationships formed by indirect investment chains between enterprises. After completing the above entity relationship extraction, a heterogeneous graph is constructed using corporate entities and person entities as nodes and the extracted relationships as edges, defined as the corporate-person knowledge graph. The type and direction of the edges in this graph can be labeled to support subsequent relationship weight calculations.

[0138] Here, after receiving the target query term, entity extraction is performed on the query term to identify the request entities that may be involved, such as company names, natural person names, abbreviations, or aliases. By matching them with nodes in the knowledge graph, it is determined whether there are corresponding company or person nodes. If a match is successful, the neighboring nodes of the matching node are further extracted, that is, other entity nodes that have direct association edges with that node.

[0139] Specifically, for each edge between the matching node (denoted as the master node) and its neighboring nodes, the system configures a corresponding relationship weight based on the relationship type and importance of the edge. The specific weight can be determined based on the following dimensions: the trust strength of the relationship category (e.g., "legal representative" has a higher weight than "ordinary employee"); the frequency of data occurrence (e.g., someone has held multiple positions in the company); the relationship's time span or the most recent update date (the newer the relationship, the higher the weight); and the confidence of multi-hop paths (e.g., the multi-hop weight decay mechanism caused by indirect holding chains).

[0140] Then, the enterprise or person entity represented by the matching node is weighted and integrated with the enterprise or person entities corresponding to the neighboring nodes according to the above relationship weights to form a set of extended entities with strong semantic association.

[0141] Furthermore, the entities corresponding to the neighbor nodes obtained through graph reasoning are added to the existing candidate name set as semantically enhanced candidate names that are related to the target query term. In the enterprise-person knowledge graph, identity information attribute tags are added to person entities, and enterprise status attribute tags are added to enterprise entities; the attribute information corresponding to the request entity is determined, and matching nodes are extracted according to the preset attribute priority.

[0142] In practical implementation, for individual entity nodes in the graph, the system automatically identifies and adds identity attribute tags based on business registration data. Identity information includes, but is not limited to, tags for legal representative, actual controller, investor / shareholder, etc. Each individual entity can carry multiple identity tags and records additional information such as its corresponding company, period of employment, and data source. For company entities, the system extracts the following company status attribute tags from business registration information: existing, operating, revoked, deregistered, etc.; whether it is a listed company; whether it is a suspected related company (e.g., derived through address or legal representative reverse mapping); and auxiliary tags such as industry type, registered capital, and years of operation.

[0143] Here, once the requested entity is identified from the target query term, it is not only compared with the entity node in the graph based on its name, but also attempts to analyze the attribute tendencies that the requested entity may imply (for example, "finding a company where someone works" implies matching entities with the "executive" label). The attribute dimensions that the user is interested in can be inferred through contextual understanding (such as query intent classification, natural language question structure analysis, etc.) or structured prompts (such as selecting parameters such as "legal person priority" or "listed company priority").

[0144] When multiple nodes with matching names are found in the graph, nodes with higher attribute priority are retained first. For example, if a person holds positions in multiple companies, the companies where they serve as the "legal representative" or "actual controller" are returned first; if multiple companies with the same name are matched, the companies with a status of "existing" or "operating" are returned first.

[0145] Those skilled in the art will understand that, in the above-described method of the specific implementation, the order in which each step is written does not imply a strict execution order and does not constitute any limitation on the implementation process. The specific execution order of each step should be determined by its function and possible internal logic.

[0146] This disclosure provides another entity name retrieval method, which extracts the association between enterprise entities and person entities based on publicly available business registration data; establishes an enterprise-person knowledge graph with the enterprise entities and person entities as nodes and the association relationships as connecting edges; extracts the request entity appearing in the target query term; when the request entity matches an enterprise node or person node in the enterprise-person knowledge graph, extracts the neighbor nodes corresponding to the matching node; configures corresponding relationship weights between the enterprise entity or person entity corresponding to the matching node and the enterprise entity or person entity corresponding to the neighbor node; and supplements the candidate names by weighting and integrating the enterprise entity or person entity corresponding to the matching node and the enterprise entity or person entity corresponding to the neighbor node according to the relationship weights. By constructing a heterogeneous enterprise-person graph and integrating a structured graph reasoning mechanism, this method provides a relationship-enhanced completion path for entity recognition and matching, which has significant value in application scenarios such as judicial risk screening, business intelligence mining, and financial compliance monitoring.

[0147] Based on the same inventive concept, this disclosure also provides an entity name retrieval device corresponding to the entity name retrieval method. Since the principle of the device in this disclosure for solving the problem is similar to the entity name retrieval method described above in this disclosure, the implementation of the device can refer to the implementation of the method, and the repeated parts will not be described again.

[0148] Please see Figure 3 , Figure 3 This is a schematic diagram of an entity name retrieval device provided in an embodiment of this disclosure. Figure 3 As shown in the embodiments of this disclosure, the entity name retrieval device 300 includes:

[0149] The feature extraction module 310 is used to acquire the target query term, determine the pronunciation features and spelling features corresponding to the target query term, and split the target query term into multiple query sub-words.

[0150] The first recall module 320 is used to retrieve, in a preset name database, matching pronunciation-similar names and spelling-similar names based on the pronunciation features and spelling features, respectively.

[0151] The second recall module 330 is used to retrieve matching associated names in the preset name database using the query sub-word as the keyword, and to retrieve matching variant names in the preset variant name database.

[0152] The similarity determination module 340 is used to determine the matching similarity between each of the candidate names and the target query term by taking the pronunciation similar name, the spelling similar name, the associated name and the variant name as candidate names.

[0153] The sorting module 350 is used to sort the candidate names according to the matching similarity and output the results as name retrieval results.

[0154] The processing flow of each module in the device and the interaction flow between each module can be referred to the relevant descriptions in the above method embodiments, and will not be detailed here.

[0155] This disclosure provides an entity name retrieval device that acquires a target query term, determines its corresponding pronunciation and spelling features, and splits the target query term into multiple query sub-words. It then retrieves matching pronunciation-similar names and spelling-similar names from a preset name database based on the pronunciation and spelling features, respectively. Next, it retrieves matching associated names from the preset name database using the query sub-words as keywords, and retrieves matching variant names from a preset variant name database. Finally, it uses the pronunciation-similar names, spelling-similar names, associated names, and variant names as candidate names, determines the matching similarity between each candidate name and the target query term, and sorts the candidate names according to the matching similarity, outputting the results as name retrieval results. This method can improve recall coverage and ranking efficiency, enhance robustness to entity alias ambiguity, and ensure the accuracy and consistency of the search experience.

[0156] Corresponding to Figure 1 and Figure 2 In addition to the entity name retrieval method in this disclosure, this embodiment also provides an electronic device 400, such as... Figure 4 The diagram shown is a structural schematic of an electronic device 400 provided in an embodiment of this disclosure, including:

[0157] Processor 41, memory 42, and bus 43; memory 42 is used to store execution instructions, including main memory 421 and external memory 422; the main memory 421, also called internal memory, is used to temporarily store the computational data in processor 41, as well as the data exchanged with external memory 422 such as hard disk. Processor 41 exchanges data with external memory 422 through main memory 421. When the electronic device 400 is running, processor 41 and memory 42 communicate through bus 43, enabling processor 41 to execute... Figure 1 and Figure 2 The steps of the entity name retrieval method in [the document / platform].

[0158] This disclosure also provides a computer-readable storage medium storing a computer program that, when executed by a processor, performs the steps of the entity name retrieval method described in the above-described method embodiments. The storage medium may be a volatile or non-volatile computer-readable storage medium.

[0159] This disclosure also provides a computer program product, which includes computer instructions. When the computer instructions are executed by a processor, they can perform the steps of the entity name retrieval method described in the above method embodiments. For details, please refer to the above method embodiments, which will not be repeated here.

[0160] The aforementioned computer program product can be implemented through hardware, software, or a combination thereof. In one optional embodiment, the computer program product is specifically embodied in a computer storage medium; in another optional embodiment, the computer program product is specifically embodied in a software product, such as a software development kit (SDK), etc.

[0161] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the device described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. In the several embodiments provided in this disclosure, it should be understood that the disclosed device and method can be implemented in other ways. The device embodiments described above are merely illustrative. For example, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. Furthermore, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Another point is that the displayed or discussed mutual coupling or direct coupling or communication connection may be through some communication interfaces; the indirect coupling or communication connection of devices or units may be electrical, mechanical, or other forms.

[0162] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0163] In addition, the functional units in the various embodiments of this disclosure can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0164] If the aforementioned functions are implemented as software functional units and sold or used as independent products, they can be stored in a processor-executable, non-volatile, computer-readable storage medium. Based on this understanding, the technical solution of this disclosure, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this disclosure. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0165] Finally, it should be noted that the above-described embodiments are merely specific implementations of this disclosure, used to illustrate the technical solutions of this disclosure, and not to limit it. The protection scope of this disclosure is not limited thereto. Although this disclosure has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments, or make equivalent substitutions for some of the technical features, within the scope of the technology disclosed in this disclosure. Such modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this disclosure, and should all be covered within the protection scope of this disclosure. Therefore, the protection scope of this disclosure should be determined by the protection scope of the claims.

Claims

1. A method for retrieving entity names, characterized in that, include: Obtain the target query term, determine the pronunciation and spelling features corresponding to the target query term, and split the target query term into multiple query sub-words; In the preset name database, similar names with similar pronunciations and similar names with similar spellings are retrieved based on the pronunciation features and spelling features, respectively. The system retrieves matching related names in the preset name database using the query sub-word as the keyword, and retrieves matching variant names in the preset variant name database. The similar-sounding names, similar-spelling names, related names, and variant names are used as candidate names to determine the matching similarity between each candidate name and the target query term; The candidate names are sorted according to the matching similarity and then output as the name search results.

2. The method according to claim 1, characterized in that, Based on the following steps, similar-sounding names are retrieved: The target query term is converted into a corresponding query term phoneme sequence, and the entity name stored in the preset name database is converted into a corresponding entity name phoneme sequence; Determine the phonetic edit distance between the query word phoneme sequence and the entity name phoneme sequence, or encode both the query word phoneme sequence and the entity name phoneme sequence into corresponding phonetic codes, and calculate the encoding similarity between the phonetic codes; The entity names that are greater than the corresponding preset pronunciation similarity threshold in the preset name database are selected as the pronunciation similar names.

3. The method according to claim 1, characterized in that, The following steps are used to retrieve similarly spelled names: The entity names stored in the preset name database are converted into the corresponding entity pinyin, the entity initials abbreviations corresponding to the entity pinyin are extracted, and an entity index table is constructed based on the entity initials abbreviations. The target query term is converted into its corresponding pinyin, and the initial letter abbreviation of the pinyin is extracted. Based on the entity index table, select spelling similar names from the entity initials that have a spelling similarity greater than a preset spelling similarity threshold with the query initials.

4. The method according to claim 1, characterized in that, The associated name is retrieved based on the following steps: The entity names stored in the preset name database are divided into corresponding entity sub-words; An inverted index is built based on the entity sub-words, and candidate sub-words that match the entity sub-words are retrieved from the preset name database based on the inverted index; Based on the position of the candidate word in the corresponding entity name, the part of speech of the candidate word, and the importance of the candidate word, a weight score is generated for each candidate word; The entity name to which the candidate sub-words with a weight score greater than a preset weight score threshold belong is selected as the associated name.

5. The method according to claim 1, characterized in that, The variant name is retrieved based on the following steps: For each entity name stored in the preset name database, the corresponding variant forms of the entity name are collected to construct the preset variant name database. The variant forms include at least abbreviations, acronyms, homophone substitutions, and English translations. After standardizing the target query term, the variant similarity between it and the variant form in the preset variant name database is determined. A fuzzy matching algorithm is used to match the variants based on the variant similarity, and the variant forms with a variant similarity greater than a preset variant threshold are selected as the variant names.

6. The method according to claim 1, characterized in that, Determining the matching similarity between each candidate name and the target query term specifically includes: Determine the string edit distance and Jaro-Winkler similarity between the candidate name and the target query term; The pre-trained text embedding model converts the entity names stored in the preset name database into corresponding entity vectors and converts the target query terms into corresponding query vectors. Calculate the vector similarity between the entity vector and the query vector; The string edit distance, the Jaro-Winkler similarity, and the vector similarity are determined as the matching similarity.

7. The method according to claim 6, characterized in that, The candidate names are sorted according to the matching similarity, specifically including: The string edit distance, Jaro-Winkler similarity, and vector similarity are weighted and integrated using a Softmax function carrying a preset temperature coefficient to determine the comprehensive matching score corresponding to each candidate name; Based on the comprehensive matching score from largest to smallest, the candidate names are sorted to generate a sequence of name retrieval results.

8. The method according to claim 1, characterized in that, The method further includes: The relationships between corporate entities and individuals are extracted based on publicly available business registration data; Establish an enterprise-person knowledge graph with the enterprise entities and person entities as nodes and the association relationships as connecting edges; Extract the request entity that appears in the target query term. When the request entity matches the enterprise node or person node in the enterprise-person knowledge graph, extract the neighbor node corresponding to the matching node. Configure corresponding relationship weights between the enterprise entity or person entity corresponding to the matching node and the enterprise entity or person entity corresponding to the neighbor node; The enterprise entity or person entity corresponding to the matching node and the enterprise entity or person entity corresponding to the neighbor node are weighted and integrated according to the relationship weight, and then added to the candidate name.

9. The method according to claim 8, characterized in that, The method further includes: In the enterprise-person knowledge graph, identity information attribute tags are added to the person entities, and enterprise status attribute tags are added to the enterprise entities; Determine the attribute information corresponding to the request entity, and extract the matching node according to the preset attribute priority.

10. An entity name retrieval device, characterized in that, include: The feature extraction module is used to acquire the target query term, determine the pronunciation features and spelling features corresponding to the target query term, and split the target query term into multiple query sub-words; The first recall module is used to retrieve, in a preset name database, matching pronunciation-similar names and spelling-similar names based on the pronunciation features and spelling features, respectively. The second recall module is used to retrieve matching associated names in the preset name database using the query sub-word as the keyword, and to retrieve matching variant names in the preset variant name database. The similarity determination module is used to determine the matching similarity between each of the candidate names and the target query term by taking the pronunciation-similar names, the spelling-similar names, the associated names and the variant names as candidate names. The sorting module is used to sort the candidate names according to the matching similarity and output the results as name retrieval results.

11. An electronic device, characterized in that, include: The device includes a processor, a memory, and a bus. The memory stores machine-readable instructions executable by the processor. When the electronic device is running, the processor communicates with the memory via the bus. When the machine-readable instructions are executed by the processor, they perform the steps of the entity name retrieval method as described in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, performs the steps of the entity name retrieval method as described in any one of claims 1 to 9.

Images