A method and system for matching alumni referral focus

Abstract

The application discloses a kind of key alumni recommendation matching method and system, it is related to data mining and intelligent recommendation technical field, this method includes: the initial collection of various public lists is formed initial collection list, after combination keyword search, multi-dimensional data deduplication, webpage text standardization cleaning, structured information of person is extracted by feature extraction model, the name of school and time verification are completed by combining the time pedigree of school evolution, finally generate key alumni recommendation list, the timeline conflict of static matching is corrected by the time verification of school evolution in the application, the accuracy of alumni determination is improved;Through multi-source confidence weighted distinction information efficiency, make result credibility level clear, it is convenient for subsequent review application.

Description

Technical Field

[0001] This invention relates to the field of data mining and intelligent recommendation technology, specifically to a method and system for recommending and matching key alumni. Background Technology

[0002] Alumni resources are a crucial foundation for universities to conduct talent recruitment, research collaborations, and social services. Key alumni who have received various honors and awards, or who hold important positions or have participated in research projects, are of practical value to the university's work. However, with the increasing abundance of publicly available information online, various honor announcements, research awards, and talent selection lists are scattered across different platforms, making traditional manual processing methods insufficient to meet the needs of large-scale identification and matching of key alumni.

[0003] Current methods for identifying key alumni have several shortcomings. First, data collection relies on manual processing, with information sources scattered and inconsistent in format, resulting in low processing efficiency and difficulty in covering the massive amount of publicly available information. Second, the matching of institution names often uses static alias mapping, without verifying against historical timelines of institution name changes or mergers, easily leading to erroneous matches where the individual's enrollment date does not match the institution's timeline. Third, traditional keyword matching methods have limited accuracy in extracting information from unstructured text, making it difficult to effectively distinguish individuals with the same name, and they do not differentiate the validity of information from different sources, making it difficult to guarantee the reliability of the final identification results. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and provide a key alumni recommendation and matching method. By building a fully automated processing chain from list collection, retrieval and collection, deduplication and cleaning to structured extraction and institution matching, and introducing an institution evolution time verification mechanism and a multi-source information confidence weighting rule, this invention solves the problems of low efficiency in traditional manual processing, timeline conflicts easily occurring in static name matching, and inability to distinguish the validity of information from different sources. It achieves automated identification and accurate matching of key alumni in massive amounts of public information, improving processing efficiency while ensuring the accuracy and credibility of alumni identity determination.

[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0006] A method for recommending and matching key alumni includes the following steps:

[0007] S1: Collect publicly available lists of various honors, awards, and talent announcements to form an initial list of candidates to be collected, including the names of individuals, information sources, and list categories.

[0008] S2: Combine the names and employers of individuals in the initial list of entities to be collected with alumni information to form multiple sets of search keywords, then perform batch searches and save the corresponding webpage links and original text as the initial dataset.

[0009] S3: Based on the initial dataset, duplicate data is removed by judging the similarity of records. The complete data of the details page is crawled for the deduplicated information source to build a deduplicated dataset.

[0010] S4: Perform standardization cleaning on the original text of the web pages collected from the deduplication dataset to obtain valid text.

[0011] S5: Input the effective text into the feature extraction model to extract structured data.

[0012] S6: Standardize the matching of school names in the extracted structured data, perform time verification in conjunction with the timeline of school history, output the standard school names, complete the alumni identity determination, and obtain the school matching results.

[0013] S7: By combining the results of matching institutions with the structured data of individuals' honor levels and job titles, a list of recommended key alumni is generated.

[0014] Preferably, in step S3, the record similarity is calculated using a similarity function. When the calculated record similarity value is greater than a preset deduplication threshold, the corresponding two records are determined to be duplicate data and are merged. The expression for the similarity function is:

[0015] In the formula, To record similarity, the physical meaning of recording similarity is to record... With records The similarity calculation value; This is an indicator function; it takes the value 1 when the condition inside the parentheses is true and 0 when it is false. , These are the names of the people corresponding to the two records; , These are the employers of the individuals corresponding to the two records; , These are the information sources corresponding to the two records; , , The weighting coefficients for the name dimension, employer dimension, and information source dimension are respectively, satisfying... .

[0016] Preferably, the standardized matching in step S6 uses a comprehensive matching score calculation. The university name with the highest comprehensive matching score is selected as the candidate matching result. When the comprehensive matching score is greater than a preset matching threshold and the time verification passes, the university is considered to have matched successfully. The formula for calculating the comprehensive matching score is:

[0017] In the formula, School names to be matched With standard institution name The overall matching score; The edit distance between the school name to be matched and the standard school name; The length of the school name to be matched; The character length of a standard institution's name; This represents the semantic vector similarity between the school name to be matched and the standard school name; For the matching degree of the level name; For core field matching degree; , , , These are the weight coefficients for the edit distance similarity term, semantic vector similarity term, category name matching term, and core field matching term, respectively, satisfying... .

[0018] Preferably, in step S2, when saving the corresponding webpage link and original text, for webpages dynamically loaded by JavaScript, the text content is collected after page rendering; for static webpages, the text content is collected using a direct HTTP request method.

[0019] Preferably, in step S5, the temperature parameter of the feature extraction model is set to 0, and the thinking mode is turned off. The structured information extraction process is implemented using an asynchronous concurrent framework, controlling the preset number of concurrent processes through semaphores, and processing data in batches according to a preset batch size. Before extraction, a prompt word template containing role definitions, core extraction instructions, field definition rules, and output format requirements is constructed, and the prompt words are concatenated with valid text before being input into the large model. During processing, call exceptions are logged, and a preset timeout is set for each model call.

[0020] Preferably, the specific process of school matching and time verification in step S6 is as follows:

[0021] S601: Pre-build a genealogy database of institutional history, storing the standard name, alias, former name of each institution, as well as the time nodes corresponding to institution name changes, mergers, and splits.

[0022] S602: Calculate the comprehensive matching score between the school name to be matched and each standard institution name, and select the standard institution name with the highest score as the candidate matching result.

[0023] S603: Extract the year of enrollment or graduation of the corresponding person and compare it with the timeline of the candidate school. If the person's enrollment time is earlier than the time of the school's name change and the candidate matching result is the new name of the school, then switch the matching result to the former name of the corresponding period and retain the standard school association relationship.

[0024] S604: When the overall matching score is greater than the preset matching threshold and the time verification is passed, the school matching is considered successful.

[0025] Preferably, in step S7, the sources are pre-categorized according to their authority level and corresponding confidence weights are set. Sources from official announcements have a higher weight than those from official organization websites, and those from official organization websites have a higher weight than those from commercial news sources. The matching results from various supporting sources for the same person are weighted and summed to obtain the comprehensive confidence level of the alumni identity. The list of recommended key alumni is then sorted and output according to the comprehensive confidence level and the level of honor.

[0026] Preferably, in step S6, when the difference in the comprehensive matching scores of the top two standard university names is less than a preset ambiguity threshold, the corresponding person's regional and industry information is retrieved, and the matching degree is calculated with the location information and advantageous discipline information of the two candidate universities respectively. The university with the higher matching degree is selected as the final matching result.

[0027] Another object of the present invention is to provide a key alumni recommendation and matching system, applicable to a key alumni recommendation and matching method, comprising:

[0028] The data collection module is used to collect various public lists and generate an initial list of lists to be collected, which includes the names of the individuals, the source of the information, and the category of the list.

[0029] The whole network search and collection module is used to combine the names and employers of people in the initial list of people to be collected with alumni information to form multiple sets of search keywords, and then perform batch search, saving the corresponding web page links and original text as the initial dataset.

[0030] The data deduplication and re-collection module is used to calculate the similarity of records based on the initial dataset through three dimensions: name, employer, and information source. It removes duplicate data and merges records from the same source. It then retrieves complete data from the details page of the deduplicated information source to construct a deduplicated dataset.

[0031] The data cleaning and noise reduction module is used to perform standardized cleaning on the original text of web pages collected in the deduplication dataset, remove irrelevant text blocks, filter text paragraphs that do not contain school names, clean HTML tags and redundant whitespace characters, and obtain valid text.

[0032] The structured information extraction module is used to input effective text into the feature extraction model and extract structured data containing basic personal information, honors and awards, work experience, and educational experience.

[0033] The alumni identity verification module is used to perform standardized matching on the extracted school names, perform time verification in conjunction with the school's historical timeline, and output the standard school name to complete the alumni identity verification.

[0034] The recommendation list generation module is used to generate a list of key alumni recommendations by combining the honor level, job level and alumni status of the individuals.

[0035] Preferably, the structured information extraction module includes a prompt word construction submodule, a model invocation submodule, a result parsing submodule, and a batch processing submodule. The prompt word construction submodule generates extraction prompt words containing field rules and output formats. The model invocation submodule invokes the large model based on the extraction prompt words and obtains the returned results. The result parsing submodule parses the returned structured data. The batch processing submodule controls the number of concurrent processes and the batch processing flow.

[0036] Compared with existing technologies, this key alumni recommendation and matching method and system has the following advantages:

[0037] I. This invention constructs a genealogical database containing time nodes of institution name changes, mergers, and splits. It adds a linkage verification step of enrollment and graduation time to the standardized matching process of institution names. This can complete the timeline consistency verification of institution names in different historical periods, identify and correct time conflicts in static alias matching, thereby reducing mismatches caused by changes in the historical names of institutions and improving the accuracy of alumni identity determination.

[0038] Second, this invention establishes a complete automated processing flow from list collection, retrieval and collection, deduplication and cleaning to structured extraction and institution matching. Combined with confidence weighting rules for different information sources, it can automatically process multi-source heterogeneous public information, replace manual screening and information extraction of massive amounts of data, and distinguish the corroborating power according to the authority of the source, so that the credibility level of the output results is clear, which facilitates subsequent manual review and graded application.

[0039] Third, this invention establishes a two-level judgment mechanism with a preset matching threshold and a preset ambiguity threshold. A successful match is determined only after the comprehensive matching score meets the standard and the time verification is passed. At the same time, ambiguous results with similar scores trigger a secondary verification based on regional and industry information, thereby forming a dual quality assurance at the end of the matching process and effectively reducing the false matching rate. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of the overall process of the key alumni recommendation and matching method of the present invention;

[0041] Figure 2 This is a flowchart illustrating the standardized matching and time verification process for institution names in this invention.

[0042] Figure 3 This is a schematic diagram of the process for extracting structured information from the AI ​​large model of the present invention;

[0043] Figure 4 This is a schematic diagram of the key alumni recommendation and matching system of the present invention. Detailed Implementation

[0044] To further illustrate the technical means and effects of the present invention in achieving its intended purpose, the following detailed description of the specific implementation methods, structures, features, and effects of the present invention, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0045] This embodiment discloses a key alumni recommendation and matching method, such as... Figure 1 As shown, the entire process can automatically complete the crawling, text cleaning, information extraction, and institution matching operations of various publicly available materials. The solution incorporates two sets of processing logics: verification of institution history dates and weighted evaluation of the credibility of information from multiple sources. This reduces errors in alumni identity verification and makes the matching results more reliable. Industry professionals familiar with data crawling and large-scale model invocation can fully reproduce the entire processing flow by following all the operational details recorded in this section, including the following steps:

[0046] S1: Collect publicly available lists of various honors, awards, and talent announcements to form an initial list of candidates to be collected, including the names of individuals, information sources, and list categories.

[0047] In some examples of this invention, publicly available list information is collected from channels such as those of science and technology management departments at all levels, academic organizations, industry regulatory bodies, and professional awards. The list types cover science and technology award announcements, talent program selection lists, academic honor lists, and industry award winner lists. The data collection process complies with relevant laws and regulations and does not involve any infringement of personal privacy or data security.

[0048] During the collection process, corresponding parsing methods were used to extract fields for tabular public notices, plain text public notices, and webpage-format lists. For each list record, three core fields were uniformly extracted: the person's name, the source of information, and the list category, which were then compiled into an initial list of lists to be collected.

[0049] During the list compilation phase, preliminary formatting was performed, including removing leading and trailing spaces and invisible characters from names, standardizing the format of employer names to remove differences between full-width and half-width characters, and recording the link to the original public announcement page and the publication date as the information source. Records of the same person appearing repeatedly on the same list were initially merged. For records where the same person appeared multiple times on the same list, the record with the most complete information was retained, and the honor category information for the remaining entries was supplemented to reduce redundant data in subsequent processing.

[0050] S2: Combine the names and employers of individuals in the initial list of entities to be collected with alumni information (words such as alumni, graduated from, and graduating institution) to form multiple sets of search keywords. Then, perform batch searches and save the corresponding web page links and original text as the initial dataset.

[0051] In some examples of this invention, after the initial list is compiled, multiple sets of search keywords are constructed for each record in the list. Based on the person's name and employer's name as search elements, these keywords are combined with three categories of terms: alumni, graduated from, and graduating institution, generating three independent sets of search keywords. This covers publicly available reports with different expression habits, improving the recall scope of the target information. For example, for the person "Zhang X", search expressions such as "Zhang X alumni", "Zhang X graduated from", and "Zhang X graduating institution" are generated; for "Zhang X" and his employer "XX Company", search expressions such as "XX Company alumni", "XX Company graduated from", and "XX Company graduating institution" are generated.

[0052] In some examples of this invention, a batch search is performed by calling a search engine application programming interface (API). After obtaining the search results, they are sorted from highest to lowest according to the matching degree returned by the search engine. A preset number of web pages at the top of the sorting are selected as candidate information sources, and the link address, page title, and original HTML text of each web page are saved. A reasonable request interval is set during the search process to avoid triggering access restrictions on the target site. Differentiated collection methods are adopted for different page types. For static content web pages, a static page request tool is used to directly crawl the page source code, resulting in faster collection speed and lower resource consumption. For web pages that use JavaScript to dynamically load content, a dynamic rendering framework is used to complete the full rendering of the page, and the text content is extracted after the page elements have finished loading, solving the problem that dynamically loaded content cannot be directly crawled. The collected raw web page data is uniformly stored in a database and associated with the basic information of the corresponding individuals.

[0053] S3: Based on the initial dataset, duplicate data is removed by judging the similarity of records. The complete data of the details page is crawled for the deduplicated information source to build a deduplicated dataset.

[0054] In some examples of this invention, record similarity is calculated using three dimensions: name, employer, and source of information, to remove duplicate data and merge records from the same source. After data collection, multi-dimensional deduplication is performed on all webpage records to avoid duplicate data of the same person occupying subsequent processing resources. The deduplication calculation is carried out around the three dimensions of name, employer, and source of information, and the degree of duplication between two records is quantified by a similarity function.

[0055] In some examples of this invention, the expression for the similarity function is:

[0056]

[0057] In the formula, To record similarity, the physical meaning of recording similarity is to record... With records The similarity calculation value; This is an indicator function; it takes the value 1 when the condition inside the parentheses is true and 0 when it is false. , These are the names of the people corresponding to the two records; , These are the employers of the individuals corresponding to the two records; , These are the information sources corresponding to the two records; , , The weighting coefficients are for the name dimension, employer dimension, and information source dimension, respectively, and the sum of the three weighting coefficients is 1.

[0058] In some examples of this invention, weights are assigned based on the ability of each dimension to distinguish a person's identity. The name, as the core identifier of a person's identity, has the highest distinguishability and corresponds to a high weight. Set to 0.5; the employer is a secondary identification field, which can effectively distinguish between different people with the same name, and its corresponding weight is... Set to 0.3; the information source is used to distinguish repeatedly collected data from the same source, with corresponding weights. The similarity threshold is set to 0.2. The preset deduplication threshold is 0.8. When the similarity calculated between two records exceeds this threshold, they are considered duplicate data and are merged. Deduplication is performed using a group comparison method to improve efficiency. All records are first grouped by name, and similarity calculations are only performed within groups with the same name, eliminating the need for pairwise comparisons of all data. When merging duplicate records, the record with the most complete text content is retained as the master record. The source links and publication times of the remaining records are added to the source list of the master record, preserving the complete information traceability path.

[0059] In some examples of this invention, after deduplication, supplementary data collection is performed on the remaining valid information sources for the detail page. Static pages use a direct request method to retrieve complete detail content, which can be achieved using Requests libraries (HTTPX, Urllib, etc.) to collect text content. Dynamically rendered pages use a page rendering method to obtain complete text, which can be achieved using the Pyppeteer framework (Selenium, Playwright, etc.) to perform page rendering followed by text content collection. The supplementary data collection process includes a two-retry mechanism: pages that fail to be retrieved on the first attempt are re-requested after a preset interval; records that fail both times are marked as failed and written to the log, without affecting subsequent processing of other data.

[0060] S4: Perform standardization cleaning on the original text of the web pages collected from the deduplication dataset to obtain valid text.

[0061] In some examples of this invention, the collected original text of the webpage is subjected to standardized cleaning to remove irrelevant text blocks, filter text paragraphs that do not contain school names, and clean up HTML tags and redundant whitespace to obtain valid text.

[0062] After the detailed data collection was completed, multiple rounds of standardized cleaning were performed on the original webpage text to remove page noise and retain valid text related to the person's information, thereby improving the accuracy of subsequent information extraction. The first round of cleaning performed page structure noise reduction, removing script tags, style tags, navigation bar elements, advertising elements, and copyright statement elements from the original HTML text, and extracting the plain text content of the main page area. The extracted plain text was then formatted, merging consecutive whitespace characters, standardizing line break formats, and removing meaningless special symbols to obtain the formatted initial text. The second round of cleaning performed content validity screening, splitting the initial text into natural paragraphs and verifying whether each paragraph contained any institution-related names. A basic dictionary of institution names was pre-loaded, containing the full names, abbreviations, and common alternative names of various institutions. Paragraphs that did not match any institution names were directly discarded, retaining only paragraphs containing institution-related statements. This step can filter out a large amount of content unrelated to the person's educational experience, reducing the amount of data processed by the subsequent model. The third round of text cleaning involves target person verification. Each retained paragraph is matched against a specific person, using precise name matching combined with contextual information to eliminate paragraphs explicitly pointing to other individuals with the same name. The judgment rule is that paragraphs mentioning the person's affiliation or research field that significantly conflict with known information about the person being processed are considered non-target content and are directly removed. The text obtained after these three rounds of cleaning is considered valid and used for subsequent structured information extraction.

[0063] S5: Input the effective text into the feature extraction model to extract structured data.

[0064] The cleaned, valid text is input into a feature extraction model to extract structured data containing basic personal information, honors and awards, work experience, and educational background.

[0065] In some examples of this invention, after obtaining valid text, it is input into a feature extraction model. The feature extraction model uses a large AI model to automatically extract the full structured information of the person. For example... Figure 3 As shown, a standardized prompt template is pre-constructed before extraction. The template consists of four parts: the first part is the role definition, which clarifies the model's task as a structured extraction tool for character information; the second part is the core extraction instructions, which require the model to strictly adhere to the original text, not to supplement information not mentioned in the original text, not to infer missing content on its own, and to leave fields with missing information blank; the third part is the field definition rules, which clarifies the meaning, extraction criteria, and format requirements of each output field; and the fourth part is the output format requirements, which specifies that the model should output the results in standard JSON format and should not add any additional explanatory text.

[0066] In some examples of this invention, a large language model that supports structured output is selected, the model temperature parameter is set to 0, and the thinking mode is turned off. Setting the temperature parameter to 0 can keep the model output results stable, and the output results of multiple calls to the same text are consistent, which is suitable for the deterministic requirements of structured extraction; turning off the thinking mode can reduce redundant content in the model output and improve processing speed.

[0067] In some examples of this invention, information extraction employs an asynchronous concurrent batch processing method, implemented based on an asynchronous input / output framework. A semaphore mechanism controls the number of concurrent model calls, preventing excessive concurrency from triggering interface rate limiting. Data is split into preset batch sizes and submitted for processing in batches. The amount of data processed per batch can be adjusted according to available computing resources to balance processing speed and system stability. A robust exception handling mechanism is implemented to uniformly capture exceptions such as model call timeouts, incorrect return formats, and missing fields. Exception records are written to the runtime log, indicating the exception type and corresponding data identifier. A preset timeout is set for each model call; after a timeout, retry is automatically performed. Records of failed retryes are marked as processing exceptions for subsequent manual review.

[0068] In some examples of this invention, the model ultimately outputs structured data in JSON format, containing four core fields. The first category is basic personal information, covering name, gender, date of birth, place of birth, nationality, ancestral home, ethnicity, political affiliation, and highest level of education. The second category is honors and awards information, stored in an array structure, with each element containing the award name and date of award. The third category is work experience information, stored in an array structure, with each element containing the employer name, position, date of employment, and date of resignation. The fourth category is educational experience information, stored in an array structure, with each element containing the school name, year of enrollment, year of graduation, major, and degree.

[0069] All fields are optional. Information not mentioned in the original text will be output as empty values ​​or empty arrays. You may not infer or supplement it yourself.

[0070] S6: Standardize the matching of school names in the extracted structured data, perform time verification in conjunction with the timeline of school history, output the standard school names, complete the alumni identity determination, and obtain the school matching results.

[0071] S601: Pre-build a genealogy database of institutional history, storing the standard name, alias, former name of each institution, as well as the time nodes corresponding to institution name changes, mergers, and splits.

[0072] In some examples of this invention, after extracting structured information, a standardized matching process is performed on all extracted school names to uniformly map school names with different expressions to a standard school name database. Simultaneously, a time verification is performed based on the school's historical timeline to correct matching results with conflicting timelines. For example... Figure 2 As shown, a database of institutional history is pre-built before matching. Each institutional record in the database contains three basic types of information: standard name, alias set, and former name set. It also records the activation and deactivation times for each former name, as well as the time points corresponding to institutional name changes, mergers, and splits. For institutions with a merger history, the database records the former names and corresponding time intervals of each institution before the merger, and the activation time of the new institution's name after the merger. For institutions with a split history, the database records the name correspondence and time points before and after the split.

[0073] In some examples of this invention, traditional institution name matching uses a static alias mapping method, which only establishes a lookup table of "alias → standard name" without considering the time factor. When encountering a scenario where "someone graduated from a certain school in a certain year," the static mapping cannot determine whether the school name for that year is correct, inevitably leading to timeline conflicts. Therefore, a time dimension is introduced into institution name matching. By constructing a genealogical database containing the time nodes of institution name changes, mergers, and splits, the traditional one-dimensional name mapping (alias → standard name) is expanded into a two-dimensional spatiotemporal mapping (alias × time interval → standard name), upgrading institution name matching from static equivalence judgment to dynamic time consistency verification. The genealogical database not only records the standard name and alias of the institution but also marks the activation and deactivation times for each former name, distinguishing between different evolution types such as name changes, mergers, and splits. Based on this, the year of enrollment or graduation of a person is used as a mandatory verification dimension and compared with the evolution time nodes of candidate institutions—if the time of enrollment of a person is inconsistent with the time validity of the candidate school name, the matching result is automatically switched to the correct school name of the corresponding historical period. This mechanism enables the present invention to identify and correct timeline conflicts that cannot be found in static alias matching, fundamentally solving the problem of mismatch caused by changes in the historical names of institutions.

[0074] By constructing a historical genealogy database containing timelines of institutional name changes, mergers, and splits, a time-dimensional verification mechanism is introduced into the standardized institutional name matching process, upgrading the traditional static alias mapping to a two-dimensional spatiotemporal mapping. Therefore, it can identify and correct timeline conflicts caused by historical name changes of institutions, avoiding mismatches of early graduates to non-existent institution names. It can handle the attribution determination of individuals in complex historical scenarios such as institutional mergers and splits, ensuring that individuals who studied at the same institution with different predecessors at different historical periods receive correct matching results. By combining the individual's enrollment year with the institutional historical timeline for dual verification, it effectively filters out candidate institutions that are chronologically impossible, reducing matching ambiguities caused by institutions with the same name but different histories.

[0075] S602: Calculate the comprehensive matching score between the school name to be matched and each standard institution name, and select the standard institution name with the highest score as the candidate matching result.

[0076] In some examples of this invention, the matching of school names adopts a comprehensive matching score calculation method, which combines the matching results of four dimensions: edit distance, semantic vector similarity, hierarchical name matching, and core field matching, and then weights them to obtain the final score. The formula for calculating the comprehensive matching score is as follows:

[0077]

[0078] In the formula, School names to be matched With standard institution name The overall matching score; This is the edit distance between the school name to be matched and the standard school name, which is the minimum number of single-character operations required to convert between the two strings. The length of the school name to be matched. The character length of the standard institution name. The semantic vector similarity between the school name to be matched and the standard school name is obtained by calculating the cosine similarity after generating text vectors through a pre-trained language model (BERT model or Word2Vec model), with a value ranging from 0 to 1. To determine the matching degree of the level names, For the core field matching degree, in some examples of this invention, the value is 1.0 when the school name to be matched belongs to the official former name of a standard institution, 0.8 when it belongs to the standardized abbreviation, 0.5 when it belongs to the unofficial alternative name, and 0 when no alternative name is found in the database. To calculate the core field matching score, the matching degree is calculated after extracting the main core fields of the school name to be matched and the standard school name. The value is 1.0 when the core field is completely matched, 0.5 when it is partially matched, and 0 when it is not matched. , , , These are the weight coefficients for the edit distance similarity term, semantic vector similarity term, category name matching term, and core field matching term, respectively, satisfying... .

[0079] In some examples of this invention, weight values ​​are set based on the matching reliability of each dimension. Semantic vector similarity can identify abbreviations and common names with significant differences in expression, and has the strongest matching generalization ability, corresponding to the weight... Set to 0.3. Edit distance can handle text input errors and differences in expression, corresponding to the weight. Set to 0.25. Tiered name matching, based on known mapping relationships, has high accuracy and distinguishes the confidence levels of different aliases, with corresponding weights. Set to 0.25. The core field matching aligns with the naming conventions of educational institutions, effectively filtering out distracting items with the same suffix but different entities, corresponding to the weight. Set to 0.2. During matching, calculate the comprehensive matching score between the school name to be matched and all names in the standard school database, and select the standard school name with the highest score as the candidate matching result. The preset matching threshold is set to 0.8. When the comprehensive matching score of the candidate result is greater than the preset matching threshold, it enters the time verification stage. If the score is lower than the threshold, it is determined that there is no matching school and it is not included in the alumni determination scope.

[0080] S603: Extract the year of enrollment or graduation of the corresponding person and compare it with the timeline of the candidate school. If the person's enrollment time is earlier than the time of the school's name change and the candidate matching result is the new name of the school, then switch the matching result to the former name of the corresponding period and retain the standard school association relationship.

[0081] In some examples of this invention, the time verification step combines the individual's enrollment date with the institution's historical timeline for consistency checks. The enrollment year corresponding to the current individual's educational experience is extracted and compared one by one with the historical timelines of candidate institutions. If the individual's enrollment date is earlier than the institution's name change date, and the candidate matching result is the institution's new name after the name change, the matching result is switched to the former name of the corresponding period, while preserving the association between the former name and the standard institution name to ensure accurate alumni attribution. If the individual's enrollment date is later than the institution's name change date, and the candidate matching result is the institution's former name, the matching result is switched to the current standard name. If the individual's enrollment date is before the institution merger date, the matching is performed to the corresponding historical institution name, synchronously associating it with the merged standard institution. When the time verification passes and the overall matching score meets the standard, the institution match is ultimately determined to be successful.

[0082] S604: When the overall matching score is greater than the preset matching threshold and the time verification is passed, the school matching is considered successful.

[0083] In some examples of this invention, when the difference in the comprehensive matching score between the top two standard university names is less than a preset ambiguity threshold, the corresponding person's regional and industry information is retrieved, and the matching degree is calculated with the location information and advantageous discipline information of the two candidate universities, respectively. The university with the higher matching degree is selected as the final matching result. That is, the corresponding person's regional and industry information is retrieved, and comprehensive matching is performed with the location information and advantageous discipline information of the two candidate universities, respectively. Where the person's regional information matches the university's location, the matching degree is increased; where the person's industry information matches the university's advantageous discipline, the matching degree is increased; and the university with the higher comprehensive matching degree is selected as the final matching result.

[0084] The matching process incorporates a secondary ambiguity check mechanism. When the difference in the overall matching score between the top two candidate universities is less than a preset ambiguity threshold (10% of the highest matching score), a secondary check is triggered. This secondary check retrieves the individual's geographic and industry information and calculates the match degree with the geographic and academic strengths of the two candidate universities. Bonus points are awarded for matching the individual's work location with the university's location and for matching the individual's research field with the university's academic strengths. The university with the higher match degree is selected as the final matching result.

[0085] In some examples of this invention, after the matching of educational institutions is completed, the overall confidence level of alumni identity is calculated for all successfully matched individuals, and a final recommendation list is generated by combining the level of honor and job title. Sources are pre-categorized according to their level of authority and corresponding confidence weights are assigned. Official announcements have the highest authority and corresponding confidence weights. Information from official websites of organizations has the next highest credibility and corresponding weights. Information from commercial news sources has relatively low credibility and corresponding weights. The matching results of all supporting sources for the same individual are statistically analyzed, and a weighted sum is calculated according to the corresponding source weights to obtain the overall confidence level of the individual's alumni identity. The confidence level ranges from 0 to 1.

[0086] In some examples of this invention, after the overall confidence level is calculated, all individuals are ranked based on their honor level and job level. The ranking prioritizes the overall confidence level; higher confidence levels result in higher rankings. When confidence levels are similar, individuals with higher honor levels and higher job levels are ranked higher.

[0087] S7: By combining the results of matching institutions with the structured data of individuals' honor levels and job titles, a list of recommended key alumni is generated.

[0088] After sorting, a final list of key alumni recommendations is generated, which includes the alumni's name, the name of their institution, a summary of their educational background, a summary of their honors, a comprehensive confidence level, and a list of information sources.

[0089] In some examples of this invention, sources are pre-categorized according to their authority level and corresponding confidence weights are assigned. Sources with official announcements have a higher weight than those with official organization websites, and those with official organization websites have a higher weight than those with commercial news. The matching results from various supporting sources for the same person are weighted and summed to obtain the overall confidence level of the alumni identity. A list of recommended key alumni is then output, sorted by overall confidence level and honor level.

[0090] In another embodiment of the present invention, a key alumni recommendation and matching system is provided to be applicable to the above-described key alumni recommendation and matching method, such as... Figure 4 As shown, it includes a data collection module, a full-network search and collection module, a data deduplication and supplementary collection module, a data cleaning and noise reduction module, a structured information extraction module, an alumni identity determination module, and a recommendation list generation module: Among them:

[0091] The data collection module is used to collect various publicly available lists and generate an initial list of lists to be collected, containing the names of individuals, information sources, and list categories. The data collection module is responsible for summarizing various publicly available lists and organizing them to generate the initial list of lists to be collected.

[0092] The comprehensive web search and collection module is used to combine the names and employers of individuals in the initial list of entities to be collected with alumni information to form multiple sets of search keywords. Then, it performs batch searches and saves the corresponding webpage links and original text as the initial dataset. The comprehensive web search and collection module is responsible for generating search keywords, calling search engine interfaces to crawl webpage data, and differentiating between static and dynamic pages using different collection methods.

[0093] The data deduplication and re-collection module is used to calculate record similarity based on the initial dataset using three dimensions: name, employer, and information source. It removes duplicate data and merges records from the same source. For each deduplicated information source, it retrieves complete data from the details page to construct a deduplicated dataset. The module is responsible for deduplicating records according to multi-dimensional similarity rules and re-collecting details page data for each deduplicated information source.

[0094] The data cleaning and denoising module is used to perform standardized cleaning on the original webpage text collected in the deduplication dataset. It removes irrelevant text blocks, filters text paragraphs that do not contain school names, and cleans HTML tags and redundant whitespace to obtain valid text. The data cleaning and denoising module is responsible for performing multiple rounds of cleaning on the original webpage text, outputting valid text that can be used for extraction.

[0095] The structured information extraction module is used to input valid text into the feature extraction model and extract structured data containing basic personal information, honors and awards, work experience, and educational background. The structured information extraction module includes four sub-modules: prompt word construction, model invocation, result parsing, and batch processing, responsible for extracting structured information from valid text.

[0096] In some examples of this invention, the structured information extraction module includes a prompt word construction submodule, a model invocation submodule, a result parsing submodule, and a batch processing submodule. The prompt word construction submodule generates extraction prompt words containing field rules and output formats. The model invocation submodule invokes a large model based on the extraction prompt words and obtains the returned results. The result parsing submodule parses the returned structured data. The batch processing submodule controls the number of concurrent processes and the batch processing flow.

[0097] The alumni identity verification module is used to perform standardized matching on the extracted school names, and after performing time verification in conjunction with the school's historical timeline, outputs the standard school name to complete the alumni identity verification. The alumni identity verification module has built-in score calculation submodule and school history time verification submodule, which are responsible for completing the comprehensive matching and time verification of school names and outputting the final matching result.

[0098] The recommendation list generation module is used to generate a list of key alumni recommendations by combining the individual's honor level, job level, and alumni status.

[0099] In some examples of the present invention, a database storage module is also included, which is responsible for storing the original data, intermediate data and final processing results of each stage.

[0100] The modules are connected sequentially according to the aforementioned method and process, and data is transferred between modules in the order of processing to complete the automated processing of the entire process.

[0101] This invention establishes a fully automated processing chain from list collection, retrieval and collection, deduplication and cleaning to structured extraction and institution matching. It introduces an institution evolution time verification mechanism and a multi-source information confidence weighting rule to solve the problems of low efficiency in traditional manual processing, timeline conflicts easily occurring in static name matching, and inability to distinguish the validity of information from different sources. It achieves automated identification and accurate matching of key alumni in massive amounts of public information. While improving processing efficiency, it ensures the accuracy and credibility of alumni identity determination. This invention can be applied to scenarios such as university alumni management, talent resource management, scientific research cooperation, and think tank resource construction.

[0102] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for recommending and matching key alumni, characterized in that, Includes the following steps: S1: Collect publicly available lists of various honors, awards, and talent announcements to form an initial list of candidates to be collected, including the names of individuals, information sources, and list categories. S2: Combine the names and employers of the individuals in the initial list of individuals to be collected with alumni information to form multiple sets of search keywords, then perform batch search, and save the corresponding web page links and original text as the initial dataset; S3: Based on the initial dataset, duplicate data is removed by judging the similarity of records. Then, the complete data of the details page is crawled from the deduplicated information source to construct a deduplicated dataset. S4: Perform standardization cleaning on the original text of the web pages collected from the deduplication dataset to obtain valid text; S5: Input the effective text into the feature extraction model to extract structured data; S6: Standardize the matching of school names in the extracted structured data, perform time verification in conjunction with the timeline of school evolution, output the standard school names, complete the alumni identity determination, and obtain the school matching results; S7: By combining the results of matching institutions with the structured data of individuals' honor levels and job titles, a list of recommended key alumni is generated.

2. The key alumni recommendation and matching method according to claim 1, characterized in that, In step S3, the record similarity is calculated using a similarity function. When the calculated record similarity value is greater than a preset deduplication threshold, the corresponding two records are determined to be duplicate data and are merged. The expression for the similarity function is: In the formula, To record similarity, the physical meaning of recording similarity is to record... With records The similarity calculation value; This is an indicator function; it takes the value 1 when the condition inside the parentheses is true and 0 when it is false. , These are the names of the people corresponding to the two records; , These are the employers of the individuals corresponding to the two records; , These are the information sources corresponding to the two records; , , The weighting coefficients for the name dimension, employer dimension, and information source dimension are respectively, satisfying... .

3. The key alumni recommendation and matching method according to claim 2, characterized in that, The standardized matching in step S6 uses a comprehensive matching score calculation. The name of the standard institution with the highest comprehensive matching score is selected as the candidate matching result. When the comprehensive matching score is greater than the preset matching threshold and the time verification passes, the institution is considered to have been successfully matched. The formula for calculating the comprehensive matching score is: In the formula, School names to be matched With standard institution name The overall matching score; The edit distance between the school name to be matched and the standard school name; The length of the school name to be matched; The character length of a standard institution's name; This represents the semantic vector similarity between the school name to be matched and the standard school name; For the matching degree of the level name; For core field matching degree; , , , These are the weight coefficients for the edit distance similarity term, semantic vector similarity term, category name matching term, and core field matching term, respectively, satisfying... .

4. The key alumni recommendation and matching method according to claim 3, characterized in that, In step S2, when saving the corresponding webpage link and original text, for webpages dynamically loaded with JavaScript, the text content is collected after page rendering; for static webpages, the text content is collected using a direct HTTP request method.

5. The key alumni recommendation and matching method according to claim 4, characterized in that, In step S5, the temperature parameter of the feature extraction model is set to 0, and the thinking mode is turned off; the structured information extraction process is implemented using an asynchronous concurrent framework, which controls the preset number of concurrent processes through semaphores and processes data in batches according to the preset batch size; before extraction, a prompt word template containing role definitions, core extraction instructions, field definition rules, and output format requirements is constructed, and the prompt words are concatenated with valid text and then input into the large model; during the processing, call exceptions are logged, and a preset timeout is set for each model call.

6. The key alumni recommendation and matching method according to claim 5, characterized in that, The specific process of school matching and time verification in step S6 is as follows: S601: Pre-build a genealogy database of institutional history, storing the standard name, alias, former name of each institution, as well as the time nodes corresponding to institution name changes, mergers, and splits; S602: Calculate the comprehensive matching score between the school name to be matched and each standard institution name, and select the standard institution name with the highest score as the candidate matching result; S603: Extract the year of enrollment or graduation of the corresponding person and compare it with the timeline of the candidate school. If the person's enrollment time is earlier than the time of the school's name change and the candidate matching result is the new name of the school, then switch the matching result to the former name of the corresponding period and retain the standard school association relationship. S604: When the overall matching score is greater than the preset matching threshold and the time verification is passed, the school matching is considered successful.

7. The key alumni recommendation and matching method according to claim 6, characterized in that, In step S7, the sources are pre-categorized according to their authority level and corresponding confidence weights are set. The weight of official announcements is higher than that of official websites, and the weight of official websites is higher than that of business news. The matching results of the various supporting sources for the same person are weighted and summed according to their corresponding weights to obtain the comprehensive confidence level of the alumni identity. The list of key alumni recommendations is sorted according to the comprehensive confidence level and the level of honor and then output.

8. The key alumni recommendation and matching method according to claim 7, characterized in that, In step S6, when the difference in the comprehensive matching scores of the top two standard college names is less than the preset ambiguity threshold, the corresponding person's regional and industry information is retrieved, and the matching degree is calculated with the location information and advantageous discipline information of the two candidate colleges respectively. The college with the higher matching degree is selected as the final matching result.

9. A system for executing the key alumni recommendation and matching method according to any one of claims 1-8, characterized in that, include: The data collection module is used to collect various public lists and generate an initial list of lists to be collected, which includes the names of the people, the source of the information, and the category of the list. The whole network search and collection module is used to combine the names and employers of people in the initial list of people to be collected with alumni information to form multiple sets of search keywords, and then perform batch search, saving the corresponding web page links and original text as the initial dataset; The data deduplication and re-collection module is used to calculate the similarity of records based on the initial dataset through three dimensions: name, employer, and information source. It removes duplicate data and merges records from the same source. It also retrieves complete data from the details page of the deduplicated information source to construct a deduplicated dataset. The data cleaning and noise reduction module is used to perform standardized cleaning on the original text of the web pages collected in the deduplication dataset, remove irrelevant text blocks, filter text paragraphs that do not contain school names, clean HTML tags and redundant whitespace characters, and obtain valid text. The structured information extraction module is used to input effective text into the feature extraction model and extract structured data containing basic personal information, honors and awards, work experience, and educational experience. The alumni identity verification module is used to perform standardized matching on the extracted school names, perform time verification in conjunction with the timeline of the school's history, and output the standard school name to complete the alumni identity verification. The recommendation list generation module is used to generate a list of key alumni recommendations by combining the honor level, job level and alumni status of the individuals.

10. The system according to claim 9, characterized in that, The structured information extraction module includes a prompt word construction submodule, a model calling submodule, a result parsing submodule, and a batch processing submodule; the prompt word construction submodule is used to generate extraction prompt words that include field rules and output formats; The model invocation submodule is used to invoke the large model based on the extracted prompt words and obtain the returned results; the result parsing submodule is used to parse the returned structured data; The batch processing submodule is used to control the number of concurrent processes and the batch processing flow.

Images