Intelligent matching and recommendation system for supply and demand of school-enterprise cooperation based on large model

Abstract

This invention belongs to the field of supply and demand management technology. It discloses a large-scale model-based intelligent matching and recommendation system for university-enterprise cooperation supply and demand. The system includes: collecting multi-dimensional data from a university-enterprise cooperation management platform and standardizing it to obtain a business dataset; performing entity disambiguation and discrete merging to output a unified set of entity profiles; performing semantic consistency judgment and demand semantic correction to obtain standard demand information; performing cold start compensation calculation to output a candidate entity scoring table; constructing positional decay weights and performing semantic encoding to output a sequence-sensitive feature set; calculating the dimensional matching value between the demand object and the candidate cooperation entity; performing comprehensive matching based on the dimensional matching value to output a comprehensive matching score set; ranking the candidate cooperation entities based on the comprehensive matching score set to form a cooperation recommendation sequence, and sending the cooperation recommendation sequence to the demand processing end. This improves the accuracy of supply and demand matching and the rationality of the recommendation results.

Description

Technical Field

[0001] This invention relates to the field of supply and demand management technology, and more specifically, to a large-scale model-based intelligent matching and recommendation system for supply and demand in university-enterprise cooperation. Background Technology

[0002] Intelligent matching technology driven by large models is a technical means to identify resource objects and recommend candidate results by associating and processing multi-source business data. It has broad application prospects in education management, resource scheduling and business recommendation, especially in the scenario of school-enterprise cooperation management, where it has the potential to achieve unified governance of cooperative units and accurate identification of needs. However, traditional matching systems still face many technical bottlenecks when facing various complex business scenarios.

[0003] In real-world scenarios, different colleges can enter information about collaborating units separately. The same collaborating unit is often entered repeatedly due to differences in abbreviations, full names, regional suffixes, or branch names. Existing traditional matching systems typically rely solely on literal deduplication of unit names or verification using the unified social credit code, lacking the ability to combine large-scale model semantics, historical collaboration content, and subject fields for entity disambiguation. This leads to fragmented collaborating unit profiles and distorted historical collaboration counts, thus affecting subsequent recommendation results. Furthermore, the titles and descriptions of demand postings often exhibit inconsistencies in theme, target audience, or missing supplementary requirements. Traditional matching systems typically treat the titles and descriptions as plain text for direct matching, lacking analysis of the text's collaboration intent and conflict resolution. The lack of recognition and semantic correction capabilities leads to non-standard input of requirements, resulting in chaotic matching benchmarks. Meanwhile, newly added partners, lacking historical cooperation records, often struggle to obtain reasonable initial scores in existing recommendation rankings. Traditional matching systems also lack a cold-start compensation mechanism based on the historical performance of similar active partners. Furthermore, the core technical directions in college and partner profiles are typically entered using multiple-selection fields. Traditional matching systems usually encode each technical direction option with equal weight, failing to identify the implicit hierarchical relationship between primary and secondary technical directions in the original selection order. This results in inaccurate matching results for technical directions, making it difficult for university-enterprise cooperation supply and demand recommendations to form accurate and stable recommendations in complex business scenarios.

[0004] In view of this, the present invention proposes a large-scale model-based intelligent matching and recommendation system for supply and demand in university-enterprise cooperation to solve the above problems. Summary of the Invention

[0005] To overcome the aforementioned shortcomings of existing technologies and to achieve the above objectives, this invention provides the following technical solution: a large-scale model-based intelligent matching and recommendation system for supply and demand in university-enterprise cooperation, comprising: The data acquisition module collects multidimensional data from the university-enterprise cooperation management platform and performs standardized processing to obtain a business dataset; the business dataset includes college profiles, partner unit profiles, demand release information, and historical cooperation project information. The entity disambiguation module performs entity disambiguation based on the profiles of collaborating units and historical collaborative project information, identifies redundant target collaborating unit information, and discretizes and merges the profiles of the corresponding target collaborating units based on the disambiguation results, outputting a unified set of entity profiles. The semantic determination module uses a large model to perform semantic consistency determination on the demand release information, identify conflicting target demand information, and perform demand semantic correction on the conflicting target demand information to obtain standard demand information. The startup compensation module identifies newly added cooperative units based on a unified unit profile set and historical cooperation project information, obtains the attribute information of the newly added cooperative units, performs cold start compensation calculations in conjunction with the unified unit profile set, and outputs a candidate unit scoring table. The feature encoding module extracts the core technology direction fields from the college profile and unified unit profile sets, identifies the original order of technology direction options, constructs positional decay weights based on the original order of technology direction options and performs semantic encoding, and outputs an order-sensitive feature set. The dimension matching module calculates the dimension matching value between the demand object and the candidate cooperation unit based on standard demand information, candidate unit scoring table and order-sensitive feature set; it performs comprehensive matching based on the dimension matching value and outputs a comprehensive matching score set. The sequence construction module sorts candidate partners based on the comprehensive matching score set to form a cooperation recommendation sequence, and sends the cooperation recommendation sequence to the demand processing end; the modules are connected to each other via wired and / or wireless means.

[0006] Furthermore, methods for disambiguating entity names include: Extract the unit attribute information corresponding to each cooperative unit from the cooperative unit profile to form unit identification features; use a large model to perform word segmentation on the unit name in each unit identification feature, extract the core word elements and modifier word elements in the unit name, and construct the name word element sequence; Calculate the overlap of core words and the difference of modifier words between the name word sequences corresponding to any two cooperating units; calculate name similarity based on the overlap of core words and the difference of modifier words, and select cooperating unit groups with name similarity higher than the preset name similarity threshold to form an initial set of duplicate units; Identify the credit codes of any two collaborating units in the initial set of duplicate units, and identify the organization code field in the credit codes; if the organization code fields of the two collaborating units are the same, mark the group of collaborating units as high-confidence duplicate units; Extract the cooperation project options from the historical cooperation project information of any two cooperating units, calculate the ratio of the number of intersections to the number of unions of the cooperation project options, and obtain the content overlap coefficient; The name similarity and content overlap coefficient are weighted and fused to obtain the comprehensive duplication confidence score of any group of cooperating units; cooperating units with a comprehensive duplication confidence score higher than the preset duplication threshold are selected and merged with high-confidence duplication units as the disambiguation result.

[0007] Furthermore, methods for performing discrete merging include: Identify the initial and latest timestamps in the profile of each collaborating unit after disambiguation, calculate the time interval between the initial and latest timestamps to obtain the profile activity period; take the profile of the collaborating unit with the longest activity period as the master record, and the profiles of the remaining collaborating units as slave records. Extract the unit attribute field from the master record and slave record and compare them field by field; If the primary record contains an empty field and the secondary record contains a non-empty field, the non-empty field in the secondary record will be filled into the corresponding empty field in the primary record; if both the primary and secondary records contain non-empty fields but are different, the latest timestamps of the corresponding fields in the two records will be compared, and the field with the latest timestamp will be retained. Extract the core technology direction field from the master record and slave record, and identify the original position of the option corresponding to each core technology direction field; add the option corresponding to the core technology direction field that exists in the slave record but not in the master record to the end of the core technology direction field of the master record; The project numbers of all high-confidence duplicate units in the disambiguation results are deduplicated and merged. The number of duplicated and merged projects is counted, and the historical collaboration count field of the master record is updated. The merged master record is used as the master unit profile, and the derived records are deleted to obtain a unified set of unit profiles.

[0008] Furthermore, the methods for performing semantic consistency determination include: Extract the requirement title and requirement description from the requirement release information to obtain the title text and description text; The large model is used to perform intent parsing on the title text and description text respectively, to obtain the title cooperation type and title action intent in the title text, and the description cooperation type and description action intent in the description text; If the title's cooperation type and the description's cooperation type do not belong to the same cooperation type category, it is marked as a type conflict; if the title's action intent and the description's action intent correspond to different cooperation object categories, it is marked as an object conflict. The descriptive text is segmented into sentences to identify the cooperation intention of the first sentence and the cooperation intention of the last sentence. If the cooperation type of the title is different from the cooperation intention of the first sentence, it is marked as a deviation of the first sentence; if the cooperation type of the title is different from the cooperation intention of the last sentence, it is marked as a deviation of the last sentence. The total number of conflict tags is obtained by counting all tags. Requests with a total number of conflict tags greater than a preset conflict threshold are marked as conflict target request information, and the corresponding conflict type tags are recorded.

[0009] Furthermore, methods for semantic correction of demand include: Identify the conflict type markers in the conflict target requirement information. If there is a type conflict in the conflict type markers, extract the cooperation type that appears most frequently in the description text as the corrected cooperation type and replace the original cooperation type in the title text. If the conflict type marker has an object conflict, the core cooperation object in the description text is extracted as the corrected cooperation object and added to the end of the requirement title; if the conflict type marker has a first sentence deviation, the first sentence in the requirement description is extracted and semantic intent compression is performed to generate a first sentence intent summary, which replaces the original requirement title; if the conflict type marker has a last sentence deviation, the last sentence in the requirement description is extracted and the supplementary cooperation intent in the last sentence is identified, and the supplementary cooperation intent is attached as a subtitle to the corrected requirement title. The revised requirement titles and requirement descriptions are linked to form standard requirement information.

[0010] Furthermore, the methods for performing cold start compensation calculations include: The number of cooperative projects for each cooperating unit in the unified unit profile set in the historical cooperative project information is counted, and cooperating units with zero cooperative projects are selected as new cooperating units to be added to the database. Extract the attribute fields of newly added partners to form a new partner description group; extract partners from the unified partner profile set whose number of partner projects exceeds a preset activity threshold, mark them as active partners, and extract the attribute fields of each active partner to form an active partner description group. Encode the new unit description group and the active unit description group into a new unit vector and an active unit vector; Calculate the similarity between the new unit vector and each active unit vector, and select the most similar active cooperative units of a predetermined number to form a set of similar active units; Extract the performance indicators of each active collaborator from the historical collaborative project information in the set of similar active units and identify the corresponding performance level; calculate the migration performance benchmark by weighted average of the performance levels of all active collaborators; The migration performance benchmark is used as the cold start compensation value for newly added partners and written into the attribute fields of the newly added partners. The performance levels of all partners are integrated to form a candidate partner scoring table.

[0011] Furthermore, the methods for constructing positional decay weights and performing semantic encoding include: Retrieve all the technology direction options in each core technology direction field and their corresponding order to form the original order of the technology direction options; All technical direction options in the original order are assigned a position number based on their position. The position distance between each position number and the preset benchmark position is calculated. The position attenuation weight of the corresponding technical direction option is then calculated based on the position distance. Construct the semantic vector of each technical direction option, multiply the semantic vector of the technical direction with the corresponding position attenuation weight, and output the attenuation semantic vector of each technical direction option. Arrange the attenuation semantic vectors according to the option position to form an ordered attenuation vector sequence. Summing up all the attenuation semantic vectors in the sequence yields the ordered aggregated vector corresponding to the core technology direction field. By integrating the sequential aggregation vectors corresponding to all core technology fields, a sequence-sensitive feature set is obtained.

[0012] Furthermore, the dimension matching value is calculated as follows: The cooperation type, technology direction tag, geographical location constraint, and effective period information extracted from the standard demand information constitute a demand constraint group; Extract the effectiveness level and cooperation content-related attribute information of each candidate partner from the candidate partner rating table to form a candidate partner attribute group; Obtain the sequential aggregation vectors corresponding to candidate collaborators in the sequence-sensitive feature set, calculate the vector distance between the sequential aggregation vectors of any demand object and the candidate collaborators, and obtain the technology direction matching value; Compare the cooperation type with the cooperation content-related attribute information in the candidate unit attribute group, and count the number of overlapping matching options to obtain the cooperation matching value; Match geographical location constraints with the relevant attribute information of the cooperation content in the candidate unit attribute group to the geographical location of the cooperation, and identify the geographical location matching level as the geographical location matching value; Based on the effective period information and the historical number of collaborations of the corresponding collaborators in the historical cooperation project information, the timeliness response value of the candidate collaborators is calculated. By combining the matching values ​​of technical direction, cooperation, geographical location, and timeliness, a dimensional matching value can be formed between any demand object and a candidate cooperation unit.

[0013] Furthermore, the methods for performing comprehensive matching include: Calculate the product of the dimension matching value and the normalized performance level for each candidate partner to obtain the initial matching value; Identify the ranking of each candidate partner's technology direction matching value among all candidate partners, and mark the candidate partners whose ranking is among the top preset number as technology priority partners; A preset gain coefficient is added to the initial matching value of the technology priority unit to obtain the gain matching value, while the initial matching value of the non-technology priority unit remains unchanged; The gain matching values ​​and initial matching values ​​of all candidate cooperative units are normalized to obtain the comprehensive matching score. The comprehensive matching scores of all candidate cooperative units are then integrated to obtain the comprehensive matching score set.

[0014] Furthermore, the ways to construct cooperative recommendation sequences include: All candidate partners in the comprehensive matching score set are sorted in descending order according to their comprehensive matching scores. The unit identifiers of the candidate partners whose ranking ranks are within a preset priority number are extracted and sorted into a cooperative recommendation sequence.

[0015] The technical effects and advantages of the university-enterprise cooperation supply and demand intelligent matching and recommendation system based on a large model, as described in this invention: By collecting multidimensional data from the university-enterprise cooperation management platform and standardizing it, a business dataset was obtained. Based on this business dataset, unit entity disambiguation and profile merging, demand semantic judgment and correction, cold start compensation calculation, order-sensitive coding, and dimension matching calculation were successively performed, realizing integrated processing of intelligent matching and recommendation of supply and demand in university-enterprise cooperation. Compared with existing technologies, by extracting core and modifier words of unit names and combining them with organizational codes and content overlap coefficients for disambiguation judgment, the fragmentation problem of profiles caused by multiple colleges repeatedly entering the same cooperative unit was optimized. By performing intent parsing and performing classification labeling and differential correction on demand titles and descriptions respectively, the problem of matching benchmark confusion caused by semantic deviation between demand titles and descriptions was solved. By identifying newly added cooperative units and performing migration compensation based on the historical performance of similar active units, the cold start problem of newly added cooperative units being at a long-term ranking disadvantage due to lack of historical records was avoided. By constructing position decay weights and performing order-sensitive coding, the problem of dilution of main technical direction information caused by equal weighting of multiple selection fields was avoided. By achieving comprehensive ranking through multi-dimensional matching and technology priority gain, the accuracy of supply and demand matching and the rationality of recommendation results were improved. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the intelligent matching and recommendation system for supply and demand in university-enterprise cooperation based on a large model, as described in this invention. Detailed Implementation

[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0018] Example 1 Please see Figure 1 As shown in this embodiment, the intelligent matching and recommendation system for supply and demand in university-enterprise cooperation based on a large model includes: The data acquisition module collects multidimensional data from the university-enterprise cooperation management platform and performs standardized processing to obtain a business dataset; the business dataset includes college profiles, partner unit profiles, demand release information, and historical cooperation project information. The entity disambiguation module performs entity disambiguation based on the profiles of collaborating units and historical collaborative project information, identifies redundant target collaborating unit information, and discretizes and merges the profiles of the corresponding target collaborating units based on the disambiguation results, outputting a unified set of entity profiles. The semantic determination module uses a large model to perform semantic consistency determination on the demand release information, identify conflicting target demand information, and perform demand semantic correction on the conflicting target demand information to obtain standard demand information. The startup compensation module identifies newly added cooperative units based on a unified unit profile set and historical cooperation project information, obtains the attribute information of the newly added cooperative units, performs cold start compensation calculations in conjunction with the unified unit profile set, and outputs a candidate unit scoring table. The feature encoding module extracts the core technology direction fields from the college profile and unified unit profile sets, identifies the original order of technology direction options, constructs positional decay weights based on the original order of technology direction options and performs semantic encoding, and outputs an order-sensitive feature set. The dimension matching module calculates the dimension matching value between the demand object and the candidate cooperation unit based on standard demand information, candidate unit scoring table and order-sensitive feature set; it performs comprehensive matching based on the dimension matching value and outputs a comprehensive matching score set. The sequence construction module sorts candidate partners based on the comprehensive matching score set to form a cooperation recommendation sequence, and sends the cooperation recommendation sequence to the demand processing end; the modules are connected to each other via wired and / or wireless means.

[0019] In this embodiment, a business dataset is obtained by collecting multi-dimensional business information from the university-enterprise cooperation management platform and performing filtering, noise reduction, and missing value imputation. The college profile includes attributes related to the college, such as its field of study, core technology direction, available cooperation content, college introduction, number of students, number of faculty and staff, number of full-time teachers, university-enterprise cooperation contact person, and university-enterprise cooperation system and mechanism. The partner unit profile includes information describing the unit's attributes, such as the partner unit name, unified social credit code, unit type, unit nature, unit level, location, unit introduction, available cooperation content, core technology direction, and cooperation capability description. The demand posting information includes, for example, the type of poster, the poster name, and the... The data reflects the content of the cooperation, including the content of the work, the text of the requirements, the relevant field, the technical direction, the location of cooperation, the effective deadline of the requirements, and the contact person information. Historical cooperation project information includes historical data such as the identity of the cooperating unit, historical cooperation content, historical contract association information, historical project name, historical project start and end time, historical project status, historical cooperation results, and historical cooperation project number. In this embodiment, a large model is pre-built, and all technical features of this embodiment are driven by the large model. This large model is a collection of multiple sub-models. For example, the language processing model in the large model is used for processing in the subsequent entity disambiguation module.

[0020] Methods for disambiguating entity entities include: Extract the unit attribute information corresponding to each cooperative unit from the cooperative unit profile to form unit identification features.

[0021] The unit attribute information includes information such as the name of the cooperating unit, the unified social credit code, the unit type, the unit nature, and the unit level. By combining these into unit identification features, various information entered by the same cooperating unit in different colleges are included in the same identification set, providing a data foundation for subsequent operations.

[0022] The large model is used to segment the unit name in the unit identifier features, extract the core words and modifier words in the unit name, and construct the name word sequence.

[0023] The language processing model in the large model is used to segment the unit name in each unit identifier feature. For example, name components that represent the subject, such as “XX”, “technology” and “Limited Company”, are identified as core word elements, and name components that represent the subordinate relationship of the organization, such as “Shenzhen Headquarters” and “Changchun Branch”, are identified as modifier word elements. The name word element sequence is constructed according to the order of appearance of various word elements in the original name.

[0024] Calculate the overlap of core terms and the difference of modifier terms between the name term sequences corresponding to any two cooperating units.

[0025] This involves extracting core word sets and modifier word sets from the name word sequences of any two collaborating units, calculating the proportion of identical word sets in the core word set to the total number of core word sets as the core word set overlap, and calculating the proportion of different word sets in the modifier word set to the total number of modifier word sets as the modifier word set difference. This allows for the differentiation between collaborating unit names that have the same subject but different subordinate descriptions and collaborating unit names that have different subjects.

[0026] Name similarity is calculated based on the overlap of core lexical units and the difference of modifier lexical units. Collaborative unit groups with name similarity higher than the preset name similarity threshold are selected to form an initial set of duplicate units.

[0027] The formula for calculating name similarity is as follows: ;in Indicates the similarity between the names of any two collaborating entities; This indicates the degree of overlap in the core terms of the two collaborating units mentioned above; This indicates the degree of difference in modifiers between the two cooperating units mentioned above; and These represent the weights of core word overlap and modifier word difference, respectively, and the sum of the two weights equals 1. Furthermore, in this embodiment, since core words have a greater impact on the identification of the main body of the cooperating unit's name than modifier words, it is necessary to ensure... The specific values ​​are set based on the specific operating conditions.

[0028] Identify the credit codes of any two cooperating units in the initial screening set of duplicate units, and identify the organization code field in the credit codes.

[0029] The credit code refers to the unified social credit code, and the organization code field refers to the code field in the unified social credit code used to represent the identity of the entity of the organization composed of cooperating units.

[0030] If two collaborating entities have the same organization code field, then the collaborating entities in that group are marked as high-confidence duplicate entities.

[0031] If two collaborating entities have the same organization code field, it means that the probability of two records related to the collaborating entity corresponding to the same organizational entity is significantly higher than that of the case where only name similarity is used for judgment. Therefore, this group of collaborating entities is marked as a high-confidence duplicate entity.

[0032] Extract the cooperation project options from the historical cooperation project information of any two cooperating units, calculate the ratio of the number of intersections to the number of unions of the cooperation project options, and obtain the content overlap coefficient.

[0033] The collaborative project options refer to the project content options in the historical collaborative project information that indicate the direction of cooperation, including collaborative content such as curriculum development or textbook development; the content overlap coefficient refers to the degree of similarity between the two collaborating units in the direction of cooperation in historical collaborative projects.

[0034] By weighting and fusing name similarity and content overlap coefficients, a comprehensive repeatability confidence score for any group of collaborating units can be obtained.

[0035] The comprehensive repetition confidence score of a group of cooperating units is obtained by weighting and summing the name similarity and content overlap coefficients by preset weights. The preset weights are weights set based on historical disambiguation experience to ensure that they correspond to the information contribution of name similarity and content overlap coefficients in historical records.

[0036] Collaborating units with a comprehensive repeat confidence level higher than a preset repeat threshold are selected and merged with high-confidence repeating units as the disambiguation result.

[0037] The specification for record merging sets a duplication threshold. If the overall duplication confidence is higher than the duplication threshold, it means that the names of the cooperating units in the corresponding group are highly similar and the historical cooperation content is highly overlapping. Such cooperating units and the names of the same cooperating units in the high-confidence duplication units are merged as the disambiguation result, which serves as the data basis for subsequent discrete merging.

[0038] Methods for performing discrete merging include: Identify the initial and latest timestamps in the profile of each collaborating unit after disambiguation, calculate the time interval between the initial and latest timestamps, and obtain the profile activity cycle.

[0039] The initial timestamp refers to the time when the profile of a partner unit was first entered, the latest timestamp refers to the time when the profile of a partner unit was last adjusted, and the profile activity period refers to the continuous active time span of the same profile of a partner unit from its initial construction to its most recent adjustment.

[0040] The profile of the partner unit with the longest active profile period is used as the master record, and the profiles of the other partner units are used as slave records.

[0041] In this embodiment, since portraits with long active periods undergo more maintenance and supplementation, their content is more complete and they are more likely to retain information on recent changes in the attributes of collaborating units, such portraits are used as the main records.

[0042] Extract the unit attribute field from the master record and slave record and compare them field by field.

[0043] The unit attribute field refers to the field used to describe the basic attributes and cooperation capabilities of the cooperating unit, such as unit type, unit nature, and unit level.

[0044] If a field in the primary record is empty and a field in the secondary record is not empty, then the non-empty field in the secondary record will be filled into the corresponding empty field in the primary record.

[0045] In the master record, an empty field means that the cooperating unit has not filled in valid information in a certain corresponding field, while a non-empty field in the slave record means that valid information has been filled in a field of the same type as the master record. If there is an empty field in the master record and the corresponding field in the slave record is not empty, the non-empty field in the slave record is used to directly fill the corresponding empty field in the master record.

[0046] If both the primary and secondary records contain non-empty fields but are different, then compare the latest timestamps of the corresponding fields in the two records and retain the field with the latest timestamp.

[0047] When the same field exists in both the master record and the slave record but the field content is different, it indicates that the same collaborating unit has experienced inconsistent updates during its cooperation with different colleges. Therefore, retaining the field value with the newer timestamp makes the profile more reflective of the collaborating unit's attribute information in the current state.

[0048] Extract the core technology direction fields from the master record and slave record, and identify the original position of the options corresponding to each core technology direction field.

[0049] The core technology direction field refers to the multi-select field in the profile of the cooperating unit used to describe the direction of technical capabilities. The original position refers to the selection order of each technical direction option during the entry. In this embodiment, the core technology direction field is extracted from the main record and the slave record respectively, and the order of each technical direction option in the original entry is read as the input basis for subsequent order-sensitive coding. By retaining the original position information, the implicit judgment of different colleges on the main technical direction and auxiliary technical direction of the cooperating unit during the profile merging process can be preserved as much as possible.

[0050] The corresponding options for the core technology direction field that exist in the record but not in the main record will be added to the end of the core technology direction field in the main record.

[0051] If there are core technology direction options in the records that are not included in the main record, it means that the same collaborating unit has supplemented and recorded different technology directions in different college scenarios. Add such different technology direction options to the end of the core technology direction field of the main record without changing the original order of the options in the main record, and preserve the primary and secondary relationship of the original technology directions in the main record.

[0052] The project IDs of all high-confidence duplicate units in the disambiguation results are deduplicated and merged. The number of duplicated and merged projects is counted, and the historical collaboration count field of the master record is updated.

[0053] The project number is a field in the historical project information corresponding to the project record. All project numbers associated with high-confidence duplicate units are merged to avoid the falsely high number of collaborations caused by the same project being counted repeatedly in multiple duplicate collaboration unit records. This ensures the accurate count of the number of valid collaborations and updates the historical collaboration count field in the main record.

[0054] The merged master record is used as the master unit profile, and the derived records are deleted to obtain a unified set of unit profiles.

[0055] The final master record is used as the only retained master unit profile, while the corresponding slave records are deleted to eliminate redundant fragmented records of the same cooperative unit in the platform, so that each cooperative unit corresponds to only one unified profile object, avoiding fragmentation of unit profiles caused by repeated entry.

[0056] The methods for performing semantic consistency determination include: Extract the title and description of the requirements from the requirement release information to obtain the title text and description text.

[0057] The requirement title refers to the text field in the requirement release information that summarizes the content of the cooperation requirement, while the requirement description refers to the text field in the requirement release information that describes the content of the cooperation requirement in detail.

[0058] The large model is used to perform intent parsing on the title text and description text respectively, to obtain the title cooperation type and title action intent in the title text, and the description cooperation type and description action intent in the description text.

[0059] The language processing model in the large model is used to process the language of the title text and the description text. The corresponding tags of cooperation type in the title text and description text, as well as certain words that indicate action intention such as "build", "develop" and "provide", are identified as the title action intention or description action intention.

[0060] If the title's collaboration type and the description's collaboration type do not belong to the same collaboration type category, it is marked as a type conflict. If the title's action intent and the description's action intent correspond to different collaboration object categories, it is marked as an object conflict.

[0061] The method involves comparing the tags corresponding to the title and description of the cooperation type with a terminology database based on the specific cooperation plan to determine whether the title and description texts correspond to the same cooperation type. If they are different, it is determined to be a type conflict. If the action intent of the title and the action intent of the description point to different cooperation objects, it is determined to be an object conflict.

[0062] The descriptive text is segmented into sentences to identify the cooperative intentions of the first sentence and the last sentence.

[0063] The text is segmented using a language processing model to extract the first and last sentences. The cooperation type labels in the first and last sentences are identified to form the cooperation intentions of the first and last sentences, respectively.

[0064] If the title's collaboration type differs from the collaboration intention of the first sentence, it is marked as "first sentence deviation". If the title's collaboration type differs from the collaboration intention of the last sentence, it is marked as "last sentence deviation".

[0065] Since the first sentence usually reflects the core requirements of the collaboration, and the last sentence usually indicates supplementary requirements or a summary of the requirements, if the collaboration type in the title is different from the collaboration intention in the first and last sentences, it will be judged as deviation in the first sentence and deviation in the last sentence, respectively, to distinguish between two different situations: the requirement title deviates from the core theme and the requirement title omits supplementary themes.

[0066] The total number of conflict tags is obtained by counting all tags. Requests with a total number of conflict tags greater than a preset conflict threshold are marked as conflict target request information, and the corresponding conflict type tags are recorded.

[0067] The system sets a reasonable conflict threshold based on the specific collaborative content plan, which is used to distinguish between slightly deviating requirements and obviously conflicting requirements. If the total number of conflict markers in the requirement release information corresponding to a certain collaborative content is greater than the preset conflict threshold, then the requirement release information corresponding to that collaborative content is regarded as the conflict target requirement information, and the conflict markers corresponding to each conflict are recorded.

[0068] Methods for semantic correction of requirements include: Identify the conflict type markers in the conflict target requirement information. If there is a type conflict in the conflict type markers, extract the cooperation type that appears most frequently in the description text as the corrected cooperation type and replace the original cooperation type in the title text.

[0069] This involves extracting the tags corresponding to the most frequently occurring cooperation types as the corrected cooperation types and replacing the original cooperation type tags in the corresponding positions, making the descriptive text closer to the actual cooperation theme corresponding to the cooperation content.

[0070] If there is an object conflict in the conflict type marker, the core collaborating object in the description text is extracted as the correcting collaborating object and added to the end of the requirement title.

[0071] The core collaborating object refers to the business object that directly corresponds to the action in the collaboration content in the requirement description, including objects such as courses, teaching materials, and projects. When there is an object conflict marker, the core collaborating object is added to the end of the requirement title, so that the requirement title not only reflects the type of collaboration, but also the specific object to which the collaborative action is applied, making the object corresponding to the collaboration content in the title clearer.

[0072] If the conflict type marker has a first sentence deviation, the first sentence in the requirement description is extracted and semantic intent compression is performed to generate a first sentence intent summary, which replaces the original requirement title.

[0073] Semantic intent compression refers to the process of compressing and semantically refining the original first sentence while retaining its core cooperative theme, core cooperative partner, and main action intent. This first sentence is then used to replace the original requirement title, ensuring that the revised title directly aligns with the theme of the first sentence.

[0074] If the conflict type marker has a deviation in the last sentence, the last sentence in the requirement description is extracted and the supplementary cooperation intention in the last sentence is identified. The supplementary cooperation intention is then attached as a subtitle to the revised requirement title.

[0075] Supplementary cooperation intent refers to textual information such as additional cooperation directions, supplementary cooperation goals, or secondary cooperation requirements that appear in the last sentence of the requirement description. If there is a deviation mark in the last sentence, the supplementary cooperation intent other than the core theme is extracted from the last sentence, and this supplementary cooperation intent is added as a subtitle after the revised requirement title. Without changing the core theme of the revised title, the secondary cooperation directions or additional descriptions are retained to avoid missing supplementary requirement information.

[0076] The revised requirement titles and requirement descriptions are linked to form standard requirement information.

[0077] This involves re-associating the revised original requirement titles and descriptions with the existing, normal requirement release information to obtain standard requirement information.

[0078] The methods for performing cold start compensation calculations include: The number of cooperative projects for each cooperating unit in the unified unit profile set in the historical cooperative project information is counted, and cooperating units with zero cooperative projects are selected as new cooperating units to be added to the database.

[0079] The number of cooperative projects refers to the number of cooperative project records that each cooperative unit is actually associated with in the historical cooperative project information in the unified unit profile set. Since each cooperative unit may correspond to multiple project records in the historical cooperative project information, a number of cooperative projects of 0 indicates that the corresponding cooperative unit has no historical records. Therefore, the cooperative unit is determined to be a newly added cooperative unit.

[0080] Extract the attribute fields of newly added partners to form a new unit description group.

[0081] The attribute fields refer to fields that can describe the cooperation capabilities and technical directions of newly added partners. In this embodiment, they include, for example, the same field types as those included in the aforementioned partner profile, such as company introduction, description of cooperation capabilities, cooperation content that can be undertaken, core technical directions, company type, and company level.

[0082] Extract cooperative units from the unified unit profile set whose number of cooperative projects exceeds a preset activity threshold, mark them as active cooperative units, and extract the attribute fields of each active cooperative unit to form an active unit description group.

[0083] Based on specific cooperation specifications, a preset activity threshold is set. If the number of cooperation projects of a certain cooperation unit is higher than the preset activity threshold, it indicates that the cooperation unit has rich cooperation experience and is therefore judged as an active cooperation unit. At the same time, the attribute fields of the active cooperation unit are extracted to construct an active unit description group to describe the active cooperation unit.

[0084] Encode the new unit description group and the active unit description group into a new unit vector and an active unit vector.

[0085] The language processing model is used to semantically encode the new unit description group and the active unit description group, converting them into vector form to obtain the new unit vector and the active unit vector.

[0086] Calculate the similarity between the new unit vector and each active unit vector, and select the most similar active cooperative units of a predetermined number to form a set of similar active units.

[0087] In this embodiment, the similarity is obtained by calculating the cosine similarity between the new unit vector and each active unit vector. Based on the cooperation specification, a reasonable threshold for the number of similar processing units is set, which is the preset number. All active cooperative units corresponding to the new unit vector are sorted in descending order according to the similarity. The first preset number of active cooperative units are extracted as the set of similar active units corresponding to the new cooperative units added to the database for the new unit vector.

[0088] Extract the performance indicators of each active collaborating unit in the set of similar active units from the historical collaborative project information and identify the corresponding performance level.

[0089] The effectiveness indicators refer to a combination of parameters set based on specific cooperation plans in historical cooperation project information to evaluate the quality of cooperation. At the same time, this combination of parameters is compared with a rating table set based on specific cooperation plans to output the effectiveness level of active cooperation units.

[0090] The migration performance benchmark is obtained by weighted averaging of the performance levels of all active collaborating units.

[0091] In this embodiment, the similarity between each active collaborator and the newly added collaborators is used as a weight to calculate the weighted average of the effectiveness level. The historical collaboration performance of similar active collaborators is used as the initial capability evaluation basis for the newly added collaborators, which is the migration effectiveness benchmark.

[0092] The migration performance benchmark is used as the cold start compensation value for newly added partners and written into the attribute fields of the newly added partners. The performance levels of all partners are integrated to form a candidate partner scoring table.

[0093] The cold start compensation value is a scoring field. The migration performance benchmark is used as the cold start compensation value for newly added partners and written into the scoring field of the corresponding attribute field of the newly added partner. At the same time, the original performance level of other partners with existing historical cooperation records is retained. The scoring results of all partners are integrated and constructed into a table to form a candidate partner scoring table. The migration performance benchmark enables newly added partners to have a comparable initial scoring basis in the subsequent ranking process, avoiding a long-term low position in the recommendation results due to zero historical cooperation records.

[0094] Methods for constructing position decay weights and performing semantic encoding include: Retrieve all the technology direction options in each core technology direction field and their corresponding order to form the original order of the technology direction options.

[0095] The core technology direction field is a multi-option field used to describe the technical capabilities of the college and partner institutions. The technology direction options refer to the selected technology directions in this field, and the option order refers to the order in which the technology direction options are selected in the core technology direction field.

[0096] All technical direction options in the original order are assigned a position number based on their option position. The position distance between each position number and the preset benchmark position is calculated, and the position attenuation weight of the corresponding technical direction option is calculated based on the position distance.

[0097] The position number is a sequential marker of the technical direction options in the original order. A reference position, defined by the cooperation plan, is used to measure the priority of each technical direction option; this is the preset benchmark position. The difference in position between each technical direction option and the preset benchmark position is the position distance. The formula for calculating the position attenuation weight is as follows: ;in This indicates the positional attenuation weight of a specific technology direction option. This indicates the distance between the above technical direction options and the preset benchmark position; the smaller the position distance, the closer the corresponding technical direction option is to the main technical direction, and the greater the position attenuation weight; the larger the position distance, the more the corresponding technical direction option is biased towards the auxiliary or supplementary technical direction, and the smaller the position attenuation weight.

[0098] Construct the semantic vector of each technical direction option, multiply the semantic vector of the technical direction with the corresponding position attenuation weight, and output the attenuation semantic vector of each technical direction option.

[0099] The language processing model is used to convert the field content of each technical direction option into a vector form to obtain the technical direction semantic vector. The corresponding positional attenuation weights are used to perform a weighted product of the corresponding technical direction semantic vectors to obtain the attenuated semantic vector for each technical direction option.

[0100] Arrange the attenuation semantic vectors according to the option position to form an ordered attenuation vector sequence. Summing up all the attenuation semantic vectors in the sequence yields the ordered aggregated vector corresponding to the core technology direction field.

[0101] The sequential decay vector sequence is obtained by arranging the decay semantic vectors according to the original position of the technology direction options. By sequentially weighting and accumulating the technology direction semantic vectors with different positions, the semantic information and primary and secondary order information of each technology direction option in the multi-select field are uniformly compressed into the feature representation of the same field, forming a sequential aggregation vector. This avoids the dilution of the main technology direction information caused by simple averaging or unordered encoding.

[0102] By integrating the sequential aggregation vectors corresponding to all core technology fields, a sequence-sensitive feature set is obtained.

[0103] The dimension matching value is calculated as follows: The cooperation type, technology direction tag, geographical location constraint, and effective period information extracted from the standard demand information constitute the demand constraint group.

[0104] The cooperation type refers to the tag of the category of cooperation content corresponding to the demand; the technology direction tag refers to the tag of the technology direction option corresponding to the cooperation content; the geographical location constraint refers to the location restriction in the cooperation content; and the effective period information refers to the time constraint formed by the deadline of the cooperation content plan and the current latest time.

[0105] Extract the effectiveness level and cooperation content-related attribute information of each candidate partner from the candidate partner rating table to form a candidate partner attribute group.

[0106] The relevant attribute information of the cooperation content refers to the attribute fields of the cooperation units that are directly related to the project of the cooperation content, as well as the attribute fields that describe the cooperation content, such as the technical direction or other requirements. By constructing candidate unit attribute groups, the information of candidate cooperation units in terms of cooperation capabilities, technical direction and geographical distribution is unified, providing a data foundation for subsequent comparisons.

[0107] Obtain the sequential aggregation vectors corresponding to candidate collaborators in the sequence-sensitive feature set, calculate the vector distance between the sequential aggregation vectors of any demand object and the candidate collaborator, and obtain the technology direction matching value.

[0108] The semantics of the demand object corresponding to the cooperation content are transformed into vector form and displayed. The vector distance between the demand object and the sequential aggregated vector of each candidate cooperation unit is calculated to obtain the semantic similarity between the two at the technical direction level, which is the technical direction matching value.

[0109] Compare the cooperation type with the cooperation content-related attribute information in the candidate unit attribute group, and count the number of overlapping matching options to obtain the cooperation matching value.

[0110] The process involves matching the cooperation type with the technology direction options in the relevant attribute information of the cooperation content, and using the number of overlapping technology direction options that match the cooperation type as the cooperation matching value.

[0111] Match the geographical location constraints with the relevant attribute information of the cooperation content in the candidate unit attribute group to the geographical location of the cooperation, and identify the geographical location matching level as the geographical location matching value.

[0112] The geographic location matching levels include, for example, province, city, and district. The geographic location matching value is obtained by matching the geographic location constraints in the matching standard requirements information with the overlapping levels of the cooperative geographic locations at the geographic location matching level. This value is used to distinguish different degrees of geographic proximity, such as different cities in the same province, different districts in the same city, or completely the same district. The specific numerical value corresponding to the geographic location matching value is set based on the standard for screening cooperative content.

[0113] Based on the effective period information and the historical number of collaborations with the corresponding collaborators in the historical collaboration project information, the timeliness response value of the candidate collaborators is calculated.

[0114] The timeliness response value is obtained by calculating the ratio of the number of historical collaborations to the time interval corresponding to the effective period information. It should be noted that this calculation formula is a dimensionless calculation. If the remaining effective time of the demand is short, it indicates a high degree of urgency. In this case, candidate partners with more historical collaborations usually have higher process familiarity and faster response capabilities, and therefore have larger corresponding timeliness response values.

[0115] By combining the matching values ​​of technical direction, cooperation, geographical location, and timeliness, a dimensional matching value can be formed between any demand object and a candidate cooperation unit.

[0116] The matching values ​​of four dimensions—technology direction matching value, cooperation matching value, geographical location matching value, and timeliness response value—are combined to form a set of evaluation indicators, which are called dimension matching values.

[0117] The methods for performing comprehensive matching include: The initial matching value is obtained by multiplying the dimensional matching value and the normalized performance level for each candidate partner.

[0118] The process begins by normalizing the performance levels, unifying all performance levels to between 0 and 1. Then, the normalized performance levels are used to weight all values ​​in the dimension matching values ​​to obtain the initial matching values.

[0119] Identify the position of each candidate partner's technology direction matching value among all candidate partners, and mark the candidate partners whose position ranks in the top preset number as technology priority partners.

[0120] The candidate partners are sorted in descending order according to their technical matching value, and the top 100 candidate partners are selected as the priority partners for technology. The specific value of the top 100 is set based on the specific cooperation plan.

[0121] A preset gain coefficient is added to the initial matching value of the technology priority unit to obtain the gain matching value, while the initial matching value of the non-technology priority unit remains unchanged.

[0122] The preset gain coefficient refers to the coefficient used in this embodiment to increase the ranking weight of technology priority units. The gain matching value is obtained by weighting the initial matching values ​​of all technology priority units using the preset gain coefficient. The preset gain coefficient is set based on the specific cooperation plan requirements to apply additional gain to technology priority units, so that candidate cooperation units whose technical direction is highly consistent with the relevant requirements of the cooperation content will have more obvious priority in the subsequent comprehensive ranking.

[0123] The gain matching values ​​and initial matching values ​​of all candidate cooperative units are normalized to obtain the comprehensive matching score. The comprehensive matching scores of all candidate cooperative units are then integrated to obtain the comprehensive matching score set.

[0124] For technology-priority units, the gain matching value is used, and for other candidate cooperation units, the initial matching value is used. The comprehensive matching score of the corresponding cooperation unit is obtained by summing all values ​​in the gain matching value or the initial matching value and normalizing them. A unified scoring result is obtained under the combined influence of dimension matching value, effectiveness level and technology priority gain.

[0125] The ways to construct cooperative recommendation sequences include: All candidate partners in the comprehensive matching score set are sorted in descending order according to their comprehensive matching scores. The unit identifiers of the candidate partners whose ranking ranks are within a preset priority number are extracted and sorted into a cooperative recommendation sequence.

[0126] The plan sets a preset priority number based on the specific cooperation content, and extracts the unit identifiers corresponding to the top preset priority number of candidate cooperation units from the sequence sorted in descending order based on the comprehensive matching score to construct a cooperation recommendation sequence. The unit identifier refers to data that can uniquely identify the unit, such as unit code.

[0127] This embodiment obtains a business dataset by collecting multidimensional data from the university-enterprise cooperation management platform and standardizing it. Based on this dataset, it performs entity disambiguation and profile merging, semantic judgment and correction of demands, cold start compensation calculation, order-sensitive coding, and dimension matching calculation, achieving integrated intelligent matching and recommendation of supply and demand in university-enterprise cooperation. Compared with existing technologies, it optimizes the fragmentation problem of profiles caused by multiple colleges repeatedly entering the same cooperative unit by extracting core and modifying words from the unit name and combining them with organizational codes and content overlap coefficients for disambiguation judgment. It solves the problem of chaotic matching benchmarks caused by semantic deviation between demand titles and descriptions by performing intent parsing and performing classification labeling and differential correction on demand titles and descriptions respectively. It avoids the cold start problem of newly added cooperative units being at a long-term ranking disadvantage due to a lack of historical records by identifying new cooperative units and performing migration compensation based on the historical performance of similar active units. It avoids the problem of dilution of main technical direction information caused by equal weighting of multiple selection fields by constructing position decay weights and performing order-sensitive coding. It improves the accuracy of supply and demand matching and the rationality of recommendation results by achieving comprehensive ranking through multi-dimensional matching and technology priority gain.

[0128] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0129] All formulas in this manual are dimensionless and calculated numerically. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters and thresholds in the formulas are set by those skilled in the art according to the actual situation.

[0130] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A big model-based school-enterprise cooperation supply and demand intelligent matching and recommendation system, characterized in that, include: The data acquisition module collects multidimensional data from the university-enterprise cooperation management platform and performs standardized processing to obtain a business dataset; the business dataset includes college profiles, partner unit profiles, demand release information, and historical cooperation project information. The entity disambiguation module performs entity disambiguation based on the profiles of collaborating units and historical collaborative project information, identifies redundant target collaborating unit information, and discretizes and merges the profiles of the corresponding target collaborating units based on the disambiguation results, outputting a unified set of entity profiles. The semantic determination module uses a large model to perform semantic consistency determination on the demand release information and identify conflicting target demand information. Perform semantic correction on conflicting target requirement information to obtain standard requirement information; The startup compensation module identifies newly added cooperative units based on a unified unit profile set and historical cooperation project information, obtains the attribute information of the newly added cooperative units, performs cold start compensation calculations in conjunction with the unified unit profile set, and outputs a candidate unit scoring table. The feature encoding module extracts the core technology direction field from the college profile and unified unit profile sets, and identifies the original order of the technology direction options; Based on the original order of the technical direction options, positional decay weights are constructed and semantic encoding is performed to output an order-sensitive feature set. The dimension matching module calculates the dimension matching value between the demand object and the candidate cooperation unit based on standard demand information, candidate unit scoring table and order-sensitive feature set; Perform comprehensive matching based on dimension matching values ​​and output a comprehensive matching score set; The sequence construction module sorts candidate partners based on the comprehensive matching score set to form a cooperation recommendation sequence, and sends the cooperation recommendation sequence to the demand processing end; the modules are connected to each other via wired and / or wireless means.

2. The intelligent matching and recommendation system for supply and demand in university-enterprise cooperation based on a large model as described in claim 1, characterized in that, The methods for disambiguating unit entities include: Extract the unit attribute information corresponding to each cooperative unit from the cooperative unit profile to form unit identification features; use a large model to perform word segmentation on the unit name in each unit identification feature, extract the core word elements and modifier word elements in the unit name, and construct the name word element sequence; Calculate the overlap of core words and the difference of modifier words between the name word sequences corresponding to any two cooperating units; calculate name similarity based on the overlap of core words and the difference of modifier words, and select cooperating unit groups with name similarity higher than the preset name similarity threshold to form an initial set of duplicate units; Identify the credit codes of any two collaborating units in the initial set of duplicate units, and identify the organization code field in the credit codes; if the organization code fields of the two collaborating units are the same, mark the group of collaborating units as high-confidence duplicate units; Extract the cooperation project options from the historical cooperation project information of any two cooperating units, calculate the ratio of the number of intersections to the number of unions of the cooperation project options, and obtain the content overlap coefficient; The name similarity and content overlap coefficient are weighted and fused to obtain the comprehensive duplication confidence score of any group of cooperating units; cooperating units with a comprehensive duplication confidence score higher than the preset duplication threshold are selected and merged with high-confidence duplication units as the disambiguation result.

3. The intelligent matching and recommendation system for supply and demand in university-enterprise cooperation based on a large model as described in claim 2, characterized in that, The methods for performing discrete merging include: Identify the initial and latest timestamps in the profile of each collaborating unit after disambiguation, calculate the time interval between the initial and latest timestamps to obtain the profile activity period; take the profile of the collaborating unit with the longest activity period as the master record, and the profiles of the remaining collaborating units as slave records. Extract the unit attribute field from the master record and slave record and compare them field by field; If the primary record contains an empty field and the secondary record contains a non-empty field, the non-empty field in the secondary record will be filled into the corresponding empty field in the primary record; if both the primary and secondary records contain non-empty fields but are different, the latest timestamps of the corresponding fields in the two records will be compared, and the field with the latest timestamp will be retained. Extract the core technology direction field from the master record and slave record, and identify the original position of the option corresponding to each core technology direction field; add the option corresponding to the core technology direction field that exists in the slave record but not in the master record to the end of the core technology direction field of the master record; The project numbers of all high-confidence duplicate units in the disambiguation results are deduplicated and merged. The number of duplicated and merged projects is counted, and the historical collaboration count field of the master record is updated. The merged master record is used as the master unit profile, and the derived records are deleted to obtain a unified set of unit profiles.

4. The intelligent matching and recommendation system for supply and demand in university-enterprise cooperation based on a large model as described in claim 3, characterized in that, The methods for performing semantic consistency determination include: Extract the requirement title and requirement description from the requirement release information to obtain the title text and description text; The large model is used to perform intent parsing on the title text and description text respectively, to obtain the title cooperation type and title action intent in the title text, and the description cooperation type and description action intent in the description text; If the title's cooperation type and the description's cooperation type do not belong to the same cooperation type category, it is marked as a type conflict; if the title's action intent and the description's action intent correspond to different cooperation object categories, it is marked as an object conflict. The descriptive text is segmented into sentences to identify the cooperative intent of the first sentence and the cooperative intent of the last sentence. If the cooperative type of the title is different from the cooperative intent of the first sentence, it is marked as a deviation of the first sentence; if the cooperative type of the title is different from the cooperative intent of the last sentence, it is marked as a deviation of the last sentence. The total number of conflict tags is obtained by counting all tags. Requests with a total number of conflict tags greater than a preset conflict threshold are marked as conflict target request information, and the corresponding conflict type tags are recorded.

5. The intelligent matching and recommendation system for supply and demand in university-enterprise cooperation based on a large model as described in claim 4, characterized in that, The methods for performing semantic correction of demand include: Identify the conflict type markers in the conflict target requirement information. If there is a type conflict in the conflict type markers, extract the cooperation type that appears most frequently in the description text as the corrected cooperation type and replace the original cooperation type in the title text. If the conflict type marker has an object conflict, the core cooperation object in the description text is extracted as the corrected cooperation object and added to the end of the requirement title; if the conflict type marker has a first sentence deviation, the first sentence in the requirement description is extracted and semantic intent compression is performed to generate a first sentence intent summary, which replaces the original requirement title; if the conflict type marker has a last sentence deviation, the last sentence in the requirement description is extracted and the supplementary cooperation intent in the last sentence is identified, and the supplementary cooperation intent is attached as a subtitle to the corrected requirement title. The revised requirement titles and requirement descriptions are linked to form standard requirement information.

6. The intelligent matching and recommendation system for supply and demand in university-enterprise cooperation based on a large model as described in claim 5, characterized in that, The methods for performing cold start compensation calculations include: The number of cooperative projects for each cooperating unit in the unified unit profile set in the historical cooperative project information is counted, and cooperating units with zero cooperative projects are selected as new cooperating units to be added to the database. Extract the attribute fields of newly added partners to form a new partner description group; extract partners from the unified partner profile set whose number of partner projects exceeds a preset activity threshold, mark them as active partners, and extract the attribute fields of each active partner to form an active partner description group. Encode the new unit description group and the active unit description group into a new unit vector and an active unit vector; Calculate the similarity between the new unit vector and each active unit vector, and select the most similar active cooperative units of a predetermined number to form a set of similar active units; Extract the performance indicators of each active collaborator from the historical collaborative project information in the set of similar active units and identify the corresponding performance level; calculate the migration performance benchmark by weighted average of the performance levels of all active collaborators. The migration performance benchmark is used as the cold start compensation value for newly added partners and written into the attribute fields of the newly added partners. The performance levels of all partners are integrated to form a candidate partner scoring table.

7. The intelligent matching and recommendation system for supply and demand in university-enterprise cooperation based on a large model as described in claim 6, characterized in that, The method of constructing position attenuation weights and performing semantic encoding includes: Retrieve all the technology direction options in each core technology direction field and their corresponding order to form the original order of the technology direction options; All technical direction options in the original order are assigned a position number based on their position. The position distance between each position number and the preset benchmark position is calculated, and the position attenuation weight of the corresponding technical direction option is calculated based on the position distance. Construct the semantic vector of each technical direction option, multiply the semantic vector of the technical direction with the corresponding position attenuation weight, and output the attenuation semantic vector of each technical direction option. Arrange the attenuation semantic vectors according to the option position to form an ordered attenuation vector sequence. Summing up all the attenuation semantic vectors in the sequence yields the ordered aggregated vector corresponding to the core technology direction field. By integrating the sequential aggregation vectors corresponding to all core technology fields, a sequence-sensitive feature set is obtained.

8. The intelligent matching and recommendation system for supply and demand in university-enterprise cooperation based on a large model as described in claim 7, characterized in that, The calculation method for the dimension matching value is as follows: The cooperation type, technology direction tag, geographical location constraint, and effective period information extracted from the standard demand information constitute a demand constraint group; Extract the effectiveness level and cooperation content-related attribute information of each candidate partner from the candidate partner rating table to form a candidate partner attribute group; Obtain the sequential aggregation vectors corresponding to candidate collaborators in the sequence-sensitive feature set, calculate the vector distance between the sequential aggregation vectors of any demand object and the candidate collaborators, and obtain the technology direction matching value; Compare the cooperation type with the cooperation content-related attribute information in the candidate unit attribute group, and count the number of overlapping matching options to obtain the cooperation matching value; Match geographical location constraints with the relevant attribute information of the cooperation content in the candidate unit attribute group to the geographical location of the cooperation, and identify the geographical location matching level as the geographical location matching value; Based on the effective period information and the historical number of collaborations of the corresponding collaborators in the historical cooperation project information, the timeliness response value of the candidate collaborators is calculated. By combining the matching values ​​of technical direction, cooperation, geographical location, and timeliness, a dimensional matching value can be formed between any demand object and a candidate cooperation unit.

9. The intelligent matching and recommendation system for supply and demand in university-enterprise cooperation based on a large model as described in claim 8, characterized in that, The methods for performing comprehensive matching include: Calculate the product of the dimension matching value and the normalized performance level for each candidate partner to obtain the initial matching value; Identify the ranking of each candidate partner's technology direction matching value among all candidate partners, and mark the candidate partners whose ranking is among the top preset number as technology priority partners; A preset gain coefficient is added to the initial matching value of the technology priority unit to obtain the gain matching value, while the initial matching value of the non-technology priority unit remains unchanged; The gain matching values ​​and initial matching values ​​of all candidate cooperative units are normalized to obtain the comprehensive matching score. The comprehensive matching scores of all candidate cooperative units are then integrated to obtain the comprehensive matching score set.

10. The intelligent matching and recommendation system for supply and demand in university-enterprise cooperation based on a large model as described in claim 9, characterized in that, The methods for constructing the cooperative recommendation sequence include: All candidate partners in the comprehensive matching score set are sorted in descending order according to their comprehensive matching scores. The unit identifiers of the candidate partners whose ranking ranks are within a preset priority number are extracted and sorted into a cooperative recommendation sequence.

Images