Scientific research review portrait construction method based on large language model and related equipment

Abstract

The application discloses a scientific research review portrait construction method based on a large language model and related equipment, which can be applied to the technical field of natural language processing and artificial intelligence. The application divides the preprocessed structured data set based on a preset label system to obtain data subsets corresponding to different labels; then inputs the data subsets and hierarchical prompt word templates into a large language model to obtain candidate labels of each dimension, labels support data associated with the candidate labels in the data subsets, and then performs confidence verification on the candidate labels labeled with the support data to obtain target portrait labels, so that high-quality, structured and multi-dimensional portraits can be obtained, the accuracy of the assignment of scientific research review experts is improved, the target portrait labels and the support data are stored in a traceable manner, and reliable portrait data can be provided in an actual application scenario, and the accuracy of the assignment of scientific research review experts is further improved.

Description

Technical Field

[0001] This application relates to the fields of natural language processing and artificial intelligence, and in particular to a method and related equipment for constructing a research review profile based on a large language model. Background Technology

[0002] In related technologies, the accurate assignment of review experts is a core application in scientific research management. Its effectiveness essentially depends on the accurate characterization of the deep features of experts and research projects, that is, building high-quality, multi-dimensional profiles of experts and research projects. However, in existing scientific research management systems, the relevant data of experts and projects are scattered across different formats such as PDFs, Word documents, and structured database tables. Relying solely on superficial keywords such as disciplines and technical fields to construct profiles of experts and projects may result in low accuracy. Although Large Language Models (LLMs) can be used for profile construction, the large amount of redundant and noisy information in the original data severely interferes with the model's learning and focusing on key features. For example, in expert review texts, most interactions between entities do not contain effective relationships, and only a small portion of interactions reflect valuable review logic or domain knowledge. This results in low accuracy of expert profiles obtained through LLM processing, thereby reducing the accuracy of scientific research review expert assignment.

[0003] In summary, the technical problems existing in the relevant technologies need to be improved. Summary of the Invention

[0004] The main objective of this application is to propose a method and related equipment for constructing a research review profile based on a large language model, which can effectively improve the accuracy of research review expert assignment.

[0005] To achieve the above objectives, one aspect of this application proposes a method for constructing a research review profile based on a large language model, the method comprising the following steps:

[0006] Acquire raw data to be processed, including expert data or project data;

[0007] The raw data to be processed is preprocessed to obtain a structured dataset;

[0008] The structured dataset is divided based on a preset label system to obtain data subsets corresponding to different labels;

[0009] The data subset and hierarchical prompt word template are input into the large language model to obtain candidate labels for each dimension;

[0010] In the data subset, supporting data corresponding to the candidate label is labeled, and each supporting data corresponds to a timestamp data.

[0011] The confidence level of the candidate labels labeled with the supporting data is verified to obtain the target profile labels, which include expert profile labels and project profile labels.

[0012] The target profile tags and supporting data are stored in a traceable manner.

[0013] In some embodiments, the preprocessing of the raw data to be processed to obtain a structured dataset includes:

[0014] Extract the unstructured data and the first structured data from the original data to be processed;

[0015] Construct a first prompt word template based on the first structured data;

[0016] Based on the first prompt word template, a set of candidate entities is extracted from the unstructured data using a large language model, and a set of candidate entity pairs is generated based on the set of candidate entities.

[0017] Calculate the confidence score of the candidate entity pairs in the candidate entity pair set;

[0018] The target entity pair set is obtained by filtering from the candidate entity pair set based on the confidence score and the first confidence threshold;

[0019] The target entity pair set and the first structured data are combined to form the structured dataset.

[0020] In some embodiments, calculating the confidence score of candidate entity pairs in the candidate entity pair set includes:

[0021] Extract the context data of each entity in each candidate entity pair from the raw data to be processed;

[0022] The two context data corresponding to each entity in each candidate entity pair are input into a preset encoder to obtain the entity context embedding representation;

[0023] The two entity context embedding representations corresponding to each entity in each candidate entity pair are fused to obtain a fused context embedding representation.

[0024] The fused context embedding representation is input into the classifier network to obtain the confidence score of the candidate entity pair. The confidence score is used to characterize the semantic relationship between the context data corresponding to the two entities in the candidate entity pair. The classifier network includes a fully connected layer and a preset activation function layer.

[0025] In some embodiments, the confidence score of the candidate entity pair is calculated using the following formula:

[0026] ;

[0027] In the formula, This represents the confidence score for candidate entity pairs; Indicates the preset activation function layer; Indicates the bias value; Represents the learnable weights; This represents the output of the fully connected layer. The calculation formula is as follows:

[0028] ;

[0029] In the formula, Indicates a fully connected layer; Indicates the bias value; Represents the learnable weights; This represents the fusion context embedding representation; and This represents the two entity context embeddings corresponding to each entity in the candidate entity pair.

[0030] In some embodiments, the hierarchical prompt word template includes a system-level prompt word template, a task-level prompt word template, and an instance-level prompt word template;

[0031] The system-level prompt word template is used to define global constraints and output specifications, and to set model roles and task boundaries;

[0032] The task-level prompt word template is used for professional design of profile tag dimensions;

[0033] The instance-level prompt word templates are used to construct output instance samples for large language models.

[0034] In some embodiments, the step of performing confidence verification on candidate labels labeled with the supporting data to obtain target profile labels includes:

[0035] Calculate the time decay factor of the supporting data based on the timestamp data corresponding to the supporting data;

[0036] Multiply the time decay factor by the embedded representation of the supporting data to obtain the weighted embedded representation of the supporting data;

[0037] Calculate the average similarity between the weighted embedding representation of the supporting data and the embedding representation of the candidate labels;

[0038] The candidate labels are filtered based on the average similarity and the second confidence threshold to obtain the target profile labels.

[0039] In some embodiments, the formula for calculating the average similarity is as follows:

[0040] ;

[0041] In the formula, Indicates average similarity; Indicates the first The time decay factor of the supporting data; The embedding representation of the candidate label y; Let F be the norm of the embedding representation of the candidate label y; This indicates the first candidate label y. Supporting data items; This indicates the first candidate label y. Embedded representation of supporting data; This represents the total number of supporting data for candidate label y.

[0042] To achieve the above objectives, another aspect of this application proposes a research review profile construction device based on a large language model, the device comprising:

[0043] The acquisition module is used to acquire raw data to be processed, including expert data or project data.

[0044] The preprocessing module is used to preprocess the raw data to be processed to obtain a structured dataset;

[0045] The partitioning module is used to partition the structured dataset based on a preset label system to obtain data subsets corresponding to different labels;

[0046] The tag generation module is used to input the data subset and hierarchical prompt word template into the large language model to obtain candidate tags in various dimensions;

[0047] The annotation module is used to annotate the supporting data corresponding to the candidate label in the data subset, wherein the supporting data corresponds to a timestamp data.

[0048] The verification module is used to perform confidence verification on the candidate labels labeled with the supporting data to obtain the target profile labels, which include expert profile labels and project profile labels.

[0049] The storage module is used to store the target image tags and the supporting data in a traceable manner.

[0050] To achieve the above objectives, another aspect of this application provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described above.

[0051] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer-readable storage medium storing a computer program that, when executed by a processor, implements the methods described above.

[0052] To achieve the above objectives, another aspect of the embodiments of this application proposes a computer program product, including a computer program that, when executed by a processor, implements the aforementioned method.

[0053] The embodiments of this application include at least the following beneficial effects: This application provides a method and related equipment for constructing scientific research review profiles based on a large language model. This scheme preprocesses the raw data to be processed, including expert data or project data, to obtain a structured dataset. Then, based on a preset labeling system, the structured dataset is divided into data subsets corresponding to different labels. Next, the data subsets and hierarchical prompt word templates are input into a large language model to obtain candidate labels for each dimension. Supporting data with timestamps associated with the candidate labels is then labeled in the data subsets. Next, the confidence level of the candidate labels with supporting data is verified to obtain target profile labels. Finally, the target profile labels and supporting data are stored traceably. This embodiment utilizes the powerful semantic understanding capabilities of a large language model to mine the semantic features of experts and projects from multi-source heterogeneous raw data to be processed, thereby constructing high-quality, structured, and multi-dimensional profiles, improving the accuracy of scientific research review expert assignment. Simultaneously, the traceable storage of target profile labels and supporting data provides reliable profile data in practical application scenarios, further improving the accuracy of scientific research review expert assignment. Attached Figure Description

[0054] Figure 1 This is a flowchart of the research review profile construction method based on a large language model provided in the embodiments of this application;

[0055] Figure 2 This is an overall schematic diagram of the scientific research review profile construction method based on a large language model provided in the embodiments of this application;

[0056] Figure 3 This is a schematic diagram illustrating the processing of candidate entity pairs provided in an embodiment of this application;

[0057] Figure 4 This is a schematic diagram illustrating the label processing of a data subset based on a large language model, provided in an embodiment of this application.

[0058] Figure 5 This is a schematic diagram of the structure of the scientific research review profile construction device based on a large language model provided in the embodiments of this application. Detailed Implementation

[0059] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of this application and are not intended to limit it. In the following description, when referring to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with those of this application; they are merely examples of apparatuses and methods consistent with some aspects of the embodiments of this application as detailed in the appended claims.

[0060] It is understood that the terms “first,” “second,” etc., used in this application may be used herein to describe various concepts, but unless otherwise stated, these concepts are not limited by these terms. These terms are only used to distinguish one concept from another. For example, without departing from the scope of the embodiments of this application, first information may also be referred to as second information, and similarly, second information may also be referred to as first information. Depending on the context, the words “if,” “when,” or “in response to a determination” as used herein may be interpreted as “when…” or “when…” or “in response to a determination.”

[0061] As used in this application, the terms "at least one", "multiple", "each", "any", etc., "at least one" includes one, two or more, "multiple" includes two or more, "each" refers to each of the corresponding multiples, and "any" refers to any one of the multiples.

[0062] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. The terminology used herein is for the purpose of describing embodiments of this application only and is not intended to limit this application.

[0063] Before providing a detailed description of the embodiments of this application, some of the nouns and terms used in the embodiments of this application will be explained first. The nouns and terms used in the embodiments of this application shall be interpreted as follows:

[0064] Project profiling is a multi-dimensional, structured descriptive tool based on data throughout the entire project lifecycle. It integrates fundamental project attributes (such as project name, code, company / department, and project nature), business characteristics (such as target users, task environment, application scenarios, core content, and research direction), management elements (such as schedule status, risk control, cost budget, and quality requirements), and value objectives (such as expected outcomes, acceptance criteria, and supporting conditions). This information is presented visually or in a tagged format, intuitively showcasing the project's key attributes and development trajectory. Its core objective is to eliminate fragmented information in project descriptions, providing precise data for project decision-making, resource allocation, and process control.

[0065] Expert profiling is a visualized and tagged expert characteristic model formed by integrating and analyzing multi-source data, including experts' personal information (such as education, professional title, and work experience), professional abilities (such as research fields, academic achievements, and technical expertise), behavioral characteristics (such as review participation, scoring preferences, and response speed), and value contributions (such as industry influence, technology transfer rate, and talent cultivation). Its essence is to replace subjective judgment with objective data to achieve accurate identification and matching of experts' abilities and value.

[0066] In related technologies, the accurate assignment of review experts is a core application in scientific research management. Its effectiveness essentially depends on the accurate characterization of the deep features of experts and research projects, that is, building high-quality, multi-dimensional profiles of experts and research projects. However, in existing scientific research management systems, the relevant data of experts and projects are scattered across different formats such as PDFs, Word documents, and structured database tables. Relying solely on superficial keywords such as disciplines and technical fields to construct profiles of experts and projects may result in low accuracy. Although Large Language Models (LLMs) can be used for profile construction, the large amount of redundant and noisy information in the original data severely interferes with the model's learning and focusing on key features. For example, in expert review texts, most interactions between entities do not contain effective relationships, and only a small portion of interactions reflect valuable review logic or domain knowledge. This results in low accuracy of expert profiles obtained through LLM processing, thereby reducing the accuracy of scientific research review expert assignment.

[0067] In view of this, this application provides a method and related equipment for constructing a research review profile based on a large language model, which can effectively improve the accuracy of research review expert assignment.

[0068] The research review profile construction method based on a large language model provided in this application relates to the fields of natural language processing and artificial intelligence. This method can be applied to a terminal, a server, or software running on either a terminal or a server. In some embodiments, the terminal can be a smartphone, tablet, laptop, desktop computer, etc., but is not limited thereto; the server can be configured as an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing basic cloud computing services such as cloud services, cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, CDN, and big data and artificial intelligence platforms. The server can also be a node server in a blockchain network; the software can be an application implementing the research review profile construction method based on a large language model, etc., but is not limited to the above forms.

[0069] This application can be used in a wide variety of general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, multiprocessor systems, microprocessor-based systems, programmable consumer electronics, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices. This application can be described in the general context of computer-executable instructions executed by a computer, such as program modules. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform specific tasks or implement specific abstract data types. This application can also be practiced in distributed computing environments where tasks are performed by remote processing devices connected via a communication network. In distributed computing environments, program modules can reside in local and remote computer storage media, including storage devices.

[0070] The embodiments of this application will be described in detail below with reference to the accompanying drawings:

[0071] Figure 1 This is an optional flowchart of the research review profile construction method based on a large language model provided in the embodiments of this application. Figure 1 The method may include, but is not limited to, steps S110 to S170:

[0072] Step S110: Obtain the raw data to be processed, wherein the raw data to be processed includes expert data or project data;

[0073] Step S120: Preprocess the raw data to be processed to obtain a structured dataset;

[0074] Step S130: Divide the structured dataset based on the preset label system to obtain data subsets corresponding to different labels;

[0075] Step S140: Input the data subset and hierarchical prompt word template into the large language model to obtain candidate labels for each dimension;

[0076] Step S150: Label the supporting data corresponding to the candidate labels in the data subset, wherein each supporting data corresponds to a timestamp data.

[0077] Step S160: Validate the confidence of candidate labels labeled with supporting data to obtain target profile labels, wherein the target profile labels include expert profile labels and project profile labels.

[0078] Step S170: Store the target profile tags and supporting data in a traceable manner.

[0079] It is understandable that expert data or project data includes both structured and unstructured data. Unstructured data contains redundancy and noise; therefore, such as... Figure 2 As shown, this embodiment preprocesses unstructured data and then combines it with the first structured data to form a structured dataset. Specifically, in the preprocessing process, this embodiment extracts unstructured data and the first structured data from the original data to be processed. A first prompt word template is then constructed based on the first structured data. Next, based on the first prompt word template, a large language model is used to extract a set of candidate entities from the unstructured data, and a set of candidate entity pairs is generated. The confidence scores of the candidate entity pairs in the candidate entity pair set are then calculated. Based on the confidence scores and a first confidence threshold, a set of target entity pairs is selected from the candidate entity pair set. Finally, the set of target entity pairs is combined with the first structured data to form a structured dataset.

[0080] Specifically, in this embodiment, for unstructured data, a first prompt word template is designed based on known first structured data, and then a pre-trained Large Language Model (LLM) is used to extract a candidate entity set from the unstructured data. ,in, Represents a set of candidate entities, containing entities from... arrive n candidate entities, and a set of candidate entity pairs containing m candidate entity pairs. ,in Let k represent the k-th entity pair, and It is its corresponding (entity, relation, entity) triple. The first i , j One entity, This refers to the relationships between entities.

[0081] It is understandable that, in this embodiment, after obtaining the set of candidate entity pairs, as follows: Figure 3 As shown, context data for each entity in each candidate entity pair is extracted from the original data to be processed. The two context data points corresponding to each entity in each candidate entity pair are then input into a pre-defined encoder to obtain an entity context embedding representation. The two entity context embedding representations corresponding to each entity in each candidate entity pair are then fused to obtain a fused context embedding representation. This fused context embedding representation is then input into a classifier network to obtain a confidence score for the candidate entity pair. The confidence score is used to characterize the semantic relationship between the context data corresponding to the two entities in the candidate entity pair. The classifier network includes fully connected layers and pre-defined activation function layers.

[0082] Specifically, for candidate entity pairs This embodiment extracts candidate entities from the raw data to be processed. With candidate entities The context data, which can be the text data of the candidate entity from its first appearance to its last appearance, is input into the BERT encoder to obtain the entity context embedding representation. and Then these two entity context embeddings represent and By concatenating and fusing the data, we obtain the fused context embedding representation of the candidate entity pair. Then, this fusion context is embedded in the representation. The input classifier network calculates the probability that there is a valid linguistic relationship in the context of two entities, which is used as the confidence score.

[0083] The confidence score for candidate entity pairs is calculated using the following formula:

[0084] ;

[0085] In the formula, This represents the confidence score for candidate entity pairs; Indicates the preset activation function layer; Indicates the bias value; Represents the learnable weights; This represents the output of the fully connected layer. The calculation formula is as follows:

[0086] ;

[0087] In the formula, Indicates a fully connected layer; Indicates the bias value; Represents the learnable weights; This represents the fusion context embedding representation; and This represents the two entity context embeddings corresponding to each entity in the candidate entity pair.

[0088] Understandably, in this embodiment, after calculating the confidence score for each entity pair, a first confidence threshold between 0 and 1 is set for the confidence score. If the confidence score of the candidate entity pair is greater than the threshold, the candidate entity pair is considered a valuable entity relationship pair and is retained; otherwise, it is considered a worthless candidate entity pair and is discarded. Finally, the selected valuable entity relationship pairs are combined with the first structured data in the original data to form a structured dataset.

[0089] It is understood that this embodiment can divide the structured dataset into different data subsets corresponding to different tags by using a pre-designed preset tag system. This allows each data subset to have a strong correlation with the profile tag dimension to be analyzed, so that when performing tag analysis using a large language model, only the data related to that tag can be focused on. For example, the data subset used to analyze expert research direction tags should include the names and abstracts of the expert's past published papers, patent applications, and research projects participated in; the data subset used to analyze expert academic influence tags should include the number of citations of the expert's past published papers, the names of the journals or conferences in which the papers were published, the level and division of labor in the research projects participated in, and the awards obtained; the data subset used to analyze expert collaboration network tags should include the data of the expert's published papers, patent applications, co-authors and co-author affiliations in research projects; the data subset used to analyze expert review performance tags should include the expert's past review records; the data subset used to analyze the project's technical content should include the project's discipline, research content, technical route, and innovative points; the data subset used to analyze project performance results should include the project's industry sector, expected results, and socio-economic benefits; and the data subset used to analyze the project's work foundation should include information on the project team composition and the past achievements of team members.

[0090] It is understood that this embodiment also includes a hierarchical prompt word template referenced by the large language model when generating profile tags. Specifically, the hierarchical prompt word template includes a system-level prompt word template, a task-level prompt word template, and an instance-level prompt word template. The system-level prompt word template is used to define global constraints and output specifications, setting model roles, output formats, and task boundaries; the task-level prompt word template is used for specialized design for the profile tag dimension; and the instance-level prompt word template is used to construct output instance samples for the large language model, which can be 2 to 3 samples, to guide the large language model in generating candidate tags that meet the requirements.

[0091] It is understandable that, in this embodiment, after obtaining the data subset and the hierarchical prompt word template, such as Figure 4 As shown, the split data subset and hierarchical prompt word template are input into the large language model to obtain candidate labels for each dimension, and the supporting data corresponding to the candidate labels is labeled in the data subset. This supporting data is a set of structured data items. It is a subset of data, containing one or more structured data items, and each data item has a timestamp. q Item data, among which For the first p Supporting data The year timestamp, the current time timestamp is Then, confidence verification is performed on the candidate labels with supporting data.

[0092] Specifically, in this embodiment, when performing execution verification, the time decay factor of the supporting data is calculated based on the timestamp data corresponding to the supporting data. The calculation process is as follows:

[0093] ;

[0094] In the formula, Indicates the current time; This indicates the candidate label y The corresponding number p The timestamp attribute value of the supporting data item; The attenuation coefficient; The candidate tag corresponds to the first p The time decay factor of the supporting data.

[0095] Simultaneously, the candidate label is denoted as y, and the candidate label and its corresponding support data are input into the Bert encoder to obtain the embedding representation of the candidate label and the corresponding support data embedding representation. , Then, the time decay factor of each supporting data item is multiplied by the embedding representation of the supporting data to obtain the weighted embedding representation of the supporting data. Finally, the average similarity between the weighted embedding representation of the supporting data and the embedding representation of the candidate tags is calculated, i.e., the semantic similarity is calculated. The calculation formula is as follows:

[0096] ;

[0097] In the formula, Indicates average similarity; Indicates the first The time decay factor of the supporting data; The embedding representation of the candidate label y; Let F be the norm of the embedding representation of the candidate label y; This indicates the first candidate label y. Supporting data items; This indicates the first candidate label y. Embedded representation of supporting data; This represents the total number of supporting data for candidate label y.

[0098] Understandably, semantic similarity The result value ranges from -1 to 1. Therefore, this embodiment sets a second confidence threshold between -1 and 1. If the semantic similarity... If the value is greater than the second confidence threshold, the candidate label is considered to have sufficient and reliable numerical support; otherwise, it is considered to lack sufficient and reliable data support. Candidate labels corresponding to values ​​greater than the second confidence threshold are used as target profile labels, thus obtaining profile labels with sufficient and reliable data support.

[0099] It is understood that this embodiment, after obtaining the target profile tags and associated supporting data, designs a storage scheme based on a NoSQL database. It supports flexible storage of JSON-formatted profile tags, enabling traceable storage of the target profile tags and associated supporting data within the database. Furthermore, it establishes an inverted index structure by linking the original supporting data table through foreign keys, achieving rapid traceability from tags to data evidence. Specifically, in this embodiment, the storage process can use a document-oriented database to store the target profile information, managing each expert's or project's profile as an independent document. The document structure includes four parts: identification information, basic attributes, a tag set, and a supporting database. The tag set is organized hierarchically, with each tag recording detailed metadata such as name, dimension classification, confidence score, and generation time, and establishing a bidirectional association with the supporting data in the supporting database. The supporting evidence database stores the original data fragments upon which the tags are based, fully preserving key information such as data source, content summary, and timestamps. Simultaneously, within the document, this embodiment achieves direct association between tags and supporting data through embedded pointers; at the system level, a global inverted index is established to support rapid retrieval based on tag content. When users need to verify the credibility of tags, the system can automatically locate relevant documents and extract complete tag evidence.

[0100] As described above, the method in this embodiment leverages the powerful semantic understanding of large language models to deeply mine the semantic features of experts and projects from multi-source heterogeneous data, thereby constructing high-quality, structured, and multi-dimensional profiles. This improves the accuracy and efficiency of expert assignment for research reviews. Furthermore, by designing a systematic data preprocessing and prompt word engineering approach, it effectively overcomes noise and redundancy issues in the original data, thus enhancing the reliability and usability of profile generation. Simultaneously, this embodiment also supports interpretable source tracing of profiles, possessing good practicality and wide applicability in scenarios such as research fund management and university research management, demonstrating significant practical application value and promising prospects for wider adoption.

[0101] Please see Figure 5 This application also provides a device for constructing a research review profile based on a large language model. The device includes:

[0102] The acquisition module is used to acquire the raw data to be processed, which includes expert data or project data.

[0103] The preprocessing module is used to preprocess the raw data to obtain a structured dataset;

[0104] The partitioning module is used to partition the structured dataset based on a preset label system to obtain data subsets corresponding to different labels;

[0105] The tag generation module is used to input a subset of data and hierarchical prompt word templates into a large language model to obtain candidate tags in various dimensions;

[0106] The annotation module is used to annotate the supporting data corresponding to the candidate label in the data subset, wherein the supporting data corresponds to a timestamp data;

[0107] The verification module is used to verify the confidence of candidate labels labeled with supporting data to obtain target profile labels, which include expert profile labels and project profile labels.

[0108] The storage module is used to store target profile tags and supporting data in a traceable manner.

[0109] It is understood that the content of the above method embodiments is applicable to the present device embodiments. The specific functions implemented by the present device embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0110] This application also provides an electronic device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the above-described method. This electronic device can be any smart terminal, including tablet computers, in-vehicle computers, etc.

[0111] It is understood that the content of the above method embodiments is applicable to this device embodiment. The specific functions implemented by this device embodiment are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0112] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the above-described method.

[0113] It is understood that the content of the above method embodiments is applicable to this storage medium embodiment. The specific functions implemented in this storage medium embodiment are the same as those in the above method embodiments, and the beneficial effects achieved are also the same as those achieved in the above method embodiments.

[0114] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0115] It is understood that the content of the above method embodiments is applicable to the embodiments of this program product. The specific functions implemented by the embodiments of this program product are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0116] The embodiments described in this application are for the purpose of more clearly illustrating the technical solutions of the embodiments of this application, and do not constitute a limitation on the technical solutions provided by the embodiments of this application. As those skilled in the art will know, with the evolution of technology and the emergence of new application scenarios, the technical solutions provided by the embodiments of this application are also applicable to similar technical problems.

[0117] Those skilled in the art will understand that the technical solutions shown in the figures do not constitute a limitation on the embodiments of this application, and may include more or fewer steps than shown, or combine certain steps, or different steps.

[0118] The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate; that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0119] Those skilled in the art will understand that all or some of the steps in the methods disclosed above, as well as the functional modules / units in the systems and devices, can be implemented as software, firmware, hardware, or suitable combinations thereof.

[0120] The terms “first,” “second,” “third,” “fourth,” etc. (if present) in the specification and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms “comprising” and “having,” and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0121] It should be understood that in this application, "at least one (item)" means one or more, and "more than" means two or more. "And / or" is used to describe the relationship between related objects, indicating that three relationships can exist. For example, "A and / or B" can represent three cases: only A exists, only B exists, and both A and B exist simultaneously, where A and B can be singular or plural. The character " / " generally indicates that the preceding and following related objects are in an "or" relationship. "At least one (item) of the following" or similar expressions refer to any combination of these items, including any combination of single or plural items. For example, at least one (item) of a, b, or c can represent: a, b, c, "a and b", "a and c", "b and c", or "a and b and c", where a, b, and c can be single or multiple.

[0122] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of the units described above is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0123] The units described above as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0124] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0125] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes multiple instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing programs, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0126] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this does not limit the scope of the claims of the present application. Any modifications, equivalent substitutions, and improvements made by those skilled in the art without departing from the scope and substance of the embodiments of the present application shall be within the scope of the claims of the present application.

Claims

1. A method for constructing a research review profile based on a large language model, characterized in that, The method includes the following steps: Acquire raw data to be processed, which includes expert data and project data; The raw data to be processed is preprocessed to obtain a structured dataset; The structured dataset is divided based on a preset label system to obtain data subsets corresponding to different labels; The data subset and hierarchical prompt word templates are input into the large language model to obtain candidate labels for each dimension. The hierarchical prompt word templates include system-level prompt word templates, task-level prompt word templates, and instance-level prompt word templates. The system-level prompt word templates are used to define global constraints and output specifications, and to set model roles and task boundaries. The task-level prompt word templates are used for specialized design for the profile label dimension. The instance-level prompt word templates are used to construct output instance samples for the large language model. In the data subset, supporting data corresponding to the candidate label is labeled, and each supporting data corresponds to a timestamp data. The confidence level of the candidate labels labeled with the supporting data is verified to obtain the target profile labels, which include expert profile labels and project profile labels. The target profile tags and supporting data are stored in a traceable manner. The step of preprocessing the original data to obtain a structured dataset includes: Extract the unstructured data and the first structured data from the original data to be processed; Construct a first prompt word template based on the first structured data; Based on the first prompt word template, a set of candidate entities is extracted from the unstructured data using a large language model, and a set of candidate entity pairs is generated based on the set of candidate entities. Calculate the confidence score of the candidate entity pairs in the candidate entity pair set; The target entity pair set is obtained by filtering from the candidate entity pair set based on the confidence score and the first confidence threshold; The target entity pair set and the first structured data are combined to form the structured dataset; The step of performing confidence verification on candidate labels labeled with the supporting data to obtain target profile labels includes: Calculate the time decay factor of the supporting data based on the timestamp data corresponding to the supporting data; Multiply the time decay factor by the embedded representation of the supporting data to obtain the weighted embedded representation of the supporting data; Calculate the average similarity between the weighted embedding representation of the supporting data and the embedding representation of the candidate labels; The candidate labels are filtered based on the average similarity and the second confidence threshold to obtain the target profile labels; The time decay factor of the support data is calculated based on the timestamp data corresponding to the support data. The calculation process is as follows: ； In the formula, Indicates the current time; This indicates the candidate label y The corresponding number p The timestamp attribute value of the supporting data item; The attenuation coefficient; The candidate tag corresponds to the first p The time decay factor of the supporting data.

2. The method according to claim 1, characterized in that, The calculation of the confidence score for candidate entity pairs in the candidate entity pair set includes: Extract the context data of each entity in each candidate entity pair from the raw data to be processed; The two context data corresponding to each entity in each candidate entity pair are input into a preset encoder to obtain the entity context embedding representation; The two entity context embedding representations corresponding to each entity in each candidate entity pair are fused to obtain a fused context embedding representation. The fused context embedding representation is input into the classifier network to obtain the confidence score of the candidate entity pair. The confidence score is used to characterize the semantic relationship between the context data corresponding to the two entities in the candidate entity pair. The classifier network includes a fully connected layer and a preset activation function layer.

3. The method according to claim 2, characterized in that, The confidence score of the candidate entity pair is calculated using the following formula: ； In the formula, This represents the confidence score for candidate entity pairs; Indicates the preset activation function layer; Indicates the bias value; Represents the learnable weights; This represents the output of the fully connected layer. The calculation formula is as follows: ； In the formula, Indicates a fully connected layer; Indicates the bias value; Represents the learnable weights; This represents the fusion context embedding representation; and This represents the two entity context embeddings corresponding to each entity in the candidate entity pair.

4. The method according to claim 1, characterized in that, The formula for calculating the average similarity is as follows: ； In the formula, Indicates average similarity; Indicates the first The time decay factor of the supporting data; The embedding representation of the candidate label y; Let F be the norm of the embedding representation of the candidate label y; This indicates the first candidate label y. Supporting data items; This indicates the first candidate label y. Embedded representation of supporting data; This represents the total number of supporting data for candidate label y.

5. A research review profile construction device based on a large language model, characterized in that, The device includes: The acquisition module is used to acquire raw data to be processed, which includes expert data and project data. The preprocessing module is used to preprocess the raw data to be processed to obtain a structured dataset; The partitioning module is used to partition the structured dataset based on a preset label system to obtain data subsets corresponding to different labels; The tag generation module is used to input the data subset and hierarchical prompt word templates into the large language model to obtain candidate tags for each dimension. The hierarchical prompt word templates include system-level prompt word templates, task-level prompt word templates, and instance-level prompt word templates. The system-level prompt word templates are used to define global constraints and output specifications, and to set model roles and task boundaries. The task-level prompt word templates are used for specialized design for the profile tag dimension. The instance-level prompt word templates are used to construct output instance samples for the large language model. The annotation module is used to annotate the supporting data corresponding to the candidate label in the data subset, wherein the supporting data corresponds to a timestamp data. The verification module is used to perform confidence verification on the candidate labels labeled with the supporting data to obtain the target profile labels, which include expert profile labels and project profile labels. The storage module is used to store the target image tags and the supporting data in a traceable manner; The step of preprocessing the original data to obtain a structured dataset includes: Extract the unstructured data and the first structured data from the original data to be processed; Construct a first prompt word template based on the first structured data; Based on the first prompt word template, a set of candidate entities is extracted from the unstructured data using a large language model, and a set of candidate entity pairs is generated based on the set of candidate entities. Calculate the confidence score of the candidate entity pairs in the candidate entity pair set; The target entity pair set is obtained by filtering from the candidate entity pair set based on the confidence score and the first confidence threshold; The target entity pair set and the first structured data are combined to form the structured dataset; The step of performing confidence verification on candidate labels labeled with the supporting data to obtain target profile labels includes: Calculate the time decay factor of the supporting data based on the timestamp data corresponding to the supporting data; Multiply the time decay factor by the embedded representation of the supporting data to obtain the weighted embedded representation of the supporting data; Calculate the average similarity between the weighted embedding representation of the supporting data and the embedding representation of the candidate labels; The candidate labels are filtered based on the average similarity and the second confidence threshold to obtain the target profile labels; The time decay factor of the support data is calculated based on the timestamp data corresponding to the support data. The calculation process is as follows: ； In the formula, Indicates the current time; This indicates the candidate label y The corresponding number p The timestamp attribute value of the supporting data item; The attenuation coefficient; The candidate tag corresponds to the first p The time decay factor of the supporting data.

6. An electronic device, characterized in that, include: At least one processor; At least one memory for storing at least one program; When the at least one program is executed by the at least one processor, the at least one processor implements the method as described in any one of claims 1 to 4.

7. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by a processor, it implements the method of any one of claims 1 to 4.

Images