A literature review generation method, system and electronic device

Abstract

The application discloses a literature review generation method and system and an electronic device. When a literature review generation request is obtained, matched literature information and review structure information are determined to ensure that a target literature review finally generated can have a structure conforming to the review structure information, so that the target literature review information generated is concise and logically clear. A matched model is selected from multiple models to generate the target literature review, so that the review generation can be performed by using a more matched model, and the accuracy of the target literature review generated is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of computer, in particular to a literature review generation method and system and electronic equipment. BACKGROUND

[0002] With the deepening of medical research, the number of scientific research literature is increasing, and in the face of a large number of literature, scientific researchers usually face the problem of information overload. For this problem, the way of literature review can be used for literature screening, content extraction and correlation analysis.

[0003] However, the traditional literature review generation scheme generally has problems such as information redundancy and unclear logic. SUMMARY

[0004] Therefore, the present application provides a literature review generation method, system and electronic equipment, and the specific scheme is as follows:

[0005] A literature review generation method comprises:

[0006] Obtaining a literature review generation request, the literature review generation request is used to generate a target literature review;

[0007] Determining literature information and review structure information matched with the target literature review based on the literature review generation request;

[0008] Selecting a target model matched with the review structure information and target model parameters of the target model from at least two models, different models are used to generate literature reviews with different structural characteristic paragraphs;

[0009] Inputting the literature information and review structure information into a target model with model parameters set to the target model parameters, and obtaining the target literature review output by the target model.

[0010] Further, the determining of the literature information and review structure information matched with the target literature review based on the literature review generation request comprises:

[0011] Determining user intention prompt words and structure prompt words based on the literature review generation request;

[0012] Determining a target number of literatures matched with the user intention prompt words from a literature library;

[0013] Determining literature prompt words of the target number of literatures, and determining the user intention prompt words and the literature prompt words as the literature information matched with the target literature review;

[0014] determine structured chapter information of the target literature review based on the structure prompt word and the user intention prompt word, and determine the structured chapter information as review structure information matched with the target literature review.

[0015] Further, the structure prompt word is determined, including:

[0016] obtain prompt information of user input with target authority;

[0017] configure the prompt information according to a standardized structure to obtain the structure prompt word.

[0018] Further, the selection of the target model matched with the review structure information from the at least two models and the target model parameter of the target model include:

[0019] obtain model information and model temperature parameter determined based on the review structure information representing structured chapter information, and a three-dimensional scheduling matrix, dimensions of the three-dimensional scheduling matrix including: structured chapter dimension, model dimension and model temperature parameter dimension;

[0020] determine a target model from the at least two models based on the model information;

[0021] set a temperature parameter of the target model based on the model temperature parameter to obtain a target model with the temperature parameter set as the model temperature parameter.

[0022] Further, it further includes:

[0023] score each paragraph content in the target literature review output by the target model based on a specific evaluation index to obtain a scoring result of each paragraph content.

[0024] Further, it further includes:

[0025] if the scoring result of the first paragraph content in the target literature review represents that the first paragraph content meets a target condition, determine the first paragraph content as a first paragraph content in a final version of the target literature review;

[0026] if the scoring result of the second paragraph content in the target literature review represents that the second paragraph content does not meet the target condition, regenerate the second paragraph content.

[0027] Further, if the scoring result of the second paragraph content in the target literature review represents that the second paragraph content does not meet the target condition, the second paragraph content is regenerated, including:

[0028] if the score result of the second paragraph content in the target literature review represents that the second paragraph content does not meet the target condition, output and display the second paragraph content;

[0029] obtain a manual score of the second paragraph content output and displayed by a user with a target authority;

[0030] regenerate the second paragraph content based on the manual score.

[0031] Further, the method further comprises at least one of the following:

[0032] obtain literature information and review structure information input by a user with a target authority and matched with the target literature review;

[0033] determine a target model input by a user with a target authority.

[0034] A literature review generation system comprises:

[0035] a first obtaining unit configured to obtain a literature review generation request, the literature review generation request being used to generate a target literature review;

[0036] a determining unit configured to determine literature information and review structure information matched with the target literature review based on the literature review generation request;

[0037] a selecting unit configured to select a target model matched with the review structure information and target model parameters of the target model from at least two models, different models being used to generate literature reviews with different structural characteristic paragraphs;

[0038] a second obtaining unit configured to input the literature information and review structure information to a target model with model parameters set as the target model parameters, and obtain the target literature review output by the target model.

[0039] An electronic device comprises:

[0040] a processor configured to obtain a literature review generation request, the literature review generation request being used to generate a target literature review; determine literature information and review structure information matched with the target literature review based on the literature review generation request; select a target model matched with the review structure information and target model parameters of the target model from at least two models, different models being used to generate literature reviews with different structural characteristic paragraphs; input the literature information and review structure information to a target model with model parameters set as the target model parameters, and obtain the target literature review output by the target model;

[0041] A memory is configured to store programs required by the processor to perform the above-mentioned processing process.

[0042] A readable storage medium is configured to store at least a set of instruction sets.

[0043] The instruction sets are configured to be invoked and at least perform the literature review generation method according to any one of the above.

[0044] From the above technical solutions, it can be seen that the literature review generation method, system and electronic device disclosed by the present application obtain a literature review generation request, the literature review generation request is used to generate a target literature review; determine literature information and review structure information matched with the target literature review based on the literature review generation request; select a target model matched with the total structure information and target model parameters of the target model from at least two models, different models are used to generate literature reviews with different structural characteristics; input the literature information and the review structure information into the target model with the model parameters set to the target model parameters, and obtain the target literature review output by the target model. The present application determines the matched literature information and review structure information when obtaining the literature review generation request, so as to ensure that the target literature review generated finally can have a structure conforming to the review structure information, thereby ensuring that the target literature review information generated is concise and the logic is clear; and the matched model is selected from multiple models to generate the target literature review, so as to ensure that the review generation can be performed by using a more matched model, and the accuracy of the generated target literature review is improved. BRIEF DESCRIPTION OF DRAWINGS

[0045] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the related art, the drawings needed to be used in the embodiments or the related art description will be briefly introduced. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0046] Figure 1 A flowchart of a literature review generation method disclosed by an embodiment of the present application;

[0047] Figure 2 A flowchart of a literature review generation method disclosed by an embodiment of the present application;

[0048] Figure 3 A schematic diagram of a literature review generation method disclosed by an embodiment of the present application;

[0049] Figure 4 A schematic diagram of determination of structured chapter information disclosed by an embodiment of the present application;

[0050] Figure 5A schematic diagram of a literature review generation method disclosed in an embodiment of the present application;

[0051] Figure 6 A flowchart of a literature review generation method disclosed in an embodiment of the present application;

[0052] Figure 7 A flowchart of a literature review generation method disclosed in an embodiment of the present application;

[0053] Figure 8 A schematic diagram of a literature review generation method disclosed in an embodiment of the present application;

[0054] Figure 9 A schematic diagram of a user with target permissions for configuration disclosed in an embodiment of the present application;

[0055] Figure 10 A structural schematic diagram of a literature review generation system disclosed in an embodiment of the present application;

[0056] Figure 11 A structural schematic diagram of an electronic device disclosed in an embodiment of the present application. DETAILED DESCRIPTION

[0057] The embodiments of the present application are described below in conjunction with the accompanying drawings. The terms used in the implementation part of the present application are only used to explain the specific embodiments of the present application, and are not intended to limit the present application.

[0058] The embodiments of the present application are described below in conjunction with the accompanying drawings. It is known to those of ordinary skill in the art that, as technology develops and new scenarios appear, the technical solutions provided by the embodiments of the present application are also applicable to similar technical problems.

[0059] The terms "first", "second", and the like in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and do not necessarily describe a specific order or sequence. It should be understood that the terms used in this way can be interchanged under appropriate circumstances, and this is only a distinguishing way used in the description of the embodiments of the present application to describe the objects with the same attributes. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion, so that the processes, methods, systems, products or devices containing a series of units do not necessarily limit to those units, but can include other units not clearly listed or inherent to these processes, methods, products or devices.

[0060] The present application discloses a literature review generation method, a flowchart thereof is shown as Figure 1 , including:

[0061] Step S11, obtain a literature review generation request, the literature review generation request is used for generating a target literature review;

[0062] Step S12, determine literature information and review structure information matched with the target literature review based on the literature review generation request;

[0063] Step S13, select a target model matched with the review structure information and target model parameters of the target model from at least two models, different models are used for generating literature reviews with different structural characteristics;

[0064] Step S14, input the literature information and the review structure information into the target model with the model parameters set as the target model parameters, and obtain a target literature review output by the target model.

[0065] With the deepening of medical research, the number of scientific research literature is increasing, and in the face of a large number of literature, scientific researchers usually face the problem of information overload, and for this problem, the literature review method can be used for literature screening, content extraction and correlation analysis.

[0066] Among them, the literature review is a kind of academic paper that collects a large number of related materials on a certain field, a certain profession or a certain aspect of the subject, problem or research topic, and makes a comprehensive introduction and exposition on the latest progress, academic insights or suggestions of the subject, problem or research topic through reading, analysis, induction and arrangement.

[0067] However, the traditional literature review generation scheme generally has problems of information redundancy and unclear logic.

[0068] Based on this, the present scheme discloses a literature review generation method, after obtaining a literature review generation request, the literature information and the review structure information are determined, and the matched target model and the target model parameters of the target model are selected based on this, so as to generate the target literature review by using the target model with the model parameters set as the target model parameters, by pre-determining the review structure information and selecting the corresponding model and model parameters, so that the target literature review finally generated by the selected model and model parameters can conform to the review structure information, and the structure of the target literature review is clear and the logic is clear.

[0069] The literature review generation method disclosed in this embodiment is applied to a literature review generation system. When a literature review needs to be generated, a user can input a literature review generation request to the literature review generation system, so that the system can generate a target literature review required by the user based on the request. For example: the user inputs "please generate a literature review related to XX", or "please generate a literature review of X direction in XXX field".

[0070] After obtaining the literature review generation request, the literature review generation request can be analyzed to determine that the literature review generation request requests to generate a literature review of a certain specific type or topic, which can be determined as the target literature review;

[0071] Based on the literature review generation request, the literature information and review structure information matching the target literature review are determined, i.e. the relevant information of the literature corresponding to the current literature review of the specific type or topic to be generated (target literature review), such as: what are the specific type or topic of literature, what is the main content, what is the topic, etc. The review structure information determined based on the literature review generation request is the structure usually used in the literature review of the specific type or topic, and the structure can be: the chapter structure of the review, such as: primary topic, secondary topic, etc.

[0072] In addition, the system corresponding to the literature review generation method disclosed in the embodiment can be provided with multiple models, each of which is used to generate a literature review. Different models can be used to generate literature reviews with different structural feature paragraphs. Different models can be selected for literature reviews with different structural feature paragraphs, i.e. each model can be used to generate a complete literature review. Different models are used to generate literature reviews with different structural feature paragraphs, or each model is used to generate a paragraph in a literature review, and different models are used to generate paragraphs with different structural features in a literature review.

[0073] Then, after determining the review structure information matching the target literature review, one or more models can be selected as target models based on the review structure information, and the target literature review can be generated using the target models to ensure that the target literature review generated can meet the type and structure of the review requested in the literature review generation request.

[0074] Of course, when multiple models are provided in the system, multiple models can also be used to generate literature reviews at the same time, and then one of the multiple literature reviews generated by the multiple models can be selected as the target literature review. The selection process can be automatically scored by the system or manually selected to ensure that the target literature review finally determined is the literature review that meets the demand more among the multiple literature reviews.

[0075] In addition, the target model parameters of the target model can also be determined based on the review structure information. Different review structure information can correspond to different values of model parameters in the model, or different positions or chapters in the same review structure can correspond to different values of model parameters in the model.

[0076] After determining the target model and the target model parameters of the target model, the parameters in the target model are configured according to the target model parameters, and a completed target model is obtained. Then, the literature information and the summary structure information determined based on the literature review generation request are input into the completed target model, so that the target model can generate a target literature review based on the literature information and the summary structure information. The content in the target literature review is matched with the literature information, and the structure in the target literature review is matched with the summary structure information, so as to ensure that the obtained target literature review has clear structure and clear content.

[0077] The literature review generation method disclosed in the embodiment obtains a literature review generation request, the literature review generation request being used for generating a target literature review. Literature information and summary structure information matched with the target literature review are determined based on the literature review generation request. A target model and target model parameters of the target model matched with the summary structure information are selected from at least two models. Different models are used for generating literature reviews of different structure characteristic paragraphs. The literature information and the summary structure information are input into the target model with the model parameters set as the target model parameters, and a target literature review output by the target model is obtained. In this scheme, the matched literature information and the summary structure information are determined when the literature review generation request is obtained, so as to ensure that the target literature review generated finally has a structure conforming to the summary structure information, thereby ensuring that the target literature review generated has concise information and clear logic. The matched model is selected from multiple models for generating the target literature review, so as to ensure that the model more matched can be used for generating the review, and the accuracy of the target literature review generated is improved.

[0078] The literature review generation method disclosed in the embodiment has a flowchart as shown in Figure 2 The literature review generation method disclosed in the embodiment has a flowchart as shown in

[0079] In step S21, a literature review generation request is obtained, the literature review generation request being used for generating a target literature review.

[0080] In step S22, user intention prompt words and structure prompt words are determined based on the literature review generation request.

[0081] In step S23, a target number of literatures matched with the user intention prompt words are determined from a literature database.

[0082] In step S24, literature prompt words of the target number of literatures are determined, and the user intention prompt words and the literature prompt words are determined as literature information matched with the target literature review.

[0083] In step S25, structured chapter information of the target literature review is determined based on the structure prompt words and the user intention prompt words, and the structured chapter information is determined as summary structure information matched with the target literature review.

[0084] Step S26, selecting a target model matching the review structure information from at least two models, and target model parameters of the target model, different models are used to generate literature reviews of different structural characteristics paragraphs;

[0085] Step S27, inputting the literature information and the review structure information into the target model with the model parameters set as the target model parameters, and obtaining the target literature review output by the target model.

[0086] After obtaining the literature review generation request, determining the literature information and the review structure information matching the target literature review based on the literature review generation request, the matched target model and target model parameters can be determined based on the review structure information, and the literature information and the review structure information are input into the target model set as the target model parameters, so as to finally obtain the target literature review output by the target model, so that the output target literature review can have clear logic and structure.

[0087] Wherein, when determining the literature information and the review structure information matching the target literature review based on the literature review generation request, it can be determined by prompting words. Specifically, based on the literature review generation request, the user intention prompt word and the structure prompt word are determined, the target number of literatures matching the user intention prompt word is determined from the literature library, the literature prompt word of the target number of literatures is determined, and the user intention prompt word and the literature prompt word are determined as the literature information matching the target literature review. The structured chapter information of the target literature review is determined based on the structure prompt word and the user intention prompt word, and is determined as the review structure information.

[0088] After obtaining the literature review generation request, through the analysis of the literature review generation request, it can be determined what the user wants, which can include literature information and review structure information.

[0089] Wherein, the literature information can at least include: user intention and literature needed, at this time, both the user intention and the literature needed can be displayed in the form of prompt words, that is, user intention prompt word and literature prompt word, so as to directly input the prompt word into the target model, so that the target model can directly generate the literature review based on the prompt word.

[0090] Specifically, the schematic diagram of the literature review generation method disclosed in the embodiment can be as shown in Figure 3 First, the user inquiry is carried out, that is, the user inputs the literature review generation request, then the user intention prompt word, the literature prompt word and the structured chapter information are obtained through three branches based on the literature review generation request, and finally the determined user intention prompt word, the literature prompt word and the structured chapter information are input into the selected target model, and the weighted comparison is carried out, and finally the target literature review is output.

[0091] The user intention prompt word can be determined by semantic analysis of the literature review generation request, that is, receiving the user input literature review generation request through the user interface, parsing the keywords through the system, that is, obtaining the user intention prompt word. Specifically, the user input literature review generation request can be analyzed by a semantic analysis model to obtain the user intention prompt word. Specifically, the semantic analysis model can be an embedding model (Embedding Model). The embedding model outputs an embedding vector of the literature review generation request. Based on the embedding vector, the user intention prompt word is obtained.

[0092] The literature prompt word can be obtained by calling the API interface of the system to crawl the related literature in the academic database (PubMed) in real time, and then vectorizing the literature title, abstract, structure, etc. by the embedding model (Embedding Model) to construct a semantic index library (open source vector retrieval library FAISS or open source vector database Milvus). The embedding vector obtained when determining the user intention prompt word is matched with the vector in the semantic index library, so as to determine the matched literature from the literature library, and then generate the literature vector of the determined matched literature. When the number of determined matched literature is large, the determined matched literature is sorted according to the similarity, and the target number of literature with high ranking is determined as the final selected literature, and the literature vector of the final selected literature is determined.

[0093] The structured chapter information can be determined by the user input literature review generation request, that is, determining the user intention prompt word and the structure prompt word based on the literature review generation request, and then determining the structured chapter information based on the structure prompt word and the user intention prompt word. Alternatively, the structured chapter information can be determined by the review structure requirement input by the user. If the structured chapter information is determined by the user input literature review generation request, the semantic analysis of the literature review generation request is also needed to determine the semantic content related to the structure, and the embedding vector of the structure is determined to finally obtain the structured chapter information, that is, the structure information of the target literature review refined to the chapter (such as primary theme, secondary theme, etc., which represents different levels of chapters through different levels of themes). If the structured chapter information is determined by the review structure requirement input by the user, the review structure requirement can be analyzed to determine the structure requirement and further determine the embedding vector of the structure, and finally obtain the structured chapter information.

[0094] Further, the determination of the structure prompt word can be specifically as follows:

[0095] The prompt information input by the user with the target permission is obtained, and the prompt information is configured according to the standardized structure to obtain the structure prompt word.

[0096] The user with target permission (such as an expert) directly inputs prompt information, analyzes the prompt information and extracts keywords, and then configures the extracted keywords according to a standardized structure (such as a YAML structure) to obtain a configured structure prompt word, that is, the related information of the structure prompt word is directly input by the expert to ensure the accuracy and effectiveness of the structure prompt word.

[0097] The structured chapter information can also have a nested calling mechanism, each chapter can call upstream semantic output as prior input to realize logical transmission of context and ensure the coherence between chapters in the final output of the target literature review.

[0098] In addition, it should be noted that the embedding vector of the structure can also be referred to when determining the user intention prompt word, that is, different structures can correspond to different user intentions, and each structure can correspond to one or more user intentions. The embedding vector of the structure can be used as a reference parameter for the user intention prompt word.

[0099] In the literature review generation method disclosed in the embodiment, the content coherence, fact support and structure consistency of the final generated target literature review are systematically guaranteed through the multiple coordination mechanism of the prompt word; the scheme disclosed in the embodiment supports the nested calling of the prompt word and the chapter semantic inheritance mechanism, combines the structured chapter information, and generates the literature review by fusing real-time literature information. For literature prompt words, based on the PubMed database, the Embedding database and the semantic recall technology are built to make the model calling content not only meet the semantic matching, but also have timeliness and data support capability.

[0100] The specific process of determining the structured chapter information can be as shown in Figure 4 Figure 4 Taking the disease A related literature as an example, after obtaining the user input literature review generation request or review structure requirement, the retrieval keywords can be determined, and the retrieval keywords are used to retrieve from a large number of literatures. The retrieval results are preprocessed (such as: removing duplicate literatures; deleting literatures not belonging to the target type, the target type can be a research article or a review; deleting literatures in which the retrieval keywords are located in non-target parts, the target part can be the title and abstract, that is, only literatures in which the retrieval keywords are located in the title and abstract are retained), to obtain a data set. After removing the stop words in the data set, the literature analysis is performed, that is, the keyword co-occurrence analysis, timeline visualization and salience detection analysis are performed, and then the structured chapter information is determined by comparing the MeSH subject heading table and combining the expert opinions.

[0101] ​Specifically, this involves a systematic investigation of the etiology, clinical pathways, and public health control strategies for disease A. Bibliometric analysis is used to mine high-frequency keywords in PubMed and CNKI databases over a specific period. Based on the MeSH thesaurus and expert opinions, four primary themes are identified: clinical research, basic research, public health research, and epidemiological and etiological research. Guided by expert consensus (two rounds of questionnaires using the Delphi method), secondary themes are further refined based on clinical diagnosis and scientific research pathways, such as early screening and diagnosis of disease A, molecular biomarkers, and early diagnosis. A structural knowledge graph can be built using the Neo4j graph database or the OWL ontology management tool. This structural knowledge graph is at least related to structured chapter information. Graph nodes can include: chapter names, keywords, and DOI links to typical literature. Each node can be bound to a structural tag (e.g., primary theme, secondary theme, etc.) as the basis for constructing prompt word templates when generating literature reviews.

[0102] This embodiment, when determining structured chapter information, drives information organization based on real-world diagnostic and treatment processes, incorporates the Delphi method for expert consensus revision, and embeds structural tags into the generation process. This ensures that model generation always revolves around key knowledge points, significantly enhancing the structure and professional credibility of the content. Taking disease A as an example, based on the clear etiology and three-tiered prevention system of disease A, a structured outline covering "etiology—screening—diagnosis—treatment—prognosis—policy" is constructed, supplemented by 19 sub-topics, such as... Figure 5 As shown, there are 19 secondary themes under each primary theme in clinical research, basic research, public health, and epidemiology. This structure closely aligns with clinical research practice, making the generated text logic clearer and the knowledge hierarchy more reasonable. Through expert consultation and multiple rounds of Delphi review to reach a consensus on the outline, the scientific and academic rigor of the structure is further guaranteed. This closed-loop method based on "clinical pathway—information organization—generation control" solves problems such as structural generalization, logical fragmentation, and lack of clinical value in existing automated review technologies, providing a replicable template for future automated medical review generation. It achieves a structure map-controlled generation mechanism for medical review writing, breaking the limitations of low controllability in traditional models that rely on spontaneous content organization, ensuring that the generated text has a complete logical structure, semantic consistency, and content aligned with the main lines of medical knowledge. Of course, the literature review generation method disclosed in this embodiment can still be used for generating literature reviews in non-medical fields.

[0103] The literature review generation method disclosed in the embodiment discloses a literature review generation method. After obtaining a literature review generation request, user intention prompt words and structure prompt words are determined based on the literature review generation request. A target number of literatures matching the user intention prompt words are determined from a literature library, and literature prompt words corresponding to the target number of literatures are determined. The user intention prompt words and the literature prompt words are determined as literature information matching the target literature review. The structured chapter information of the target literature review is determined based on the structure prompt words and the user intention prompt words, and is determined as review structure information matching the target literature review. The matched target model is determined by using the literature information and the review structure information, and the target literature review is output by using the target model. In this scheme, the user intention prompt words, the literature prompt words, and the structured chapter information are cooperatively input to finally output the target literature review, so that the output target literature review not only meets the user intention, but also meets the related information of the literature in the literature library, and meets the structured chapter information, thereby improving the accuracy of the determined target literature review.

[0104] The literature review generation method disclosed in the embodiment has a flowchart as shown in Figure 6 The literature review generation method disclosed in the embodiment has a flowchart as shown in

[0105] Step S61, obtaining a literature review generation request, the literature review generation request being used to generate a target literature review;

[0106] Step S62, determining literature information and review structure information matching the target literature review based on the literature review generation request;

[0107] Step S63, obtaining model information and model temperature parameters determined based on the review structure information representing the structured chapter information, the dimensions of the three-dimensional scheduling matrix including: the structured chapter dimension, the model dimension, and the model temperature parameter dimension;

[0108] Step S64, determining a target model from at least two models based on the model information;

[0109] Step S65, setting the temperature parameter of the target model based on the model temperature parameter, to obtain the target model with the temperature parameter set to the model temperature parameter;

[0110] Step S66, inputting the literature information and the review structure information into the target model with the model parameter set to the target model parameter, and obtaining the target literature review output by the target model.

[0111] After obtaining the literature review generation request, the matched literature information and review structure information are determined, and then the matched target model and the target model parameter of the target model are selected from a plurality of models based on the review structure information, so that the model parameter of the target model is set to the target model parameter, and the target literature review is output by using it.

[0112] The summary structure information can be specifically structured chapter information, that is, structure information including at least a chapter structure of the summary.

[0113] When the target model and the target model parameter are selected, the three-dimensional scheduling matrix can be set. The three-dimensional scheduling matrix is a data structure or a conceptual model for mathematical modeling of a scheduling problem, which integrates three most critical dimensions in the scheduling problem into a three-dimensional matrix, so as to comprehensively describe, analyze and optimize the resource allocation scheme. In the embodiment, the three dimensions of the three-dimensional scheduling matrix are: structured chapter dimension, model dimension and model temperature parameter dimension. The three-dimensional scheduling matrix takes the structured chapter information as an input variable, and outputs the model information and the model temperature parameter, so as to realize dynamic control and fine adaptation of the model behavior.

[0114] After obtaining the structured chapter information, the labels of the content of each chapter are determined by automatic keyword recognition and rule setting. The labels can include descriptive content, methodological content, explanatory content, predictive / prospective content and the like. The labels of each chapter are mapped to a plurality of models in the three-dimensional scheduling matrix, and the most suitable target model is determined, so as to ensure the accuracy and rigor of the content of the finally generated target literature review.

[0115] The model temperature parameter is a parameter directly affecting the determinacy and diversity of the generated model. The model generates logits of the original probability distribution when each word is generated. The model temperature parameter affects the original probability distribution. For example, when the temperature approaches 0 (low temperature), the weight of the high probability word is amplified, and the output tends to be determined. When the temperature is equal to 1 (medium temperature), the original probability distribution is maintained. When the temperature is greater than 1 (high temperature), the probability difference is compressed, and the low probability word is more likely to be selected.

[0116] In the embodiment, the method and conclusion paragraphs can use low temperature (such as 0.2-0.4) to ensure the rigor of expression. The discussion and trend paragraphs use medium temperature (such as 0.5-0.7) to improve the openness of language expression.

[0117] Specifically, the correspondence between the model corresponding to the disease A and the model temperature parameter can be as shown in Table 1.

[0118]

[0119] Table 1

[0120] Through the three-dimensional scheduling matrix of chapter-model-temperature control, each paragraph generation process is clear in task awareness, reasonable in parameter setting, and unified in expression style, thereby improving the term accuracy, paragraph logic, and language naturalness in the output target literature review; the three-dimensional scheduling matrix is constructed based on semantics, so that the generated content not only matches the professionalism of the model, but also has language style and logical consistency, greatly improving the coherence and semantic rigor inside and outside the paragraph.

[0121] By determining the content attributes of each chapter (such as clinical methods, basic research, policy review, etc.), the system can automatically select the most suitable language model (such as ChatGPT-4, MedPaLM2, MediGPT, etc.) for calling. At the same time, the model temperature parameter is adjusted according to the chapter task style for generating discussion and flexible paragraphs. This mechanism not only improves the style consistency and accuracy of the generated text, but also significantly reduces the cost of manual post-processing. Unlike existing multi-model calling mechanisms that lack collaborative scheduling logic, this mechanism embeds joint scheduling levels into the automatic generation chain, which is a key module for improving the quality of language generation in vertical fields. It establishes a parameter-adaptive chapter-level generation control strategy, greatly improving term precision, style consistency, and expression stability.

[0122] The literature review generation method disclosed in the embodiment, after obtaining a literature review generation request and determining matching literature information and review structure information, obtains model information and model temperature parameters determined based on the review structure information representing structured chapter information from a three-dimensional scheduling matrix, determines a target model from at least two models based on the model information, sets the temperature parameter of the target model based on the model temperature parameter, to obtain the target model with the temperature parameter set to the model temperature parameter, and outputs the target literature review based on the target model. The scheme determines the target model and the model temperature parameter through the three-dimensional scheduling matrix, ensures the accuracy of the determined target model and the model temperature parameter, and sets the corresponding temperature parameter for the target model, ensuring the accuracy of the target literature review generated by the target model.

[0123] The embodiment discloses a literature review generation method, and a flowchart thereof is as shown in Figure 7 The embodiment discloses a literature review generation method, and a flowchart thereof is as shown in

[0124] Step S71, obtaining a literature review generation request, the literature review generation request being used for generating a target literature review;

[0125] Step S72, determining literature information and review structure information matched with the target literature review based on the literature review generation request;

[0126] Step S73, selecting a target model matching the review structure information from at least two models, and target model parameters of the target model, different models being used to generate literature reviews of different structural characteristics paragraphs;

[0127] Step S74, inputting the literature information and the review structure information into the target model with the model parameters set as the target model parameters, to obtain a target literature review output by the target model;

[0128] Step S75, scoring each paragraph content in the target literature review output by the target model based on a specific evaluation index, to obtain a scoring result of each paragraph content.

[0129] After obtaining the literature review generation request, the matching literature information and review structure information are determined, and then the matching target model and target model parameters of the target model are selected from multiple models based on the review structure information, so as to set the model parameters of the target model as the target model parameters, and output the target literature review.

[0130] In this embodiment, after outputting the target literature review by the target model, each paragraph content in the target literature review needs to be scored to determine the scoring result of each paragraph in the target literature review output by the target model, so as to facilitate subsequent improvement of the target model, or to obtain a target literature review in which each paragraph meets the scoring standard.

[0131] Each paragraph content in the target literature review can be scored by a specific evaluation index, which can include a bilingual evaluation tool BLEU, a recall-based summary evaluation tool ROUGE-L, and a term frequency-inverse document frequency TF-IDF. The bilingual evaluation tool BLEU is mainly used to evaluate the quality of machine translation, and the translation accuracy is measured by calculating the n-gram matching degree of the generated text and the reference text. The recall-based summary evaluation tool ROUGE-L is an index for evaluating text generation tasks, and the similarity of the longest common subsequence between the generated text and the reference text is calculated to measure the coherence and consistency of the generated content and human text. The term frequency-inverse document frequency TF-IDF is a basic tool in text analysis, commonly used for feature extraction and document classification, and by calculating the inverse relationship between term frequency and document set, it identifies the word combination with discriminability.

[0132] Further, scoring each paragraph in the target literature review can obtain a scoring result of each paragraph. If the scoring result of the first paragraph content in the target literature review indicates that the first paragraph content meets the target condition, the first paragraph content is determined as the first paragraph content in the final draft of the target literature review. If the scoring result of the second paragraph content in the target literature review indicates that the second paragraph content does not meet the target condition, the second paragraph content is regenerated.

[0133] That is, after the target model outputs the target literature review, the target literature review is scored so as to determine the final version of the target literature review based on the scoring results, and each paragraph content in the final version meets the scoring standards, that is, meets the target conditions. Therefore, when the scoring result of scoring a certain paragraph content indicates that it meets the scoring standards (i.e., meets the target conditions), the paragraph content can be directly determined as one of the paragraph contents in the final version of the target literature review; when the scoring result of scoring a certain paragraph content indicates that it does not meet the scoring standards (i.e., does not meet the target conditions), the paragraph content cannot be directly used as a paragraph content in the final version of the target literature review, and the paragraph content needs to be regenerated.

[0134] Further, for regenerating a certain paragraph content (such as the second paragraph content), it can be specifically: if the scoring result of the second paragraph content in the target literature review indicates that the second paragraph content does not meet the target conditions, the second paragraph content is output and displayed; obtaining the artificial score of the user with target authority for the output and displayed second paragraph content; and regenerating the second paragraph content based on the artificial score.

[0135] That is, if a certain paragraph content in the target literature review is scored by a specific evaluation index, and the automatic scoring result indicates that it does not meet the target conditions, then the re-scoring can be continued by the artificial scoring method to avoid the error of the automatic scoring result.

[0136] Artificial scoring, that is, a user with target authority (such as an expert) scores a certain paragraph content. If the scoring result of the second paragraph content in the automatic scoring process indicates that it does not meet the target conditions, the second paragraph content is displayed so that the expert can see the second paragraph content through the display screen and artificially score it. If the artificial scoring result indicates that the second paragraph content meets the target conditions, the second paragraph content can be determined as the second paragraph content in the final version of the target literature review, and the second paragraph content does not need to be regenerated; if the artificial scoring result indicates that the second paragraph content does not meet the target conditions, the second paragraph content is regenerated.

[0137] Among them, for the paragraph content whose scoring result indicates that it does not meet the target conditions, the generation path (including the used model, prompt word content and model temperature parameter information) when generating the target literature review can be retained, and the multi-path backtracking generation process is automatically started. At this time, the structured chapter information can be replaced or the domain prompt (such as guiding the model to refer to authoritative guidelines) can be added; or a more professional model can be switched to, such as switching from ChatGPT-4 to MedPaLM2 or MediGPT; or the model temperature parameter can be adjusted, such as adjusting from 0.6 to 0.3 to enhance the stability of expression.

[0138] The plurality of regenerated versions can be generated in the above manner, and a paragraph version tree is constructed, each version node records a source path and a score result, the paragraph version tree adopts a hierarchical identifier and a score index structure, and path tracing and version rollback are supported, so as to finally select the version with the highest score as the final draft, and realize semantic optimization and structural stability improvement of the generated target literature review result.

[0139] The schematic diagram of the literature review generation method disclosed in the embodiment can be as shown in Figure 8 After automatic scoring, the paragraph content that does not meet the target condition is manually scored, and the paragraph content that still does not meet the target condition after manual scoring is regenerated. In the regeneration process, model switching, model temperature parameter switching, and structured chapter information switching can be involved. Finally, the paragraph version tree is obtained, and the optimal paragraph content is selected to achieve the purpose of optimizing the target literature review generated by the target model.

[0140] In the embodiment, the two scoring methods of automatic evaluation and manual scoring are used to avoid scoring errors, improve the accuracy of scoring, and ensure the accuracy of the content in the final draft of the target literature review.

[0141] The literature review generation method disclosed in the embodiment obtains a literature review generation request, determines matching literature information and review structure information, selects a matching target model and target model parameters from at least two models, and outputs a target literature review using the target model set as the target model parameters. Then, each paragraph content in the target literature review needs to be scored to obtain the score result of each paragraph content, so as to determine whether each paragraph in the target literature review generated by the model meets the requirements, so as to adjust the target model based on the score result, thereby ensuring the accuracy of the finally output target literature review.

[0142] Further, the literature review method disclosed in the embodiment can further include at least one of the following:

[0143] Obtaining literature information and review structure information matching the target literature review input by a user with target permission; determining a target model input by a user with target permission.

[0144] That is, the document information and the review structure information, or the target model is not directly or indirectly determined by the document review generation request, but is directly input by a user (such as an expert) with target authority, to ensure the professional reliability and low threshold adaptability of the system, and to enable the user with target authority to participate in the process of generating the target document review, and to perform soft control on the generation behavior of the model in a structured manner.

[0145] Specifically, as shown in Figure 9 , the user (such as an expert) with target authority can configure prompt words, control models, and give score suggestions, so as to output the target document review with expert parameters.

[0146] Among them, the information configured by the user with target authority in a structured manner can include: semantic tags corresponding to chapter structures, recommended term sets, document source preferences, generation styles corresponding to chapters, etc. The configuration will be automatically mapped into the system prompt word logic, and support priority ordering and conditional triggering, to realize the structural control of the expert intention; in addition, the system establishes a generation scheduling system driven by expert weights, and the expert can set weight suggestions for prompt words, models, and model temperature parameters. The system dynamically rearranges the generation path priority in the generation process combined with score feedback and weight levels, and tracks the influence of expert suggestions on the results to form a feedback loop; finally, the system designs a unified API adaptation layer, supports remote or local calling of models such as OpenAI, Google MedPaLM, and DeepSeek, supports intelligent matching of models according to chapter tasks and resource states, and realizes the task-aware model adaptation strategy. It realizes the transformation from “expert review” to “expert regulation”, enhances the professional response ability and control accuracy of the system, and is a key technical path to improve the controllability and reliability of the language generation system in the vertical field.

[0147] The embodiment discloses a document review generation system, and a structural schematic diagram thereof is as shown in Figure 10 , which comprises:

[0148] The first obtaining unit 101, the determining unit 102, the selecting unit 103, and the second obtaining unit 104.

[0149] Among them, the first obtaining unit 101 is used for obtaining a document review generation request, and the document review generation request is used for generating a target document review;

[0150] The determining unit 102 is used for determining document information and review structure information matched with the target document review based on the document review generation request;

[0151] The selecting unit 103 is used for selecting a target model matched with the review structure information and a target model parameter of the target model from at least two models, and different models are used for generating document reviews with different structural characteristics.

[0152] The second obtaining unit 104 is configured to input the literature information and the review structure information into a target model with model parameters set as target model parameters, and obtain a target literature review output by the target model.

[0153] Further, the determining unit is configured to:

[0154] determine a user intention prompt word and a structure prompt word based on the literature review generation request, determine a target number of literatures matching the user intention prompt word from the literature library, determine literature prompt words of the target number of literatures, and determine the user intention prompt word and the literature prompt words as literature information matching the target literature review; determine structured chapter information of the target literature review based on the structure prompt word and the user intention prompt word, and determine the structured chapter information as review structure information matching the target literature review.

[0155] Further, the selecting unit is configured to:

[0156] obtain a three-dimensional scheduling matrix based on model information and model temperature parameters determined based on the review structure information representing the structured chapter information, the dimensions of the three-dimensional scheduling matrix including a structured chapter dimension, a model dimension, and a model temperature parameter dimension; determine a target model from at least two models based on the model information; and set a temperature parameter of the target model based on the model temperature parameters to obtain a target model with the temperature parameter set as the model temperature parameter.

[0157] Further, the literature review generation system disclosed in the embodiment can further include:

[0158] The scoring unit is configured to score each paragraph content in the target literature review output by the target model based on a specific evaluation index to obtain a scoring result of each paragraph content.

[0159] Further, the scoring unit is configured to:

[0160] If the scoring result of the first paragraph content in the target literature review indicates that the first paragraph content meets a target condition, the first paragraph content is determined as the first paragraph content in the final draft of the target literature review; and if the scoring result of the second paragraph content in the target literature review indicates that the second paragraph content does not meet the target condition, the second paragraph content is regenerated.

[0161] Further, the scoring unit is configured to:

[0162] If the scoring result of the second paragraph content in the target literature review indicates that the second paragraph content does not meet the target condition, the second paragraph content is output and displayed; an artificial score of the output and displayed second paragraph content by a user with a target authority is obtained; and the second paragraph content is regenerated based on the artificial score.

[0163] Further, the literature review generation system disclosed in the embodiment further comprises:

[0164] a third obtaining unit configured to obtain literature information and review structure information input by the user with the target permission and matched with the target literature review, and / or determine a target model input by the user with the target permission.

[0165] The literature review generation system disclosed in the embodiment is realized based on the literature review generation method disclosed in the above embodiment, and thus will not be described here.

[0166] The literature review generation system disclosed in the embodiment obtains a literature review generation request, the literature review generation request being used to generate a target literature review; determines literature information and review structure information matched with the target literature review based on the literature review generation request; selects a target model matched with the review structure information and target model parameters of the target model from at least two models, different models being used to generate literature reviews with different structural characteristic paragraphs; inputs the literature information and the review structure information into the target model with the model parameters set as the target model parameters, and obtains a target literature review output by the target model. The scheme determines the matched literature information and review structure information when obtaining the literature review generation request, so as to ensure that the finally generated target literature review can have a structure conforming to the review structure information, thereby ensuring that the generated target literature review information is concise and logically clear; and selects the matched model from the multiple models to generate the target literature review, so as to ensure that the review generation can be performed by using a more matched model, and the accuracy of the generated target literature review is improved.

[0167] The embodiment discloses an electronic device, a structural schematic diagram of which is shown in Figure 11 The electronic device comprises:

[0168] a processor 111 and a memory 112.

[0169] The processor 111 is configured to obtain a literature review generation request, the literature review generation request being used to generate a target literature review; determine literature information and review structure information matched with the target literature review based on the literature review generation request; select a target model matched with the review structure information and target model parameters of the target model from at least two models, different models being used to generate literature reviews with different structural characteristic paragraphs; input the literature information and the review structure information into the target model with the model parameters set as the target model parameters, and obtain a target literature review output by the target model.

[0170] The memory 112 is configured to store programs required by the processor to perform the above processing process.

[0171] The electronic device disclosed in the embodiment is implemented based on the literature review generation method disclosed in the above embodiment, and will not be described herein.

[0172] The electronic device disclosed in the embodiment obtains a literature review generation request, the literature review generation request being used for generating a target literature review; determines literature information and review structure information matched with the target literature review based on the literature review generation request; selects a target model matched with the total structure information and target model parameters of the target model from at least two models, different models being used for generating literature reviews with different structure characteristics; inputs the literature information and the review structure information into the target model with the model parameters being set as the target model parameters, and obtains a target literature review output by the target model. The scheme determines the matched literature information and the review structure information when the literature review generation request is obtained, so as to ensure that the target literature review generated finally can have a structure conforming to the review structure information, thereby ensuring that the target literature review information generated is concise and logically clear; and the matched model is selected from the multiple models for generating the target literature review, so as to ensure that the generation of the review can be performed by using a more matched model, and the accuracy of the target literature review generated is improved.

[0173] The embodiment of the present application further provides a readable storage medium, which has a computer program stored thereon, the computer program is loaded and executed by a processor, and each step of the above-mentioned literature review generation method is implemented, and the specific implementation process can be referred to the description of the corresponding part of the above-mentioned embodiment, and the embodiment will not be described herein.

[0174] The present application further provides a computer program product or a computer program, the computer program product or the computer program comprising computer instructions stored in a computer readable storage medium. The processor of the electronic device reads the computer instructions from the computer readable storage medium, and the processor executes the computer instructions, so that the electronic device executes the method provided in the various optional implementation manners of the above-mentioned literature review generation method aspect or the literature review generation system aspect, and the specific implementation process can be referred to the description of the above-mentioned corresponding embodiment, and will not be described herein.

[0175] In addition, it should be noted that the apparatus embodiments described above are only schematic, wherein the units as described in the units can or can not be physically separate, and the units as shown can or can not be physical units, that is, they can be located in one place, or can be distributed on multiple network units. Part or all of the modules can be selected according to actual needs to achieve the purpose of the embodiment. In addition, the connection relationship between the modules in the apparatus embodiment provided by the present application indicates that there is a communication connection between them, which can be realized as one or more communication buses or signal lines.

[0176] Those skilled in the art can clearly understand that the application can be implemented by means of software plus necessary universal hardware, and of course can also be implemented by means of dedicated hardware including special integrated circuit, special CPU, special memory, special component, etc. Generally, any function completed by computer program can be easily implemented by corresponding hardware, and the specific hardware structure for implementing the same function can also be various, such as analog circuit, digital circuit or special circuit, etc. However, for the application, software program implementation is a better embodiment. Based on such understanding, the technical solution of the application or the part of the application which makes contribution to the prior art can be embodied in the form of software product, which is stored in readable storage medium, such as computer floppy disk, U disk, mobile hard disk, ROM, RAM, magnetic disk or optical disk, etc., and includes a plurality of instructions for making a computer device (which can be personal computer, training device or network device, etc.) execute the method described in various embodiments of the application.

[0177] In the above embodiments, the implementation can be achieved by software, hardware, firmware or any combination thereof, entirely or partially. When implemented by software, the implementation can be achieved in the form of a computer program product, entirely or partially.

[0178] The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, the flow or function described in the embodiments of the application is generated entirely or partially. The computer can be a general-purpose computer, a special-purpose computer, a computer network or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another, for example, the computer instructions can be transmitted from one website, computer, training device or data center to another website, computer, training device or data center through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) mode. The computer-readable storage medium can be any available medium that can be stored by a computer or a data storage device such as a training device, a data center, etc. integrated with one or more available media sets. The available medium can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)), etc.

Claims

1. A method of generating a literature review, the method comprising: The method comprises the following steps: obtaining a literature review generation request for generating a target literature review; determining literature information and review structure information matched with the target literature review based on the literature review generation request; selecting a target model matched with the review structure information and target model parameters of the target model from at least two models, wherein different models are used to generate literature reviews with different structural feature paragraphs; inputting the literature information and review structure information into a target model with model parameters set as the target model parameters, and obtaining the target literature review output by the target model; wherein the step of selecting a target model matched with the review structure information and target model parameters of the target model from at least two models comprises the following steps: obtaining model information and model temperature parameters determined based on the review structure information representing structured chapter information; the dimension of the three-dimensional scheduling matrix comprises: structured chapter dimension, model dimension and model temperature parameter dimension; determining a target model from the at least two models based on the model information; and setting the temperature parameter of the target model based on the model temperature parameter to obtain a target model with temperature parameters set as the model temperature parameters.

2. The method of claim 1, wherein, The step of determining literature information and review structure information matched with the target literature review based on the literature review generation request comprises the following steps: determining user intention prompt words and structure prompt words based on the literature review generation request; determining a target number of literatures matched with the user intention prompt words from a literature library; determining literature prompt words of the target number of literatures, and determining the user intention prompt words and the literature prompt words as literature information matched with the target literature review; determining structured chapter information of the target literature review based on the structure prompt words and the user intention prompt words, and determining the structured chapter information as review structure information matched with the target literature review.

3. The method of claim 2, wherein, The step of determining structure prompt words comprises the following steps: obtaining prompt information input by a user with target authority; configuring the prompt information according to a standardized structure to obtain the structure prompt words.

4. The method of claim 1, wherein, The method further comprises the following steps: scoring each paragraph content in the target literature review output by the target model based on a specific evaluation index to obtain a scoring result of each paragraph content.

5. The method of claim 4, wherein, The method further comprises the following steps: if the scoring result of the first paragraph content in the target literature review indicates that the first paragraph content meets a target condition, determining the first paragraph content as the first paragraph content in a final version of the target literature review; if the scoring result of the second paragraph content in the target literature review indicates that the second paragraph content does not meet the target condition, regenerating the second paragraph content.

6. The method of claim 5, wherein, The step of regenerating the second paragraph content if the scoring result of the second paragraph content in the target literature review indicates that the second paragraph content does not meet the target condition comprises the following steps: if the score result of the second paragraph content in the target literature review represents that the second paragraph content does not meet the target condition, output and display the second paragraph content; obtain a manual score of the second paragraph content output and displayed by a user with a target authority; regenerate the second paragraph content based on the manual score.

7. The method of claim 1, wherein, The method further comprises at least one of the following: obtain literature information and review structure information input by a user with a target authority and matched with the target literature review; determine a target model input by a user with a target authority.

8. A literature review generation system characterized by, Comprise: a first obtaining unit, configured to obtain a literature review generation request, the literature review generation request being used to generate a target literature review; a determining unit, configured to determine literature information and review structure information matched with the target literature review based on the literature review generation request; a selecting unit, configured to select a target model matched with the review structure information and a target model parameter of the target model from at least two models, different models being used to generate literature reviews of different structural characteristic paragraphs; a second obtaining unit, configured to input the literature information and the review structure information into a target model with the model parameter set as the target model parameter, and obtain the target literature review output by the target model; wherein the selecting unit is configured to: obtain model information and model temperature parameters determined based on the review structure information representing structured chapter information; a three-dimensional scheduling matrix has dimensions including: a structured chapter dimension, a model dimension, and a model temperature parameter dimension; determine a target model from the at least two models based on the model information; and set a temperature parameter of the target model based on the model temperature parameter, to obtain a target model with the temperature parameter set as the model temperature parameter.

9. An electronic device, comprising: Comprise: a processor, configured to obtain a literature review generation request, the literature review generation request being used to generate a target literature review; determine literature information and review structure information matched with the target literature review based on the literature review generation request; select a target model matched with the review structure information and a target model parameter of the target model from at least two models, different models being used to generate literature reviews of different structural characteristic paragraphs; input the literature information and the review structure information into a target model with the model parameter set as the target model parameter, and obtain the target literature review output by the target model; wherein the selecting a target model matched with the review structure information and a target model parameter of the target model from at least two models specifically comprises: obtaining model information and model temperature parameters determined based on the review structure information representing structured chapter information; a three-dimensional scheduling matrix has dimensions including: a structured chapter dimension, a model dimension, and a model temperature parameter dimension; determining a target model from the at least two models based on the model information; and setting a temperature parameter of the target model based on the model temperature parameter, to obtain a target model with the temperature parameter set as the model temperature parameter. a memory for storing programs required by the processor to perform the above-mentioned processing process.

Images