An intelligent document assisted screening method and system based on human-computer collaborative iteration

Abstract

The application discloses a kind of intelligent document assisted screening method and system based on man-machine cooperation iteration, which comprises: obtaining target research theme and initial search condition, generating initial search strategy and obtaining candidate document from document library;Multi-dimensional feature analysis is carried out on the candidate document, the attribute is extracted to construct comprehensive evaluation model and is sorted, and the output is to be screened set;Through user interactive screening, according to feedback, screening intention is constructed and screening strategy is updated;Analyze positive and negative literature text features, calculate candidate keyword discrimination, and generate exclusion word set;After optimizing the search strategy, the search, screening, feedback and optimization are cycled until the termination condition is met, and the target document set is output.The application can improve the accuracy and efficiency of document screening and enhance the processing capacity when facing large-scale document data.The application can be widely applied in the field of information retrieval technology as an intelligent document assisted screening method and system based on man-machine cooperation iteration.

Description

Technical Field

[0001] This invention relates to the field of information retrieval technology, and in particular to an intelligent document-assisted screening method and system based on human-computer collaborative iteration. Background Technology

[0002] PubMed, the world's largest biomedical literature database, contains over 36 million articles and supports standardized search syntax such as Boolean logic operations, field limiting, and truncation search, making it the preferred tool for biomedical researchers to obtain literature information. However, traditional PubMed literature retrieval and screening methods still suffer from technical shortcomings such as redundant search results, low search efficiency, and unclear recommendations of relevant literature. In recent years, generative AI models (such as the GPT series and Claude) have been attempted to optimize literature retrieval and review generation. Analysis of existing patent and non-patent literature reveals that existing literature-assisted screening technologies still have the following common shortcomings, such as:

[0003] 1) Lack of specialized design for the characteristics of the biomedical literature field, and no optimization or adaptation for the biomedical field's unique knowledge system such as MeSH thesaurus, PubMed search syntax, and PICO structure;

[0004] 2) The retrieval process is a one-way linear model, which does not form an iterative closed loop of "retrieval-filtering-feedback-optimization" and lacks a mechanism to automatically adjust the retrieval strategy based on user feedback;

[0005] 3) Existing AI collaboration models rely too heavily on AI, lack sufficient user control, have a simplistic interaction feedback mechanism, and lack a multi-dimensional quantitative evaluation system, making it difficult to accurately capture users' selection and judgment information.

[0006] 4) Insufficient retrieval optimization capabilities; unable to automatically generate exclusion words based on user filtering behavior; lacks intelligent optimization functions to convert positive and negative semantics into Boolean queries.

[0007] 5) There is a lack of systematic and quantitative evaluation indicators for screening efficiency, making it impossible to objectively measure the performance differences between different retrieval and screening methods;

[0008] 6) There is a technological gap in the automatic discovery of exclusion words, and no mechanism is involved in automatically mining and recommending exclusion words from document features. Summary of the Invention

[0009] To address the aforementioned technical problems, the present invention aims to provide an intelligent literature screening method and system based on human-machine collaborative iteration, which can improve the accuracy and efficiency of literature screening and enhance the processing capability when faced with large-scale literature data.

[0010] The first technical solution adopted in this invention is: an intelligent literature-assisted screening method based on human-machine collaborative iteration, comprising the following steps:

[0011] The system obtains the target research topic information and initial literature search conditions input by the user, generates an initial search strategy based on the target research topic information and initial literature search conditions, and obtains candidate literature data from the literature database based on the initial search strategy.

[0012] Multidimensional feature analysis is performed on the candidate literature data to extract the literature attribute information in the candidate literature data, and a comprehensive literature evaluation model is constructed based on the literature attribute information. As needed, the candidate literature can be sorted by randomness, publication time and comprehensive effectiveness, and the set of literature to be screened is output.

[0013] The set of literature to be screened is provided to users for interactive screening. User feedback information on different literature inputs is obtained. User screening intent data is constructed based on the screening feedback information, and the literature screening strategy is updated based on the user screening intent data.

[0014] Text feature analysis is performed on the positive and negative literature sets corresponding to the screening feedback information to extract candidate keywords from the literature. The discriminative power of the candidate keywords is calculated based on the difference between the positive and negative literature sets. Based on the discriminative power, the candidate keywords are displayed in order for users to subjectively screen and confirm according to their specific research objectives and discriminative power, and a set of excluded words confirmed by the user is generated.

[0015] The initial search strategy is optimized based on the user-confirmed exclusion term set, generating a new search strategy. Candidate document data is then re-acquired based on the new search strategy. The process of document screening, user feedback, strategy optimization, and re-searching is repeated until the preset screening termination conditions are met, and the target document set is output.

[0016] Furthermore, the step of obtaining the target research topic information and initial literature search conditions input by the user, and generating an initial search strategy based on the target research topic information and initial literature search conditions, specifically includes:

[0017] It receives user input including research topic keywords, research subjects, disease names, intervention factors, comparison factors, research results information, and biomedical terms.

[0018] Construct corresponding Boolean logic expressions based on research topic keywords, and generate initial search queries by combining field limitation rules supported by literature databases;

[0019] Send the initial search query to the literature database interface, perform a literature search operation through the database search interface, and obtain the literature identifier data that matches the initial search query;

[0020] Based on the document identification data, the document data acquisition interface is called to obtain the corresponding document title, abstract, author, journal, publication year, digital object identifier, and citation information, forming an initial candidate document dataset.

[0021] Furthermore, the step of performing multidimensional feature analysis on the candidate literature data specifically includes:

[0022] The publication time data in the candidate literature data is used as the timeliness feature input, and the timeliness evaluation value of the literature is calculated based on the time difference between the current time and the publication time of the literature.

[0023] The journal impact data corresponding to the candidate literature is used as the input of journal impact characteristics to obtain journal impact indicators;

[0024] The citation count data corresponding to the candidate literature is used as the input of the citation impact feature, and the average annual citation evaluation value is calculated based on the relationship between the citation count and the publication time.

[0025] The timeliness evaluation value, journal impact index and annual average citation evaluation value are standardized and then integrated according to preset weights to obtain the comprehensive score of the literature.

[0026] Based on your needs, you can choose to sort the candidate documents in descending order by randomness, publication time, or comprehensive document score, and output the sorted document screening list.

[0027] Furthermore, the step of providing the set of documents to be screened to the user for interactive screening specifically includes:

[0028] The collection of literature to be screened is converted into interactive literature display data, and you can choose to sort it by randomness, publication time and comprehensive effectiveness as needed;

[0029] The interactive document display shows the title, author, journal, publication year, abstract, and overall score of each document.

[0030] Receive relevance feedback information from the user regarding the currently displayed literature, wherein the relevance feedback information includes at least the status of being relevant, irrelevant, and pending determination;

[0031] Based on the relevance feedback information, establish a positive literature set, a negative literature set, and a literature set to be confirmed.

[0032] Record the identification information of the literature that the user has already filtered, and generate historical filtering record data to avoid repeatedly displaying filtered literature in subsequent iterations.

[0033] Furthermore, the step of performing text feature analysis on the positive and negative literature sets corresponding to the filtered feedback information specifically includes:

[0034] Obtain title text data and abstract text data from the negative literature collection, perform text cleaning on the title text data and abstract text data to obtain cleaned text data;

[0035] The cleaned text data is processed for terminology recognition, and candidate core words are extracted by noun recognition rules, biomedical terminology dictionary matching, and compound word recognition.

[0036] Part-of-speech tagging is performed on candidate core words to obtain corresponding unary keywords, binary keywords and ternary keywords, and keywords at each granularity are output.

[0037] Using keywords at each granularity as input data, and combining their frequency of occurrence in the negative literature set and their absence in the positive literature set, candidate exclusion words are output to characterize the irrelevance of the literature.

[0038] Furthermore, the step of calculating the discriminative power of candidate keywords based on the difference between the positive and negative document sets specifically includes:

[0039] Statistical analysis of the local frequency of each candidate keyword in the negative literature set;

[0040] Count the number of negative documents containing candidate keywords and the number of positive documents not containing candidate keywords;

[0041] The exclusion priority of candidate keywords is calculated based on their local occurrence frequency, the number of negative documents covering them, and the number of positive documents missing them.

[0042] Candidate keywords are sorted from highest to lowest according to their exclusion priority, and different levels of candidate exclusion word recommendation lists are generated according to keyword type.

[0043] Output the candidate exclusion word recommendation list to the user interface for the user to select and confirm.

[0044] Furthermore, the step of optimizing the initial search strategy based on the user-confirmed exclusion term set specifically includes:

[0045] Get the set of excluded words after user confirmation;

[0046] Convert each exclusion term in the exclusion term set into a corresponding field-qualified query expression;

[0047] Construct exclusion query conditions based on the logical relationships between multiple exclusion terms in the exclusion term set;

[0048] Combine exclusion criteria with the original search strategy to generate a new search strategy that includes exclusion logic;

[0049] Save the new search strategy generated in the current round and the historical exclusion keyword set so that the cumulative optimization results can be continuously used in subsequent iterations.

[0050] Furthermore, it also includes efficiency evaluation of the iterative screening process, specifically including:

[0051] Get the number of retrieved documents, the number of documents confirmed as relevant by users, and the number of documents confirmed as irrelevant by users for the current iteration round;

[0052] The screening efficiency index is calculated based on the relationship between the current round of data and the previous round of data.

[0053] The browsing efficiency index is calculated based on the ratio between the number of user-selected results and the number of documents viewed.

[0054] The trend of revenue changes in the current iteration process is determined based on screening efficiency indicators and browsing efficiency indicators;

[0055] When the efficiency index meets the preset convergence condition, the iteration termination instruction is output, and the final target literature set is generated.

[0056] Furthermore, each iteration specifically includes:

[0057] Obtain new candidate literature data generated in the current round;

[0058] Filter literature data that has already undergone user evaluation based on historical screening records;

[0059] The unscreened literature data were reordered according to the updated comprehensive score;

[0060] Show users new documents to be screened based on the sorting results;

[0061] We continuously acquire user feedback and use the new feedback to update the user filtering intent model, so as to continuously optimize the intelligent search strategy based on human judgment.

[0062] The second technical solution adopted in this invention is: an intelligent document-assisted screening system based on human-machine collaborative iteration, comprising:

[0063] The first module is used to obtain the target research topic information and initial literature search conditions input by the user, generate an initial search strategy based on the target research topic information and initial literature search conditions, and obtain candidate literature data from the literature database based on the initial search strategy.

[0064] The second module is used to perform multidimensional feature analysis on candidate literature data, extract literature attribute information from candidate literature data, and construct a comprehensive literature evaluation model based on literature attribute information. As needed, it can sort each candidate literature according to randomness, publication time, and comprehensive effectiveness, and output a set of literature to be screened.

[0065] The third module is used to provide the set of documents to be screened to the user for interactive screening, obtain the user's screening feedback information for different documents, construct user screening intent data based on the screening feedback information, and update the document screening strategy based on the user screening intent data.

[0066] The fourth module is used to perform text feature analysis on the positive and negative literature sets corresponding to the screening feedback information, extract candidate keywords from the literature, calculate the discrimination of candidate keywords based on the difference between the positive and negative literature sets, and display the candidate keywords in order based on the discrimination, so that users can make subjective screening and confirmation based on their specific research objectives and discrimination, and generate a set of excluded words confirmed by the user.

[0067] The fifth module is used to optimize the initial search strategy based on the user-confirmed exclusion term set, generate a new search strategy, and re-acquire candidate document data based on the new search strategy. It iteratively executes the document screening, user feedback, strategy optimization, and re-search process until the preset screening termination conditions are met, and outputs the target document set.

[0068] The beneficial effects of the method and system of this invention are as follows: By constructing a human-machine collaborative iterative screening mechanism, this invention integrates user feedback on manual screening with the intelligent literature analysis process. It continuously optimizes the literature retrieval strategy and screening model using user screening results, overcoming the problems of traditional literature screening processes that rely primarily on manual reading and judgment, resulting in low screening efficiency and difficulty in adapting to complex changes in research topics. This allows the system to dynamically adjust the direction of literature screening based on the user's actual screening intentions, improving the accuracy and adaptability of the literature screening process. Furthermore, by comprehensively analyzing the time characteristics, journal impact characteristics, citation characteristics, and textual semantic characteristics of candidate literature, a multi-dimensional literature evaluation mechanism is established to achieve intelligent ranking and priority recommendation of candidate literature. This enables users to quickly locate highly relevant literature, reduce invalid literature browsing operations, and improve literature screening efficiency. Furthermore, by analyzing the positive and negative literature sets formed by user feedback, the system automatically extracts the core features affecting literature relevance and displays candidate keywords in order based on their discriminative power. Users can then subjectively select and confirm these keywords based on their specific research objectives and discriminative power, generating a user-confirmed exclusion keyword set. The search strategy is iteratively optimized based on user confirmation results, enabling the system to gradually learn user filtering preferences and achieve dynamic convergence of the search scope, reducing the workload of repeatedly adjusting search queries manually. Therefore, this embodiment achieves intelligent processing of literature data, automated optimization of the filtering process, and continuous adaptation to user needs. It not only improves the accuracy and efficiency of literature filtering but also enhances the system's processing capabilities when facing different research topics, complex search conditions, and large-scale literature data, providing users with a more accurate, efficient, and research-oriented set of target literature. Attached Figure Description

[0069] Figure 1 This is a flowchart of the steps of an intelligent literature-assisted screening method based on human-machine collaborative iteration according to the present invention;

[0070] Figure 2 This is a structural block diagram of an intelligent document-assisted screening system based on human-machine collaborative iteration according to the present invention;

[0071] Figure 3 This is a schematic diagram of the overall system architecture provided in a specific embodiment of the present invention;

[0072] Figure 4 This is a schematic diagram of the iterative screening closed-loop process provided in a specific embodiment of the present invention;

[0073] Figure 5 This is a schematic diagram of the exclusion word discovery and retrieval optimization process provided in a specific embodiment of the present invention. Detailed Implementation

[0074] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are only for ease of explanation and do not limit the order of the steps. The execution order of each step in the embodiments can be adapted according to the understanding of those skilled in the art.

[0075] First, the technical terms used in the embodiments of this invention will be explained:

[0076] 1) PubMed E-utilities API (Entrez Utilities Application Programming Interface): A set of HTTP-based web service interfaces provided by the National Center for Biotechnology Information (NCBI) for programmatic access to all NCBI databases (including PubMed). Developers can perform operations such as searching, downloading, and linking by sending standard URL requests, without having to manually access web pages.

[0077] 2) Esearch (E-utilities Search) and PMID (PubMed ID): A core tool in the E-utilities suite, used to perform searches in a specified database (such as PubMed) and return a list of unique identifiers for records that meet the criteria. Receiving the user's search query and returning a list of matching PMIDs (i.e., document IDs) is the first step in the search process.

[0078] 3) PMID (PubMed IDentifier): A unique numerical identifier assigned by PubMed to each indexed document (e.g., 41342206). It is used to accurately identify a document, and subsequent esummary / efetch requests rely on this ID to retrieve it.

[0079] 4) Esummary (E-utilities Summary): Based on the provided PMID list, returns the abstract information for each document (including title, author, journal, publication date, DOI, etc., but not the complete abstract text). Quickly obtain summary metadata of documents for initial display or filtering of detailed data.

[0080] 5) Efetch (E-utilities Fetch): Retrieves complete literature records based on the provided PMID list. The return format can be specified (e.g., XML, MEDLINE), including all fields such as abstract, MeSH terms, and references. It provides the most comprehensive literature data for in-depth analysis (e.g., keyword extraction, score calculation).

[0081] 6) DOI (Digital Object Identifier): This is an internationally recognized permanent digital identifier used to uniquely identify digital objects (such as academic papers) on the internet. It is the "ID number" of a document on the internet and has a wider range of applications than PMID.

[0082] Furthermore, it's important to note that with the development of large language model technology, it has been widely used in the automatic retrieval, abstract generation, and review writing of biomedical literature. While such solutions improve the efficiency of literature retrieval to some extent—AI can understand significant discrepancies between the brief description initially input by the user and the user's actual needs—experts cannot effectively control the retrieval process. This leads to a large number of documents in the search results that are irrelevant or only marginally relevant to the user's actual research intent. In the subsequent automatic review generation stage, this dilutes the effective information density, producing erroneous statements or false associations that contradict the user's needs, significantly reducing the accuracy and credibility of the generated content. This fails to meet the high professionalism and accuracy requirements of the scientific research industry for literature retrieval.

[0083] Based on this, the present invention designs an iterative screening mechanism for human-computer collaboration, which enables users to continuously transmit screening intentions and expert judgments to the system in the retrieval cycle with minimal interaction cost, guiding the AI ​​retrieval to gradually converge to the core literature set that best meets expectations, thereby quickly and accurately locking in high-value literature.

[0084] Therefore, the objective of the embodiments of the present invention is:

[0085] 1) This invention addresses the problems of redundant and noisy traditional literature retrieval results, as well as the lack of a bibliometric-oriented quantitative ranking mechanism. It constructs a multi-dimensional weighted scoring system (S-score) that comprehensively considers the timeliness, impact factor, and average annual citations of literature, providing users with a quantitative ranking of literature priorities. This helps users prioritize browsing literature most likely to have high value and relevance, effectively reducing noise interference in the initial result set.

[0086] 2) This invention addresses the problem that AI-driven autonomous retrieval can easily deviate from the user's true intent, and that existing feedback mechanisms cannot dynamically optimize retrieval strategies. Through a closed-loop iterative mechanism involving human-machine collaboration, this invention automatically transforms real-time user feedback (like / dislike / pending) during browsing into retrieval optimization suggestions (covering intelligent generation of keyword weight adjustments, synonym expansion, and NOT exclusion logic), gradually converging retrieval results to reflect the user's deeper, true needs.

[0087] 3) This invention addresses the problem of relying on expert experience and manual input for exclusion word discovery, lacking automated mining methods. It employs two sets of comparative statistical methods (need group and noneed group) to automatically calculate the exclusion priority EP(w) of candidate exclusion words, intelligently recommending the most valuable exclusion words for user selection. This transforms the traditional exclusion word construction process, which relies on subjective experience, into a data-driven automated workflow.

[0088] 4) This invention addresses the problem of a lack of systematic quantitative evaluation in the screening process, making it difficult for users to determine when to terminate the iteration. It introduces quantitative indicators such as the Screening Efficiency Index (SEI) and Browsing Efficiency Ratio (BER) to provide users with data-driven decision-making for iteration termination. This allows the marginal benefits of the screening process to be calculated in real time, effectively avoiding over-screening or premature termination that could lead to missed screenings.

[0089] Reference Figure 1 and Figure 4 This invention provides an intelligent literature-assisted screening method based on human-machine collaborative iteration, which includes the following steps:

[0090] S100: Obtain the target research topic information and initial literature search conditions input by the user; generate an initial search strategy based on the target research topic information and initial literature search conditions; and obtain candidate literature data from the literature database based on the initial search strategy.

[0091] Specifically, the system receives user input including research topic keywords, research subjects, disease names, intervention factors, comparison factors, research results information, and biomedical terms. It then constructs corresponding Boolean logic expressions based on the research topic keywords and generates an initial search query by combining the field limitation rules supported by the literature database. The initial search query is sent to the literature database interface, and a literature query operation is performed through the database retrieval interface to obtain literature identifier data matching the initial search query. Based on the literature identifier data, the system calls the literature data acquisition interface to obtain the corresponding literature title, abstract, author, journal, publication year, numeric object identifier, and citation information, forming an initial candidate literature dataset.

[0092] In this embodiment, the PubMed database is interacted with via the PubMed E-utilities API. Specific functions include: receiving standard PubMed search queries input by the user; performing a search via the esearch interface to obtain a list of matching document PMIDs; obtaining complete metadata of the documents via the esummary / efetch interface, including title, abstract, author, journal, publication year, PMID, DOI, etc.; supporting Boolean logic operators (AND / OR / NOT), field qualifiers (such as [Title / Abstract]), and truncation symbols (*) for PubMed search queries; and passing the obtained document metadata to the scoring module for multi-dimensional scoring.

[0093] S200. Perform multidimensional feature analysis on the candidate literature data, extract the literature attribute information from the candidate literature data, and construct a comprehensive literature evaluation model based on the literature attribute information. As needed, sort the candidate literature according to randomness, publication time, and comprehensive effectiveness, and output the set of literature to be screened.

[0094] Specifically, the publication time data in the candidate literature data is used as the timeliness feature input, and the timeliness evaluation value of the literature is calculated based on the time difference between the current time and the publication time of the literature; the journal influence data corresponding to the candidate literature is used as the journal influence feature input to obtain the journal influence index; the citation count data corresponding to the candidate literature is used as the citation influence feature input, and the average annual citation evaluation value is calculated based on the relationship between the citation count and the publication time; the timeliness evaluation value, journal influence index, and average annual citation evaluation value are standardized respectively, and then fused according to preset weights to obtain the comprehensive literature score; the candidate literature is sorted in descending order by randomness, publication time, or comprehensive literature score according to the needs, and the sorted literature screening list is output.

[0095] In this embodiment, a comprehensive score (S-score) is calculated for each retrieved document. The scoring system includes three dimensions:

[0096] 1) Time-efficiency score ( ): Reflects the novelty of a document, and its expression is:

[0097]

[0098] in, For the current year, This refers to the year the document was published. The more recent the document, the higher the publication date. The larger the value.

[0099] 2) Impact Factor (Impact Factor): Reflects the academic influence of the journal in which the article is published. It is obtained by matching the journal name with JCR (Journal Citation Reports) data. The impact factor uses a journal-level indicator published annually by Clarivate through the Journal Citation Reports (JCR), reflecting the average frequency of citations per article in a journal in a specific year. The most recently published indicator is used; the 2026 JCR is calculated based on citation data from 2025. Its expression is:

[0100]

[0101] 3) Annual Citations per Year (ACY): Reflects the academic influence of a document, and its expression is:

[0102]

[0103] in, This represents the total number of citations for the document, obtained via the OpenAlex API.

[0104] Z-score standardization is used to make the three dimensions comparable because they have different units of measurement. Its expression is:

[0105]

[0106] in, This is the mean of all documents in this dimension. The standard deviation is denoted as .

[0107] The combined S-score is expressed as:

[0108]

[0109] The default weight is (aging), (Impact Factor) (Quote).

[0110] S300. Provide the set of documents to be screened to the user for interactive screening, obtain the user's screening feedback information for different documents, construct user screening intent data based on the screening feedback information, and update the document screening strategy based on the user screening intent data.

[0111] Specifically, the set of documents to be screened is converted into interactive document display data, which can be sorted by randomness, publication time, and comprehensive effectiveness as needed. The interactive document display data shows the title, author, journal, publication year, abstract, and comprehensive score of each document. The system receives relevance feedback from users on the currently displayed documents, which includes at least the status of relevant, irrelevant, and pending confirmation. Based on the relevance feedback, a positive document set, a negative document set, and a document set pending confirmation are established. The system records the document identification information that the user has completed screening and generates historical screening record data to avoid repeatedly displaying screened documents in subsequent iterations.

[0112] In this embodiment, providing a card-style, article-by-article browsing interface specifically includes:

[0113] 1) Arrange the reference cards in descending order of their S-scores;

[0114] 2) Each card displays the title, first author, journal, publication year, impact factor, and S-score of the document;

[0115] 3) Users can mark each document as like (relevant), dislike (irrelevant), or pending (pending);

[0116] 4) Maintain a global read tracking set (readPmids) to record the documents that users have viewed across rounds, avoiding duplicate display.

[0117] S400. Perform text feature analysis on the positive and negative literature sets corresponding to the screening feedback information, extract candidate keywords from the literature, calculate the discrimination of candidate keywords based on the difference between the positive and negative literature sets, and display the candidate keywords in order based on the discrimination, so that users can make subjective screening and confirmation based on their specific research objectives and discrimination, and generate a set of excluded words confirmed by the user.

[0118] S410. Perform text feature analysis on the positive and negative literature sets corresponding to the screening feedback information.

[0119] Specifically, the title and abstract text data from the negative literature set are acquired, and the title and abstract text data are cleaned to obtain cleaned text data. Terminology recognition processing is performed on the cleaned text data, and candidate core words are extracted using noun recognition rules, biomedical terminology dictionary matching, and combined word recognition methods. Part-of-speech tagging is performed on the candidate core words to obtain corresponding unary keywords, binary keywords, and ternary keywords, outputting keywords at each granularity. Using each granularity keyword as input data, and combining its frequency of occurrence in the negative literature set and its absence in the positive literature set, candidate exclusion word data is output to characterize the irrelevance of the literature.

[0120] In this embodiment, core biomedical terms are automatically extracted from the titles and abstracts of documents, specifically including:

[0121] 1) Text preprocessing: After concatenating the title and abstract, clean the text, remove Unicode special characters, and replace punctuation marks with spaces.

[0122] 2) Hyphen concatenation: Identifies and concatenates hyphenated terms such as “A-to-I”, “double-stranded”, “HIV-1”, etc.

[0123] 3) POS heuristic annotation: Identify noun words by matching noun suffixes (such as -tion, -sis, -ome, -ase, etc.), matching known noun dictionaries (containing 500+ known biomedical nouns), and biomedical regular expression patterns (such as uppercase abbreviations, alphanumeric combinations, etc.).

[0124] 4) N-gram extraction: Unigram: a single noun or biomedical term; Bigram: a combination of modifier and noun, such as "RNA editing"; Trigram: a combination of modifier and modifier and noun, such as "innate immune response".

[0125] 5) Strict filtering: Non-nominal words are filtered out through a predefined set of stop words (including 500+ words in 6 categories such as prepositions, conjunctions, pronouns, auxiliary verbs, report verbs, and adverbs), ensuring the professionalism and accuracy of the extraction results.

[0126] 6) Lemmatization: Supports lemmatization, allowing different inflectional forms of nouns and verbs to match each other. For example, "expression" can match variants such as "express", "expressed", and "expressing".

[0127] S420. Calculate the discriminative power of candidate keywords based on the difference between the positive and negative document sets.

[0128] Specifically, the local frequency of each candidate keyword in the negative literature set is counted; the number of negative documents containing the candidate keyword and the number of positive documents not containing the candidate keyword are counted; the exclusion priority corresponding to the candidate keyword is calculated based on the local frequency, the number of negative documents covered, and the number of positive documents missing; the candidate keywords are sorted from high to low according to the exclusion priority, and different levels of candidate exclusion word recommendation lists are generated according to the keyword type; the candidate exclusion word recommendation lists are output to the user interface for the user to select and confirm.

[0129] In this embodiment, as Figure 5 As shown, its working principle is as follows:

[0130] 1) Data grouping: The user-tagged documents are divided into two groups: the need group ( ) indicates documents marked as "like" by the user (considered relevant); the "noneed" group ( () indicates documents that users have marked as dislike (considered irrelevant).

[0131] 2) Core word extraction: For each document in the noneed group, extract its unary words, bigrams and trigrams through the core word extraction module.

[0132] 3) Calculate the frequency of vocabulary in source documents. and global specific frequency ,in:

[0133] To measure vocabulary In negative irrelevant literature Representativeness in, statistics on its Frequency of appearance in title and abstract:

[0134]

[0135] In the above formula, The "importance" of a word in negative literature is measured. A higher value indicates that it is mentioned more frequently in the excluded literature, and that it has a stronger representational power of the literature content, making it a core feature word of the negative irrelevant literature.

[0136] To measure the keywords in the complete set of candidate "negative-specific words" The degree of general absence in all positive literature is defined as the global specific frequency for candidate words. Statistics on all The total number of times words were identified as "unique" in a comparison of positive and negative literature is calculated using the following formula:

[0137]

[0138] In the above formula, The range of values ​​is , Indicates the inclusion word The number of negative literature (noneed group); Indicates words that do not contain The number of positive literature (need group). The larger the value, the more positive literature the word is absent from, the stronger its ability to distinguish between positive and negative literature, and the higher its specificity as an exclusion keyword.

[0139] Normalization:

[0140]

[0141] in The total number of positively relevant documents selected for the user. This indicates the average number of times the word appears in negative literature per positive article.

[0142] Exclusion priority calculation: For each candidate exclusion word Calculate its exclusion priority:

[0143]

[0144] in: Indicator Frequency of occurrence in the target literature (local frequency); Indicates the inclusion word The number of negative literature (noneed group); Indicates words that do not contain The number of positive literature (need group).

[0145] 4) Hierarchical Display: Candidate exclusion words are sorted in descending order of EP value and displayed hierarchically: ternary words: display the top 5; binary words: display the top 10; unary words: display... And the maximum number of words that have not appeared in the multi-word phrase is 20.

[0146] 5) Deduplication mechanism: If a binary word (such as "innate immune") or ternary word (such as "innate immuneresponse") has been listed as a candidate excluded word, its split unary words ("innate", "immune", "response") will no longer be displayed as independent unary words.

[0147] 6) Double-click drill-down: When a user double-clicks a binary or ternary word card, a pop-up window breaks the word down into individual word cards, displaying statistics for each constituent word when it is used as a unary candidate exclusion word. , (EP value). Users can directly click on the card to select it as an exclusion word.

[0148] S500: Optimize the initial search strategy based on the user-confirmed exclusion term set, generate a new search strategy, and re-acquire candidate document data based on the new search strategy. Repeatedly execute the document screening, user feedback, strategy optimization, and re-search process until the preset screening termination condition is met, and output the target document set.

[0149] Specifically, the process involves: obtaining the user-confirmed exclusion term set; converting each exclusion term in the exclusion term set into a corresponding field-qualified query expression; constructing exclusion query conditions based on the logical relationships between multiple exclusion terms in the exclusion term set; combining the exclusion query conditions with the original search strategy to generate a new search strategy containing exclusion logic; and saving the new search strategy generated in the current round and the historical exclusion term set so that the cumulative optimization results can be continuously used in subsequent iterations.

[0150] In this embodiment, the user-confirmed exclusion terms are converted into PubMed NOT queries. First, each exclusion term is constructed as a PubMed field-qualified search fragment, such as " Furthermore, multiple exclusion terms are connected using OR, and the entire term is connected to the original search query using NOT; the optimized search query format is as follows: It supports multi-round accumulation. The excluded words confirmed in each round are added to the cumulative excluded word list and continue to be effective in all subsequent rounds.

[0151] Furthermore, it also includes evaluating the efficiency of the iterative screening process, obtaining the number of retrieved documents, the number of documents confirmed as relevant by users, and the number of documents confirmed as irrelevant by users for the current iteration round; calculating the screening efficiency index based on the relationship between the data of the current round and the data of the previous round; calculating the browsing efficiency index based on the ratio between the number of user screening results and the number of documents viewed; judging the trend of the benefit change of the current iteration process based on the screening efficiency index and the browsing efficiency index; and outputting a termination instruction when the efficiency index meets the preset convergence condition, and generating the final target document set.

[0152] In each iteration, new candidate literature data generated in the current round is acquired; literature data that has been evaluated by users is filtered based on historical screening records; unscreened literature data is re-sorted according to the updated comprehensive score; new literature to be screened is displayed to users based on the sorting results; user feedback information is continuously acquired, and the user screening intent model is updated using the newly added feedback information to achieve continuous optimization of the intelligent search strategy based on human judgment information.

[0153] In this embodiment, the complete process of managing multiple iterations is as follows: maintaining the iteration state: current round, search query, exclusion list, and statistical indicators; after each iteration, re-executing the PubMed search using the optimized search query; excluding read documents (readPmids set) from the new search results to ensure that users do not browse repeatedly; and supporting users to terminate the iteration at any time or iterate indefinitely.

[0154] Calculate and quantify indicators to evaluate screening efficiency:

[0155] 1) Screening Efficiency Index (SEI):

[0156]

[0157] in This represents the number of results from the previous search. This represents the number of search results in this round. This is the iteration number for this round. The higher the SEI value, the higher the screening efficiency of this round.

[0158] 2) Browse Efficiency Ratio (BER):

[0159]

[0160] in This represents the number of like documents in this round. The iteration number, This represents the number of disliked documents in this round.

[0161] In summary, as Figure 3 As shown, the overall implementation process of this invention is as follows:

[0162] 1) Initial Search. Based on the search topic, determine the core keywords and construct the initial PubMed Boolean search expression. The user inputs the search query, and the system executes the search through the PubMed E-utilities API to obtain a list of documents and their metadata.

[0163] 2) Multidimensional scoring. The scoring module calculates three-dimensional indicators for each document. After Z-score standardization, the weighted sum is used to obtain the S-score, and the literature is arranged in descending order of S-score.

[0164] 3) Interactive browsing. Users browse document cards one by one, marking each document as like, dislike, or pending. The system records the PMIDs of read documents into the readPmids collection.

[0165] 4) Core word extraction. For each disliked document, the core word extraction module extracts unary words, bigrams, and trigrams from the title and abstract.

[0166] 5) Exclusion Term Discovery. The exclusion term discovery module calculates the exclusion priority of each candidate term through two sets of comparative statistics (need group vs. noneed group). Candidate exclusion words are recommended in descending order of EP value.

[0167] 6) Search Query Optimization. The user selects and confirms from the candidate exclusion terms. The search query optimization module constructs a NOT query from the confirmed exclusion terms and combines it with the original search query to form an optimized search query.

[0168] 7) Iterative Search. Re-execute the PubMed search using the optimized search query to obtain a new list of documents. Read documents (readPmids) are excluded from the new results.

[0169] 8) Efficiency Evaluation. The efficiency evaluation module calculates the SEI and BER metrics. If the user believes the screening results meet their requirements, the iteration terminates; otherwise, return to step 3) to continue the next iteration.

[0170] This invention also includes the participation of the entire positive literature set in the calculation during multiple rounds of screening, and the recalculation of global specific frequencies. At that time, the positively related literature set It must include all positive literature selected by the user in all historical rounds, and its expression is:

[0171]

[0172] in This is the set of documents marked as "like" in this round. This is the set of documents marked as "like" in each historical round. The system persistently stores all historical "liked" documents through the `likedPaperCache` caching mechanism to ensure... This reflects the general lack of all relevant literature, rather than just the situation in this particular round.

[0173] Finally, the embodiments of the present invention will be explained and illustrated in specific implementation scenarios:

[0174] Objective: To screen relevant literature on "A-to-I RNA editing" in the field of evolution, specifically including:

[0175] 1) Initial search: The user enters a PubMed search query. The search query is: The system performed a search using the PubMed E-utilities API and obtained 160 matching documents.

[0176] 2) Multidimensional scoring: The system calculates the S-score for 160 articles, specifically including:

[0177] Timeliness Score:

[0178]

[0179] in, For the current year, This indicates the year the document was published.

[0180] Impact Factor (IF): The IF is calculated by Clarivate using the latest Journal Citation Reports (JCR) based on citation data from 2025. Journals without citation data are uniformly recorded as 0.01.

[0181] Average annual citations:

[0182]

[0183] in This represents the total number of citations for the document.

[0184] Z-score standardization:

[0185]

[0186] in This is the mean of all documents in this dimension. The standard deviation is denoted as .

[0187] Overall S-score:

[0188]

[0189] Taking PMID32406913 as an example:

[0190] Year of publication: 2020 ;

[0191] Journal Impact Factor: ;

[0192] Total citations: 370 .

[0193] Suppose there are 160 articles: ; ; .

[0194] but:

[0195]

[0196]

[0197]

[0198]

[0199] The 160 articles are sorted in descending order of S-score, and users can start browsing them one by one.

[0200] 3) Interactive browsing: Users browsed and marked the following: 3 disliked articles with PMIDs of 32406913, 35508532, and 37857907; 3 liked articles with PMIDs of 29692414, 15772668, and 42472832; 1 pending article with PMID 35246538; and the rest were not browsed.

[0201] 4) Core word extraction: Core words were systematically extracted from the three disliked articles. Taking PMID32406913 as an example: Title: "Dynamic landscape and evolution of m6A methylation in human."; Extracted unary terms: m6A, sites, human, methylation, modification, tissues, selection, etc.; Extracted binary terms: internal modification, mRNA fate, dynamic m6A, etc.; Extracted ternary terms: prevalent internal modification, reveal dynamic m6A, etc.

[0202] 5) Exclusion term discovery: The system calculates the EP value of candidate exclusion terms (total number of liked documents is 3). Table 1 shows some of the exclusion terms selected from document 32406913:

[0203] Table 1. Some excluded words selected from document 32406913

[0204] Candidate words EP m6A 15 1 15 selection 3 1 3 tissues 3 1 3 methylation 3 1 3 modification 3 0.33 1 m6A methylation 3 1 3

[0205] 6) Search query optimization: Users select exclusion terms: m6A, m6A methylation, Cephalopods, Doryteuthis, Octopus, Euripymna, C-to-U. The system constructs optimized search queries, specifically as follows:

[0206]

[0207] 7) Iterative retrieval: The optimized retrieval formula was used to retrieve 130 documents (a reduction of 30 documents).

[0208] 8) Efficiency evaluation, specifically including:

[0209] Screening Efficiency Index (SEI):

[0210]

[0211] in This represents the number of results from the previous search. This represents the number of search results in this round. This is the iteration number for this round.

[0212]

[0213] Browse Efficiency Ratio (BER):

[0214]

[0215] in This represents the number of like documents in this round. The iteration number, Number of disliked documents in this round:

[0216]

[0217] If the user judges the filtering effect to be good, the next iteration can proceed.

[0218] The detailed calculation process for identifying excluded words is as follows:

[0219] 1) Assume that the need group has 3 liked papers (PMID: 29692414, 15772668, 42472832) and the noneed group has 3 disliked papers (PMID: 32406913, 35508532, 37857907).

[0220] 2) The core words extracted from dislike document 32406913 are: m6A, sites, human, methylation, tissues.

[0221] 3) Calculate the EP value of “m6A”: = "m6A" appears 15 times in document 32406913 = 15; M = number of noneed documents containing "m6A" = 1 (appears only in document 32406913); N = number of need documents not containing "m6A" = 3 (does not appear in documents 29692414, 15772668, and 42472832); .

[0222] 4) Calculate the EP value of "sites": = “sites” appears 6 times in document 32406913 = 6; M = number of noneed documents containing “sites” = 3; N = number of needed documents not containing “sites” = 0 (documents 29692414, 15772668, and 42472832 all contain “sites”); #imgpt106#. “sites” is filtered out as a candidate exclusion word because N=0 (appears in all liked documents).

[0223] Finally, the results of multiple iterations are shown in Table 2:

[0224] Table 2. Data on the effects of multiple iterations

[0225] round Number of search results number of likes Number of dislikes SEI BER Round 1 160 3 3 30 4.28 Round 2 130 3 5 13 2.6 Round 3 104 3 2 2.67 1 termination 96 - - - -

[0226] Trend analysis: The SEI decreased from 32 to 2.67, and the BER decreased from 4.28 to 1, indicating that the marginal effect of each round of screening diminishes with iteration. When the SEI and BER stabilize, the user can determine that the screening has converged and terminate the iteration. At this point, 9 "like" documents were marked. From the final 96 documents selected, 23 high-quality documents with a final score greater than 0 were selected, forming the high-precision document set required by the user.

[0227] Therefore, the embodiments of the present invention have the following distinguishing technical features compared to the prior art:

[0228] 1) A human-machine collaborative closed-loop iterative mechanism constructs a complete closed-loop iterative mechanism of "user tagging - system discovery of exclusion words - NOT query optimization - re-retrieval - user re-tagging". Every like / dislike tag made by the user during the browsing process is recorded and analyzed by the system, and automatically transformed into the basis for retrieval optimization. This solves the limitations of traditional retrieval that is one-time and without feedback, as well as the problem of AI autonomous retrieval deviating from the user's true intent. Through human-machine collaboration, the system can gradually learn and converge to the user's true needs, and each iteration makes the retrieval results more accurate. Users only need to tag a small number of documents (usually 10-20), and the system can automatically discover effective exclusion words, significantly reducing the number of documents that need to be manually browsed.

[0229] 2) The multidimensional Z-score weighted scoring system (S-score) proposes a weighted scoring system that integrates three dimensions: timeliness (Re), impact factor (IF), and annual citations (ACY). By standardizing the three indicators with Z-scores and then weighting and summing them, an objective quantitative ranking of document priorities is achieved. This solves the problems of opaque ranking of traditional PubMed search results and failure to comprehensively consider the multidimensional value of documents. Users can prioritize browsing the most likely relevant and high-quality documents, reducing time wasted on low-quality or outdated documents. The default weights (timeliness 40%, impact factor 40%, citations 20%) can be flexibly adjusted according to user needs.

[0230] 3) A two-group comparative automatic exclusion word discovery method is proposed, based on statistical comparison between the "need" and "noneed" groups. By calculating the exclusion priority (#imgpt107#), the value of each candidate word as an exclusion word is quantified. Simultaneously, a hierarchical display (ternary → binary → unary), deduplication mechanism, and double-click drill-down interaction are employed to improve the efficiency of user exclusion word selection. This solves the problem that exclusion word discovery relies entirely on user experience and manual input. It can automatically discover the most valuable exclusion words from user-marked disliked documents, and users can independently select and confirm from the system's recommended candidate list, greatly reducing the threshold and workload of exclusion word discovery.

[0231] 4) The system automatically converts user-confirmed exclusion terms into standard PubMed NOT query syntax, supporting multi-round accumulation. Each confirmed exclusion term is added to the accumulated exclusion term list, remaining effective in all subsequent rounds. This solves the problem of users needing to manually write and modify complex PubMed search queries. Users do not need to understand PubMed search query syntax; the system automatically handles the format conversion of exclusion terms and search query concatenation. The multi-round accumulation mechanism ensures that the filtering effect monotonically increases with the number of iterations.

[0232] 5) A quantitative efficiency indicator system is proposed, which includes two quantitative indicators: Screening Efficiency Index (SEI) and Browsing Efficiency Ratio (BER), to measure the absolute and relative screening efficiency of each iteration, respectively. This solves the problem of lacking quantitative evaluation of the screening process and users' inability to determine when to terminate iterations. Users can observe the changing trends of SEI and BER to make data-driven decisions on when to terminate iterations. When SEI and BER tend to stabilize or decrease, it indicates that the marginal benefit of further iterations is diminishing, and screening can be terminated.

[0233] In summary, the embodiments of the present invention have the following advantages over the prior art:

[0234] 1) It provides a multi-dimensional quantitative scoring system (Z-score weighted S-score, which combines timeliness, impact factor and annual citations), rather than a simple + / - mark relevance score; it also has the ability to automatically discover exclusion words (EP(w) statistics) and automatically convert NOT queries, realizing iterative optimization of the original search query itself, rather than just reordering the results on the current page.

[0235] 2) This invention not only assists in filtering and ranking through an active learning mechanism, but also further narrows the search scope and improves search accuracy from the source by identifying exclusion words and transforming NOT queries. It also closes the user feedback loop to PubMed for re-searching, achieving a substantial evolution of the search strategy, rather than simply optimizing the filtering order within a fixed result set. Furthermore, this invention provides quantitative efficiency indicators such as SEI and BER, overcoming the limitation of relying solely on recall estimation and lacking systematic evaluation.

[0236] 3) A closed-loop iterative mechanism for human-machine collaboration is introduced, incorporating real-time three-state feedback of the user's Like / Dislike / Pending into the search optimization process, realizing dynamic literature screening with deep user participation. This overcomes the shortcomings of the patent, which can only perform one-time, non-interactive automatic novelty evaluation. Moreover, this invention is specifically designed for the PubMed biomedical literature scenario, making it more domain-specific.

[0237] 4) Targeting the PubMed biomedical literature database, rather than clinical visit records; this invention automatically discovers exclusion words and generates NOT queries through user feedback, overcoming the limitations of this patent that relies solely on manual input of blocked words by users and cannot be iteratively optimized. At the same time, this invention provides multi-dimensional bibliometric scoring, rather than relying solely on the single influence indicator of the number of incoming links.

[0238] 5) A real-time user feedback calibration mechanism is introduced, overcoming the shortcomings of this patent, which relies entirely on historical combined search text and cannot respond to the dynamic intent of users; this invention actively recommends excluded words through EP(w) statistics, rather than passively relying on historical data; at the same time, this invention focuses on the biomedical literature screening scenario and provides domain-specific terminology recognition and quantitative efficiency evaluation.

[0239] 6) The embodiments of this invention focus on the relevance judgment of search results and human-computer collaborative iterative optimization, rather than only focusing on the performance of the search system (such as ES parallel query and sharding optimization); this invention provides multi-dimensional scoring, exclusion word discovery and efficiency quantification indicators, which solves the fundamental problem that the patent cannot dynamically optimize the search strategy based on user relevance feedback, and improves the optimization focus from "searching fast" to "searching accurately and filtering precisely".

[0240] Reference Figure 2 A human-machine collaborative iterative intelligent document screening system includes:

[0241] The first module 201 is used to obtain the target research topic information and initial literature search conditions input by the user, generate an initial search strategy based on the target research topic information and initial literature search conditions, and obtain candidate literature data from the literature database based on the initial search strategy.

[0242] The second module 202 is used to perform multidimensional feature analysis on candidate literature data, extract literature attribute information from candidate literature data, and construct a comprehensive literature evaluation model based on literature attribute information. As needed, it can sort each candidate literature according to randomness, publication time, and comprehensive effectiveness, and output a set of literature to be screened.

[0243] The third module 203 is used to provide the set of documents to be screened to the user for interactive screening, obtain the user's screening feedback information for different documents, construct user screening intention data based on the screening feedback information, and update the document screening strategy based on the user screening intention data.

[0244] The fourth module 204 is used to perform text feature analysis on the positive and negative literature sets corresponding to the screening feedback information, extract candidate keywords from the literature, calculate the discrimination of candidate keywords based on the difference between the positive and negative literature sets, and display the candidate keywords in order based on the discrimination, so that users can make subjective screening and confirmation based on their specific research objectives and discrimination, and generate a set of excluded words confirmed by the user.

[0245] The fifth module 205 is used to optimize the initial search strategy based on the user-confirmed exclusion term set, generate a new search strategy, and re-acquire candidate document data based on the new search strategy. It iteratively executes the document screening, user feedback, strategy optimization, and re-search process until the preset screening termination conditions are met, and outputs the target document set.

[0246] The content of the above method embodiments is applicable to this system embodiment. The specific functions implemented in this system 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.

[0247] The preferred embodiments of the present application have been described above with reference to the accompanying drawings, but this is not intended to limit the scope of the embodiments 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 embodiments of the present application.

Claims

1. A human-machine collaborative iterative intelligent document screening method, characterized in that, Includes the following steps: The system obtains the target research topic information and initial literature search conditions input by the user, generates an initial search strategy based on the target research topic information and initial literature search conditions, and obtains candidate literature data from the literature database based on the initial search strategy. Multidimensional feature analysis is performed on the candidate literature data to extract the literature attribute information in the candidate literature data, and a comprehensive literature evaluation model is constructed based on the literature attribute information. As needed, the candidate literature can be sorted by randomness, publication time and comprehensive effectiveness, and the set of literature to be screened is output. The set of literature to be screened is provided to users for interactive screening. User feedback information on different literature inputs is obtained. User screening intent data is constructed based on the screening feedback information, and the literature screening strategy is updated based on the user screening intent data. Text feature analysis is performed on the positive and negative literature sets corresponding to the screening feedback information to extract candidate keywords from the literature. The discriminative power of the candidate keywords is calculated based on the difference between the positive and negative literature sets. Based on the discriminative power, the candidate keywords are displayed in order for users to subjectively screen and confirm according to their specific research objectives and discriminative power, and a set of excluded words confirmed by the user is generated. The initial search strategy is optimized based on the user-confirmed exclusion term set, generating a new search strategy. Candidate document data is then re-acquired based on the new search strategy. The process of document screening, user feedback, strategy optimization, and re-searching is repeated until the preset screening termination conditions are met, and the target document set is output.

2. The intelligent document-assisted screening method based on human-machine collaborative iteration as described in claim 1, characterized in that, The step of obtaining the target research topic information and initial literature search conditions input by the user, and generating an initial search strategy based on the target research topic information and initial literature search conditions, specifically includes: It receives user input including research topic keywords, research subjects, disease names, intervention factors, comparison factors, research results information, and biomedical terms. Construct corresponding Boolean logic expressions based on research topic keywords, and generate initial search queries by combining field limitation rules supported by literature databases; Send the initial search query to the literature database interface, perform a literature search operation through the database search interface, and obtain the literature identifier data that matches the initial search query; Based on the document identification data, the document data acquisition interface is called to obtain the corresponding document title, abstract, author, journal, publication year, digital object identifier, and citation information, forming an initial candidate document dataset.

3. The intelligent document-assisted screening method based on human-machine collaborative iteration according to claim 2, characterized in that, The step of performing multidimensional feature analysis on candidate literature data specifically includes: The publication time data in the candidate literature data is used as the timeliness feature input, and the timeliness evaluation value of the literature is calculated based on the time difference between the current time and the publication time of the literature. The journal impact data corresponding to the candidate literature is used as the input of journal impact characteristics to obtain journal impact indicators; The citation count data corresponding to the candidate literature is used as the input of the citation impact feature, and the average annual citation evaluation value is calculated based on the relationship between the citation count and the publication time. The timeliness evaluation value, journal impact index and annual average citation evaluation value are standardized and then integrated according to preset weights to obtain the comprehensive score of the literature. Based on your needs, you can choose to sort the candidate documents in descending order by randomness, publication time, or comprehensive document score, and output the sorted document screening list.

4. The intelligent document-assisted screening method based on human-machine collaborative iteration according to claim 3, characterized in that, The step of providing the set of documents to be screened to the user for interactive screening specifically includes: The collection of literature to be screened is converted into interactive literature display data, and you can choose to sort it by randomness, publication time and comprehensive effectiveness as needed; The interactive document display shows the title, author, journal, publication year, abstract, and overall score of each document. Receive relevance feedback information from the user regarding the currently displayed literature, wherein the relevance feedback information includes at least the status of being relevant, irrelevant, and pending determination; Based on the relevance feedback information, establish a positive literature set, a negative literature set, and a literature set to be confirmed. Record the identification information of the literature that the user has already filtered, and generate historical filtering record data to avoid repeatedly displaying filtered literature in subsequent iterations.

5. The intelligent document-assisted screening method based on human-machine collaborative iteration according to claim 4, characterized in that, The step of performing text feature analysis on the positive and negative literature sets corresponding to the screening feedback information specifically includes: Obtain title text data and abstract text data from the negative literature collection, perform text cleaning on the title text data and abstract text data to obtain cleaned text data; The cleaned text data is processed for terminology recognition, and candidate core words are extracted by noun recognition rules, biomedical terminology dictionary matching, and compound word recognition. Part-of-speech tagging is performed on candidate core words to obtain corresponding unary keywords, binary keywords and ternary keywords, and keywords at each granularity are output. Using keywords at each granularity as input data, and combining their frequency of occurrence in the negative literature set and their absence in the positive literature set, candidate exclusion words are output to characterize the irrelevance of the literature.

6. The intelligent document-assisted screening method based on human-machine collaborative iteration according to claim 5, characterized in that, The step of calculating the discriminative power of candidate keywords based on the difference between the positive and negative document sets specifically includes: Statistical analysis of the local frequency of each candidate keyword in the negative literature set; Count the number of negative documents containing candidate keywords and the number of positive documents not containing candidate keywords; The exclusion priority of candidate keywords is calculated based on their local occurrence frequency, the number of negative documents covering them, and the number of positive documents missing them. Candidate keywords are sorted from highest to lowest according to their exclusion priority, and different levels of candidate exclusion word recommendation lists are generated according to keyword type. Output the candidate exclusion word recommendation list to the user interface for the user to select and confirm.

7. The intelligent document-assisted screening method based on human-machine collaborative iteration according to claim 6, characterized in that, The step of optimizing the initial search strategy based on the user-confirmed exclusion term set specifically includes: Get the set of excluded words after user confirmation; Convert each exclusion term in the exclusion term set into a corresponding field-qualified query expression; Construct exclusion query conditions based on the logical relationships between multiple exclusion terms in the exclusion term set; Combine exclusion criteria with the original search strategy to generate a new search strategy that includes exclusion logic; Save the new search strategy generated in the current round and the historical exclusion keyword set so that the cumulative optimization results can be continuously used in subsequent iterations.

8. The intelligent document-assisted screening method based on human-machine collaborative iteration according to claim 7, characterized in that, It also includes efficiency evaluation of the iterative screening process, specifically including: Get the number of retrieved documents, the number of documents confirmed as relevant by users, and the number of documents confirmed as irrelevant by users for the current iteration round; The screening efficiency index is calculated based on the relationship between the current round of data and the previous round of data. The browsing efficiency index is calculated based on the ratio between the number of user-selected results and the number of documents viewed. The trend of revenue changes in the current iteration process is determined based on screening efficiency indicators and browsing efficiency indicators; When the efficiency index meets the preset convergence condition, the iteration termination instruction is output, and the final target literature set is generated.

9. The intelligent document-assisted screening method based on human-machine collaborative iteration according to claim 8, characterized in that, Each iteration specifically includes: Obtain new candidate literature data generated in the current round; Filter literature data that has already undergone user evaluation based on historical screening records; The unscreened literature data were reordered according to the updated comprehensive score; Show users new documents to be screened based on the sorting results; We continuously acquire user feedback and use the new feedback to update the user filtering intent model, so as to continuously optimize the intelligent search strategy based on human judgment.

10. An intelligent document-assisted screening system based on human-machine collaborative iteration, characterized in that, Includes the following modules: The first module is used to obtain the target research topic information and initial literature search conditions input by the user, generate an initial search strategy based on the target research topic information and initial literature search conditions, and obtain candidate literature data from the literature database based on the initial search strategy. The second module is used to perform multidimensional feature analysis on candidate literature data, extract literature attribute information from candidate literature data, and construct a comprehensive literature evaluation model based on literature attribute information. As needed, it can sort each candidate literature according to randomness, publication time, and comprehensive effectiveness, and output a set of literature to be screened. The third module is used to provide the set of documents to be screened to the user for interactive screening, obtain the user's screening feedback information for different documents, construct user screening intent data based on the screening feedback information, and update the document screening strategy based on the user screening intent data. The fourth module is used to perform text feature analysis on the positive and negative literature sets corresponding to the screening feedback information, extract candidate keywords from the literature, calculate the discrimination of candidate keywords based on the difference between the positive and negative literature sets, and display the candidate keywords in order based on the discrimination, so that users can make subjective screening and confirmation based on their specific research objectives and discrimination, and generate a set of excluded words confirmed by the user. The fifth module is used to optimize the initial search strategy based on the user-confirmed exclusion term set, generate a new search strategy, and re-acquire candidate document data based on the new search strategy. It iteratively executes the document screening, user feedback, strategy optimization, and re-search process until the preset screening termination conditions are met, and outputs the target document set.

Images