A scientific research literature management system and method based on core literature identification and front prediction

Abstract

The application relates to the technical field of document management, in particular to a scientific research document management system and method based on core document identification and front prediction, which comprises a data acquisition and preprocessing module, a core document identification module, a front document supplement module, a trend analysis and prediction module and an error processing and adaptive engine; the data acquisition and preprocessing module is used for realizing automatic acquisition, cleaning and standardization of document data. The application combines citation network structure analysis and front semantic modeling to construct a scientific research document management system with high accuracy, interpretability and foresight, which not only improves the document screening efficiency, but also provides an intelligent, visual and predictive scientific research auxiliary tool for scientific researchers, and can be widely applied to scientific research activities such as topic design, review writing and research trend judgment.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of document management, in particular to a scientific research document management system and method based on core document identification and frontiers prediction. BACKGROUND

[0002] In the process of modern scientific research, scientific researchers often need to face a large amount of literature information, from which to screen out high-value core literature, master the development context of the research field, and identify the academic frontiers, so as to guide the research topic selection, project design and achievement writing. Therefore, the management and analysis of scientific research documents have become an important part of scientific research work.

[0003] Existing scientific research document management tools and systems, such as EndNote, Mendeley, Zotero, etc., although provide certain convenience in document storage, labeling, citation format management, etc., their main functions are still focused on static management and simple retrieval, and cannot meet the deep-seated needs of scientific researchers in high-quality literature screening, research trend identification and knowledge graph construction, etc.

[0004] At the same time, mainstream academic database platforms (such as Web of Science, Scopus, CNKI, Google Scholar) provide basic data such as citation frequency, highly cited papers, etc., but these indicators are often displayed in isolation with single documents, lacking structured relationship mining and visualization analysis between documents, and it is difficult to accurately identify research hotspots and evolution trends across time and across themes.

[0005] Specifically, the existing technology still has the following significant deficiencies:

[0006] Lack of intelligent identification of core documents: existing systems mostly rely on citation frequency ranking for judgment, without combining citation network structure, time distribution characteristics, co-citation relationships and other multi-dimensional information, making it difficult to effectively identify core documents with foundational and lasting influence in a certain field.

[0007] Lack of integration mechanism between classic documents and emerging frontiers: highly cited documents are mostly early achievements, reflecting "consensus knowledge"; while the latest documents are often ignored by the system due to low citation frequency. Existing tools cannot present both "authority" and "timeliness", which is not conducive to scientific researchers to fully grasp the research status and development trends.

[0008] Lack of automated analysis capability for research theme evolution and trends: existing systems generally do not have the modeling and prediction function for research theme evolution paths, and cannot assist scientific researchers in identifying potential growth areas or future development directions.

[0009] Highly dependent on the subjective experience of users and manual operation: the literature screening, frontier tracking, research topic selection and other links still need to rely on the personal knowledge accumulation and manual judgment of scientific research personnel, which is low in efficiency and accuracy, and is not conducive to the systematic research of junior researchers or interdisciplinary explorers.

[0010] Therefore, although the existing literature management and analysis technology can meet the basic needs, there are still obvious shortcomings in supporting deep academic insights and research decision assistance.

[0011] Based on this, the present application provides a scientific research literature management system and method based on core literature identification and frontier prediction to solve the technical problems proposed above. SUMMARY

[0012] The purpose of the present application is to provide a scientific research literature management system and method based on core literature identification and frontier prediction to solve the problems proposed in the background art.

[0013] To achieve the above purpose, the present application provides the following technical solutions:

[0014] The present application provides a scientific research literature management system based on core literature identification and frontier prediction, which comprises a data acquisition and preprocessing module, a core literature identification module, a frontier literature supplement module, a trend analysis and prediction module, and an error handling and adaptive engine. The data acquisition and preprocessing module is used to realize the automatic acquisition, cleaning and standardization of literature data. The core literature identification module is used to accurately locate the core literature with foundational and continuous influence in the field. The frontier literature supplement module is used to dynamically capture the latest research frontiers related to core knowledge. The trend analysis and prediction module is used to quantify the evolution trend of research topics and predict future directions. The error handling and adaptive engine is used to ensure the robustness of the system and dynamically optimize parameters.

[0015] Preferably, the data acquisition and preprocessing module acquires the initial literature set and citation relationship from the academic database through API or file parsing, performs cross-library literature deduplication, text cleaning, metadata extraction, handles version conflicts and marks data sources, provides clean and structured input data for the core literature identification module, and ensures the accuracy of subsequent analysis.

[0016] Preferably, the core literature identification module screens the highly cited literature set, dynamically calculates the active citation threshold, identifies the highly active citation literature set, extracts the reference literature, reuses the deduplication logic of the preprocessing module, calculates the intersection, outputs the core literature set to the frontier literature supplement module as the basis for frontier screening, and provides classic knowledge base data for the trend prediction module.

[0017] Preferably, the frontier literature supplement module screens a set of N years of literature, builds a frontier literature set through a double mechanism, the basic screening retains references ≥ 1 literature, and the enhanced screening co-citation frequency threshold or theme similarity filtering, relies on the Core_Set output by the core literature identification module, outputs the Frontier_Set to the trend prediction module, and provides timely data input.

[0018] Preferably, the trend analysis and prediction module generates input corpus by weight through text fusion, uses modeling research topics, automatically optimizes the number of topics, constructs a theme annual heat sequence, predicts future 2-5 year trends using an ARIMA model, outputs a slope label, inputs the Core_Set and the Frontier_Set relying on the core identification module, and outputs the prediction result to the visualization interface to assist scientific research decision-making.

[0019] Preferably, the error handling and adaptive engine records data parsing exceptions, skips invalid entries, dynamically calculates thresholds, adapts to different scale data sets, performs ADF stationarity test and difference order selection, provides fault tolerance support and parameter adaptive ability for the system, and ensures process continuity and result reliability.

[0020] Based on the above system, the application further proposes a scientific research literature management method based on core literature identification and frontier prediction, comprising the following steps:

[0021] S1: calling API from the academic database WebofScience / Scopus or parsing BibTeX / RIS format files uploaded by the user, obtaining an initial literature set TargetCorpus and its citation relationship data;

[0022] Then the module automatically performs cross-library literature deduplication, preferentially uses DOI alignment, uses title plus author plus year hash matching when missing, cleans the text, removes special characters and stop words, extracts standardized metadata including title, abstract, keywords and year, and handles version conflicts, preferentially selects journals or highly cited versions, and simultaneously marks the data source from_WoS. Finally, a clean and structured literature set and citation network are output, laying a data foundation for subsequent analysis;

[0023] S2: screening the literature with a citation frequency ≥ threshold of the top 5% to form a highly cited set HC_Set;

[0024] Then the active citation threshold is dynamically calculated:

[0025] activeC=μ+σ;

[0026] activeC=P90;

[0027] Identify high active papers (HA_Papers), then extract all references of HA_Papers, and generate a non-redundant reference set (Refs_from_HA) by reusing the deduplication logic of the preprocessing module;

[0028] Finally, calculate the intersection of the two:

[0029] Core_Set = HC_Set ∩ Refs_from_HA;

[0030] Ensure the uniqueness of the papers by version conflict handling rules, and output the core paper set and its metadata representing the field knowledge base;

[0031] S3: Select papers published in the last N years to form Recent_Set;

[0032] Then, based on the association of Core_Set, select papers that cite at least one core paper as the frontier candidate set;

[0033] To improve accuracy, optional enhanced strategies can be enabled:

[0034] Calculate the total frequency of co-citation of candidate papers and Core_Set or the similarity of BERT / LDA-based topic vectors, and finally output the frontier paper set Frontier_Set with timeliness and knowledge inheritance;

[0035] S4: Fuse the text content of Core_Set and Frontier_Set, and generate input corpus according to the weight: abstract 1.0, title 0.6, keyword 0.4;

[0036] Then, use LDA or BERTopic for topic modeling, automatically optimize the number of topics K based on topic consistency scores, and output paper-topic distribution and keyword set;

[0037] Finally, construct the annual heat sequence of each topic, judge the stationarity through ADF test, and if necessary, perform ≤2 order difference, search for the optimal ARIMA parameters: p, q ∈ [0, 5] with the goal of minimizing AIC;

[0038] Predict the heat trend in the next 2-5 years and output the slope label: ↑ / → / ↓, form a quantitative research evolution view, and in this step, dynamically monitor the threshold calculation and abnormal diagnosis through error handling and adaptive engine.

[0039] Compared with the prior art, the beneficial effects of the present application are:

[0040] The application provides a scientific research literature management system and method, aiming to solve the problems of insufficient accuracy and low intelligence level of existing literature management tools in core literature identification, research front tracking and trend prediction, by combining citation network structure analysis and front semantic modeling, a scientific research literature management system with high accuracy, interpretability and foresight is constructed, which not only improves the literature screening efficiency, but also provides an intelligent, visual and predictive scientific research auxiliary tool for scientific researchers, and can be widely applied to scientific research activities such as topic design, review writing and research trend judgment. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 The Core_Set identification flowchart of the application. DETAILED DESCRIPTION

[0042] The technical solutions in the embodiments of the application will be clearly and completely described below with reference to the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the application.

[0043] Embodiment 1, please refer to Figure 1 The application provides a scientific research literature management system based on core literature identification and front prediction, which comprises a data acquisition and preprocessing module, a core literature identification module, a front literature supplement module, a trend analysis and prediction module and an error processing and adaptive engine. The data acquisition and preprocessing module is used for realizing automatic acquisition, cleaning and standardization of literature data. The core literature identification module is used for accurately positioning core literature with foundational and continuous influence in the field. The front literature supplement module is used for dynamically capturing the latest research front related to core knowledge. The trend analysis and prediction module is used for quantifying research theme evolution trend and predicting future direction. The error processing and adaptive engine is used for guaranteeing system robustness and dynamic optimization of parameters.

[0044] In the embodiment, it is also necessary to explain that the data acquisition and preprocessing module acquires the initial literature set and citation relationship from the academic database through API or file parsing, performs cross-library literature deduplication, text cleaning, metadata extraction, handles version conflicts and marks data sources, provides clean and structured input data for the core literature identification module, and ensures the accuracy of subsequent analysis.

[0045] In this embodiment, it also needs to be explained that the core literature recognition module dynamically calculates the citation active threshold by screening the highly cited literature set, recognizes the highly active cited literature set, extracts the references, reuses the deduplication logic of the preprocessing module, calculates the intersection, outputs the core literature set to the frontier literature supplement module as the basis for the correlation of the frontier screening, and provides the classic knowledge base data for the trend prediction module.

[0046] In this embodiment, it also needs to be explained that the frontier literature supplement module screens the literature set in the recent N years, constructs the frontier literature set through a double mechanism, retains the reference of ≥1 literature through basic screening, enhances the co-citation frequency threshold or the theme similarity filtering through enhanced screening, depends on the Core_Set output by the core literature recognition module, and outputs the Frontier_Set to the trend prediction module to provide the timeliness data input.

[0047] In this embodiment, it also needs to be explained that the trend analysis and prediction module generates input corpus according to weight by fusing text, uses modeling research themes, automatically optimizes the number of themes, constructs theme annual heat sequence, predicts the trend in the next 2-5 years by using the ARIMA model, outputs the slope label, inputs the Core_Set and the Frontier_Set depending on the core recognition module and the frontier supplement module, and outputs the prediction result to the visualization interface to assist scientific research decision-making.

[0048] In this embodiment, it also needs to be explained that the error processing and adaptive engine records data analysis exceptions, skips invalid items, dynamically calculates thresholds, adapts to different scale data sets, performs ADF stationarity test and difference order selection, provides fault tolerance support and parameter adaptive ability for the system, and ensures the continuity of the process and the reliability of the result.

[0049] Embodiment 2, based on the above system, the present application also proposes a scientific research literature management method based on core literature recognition and frontier prediction, comprising the following steps:

[0050] S1: calling API from the academic database WebofScience / Scopus or parsing BibTeX / RIS format files uploaded by the user to obtain an initial literature set TargetCorpus and its citation relationship data;

[0051] Then the module automatically performs cross-library literature deduplication, preferentially uses DOI alignment, uses title plus author plus year hash matching when missing, cleans the text, removes special characters and stop words, extracts standardized metadata including title, abstract, keywords and year, and handles version conflicts, preferentially selects journals or highly cited versions, and at the same time marks the data source from_WoS. Finally, a clean and structured literature set and citation network are output, laying a data foundation for subsequent analysis;

[0052] S2: Screening the top 5% of papers with citation frequency ≥ threshold to form a highly cited set HC_Set;

[0053] Then dynamically calculate the active citation threshold:

[0054] activeC = μ + σ;

[0055] activeC = P90;

[0056] Identify highly active citing papers HA_Papers, then extract all references of HA_Papers, and reuse the deduplication logic of the preprocessing module to generate a non-redundant reference set Refs_from_HA;

[0057] Finally, calculate the intersection of the two:

[0058] Core_Set = HC_Set ∩ Refs_from_HA;

[0059] Ensure the uniqueness of the literature through version conflict processing rules, output the core literature set and its metadata representing the field knowledge base;

[0060] S3: Screening papers published in the last N years to form Recent_Set;

[0061] Then, based on the association of Core_Set, filter out papers that cite at least one core paper as a frontier candidate set;

[0062] To improve accuracy, you can optionally enable the enhancement strategy:

[0063] Calculate the total co-citation frequency of the candidate papers and Core_Set or the BERT / LDA-based topic vector similarity, and finally output the frontier literature set Frontier_Set with timeliness and knowledge inheritance;

[0064] S4: Merge the text content of Core_Set and Frontier_Set, and generate input corpus according to the weight: abstract 1.0, title 0.6, keyword 0.4;

[0065] Then use LDA or BERTopic for topic modeling, automatically optimize the number of topics K based on the topic consistency score, and output the literature-topic distribution and keyword set;

[0066] Finally, construct the annual heat sequence of each topic, perform ADF test to determine stationarity, and if necessary, perform ≤2 order difference, search for ARIMA optimal parameters: p, q ∈ [0, 5] with AIC minimization as the goal;

[0067] Predict the trend of the next 2-5 years and output the slope label: ↑ / → / ↓, form a quantitative research evolution view, in this step, through error handling and adaptive engine real-time monitoring threshold value calculation dynamic τ and abnormal diagnosis.

[0068] Example 3, see Figure 1 In practical applications, the method for scientific literature management based on core literature identification and frontier prediction specifically includes the following steps:

[0069] Step 1. Literature data acquisition and preprocessing;

[0070] (1.1) Target literature set construction;

[0071] First, from the literature database of a specific discipline or research topic (such as Web of Science, Scopus, etc.), through keyword, research field, author, institution or journal screening conditions, an initial target literature set (Target Corpus) is constructed as the basis for subsequent analysis of data collection;

[0072] (1.2) Citation network data acquisition;

[0073] To ensure that the system can be flexibly deployed in different application scenarios, the acquisition of citation data supports two implementation methods:

[0074] Priority recommendation method: In the process of automatic extraction of citation data, the system preferentially calls the open API interface of mainstream databases for standardized access;

[0075] For example, in the Web of Science platform, the UT (Unique Identifier) field is used as the unique identifier of the literature, the CITEDREF or CR field is called to obtain the list of cited references, and the TimesCited field is used to count the total number of citations; while in Scopus, the REF field can be used to access the reference list, and CITEDBY-COUNT provides the number of citing literature;

[0076] The system maps the returned fields of different platforms through a unified abstract data extraction interface to ensure the structured construction of the citation network;

[0077] Alternative method: Support users to upload structured literature data package (such as BibTeX, RIS, EndNote XML), which is parsed and constructed by the system to build the citation network, suitable for users in environments without API permissions;

[0078] The system integrates multi-format compatible parsers during the parsing process. If format errors or missing fields are detected, the error handling mechanism will be automatically triggered, including: ① prompting the user to upload entries with formats that do not meet the requirements; ② writing abnormal records to log files for subsequent troubleshooting; ③ skipping processing for unrecognized entries to ensure the overall citation network construction task is not interrupted;

[0079] Users can modify or supplement data based on system feedback;

[0080] The system extracts the following citation relationships based on the above methods: API or citation index function, extracts the following citation relationships:

[0081] Cited References: The references cited by the literature;

[0082] Citing Papers: Other literature that cites this literature, and if necessary, calls for extended extraction across databases;

[0083] (1.3) Data preprocessing;

[0084] To ensure data consistency and availability, all collected literature and metadata are standardized, including but not limited to:

[0085] De-duplication;

[0086] Uniform literature format (such as BibTeX / XML standard);

[0087] Text cleaning (remove special characters, stop words, etc.);

[0088] Extract title, abstract, keywords, publication year, author information, number of citations, etc.

[0089] Unique identifier alignment: The system prioritizes matching citation entries from different databases based on DOI (Digital Object Identifier) to avoid redundancy caused by format differences. If the DOI is missing or has format errors, the system will automatically enable backup matching strategies, including generating a hash value based on "title + first author + publication year" combinations for similar matching, or using fuzzy string comparison techniques (such as Jaccard similarity, cosine similarity) to identify possible duplicate literature, improving the robustness and coverage of matching;

[0090] Version conflict handling: When there are multiple versions of the same literature (such as conferences, preprints, journals), the system sets the priority to retain the official journal version, or selects the main version based on citation frequency;

[0091] Source marker: Each reference keeps its original source (e.g. from_WoS, from_Scopus) for traceability, error correction or further analysis;

[0092] Step 2. Core literature set identification;

[0093] The present application proposes a "core literature identification mechanism" combining cited influence and field recognition, the specific steps are as follows:

[0094] (2.1) High-cited literature screening (HC_Set);

[0095] From the target literature set, filter out the literature whose cited times exceed the set threshold highC (such as the top 5% or cited times ≥ 50), form the high-cited literature set HC_Set;

[0096] (2.2) High-active citing literature screening (HA_Papers)

[0097] Identify "high-active citing literature" in the target literature set, that is, the literature set whose citing number exceeds the threshold activeC: high-active citing literature set HA_Papers;

[0098] The setting of activeC can use the following dynamic calculation logic:

[0099] Method one: activeC = μ + σ, where μ is the mean of the citing number, and σ is the standard deviation;

[0100] For each literature in the target literature set, count the number of references (i.e. the number of citations);

[0101] Calculate the mean (μ) and standard deviation (σ) of all citing numbers;

[0102] Set the threshold activeC = μ + σ;

[0103] If the citing number of a literature is ≥ activeC, it is considered as "high-active citing literature";

[0104] Method two: activeC = Q 90 That is, the 90th percentile of the citing number is taken as the threshold, which automatically adapts to the change of the data set size;

[0105] Sort all literatures in ascending order of citing number;

[0106] Set a quantile threshold (such as the 90th percentile, P90);

[0107] activeC = P90;

[0108] Papers in the top 10% of the distribution are included in HA_Papers;

[0109] In addition, the system allows users to manually adjust the threshold strategy or set custom percentile values to adapt to datasets of different scales and distribution characteristics;

[0110] To enhance transparency, the system automatically displays the following after screening:

[0111] The current threshold setting method (e.g., "μ+σ" or "P90");

[0112] The set value (e.g., activeC = 38);

[0113] The proportion of HA_Papers in the target literature set (e.g., "184 high-activity citing papers are identified, accounting for 12.4% of the total");

[0114] This mechanism ensures the dynamicity and objectivity of literature screening, effectively filtering low-activity literature and focusing on truly integrated and influential literature;

[0115] (2.3) Extract the reference list of high-activity citing literature;

[0116] For each paper in the high-activity citing literature set (HA_Papers), the system extracts all the reference entries it cites, and aggregates them to form the "high-activity citing reference full set" (Refs_from_HA), which serves as the core basis for subsequent cross-comparison;

[0117] Considering that references may come from multiple databases (such as Web of Science and Scopus), and the same literature may appear in different formats or metadata forms on different platforms, the system introduces a cross-database unified deduplication mechanism during the generation of Refs_from_HA to ensure the accuracy and uniqueness of the citation relationship;

[0118] This deduplication mechanism reuses some of the consistent basic logic from the "1.3 Data Preprocessing" stage, but enhances the processing at the citation level as follows:

[0119] DOI priority deduplication principle: The system first determines whether the reference contains a DOI field;

[0120] If multiple citation entries have the same DOI, they are considered the same literature, and only one standardized record is retained;

[0121] Metadata hash deduplication (Fallback): If the DOI is missing, a hash signature is constructed based on the "title + first author + publication year" core fields, and fuzzy matching and merging are performed on structurally similar citation entries;

[0122] Source tracing markers: Retain the original source database identifiers (such as from_WoS, from_Scopus) for subsequent anomaly tracing and difference analysis;

[0123] Deduplication consistency strategy: The deduplication logic used in this stage is consistent with the deduplication of the original document records in 1.3. The system reuses a unified citation entity parsing and comparison module internally to avoid inconsistent results caused by multiple matching.

[0124] Through the above mechanism, the system ensures that Refs_from_HA has high accuracy, no duplicates, and traceability, thereby improving the scientificity and robustness of the intersection calculation with HC_Set;

[0125] (2.4) Calculate the core document set (Core_Set);

[0126] The system identifies the core document set Core_Set by performing an intersection operation between the highly cited document set HC_Set and the highly active citing reference set Refs_from_HA, i.e.: Core_Set = HC_Set ∩ Refs_from_HA;

[0127] The literature in this collection has both broad citation influence and has been recognized by active researchers in the field, representing the core knowledge base with foundational and continuous contribution value under this research topic;

[0128] To ensure the scientific rigor of the intersection calculation and the consistency of the results, the system introduces a version conflict handling mechanism during the cross-matching process, consistent with the "1.3 Data Preprocessing" stage, and adopts the following rules:

[0129] Version selection principle: When the same document exists in different versions in HC_Set and Refs_from_HA (such as conference paper, preprint, and formally published journal manuscript), the system will default to selecting the formally published journal version; if there is no formal journal version, the version with higher citation frequency will be selected.

[0130] Source citation retention: The system retains the source citation identifier and version source during the matching process to support source tracing verification and subsequent visualization analysis;

[0131] DOI alignment is the primary method, while meta-field alignment is secondary: direct matching via DOI is preferred; if DOI is missing, fuzzy comparison is performed based on the hash signature of title + author + year.

[0132] By introducing a version conflict disambiguation and unified processing mechanism, we ensure that each document record in Core_Set is unique, representative, and traceable, thereby enhancing the robustness and reliability of the core document identification process.

[0133] (2.5) output;

[0134] Output the Core Set and its metadata (title, author, keyword, cited number, cited path, etc.) for subsequent research topic exploration;

[0135] Step 3. Frontier literature supplement;

[0136] To make up for the time lag of highly cited literature, the invention introduces a "related frontier literature supplement mechanism", the specific steps are as follows:

[0137] (3.1) Screen recent literature (Recent_Set);

[0138] In the target database, filter out the literature published in the past N years (such as N = 3 years) to form the recent literature set Recent_Set;

[0139] (3.2) Correlation screening frontier literature (Frontier_Set);

[0140] To ensure that the system has real relevance to the core knowledge base when supplementing the latest literature, the invention proposes a "frontier literature screening mechanism based on core literature citation relationship", which is divided into two stages: basic screening and enhanced screening:

[0141] Basic screening conditions (Boolean matching);

[0142] The system first filters out those records in Recent_Set that cite at least one literature in Core_Set from the reference list as a candidate set of frontier literature;

[0143] This step is based on Boolean logic matching, that is:

[0144] If a literature r∈Recent_Set, and there exists c∈Core_Set such that c∈Refs(r), then r∈Frontier_Set;

[0145] This way can efficiently lock the research results that belong to the "latest publication" in time, and are directly related to the "core knowledge base" in citation relationship;

[0146] Weak reference exclusion mechanism (optional);

[0147] Considering that some of the citation relationships may only be for background description, term introduction or literature review purposes, and are difficult to reflect substantive academic inheritance, the system introduces a "weak reference" exclusion mechanism as a supplementary refinement means;

[0148] Including but not limited to the following strategies:

[0149] Citation context semantic analysis: based on natural language processing techniques, extract citation location context (such as the citation sentence, the two sentences before and after) for semantic annotation, and determine whether the citation purpose is for "background citation", "research motivation" or "method / conclusion support";

[0150] The system can set rules to filter out citations that are significantly biased towards background use;

[0151] Structural position information weighting: identify the position of the citation in the structure of the full text (such as whether it is concentrated in the introduction section), and adjust the citation score combined with the position weight;

[0152] Literature whose citations mainly appear in the "research method" and "result analysis" sections is given higher weight;

[0153] Enhanced screening strategy (optional overlay);

[0154] To further enhance the structural reliability and semantic focus of Frontier_Set, the system supports the following enhanced conditions, which can be enabled individually or in combination:

[0155] Co-citation frequency threshold (τ): for each candidate literature r, the system calculates the total number of co-citations ΣCoCite(r,c) with all literature in Core_Set, if the value ≥ τ, it is considered that the literature has a significant aggregation effect in the citation network;

[0156] The system defaults τ=2, while supporting dynamic setting of threshold τ=μ+α·σ (α is the adjustment coefficient, default α=0.5) according to the average co-citation frequency μ and standard deviation σ of all literature in the corpus to Core_Set, to adapt to different disciplines of citation patterns;

[0157] Topic vector similarity filtering: candidate literature and Core_Set literature are respectively converted into topic embedding vectors through topic modeling algorithms (such as LDA) or semantic encoding models (such as BERT), and the cosine similarity between the two is calculated;

[0158] If the average semantic similarity of the literature to Core_Set ≥ θ (default θ=0.75), it is retained as a frontier literature;

[0159] This mechanism effectively reduces the entry of "formal citation but content deviation" literature into the analysis system;

[0160] Output results:

[0161] The system finally outputs the literature set Frontier_Set that meets the basic screening conditions, and (when enhanced screening is enabled) at the same time meets the co-citation frequency or topic similarity threshold requirements;

[0162] The document set has clear "timeliness" (published in recent years), "inheritance" (core reference) and "relevance" (structure / semantic consistency), which will serve as the input basis for research direction evolution modeling and trend prediction;

[0163] (3.3) Output;

[0164] Output the frontier document set Frontier_Set and its structured metadata for joint analysis with the core document set;

[0165] Step 4. Research topic exploration and future direction prediction;

[0166] (4.1) Text modeling object preparation;

[0167] Before conducting research topic modeling, the system needs to uniformly extract and preprocess the text objects used for modeling, building a structured and semantically consistent corpus as input for the topic modeling algorithm;

[0168] Specifically, the system extracts the title (Title), abstract (Abstract), and keywords (Keywords) fields from the core document set (Core_Set) and the frontier document set (Frontier_Set);

[0169] These fields usually contain the core expressions of research topics and are the main semantic carriers for topic modeling;

[0170] During the extraction process, the unique identification code (such as DOI) and publication year of each document are also retained to support literature mapping and time series modeling of the modeling results;

[0171] When building the input corpus, the system sets weight coefficients for different source fields (such as Abstract × 1.0, Title × 0.6, Keywords × 0.4) to reflect the importance differences in identifying research topics;

[0172] Abstract (Abstract) has the highest weight: as the most concentrated and complete expression carrier of research content, the system gives the highest weight (recommended weighting coefficient is 1.0) to the abstract field when building the corpus;

[0173] Title (Title) has a medium weight: the title usually summarizes the research topic and core terms, giving it a medium weight (recommended weighting coefficient is 0.6);

[0174] Keywords (Keywords) have a lower weight: due to the repetition and subjective labelization of some keywords, their semantic integrity is relatively weak, so they are given a basic weight (recommended weighting coefficient is 0.4);

[0175] Weighting method: the system adopts a weighted splicing or weighted word frequency synthesis strategy, integrates different fields according to the set weight into a unified text vector input theme modeling algorithm, such as: T input=1.0×Abstract+0.6×Title+0.4×Keywords;

[0176] The weighting mechanism can be flexibly adjusted through configuration parameters, supports semantic weight adaptation for different research fields or language styles, and enhances the stability and interpretability of the theme modeling model;

[0177] (4.2) Research topic modeling;

[0178] The application adopts the LDA (Latent Dirichlet Allocation) topic modeling method, models and analyzes the text content extracted from the core literature set (Core_Set) and the frontier literature set (Frontier_Set), to automatically identify research topics and construct topic distribution;

[0179] The system first standardizes the input text, including character uniform coding, special symbol removal, word segmentation, stop word removal and morphological restoration;

[0180] The processed text is vectorized into a bag-of-words model (Bag-of-Words) as the input of the LDA model;

[0181] The LDA model performs unsupervised probability modeling, representing each document as a probability distribution composed of multiple latent topics, and each topic is composed of several keywords with high joint occurrence probability;

[0182] The system automatically selects the optimal number of topics K according to the evaluation indicators such as perplexity (Perplexity) or topic coherence (Coherence Score); The system uses a grid search strategy to traverse different K values within the preset topic number range [5, 50], trains the LDA model and calculates the topic coherence score (Coherence Score), and selects the K value with the highest score as the final topic number;

[0183] In order to balance the stability and semantic interpretability of modeling;

[0184] The modeling results include the following contents:

[0185] The weight distribution of each document on each topic;

[0186] The keyword set corresponding to each topic;

[0187] The appearance frequency of the topic in different time periods;

[0188] By combining the publication year of the literature with its topic weight distribution, the system can construct a "time-topic" mapping relationship, providing structured input for subsequent trend modeling and future research direction prediction;

[0189] In addition to LDA, the system can also adapt advanced topic modeling methods such as BERTopic and Neural Topic Modeling (NTM);

[0190] Among them, BERTopic combines BERT word vectors and HDBSCAN clustering, suitable for large-scale multilingual text scenarios; NTM captures the nonlinear semantic structure between topics through deep neural networks, suitable for research environments with higher model accuracy requirements;

[0191] (4.3) Future trend prediction mechanism;

[0192] The ARIMA (Autoregressive Integrated Moving Average) model is used to predict research interest trends, with good interpretability and modeling ability for medium-sized annual data, suitable for scientific research topic evolution analysis;

[0193] First, the system extracts research topics from the Core_Set and Frontier_Set using topic modeling algorithms such as LDA or BERTopic, and counts the number of literature appearing in each topic in different years to construct time series interest data as input for trend modeling;

[0194] Before modeling, the system performs Augmented Dickey-Fuller (ADF) test on the annual interest sequence of each topic to determine the stationarity of the sequence;

[0195] If the ADF test result does not pass (p>0.05), the system automatically performs difference processing until the sequence passes the ADF test;

[0196] The setting logic of the difference order ddd is as follows:

[0197] d=0: The original sequence is stationary;

[0198] d=1: Stationary after first-order difference;

[0199] d=2: Stationary after second-order difference (the system allows a maximum of two-order difference to avoid excessive smoothing and information loss);

[0200] The system determines the topic trend type based on the linear fitting slope s and the prediction confidence interval boundary: when s>δ (default δ=0.2) and the confidence interval does not contain zero, it is marked as 'rising trend', otherwise as 'declining trend', if |s|<δ it is considered'stable';

[0201] After the differential processing, for each topic heat sequence, the system searches for model parameters based on AIC (Akaike information criterion):

[0202] The parameter search range is set to p∈[0, 5], q∈[0, 5];

[0203] The system uses a grid search strategy to traverse the possible parameter combinations;

[0204] The search termination condition is: find the (p, q) combination corresponding to the current optimal AIC value, and continuously 5 candidate parameter combinations fail to significantly reduce the AIC value (decrease amplitude <0.1), that is, convergence, stop searching;

[0205] After fitting the ARIMA model under the optimal parameters, the system predicts the trend of the heat of each topic for the next 2-5 years, and outputs:

[0206] Annual predicted heat estimate;

[0207] 95% confidence interval;

[0208] Trend label (up / stable / down), label definition as follows:

[0209] Up (↑): the linear fitting slope β>δ\beta>δβ>δ in the prediction period, where δ is set to 0.05 by default (unit: number of papers / year);

[0210] Stable (→): |β|≤δ;

[0211] Down (↓): β<-δ;

[0212] This mechanism provides researchers with a quantitative view of the evolution of research topics, which helps to identify potential emerging fields and the life cycle of research hotspots.

[0213] Through the above steps, the application proposes a core literature set identification method combining cited influence and field recognition degree, constructs a "highly cited literature set (HC_Set)" and a "highly active citing literature reference set (Refs_from_HA)" through double citation dimensions, and calculates the intersection to form a "core literature set (Core_Set)".

[0214] The specific implementation includes:

[0215] Use citing frequency and cited times to set two screening thresholds (activeC and highC), see steps 2.1, 2.2;

[0216] Standardization mechanism: Introduce DOI alignment mechanism, version conflict preference rules and database source annotation mechanism, unify and clean cross-database citation records, see step 1.3;

[0217] Intersection logic: Through the set intersection logic of Core_Set=HC_Set∩Refs_from_HA, see step 2.4, ensure that the finally selected core literature has "double recognition": that is, both widely cited in the field and continuously used by high-activity researchers.

[0218] Compared with the traditional single citation sorting or co-citation clustering identification method, the method can more accurately capture the knowledge base literature with both foundation and dissemination, and significantly improve the stability and academic representativeness of literature evaluation.

[0219] The application also proposes a frontier literature screening mechanism combining citation relationship and semantic association, which is used to dynamically supplement the latest research results (Frontier_Set) under the research theme.

[0220] The specific implementation includes:

[0221] Basic screening and enhancement strategy: In (3.1), the recent literature set (Recent_Set) is constructed, and in (3.2), the basic frontier set is established by taking the citation of Core_Set literature as the access condition; On this basis, multiple enhancement strategies are introduced: including reference co-occurrence frequency screening (such as co-citation number threshold), theme vector cosine similarity calculation (≥0.75), and citation density and context weight model;

[0222] Compared with the traditional time window or simple co-citation based supplement method, the mechanism can ensure that the supplemented Frontier_Set can achieve a balance between timeliness and theme continuity, and provide semantic accurate and structure reliable input for research frontier tracking.

[0223] The application also constructs a research trend prediction method combining classical and frontier corpus, which combines theme modeling algorithm and time series analysis technology to quantitatively predict the future development direction of scientific research hotspots.

[0224] The specific process includes:

[0225] Fusion corpus modeling: see step 4.2, use LDA, BERTopic or NTM to extract research topics from Core_Set and Frontier_Set;

[0226] ARIMA parameter adaptive (d / p / q selection): see step 4.3, construct the annual heat sequence of each topic, perform unit root test (ADF) to judge stationarity, and use difference operation to process non-stationary sequence; automatically fit ARIMA model to predict the trend of topic heat in the next 2-5 years, and output including heat value, confidence interval and trend label (up, stable, down).

[0227] Compared with Prophet, LSTM and other models, ARIMA has better interpretability, low complexity and deployment flexibility, especially suitable for small and medium sample scientific corpus environment, and can provide reliable trend data support for topic layout and project guide design.

[0228] In the description of the present specification, the description referring to the terms "one embodiment", "an example", "a specific example" and the like means that the specific features, structures, materials or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of the present application. In the present specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner.

[0229] The preferred embodiments of the application disclosed above are only used to help explain the application. The preferred embodiments do not describe all the details and limit the application to the specific embodiments described. Obviously, many modifications and changes can be made according to the content of the present specification. The present specification selects and describes these embodiments in order to better explain the principles and practical applications of the application, so that those skilled in the art can well understand and utilize the application. The application is limited only by the claims and their full scope and equivalents.

Claims

1. A scientific research literature management system based on core literature identification and frontier prediction, characterized in that, It includes a data acquisition and preprocessing module, a core literature identification module, a cutting-edge literature supplementation module, a trend analysis and prediction module, and an error handling and adaptive engine; The data acquisition and preprocessing module is used to automate the acquisition, cleaning, and standardization of literature data. The core literature identification module is used to accurately locate core literature with foundational and lasting influence in the field. The cutting-edge literature supplement module is used to dynamically capture the latest research frontiers associated with core literature; The trend analysis and prediction module is used to quantify the evolution trend of the research topic and predict its future direction. The data acquisition and preprocessing module obtains the initial literature set and citation relationships from academic databases through API or file parsing, performs cross-database literature deduplication, text cleaning, metadata extraction, handles version conflicts and marks data sources, and provides clean and structured input data for the core literature identification module to ensure the accuracy of subsequent analysis. The core literature identification module first filters the highly cited literature set, then dynamically calculates the citation activity threshold, identifies the highly active citation literature set whose citation count exceeds the citation activity threshold, extracts the references from the highly active citation literature set, reuses the deduplication logic of the preprocessing module to obtain a non-redundant reference set, and finally calculates the intersection of the highly cited literature set and the non-redundant literature set to obtain the core literature set. The core literature set is output to the frontier literature supplementation module as the correlation basis for frontier selection and provides classic knowledge base data for the trend prediction module. The frontier literature supplementation module filters the literature set of the past N years and constructs the frontier literature set through a dual mechanism. In the target database, it filters the literature published in the past N years to form the recent literature set. From the recent literature set, it selects the literature that cites at least one core literature as the frontier candidate set. It filters the literature in the frontier candidate set with the core literature set based on the co-citation frequency or topic similarity between the literature in the frontier candidate set and the core literature set to obtain the frontier literature set. The core literature set and the frontier literature set are input to the trend prediction module to provide timely data input.

2. The scientific research literature management system based on core literature identification and frontier prediction according to claim 1, characterized in that, The trend analysis and prediction module generates input corpus by fusing text and assigning weights, uses modeling research topics, automatically optimizes the number of topics, constructs an annual topic popularity sequence, uses the ARIMA model to predict future trends, outputs slope labels, and outputs the prediction results to a visualization interface to assist in scientific research decision-making. The input corpus is generated by fusing text with weights of 1.0 for abstracts, 0.6 for titles, and 0.4 for keywords. The number of topics K is automatically optimized based on topic consistency scores, and the optimal ARIMA parameters are searched with the goal of minimizing AIC.

3. The scientific literature management system based on core literature identification and frontier prediction according to claim 2, characterized in that, The error handling and adaptive engine records data parsing anomalies, skips invalid entries, dynamically calculates thresholds to adapt to datasets of different sizes, performs ADF stationarity tests and selects the difference order, providing fault tolerance and parameter adaptation capabilities for the system, ensuring process continuity and result reliability.

4. A research literature management method based on core literature identification and frontier prediction, characterized in that, Includes the following steps: S1: The system calls APIs from the academic database WebofScience / Scopus or parses user-uploaded BibTeX / RIS format files to obtain the initial TargetCorpus document set and its citation relationship data. Then, it automatically performs cross-database document deduplication, uses DOI alignment, and matches missing DOIs using title, author, and year hashes. It cleans the text, removes special characters and stop words, extracts standardized metadata including title, abstract, keywords, and year, and handles version conflicts. Simultaneously, it marks the data source as from_WoS, ultimately outputting a clean and structured document set and citation network, laying the data foundation for subsequent analysis. S2: Select the high-citation set HC_Set based on the citation frequency of the literature; then dynamically calculate the citation activity threshold activeC; The system identifies highly active cited documents HA_Papers with a citation count exceeding activeC, then extracts all references from HA_Papers; deduplication is performed to generate a non-redundant reference set Refs_from_HA; finally, the intersection of the highly cited set HC_Set and the non-redundant reference set Refs_from_HA is calculated: Core_Set = HC_Set ∩ Refs_from_HA; version conflict handling rules ensure document uniqueness, and the system outputs the core document set Core_Set, representing the foundation of domain knowledge, along with its metadata. S3: Select literature published in the last N years to form Recent_Set; then, based on Core_Set, select literature from Recent_Set that cites at least one core literature as the frontier candidate set; to improve accuracy, enable the enhancement strategy: based on the total co-citation frequency of candidate literature in the frontier candidate set and Core_Set or the topic vector similarity based on BERT / LDA, finally output the frontier literature set Frontier_Set, which has both timeliness and knowledge inheritance. S4: The text content of Core_Set and Frontier_Set is merged and weighted as follows: abstract 1.0, title 0.6, keywords 0.4 to generate input corpus; then LDA or BERTopic is used for topic modeling, and the number of topics K is automatically optimized based on topic consistency score, outputting document-topic distribution and keyword set; finally, the annual popularity sequence of each topic is constructed, and the stationarity is judged by ADF test. The optimal ARIMA parameter is searched with the goal of minimizing AIC: predicting future popularity trend and outputting slope label to form a quantitative research evolution view.

Images