Two-step literature retrieval method and system based on keyword extension

Abstract

The invention relates to the technical field of literature retrieval, in particular to a two-step literature retrieval method and system based on keyword extension. Comprising the following steps: calibrating an initial word, inputting a user keyword, querying a preset high-frequency subject word bank, and executing a calibration rule; preliminary retrieval: performing retrieval in an AuthorKeywords field of a target database by using the calibrated seed keyword to obtain a preliminary retrieval result set; and extracting and purifying extended keywords. According to the two-step method provided by the invention, through automatic expansion and parallel field retrieval, wider coverage can be realized in a single process, it is expected that operation rounds required by a user for repeatedly trying different keyword combinations can be obviously reduced, and the automatic expansion keyword extraction, filtering and multi-field parallel retrieval mechanism can improve the efficiency of the user. According to the method provided by the invention, a relatively wide literature range can be covered by single execution, so that a relatively comprehensive retrieval result can be expected to be obtained more efficiently through a structured automatic process, and the time cost of repeated trial and error and screening of researchers is saved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of document retrieval, and particularly to a two-step document retrieval method and system based on keyword expansion. BACKGROUND

[0002] The prior art usually adopts a single keyword or a simple Boolean logic combination (such as AND, OR) to perform retrieval in a single field (such as title or abstract). For example, if a user inputs the keyword 'Intelligent Manufacturing', in addition to considering its common synonymous expressions such as 'Smart Manufacturing', the user also needs to pay attention to its associated broad terms such as 'Industry 4.0'. For surface variant words such as 'Industry 4' and 'Industry 40', the system needs to verify their actual relevance to the target theme in combination with the context semantics, so as to avoid introducing documents with inconsistent semantics. Single keyword retrieval is prone to miss documents that use different terms to express the same or related concepts. At the same time, searching only in a single field (such as 'Intelligent Manufacturing') may miss documents that contain 'Intelligent Manufacturing' in the title and abstract. (The present application provides a two-step document retrieval method and system based on keyword expansion, aiming to solve the problem of insufficient document recall rate caused by term diversity and single-field retrieval mentioned in the background.)

[0003] Although existing databases (such as the keyword suggestion pop-up window of Web of Science and the'related keywords' sidebar function of Scopus) provide keyword expansion tools, these functions are usually independent of the core retrieval process (requiring the user to manually click to trigger) and mainly rely on pre-defined static word libraries (such as fixed synonym tables based on historical data), lacking the ability to dynamically extract terms from the current retrieval results. For example, the 'Index Keywords' recommendation function of Scopus is only based on global word frequency statistics, and cannot dynamically optimize the expanded word set according to the preliminary retrieval results.

[0004] Based on this, the present application provides a two-step document retrieval method and system based on keyword expansion to solve the above-mentioned technical problems. SUMMARY

[0005] The present application aims to provide a two-step document retrieval method and system based on keyword expansion to solve the problems mentioned in the background.

[0006] To achieve the above-mentioned purpose, the present application provides the following technical solutions:

[0007] The application provides a two-step literature retrieval method based on keyword expansion, characterized in that the method comprises the following steps:

[0008] S1. Initial word calibration, input user keywords, query pre-set high-frequency subject keyword library, and execute calibration rules;

[0009] S2. Preliminary retrieval, use the calibrated seed keywords to perform retrieval in the AuthorKeywords field of the target database, and obtain a preliminary retrieval result set;

[0010] S3. Extraction and purification of extended keywords, extract candidate words, extract terms in the title, author keywords and extended keywords fields from the literature in the result set, and perform two-level noise filtering, keyword expansion and co-occurrence verification;

[0011] S4. User confirmation and adjustment, present the extended set for user modification, the user can delete redundant words or add missing words, and generate a final extended set;

[0012] S5. Deep retrieval, use the final extended set to perform logical or retrieval in the title, author keywords and extended keywords fields of the database in parallel, and construct a retrieval formula;

[0013] S6. Result processing, perform merging, deduplication, weighted sorting and field adaptation.

[0014] Preferably, the implementation process of step S1 is as follows:

[0015] After obtaining the literature set N in the preliminary retrieval stage, the three field information of each literature, i.e., the title, author keywords and extended keywords, is extracted in turn, and the noise filtering and semantic expansion process is entered.

[0016] Preferably, the implementation process of step S2 is as follows:

[0017] S2.1. Noise filtering:

[0018] First level: filter out non-informational words using a pre-defined stop word list;

[0019] Second level: calculate the TF-IDF value of the remaining candidate words, and set a dynamic TF-IDF threshold value, with a default TF-IDF threshold value of 0.01;

[0020] S2.2. Determine by grid search:

[0021] Select 100 literatures in each of the three fields to form a test set, compare the noise rate and recall rate of the extended keyword set in the threshold interval [0.001, 0.05], the noise rate is ≤15% and the recall rate is ≥88% when the threshold value = 0.01, the comprehensive F1 value is the highest (0.91±0.03), see formula (1);

[0022] TF-IDF(t,d,N) = tf(t,d) x idf(t,N) (1);

[0023] tf(t,d): the frequency of term 't' in document 'd';

[0024]

[0025] |N|: the total number of documents in the initial search result set;

[0026] |d e N: t e d|: the number of documents containing term 't'.

[0027] Preferably, the implementation process of step S3 is as follows:

[0028] S3.1. Expansion keyword generation: for the filtered candidate keywords, perform expansion keyword generation operation:

[0029] Obtain synonyms through predefined synonym library matching, calculate the cosine similarity of candidate keywords and seed keywords in vector space by using built-in pre-trained word vector model, retain words reaching the preset threshold (for example, > 0.7), see formula (2);

[0030] Cosine similarity calculation formula:

[0031] where, Word vector of seed keyword and candidate keyword;

[0032] S3.2. Co-occurrence analysis and confirmation:

[0033] Based on the candidate keywords and seed keywords in the initial search result set N, use the model to calculate their correlation in the semantic space, check and include candidate keywords that meet the substring inclusion relationship, to overcome the noise that may be introduced by simply relying on semantic similarity or substring relationship, further analyze the co-occurrence of candidate keywords with seed keywords in the initial search result set N;

[0034] Statistical number of documents in the entire set N that meet this co-occurrence condition, only when the statistical value reaches or exceeds the set empirical threshold, the candidate keyword is considered as an effective expansion keyword, the co-occurrence threshold is a configurable parameter, allowing users to adjust within a reasonable range according to specific needs, directly using the statistical information contained in the initial search result set N for key verification, finally generating an expansion keyword set S = {A1, A2,..., An}, the generation of this set is based on the initial search result set N, semantic resources are used to assist semantic correlation calculation and preliminary screening, and context co-occurrence verification is based on the statistical information of N to filter the candidate keywords, ensuring the actual relevance of the expansion keywords and the theme.

[0035] Preferably, the implementation process of step S4 is as follows:

[0036] S4.1. Present the generated set S to the user, and the user manually deletes irrelevant or redundant keywords based on professional knowledge, or adds important but not automatically captured keywords, forming an extended keyword set S_final for in-depth retrieval;

[0037] S4.2. If the user chooses to skip, S_final=S.

[0038] Preferably, the implementation process of step S5 is as follows: using S_final={A1, A2,..., Am}, perform parallel retrieval in the TI, AK and KP fields of the database, and use logical OR combination within the field, that is, construct the retrieval formula as (TI=(A1 OR A2 OR...OR Am)) OR (AK=(A1 OR A2 OR...OR Am)) OR (KP=(A1 OR A2 OR...OR Am)).

[0039] Preferably, the implementation process of step S6 is as follows:

[0040] S6.1. After retrieval execution, collect all documents retrieved by parallel retrieval in the three fields, merge based on the unique identifier of each document, remove duplicate documents, and form a de-duplicated in-depth retrieval result set M;

[0041] S6.2. Result sorting: calculate the comprehensive score of each document in the result set M, and arrange all documents in descending order of total score, see formula (3);

[0042]

[0043] Field weight:

[0044]

[0045] Weight basis: determined based on field semantic importance experiment (TI> KP> AK), and the weight sum is 1;

[0046] AK), and the weight sum is 1;

[0047] Adjustability: the weight supports user customization (must satisfy 'w TI +w KP +w AK =1).

[0048] The application also proposes a system of a two-step literature retrieval method based on keyword expansion, characterized by comprising:

[0049] A word library management module is configured to store a high-frequency subject word library, a professional synonym library, and a stop word list, and is further configured to provide word frequency data and synonyms required for calibration for step S1, provide synonym matching support in the expansion stage of step S3, and support periodic automatic / manual updating of the word library;

[0050] A search execution module is configured to generate a search formula, call a database interface to execute a query, and is further configured to perform the preliminary search of step S2 and perform the multi-field parallel search of step S5;

[0051] A keyword processing engine is configured to implement TF-IDF calculation, word vector similarity analysis, and co-occurrence statistics, and is further configured to perform candidate word extraction, noise filtering, and semantic expansion in the stage of step S3, and rely on a pre-trained word vector model;

[0052] A sorting and adaptation module is configured to result deduplication, weighted scoring, and cross-library field mapping, and is further configured to process the multi-field search results in the stage of step S6, generate a sorting list, and dynamically adjust weights according to a database field structure;

[0053] A user interaction module is configured to receive user input, present an expanded keyword set and search results, and is further configured to receive initial keywords in step S1, provide a manual adjustment interface in S4, and output a final sorting result.

[0054] Compared with the prior art, the method has the following beneficial effects:

[0055] The two-step method can achieve wider coverage in a single process through automatic expansion and parallel field search, and is expected to significantly reduce the number of operation rounds required by users to repeatedly try different keyword combinations. The automatic expansion keyword extraction, filtering, and multi-field parallel search mechanism enable a single execution to cover a wider range of literature. At the same time, the embodiment results show that the method significantly improves recall rate and precision rate, and the automated process of a single operation avoids time-consuming operations such as keyword adjustment and search execution by the user. Therefore, the method is expected to more efficiently obtain more comprehensive search results and save the time cost of repeated trial and error and screening by researchers through a structured automated process. BRIEF DESCRIPTION OF DRAWINGS

[0056] Figure 1 A two-step literature search method flowchart based on keyword expansion is provided. DETAILED DESCRIPTION

[0057] With reference to the embodiments of the present application, the technical solutions in the embodiments of the present application will be clearly and completely described, obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work belong to the scope of protection of the present application.

[0058] In the embodiment 1, the method of the present application is based on the system operation of the two-step literature retrieval method of keyword expansion, specifically, comprising:

[0059] The thesaurus management module is used for storing the high-frequency subject thesaurus, the professional synonym thesaurus and the stop word list, and is also used for providing the word frequency data and the synonym required for calibration for the step S1, providing the synonym matching support in the expansion stage of the step S3, and supporting the regular automatic / manual update of the thesaurus (such as every quarter);

[0060] The retrieval execution module is used for generating the retrieval formula, calling the database interface to execute the query, and is also used for executing the step S2 of the preliminary retrieval (AK field) and the step S5 of the multi-field parallel retrieval (TI / AK / KP);

[0061] The keyword processing engine is used for implementing the TF-IDF calculation, the word vector similarity analysis and the co-occurrence statistics, and is also used for executing the candidate word extraction, the noise filtering and the semantic expansion in the step S3 stage, and depends on the pre-trained word vector model (based on the historical literature abstract training);

[0062] The sorting and adaptation module is used for the result deduplication, the weighted scoring and the cross-database field mapping, and is also used for processing the multi-field retrieval result in the step S6 stage, generating the sorting list, and dynamically adjusting the weight according to the database field structure (such as Scopus→IndexKeywords mapping as KP);

[0063] The user interaction module is used for receiving the user input, presenting the expanded keyword set and the retrieval result, and is also used for receiving the initial keyword in the step S1, providing the manual adjustment interface in the S4, and outputting the final sorting result;

[0064] Please refer to Figure 1 Based on the above-mentioned system of the two-step literature retrieval method based on keyword expansion, the present application proposes a two-step literature retrieval method based on keyword expansion, comprising the following steps:

[0065] S1. Initial word calibration, inputting the user keyword, querying the preset high-frequency subject thesaurus, and executing the calibration rule;

[0066] In the embodiment, it also needs to be explained that the implementation process of the step S1 is:

[0067] After obtaining the document set N in the preliminary retrieval stage, three field information of each document is extracted in turn:

[0068] Title (Title, TI), Author Keywords (Author Keywords, AK) and Keywords Plus (Keywords Plus, KP), the terms in the three types of fields are used as candidate keywords, which enter the subsequent noise filtering and semantic expansion process;

[0069] S2. Preliminary retrieval, using the calibrated seed keywords, performing retrieval in the AuthorKeywords field of the target database to obtain the preliminary retrieval result set;

[0070] In this embodiment, it is also necessary to explain that the implementation process of step S2 is:

[0071] S2.1. Noise filtering:

[0072] First level:

[0073] Using a predefined stop word list (containing general words such as “study”, “analysis”, “method”) to filter out non-informational words;

[0074] Second level:

[0075] Calculate the TF-IDF value (Term Frequency-Inverse Document Frequency) of the remaining candidate words, and set a dynamic TF-IDF threshold (for example, ≥0.01), and the default TF-IDF threshold is 0.01;

[0076] S2.2. Determine by grid search:

[0077] In three fields (intelligent manufacturing, medical information, environmental engineering), 100 documents are selected to form a test set, and the noise rate (the proportion of manually annotated irrelevant words) and the recall rate (the proportion of covering the gold standard set) of the expanded keyword set in the threshold interval [0.001, 0.05] are compared. When the threshold value = 0.01, the noise rate ≤ 15% and the recall rate ≥ 88%, the comprehensive F1 value is the highest (0.91±0.03), see equation (1);

[0078] TF-IDF(t,d,N)=tf(t,d)×idf(t,N)(1);

[0079] In the formula, tf(t,d): the occurrence frequency of term ‘t’ in document ‘d’;

[0080]

[0081] |N|: the total number of documents in the preliminary retrieval result set;

[0082] |d∈N:t∈d| : Number of documents containing the term 't';

[0083] Threshold: default threshold 0.01, user adjustable (typical range: 0.005-0.05);

[0084] S3. Extended keyword extraction and refinement, extract candidate keywords from the title, author keywords, and extended keywords field of the documents in the result set, perform two-stage noise filtering, keyword expansion, and co-occurrence verification;

[0085] In this embodiment, it is also necessary to note that the implementation process of step S3 is:

[0086] S3.1. Extended keyword generation, perform extended keyword generation on the filtered candidate keywords:

[0087] Obtain synonyms through a predefined synonym library, calculate the cosine similarity of candidate keywords and seed keywords in the vector space using an embedded pre-trained word vector model (such as Word2Vec, GloVe) (this model is pre-trained based on large-scale field-related corpus such as target database historical literature abstracts and integrated as a component;

[0088] Load and use the pre-trained model for similarity calculation during retrieval execution), retain words that meet the preset threshold (e.g., ≥0.7), see equation (2);

[0089] Cosine similarity calculation formula:

[0090] Word vectors of seed keywords and candidate keywords (from pre-trained model);

[0091] Threshold: similarity ≥0.7 is an effective associated word (adjustable range: 0.6-0.8);

[0092] S3.2. Co-occurrence analysis and confirmation:

[0093] Based on the candidate keywords and seed keywords in the preliminary retrieval result set N, use the model to calculate their association in the semantic space, check and include candidate keywords that meet the substring inclusion relationship, to overcome the noise that may be introduced by simply relying on semantic similarity or substring relationship (such as avoiding the incorrect inclusion of 'Cloud Computing' into a retrieval unrelated to the 'Industrial Cloud' topic), further analyze the co-occurrence of candidate keywords with seed keywords in the preliminary retrieval result set N;

[0094] Here, co-occurrence is defined as:

[0095] In the same document, the candidate word and the seed keyword appear in the title (TI), author keywords (AK) or keywords plus (KP) field of the document simultaneously;

[0096] Count the number of documents in the whole set N that satisfy this co-occurrence condition (i.e. the number of documents containing the co-occurrence pair (seed keyword, candidate word));

[0097] Only when the statistical value reaches or exceeds the set empirical threshold (e.g. ≥ 3), the candidate word is considered as a valid extended keyword;

[0098] The co-occurrence threshold (default value 3) is a configurable parameter, allowing users to adjust within a reasonable range (e.g. 1 to 10) according to specific needs;

[0099] Directly use the statistical information contained in the preliminary search result set N itself for key verification;

[0100] Finally generate the extended keyword set S = {A1, A2,..., An};

[0101] The generation of this set is based on the preliminary search result set N (providing candidate word sources and co-occurrence statistical information), semantic resources (predefined synonym library, pre-trained word vector model) are used to assist semantic association calculation and preliminary screening, and context co-occurrence verification is based on the statistical information of N to filter the candidate words, ensuring the actual relevance of the extended words to the theme;

[0102] S4. User confirmation and adjustment, present the extended set for user modification, users can delete redundant words or add missing words to generate the final extended set;

[0103] In this embodiment, it should be further explained that the implementation process of step S4 is:

[0104] S4.1. Present the generated set S to the user, and the user manually deletes irrelevant or redundant keywords based on professional knowledge, or adds important but not automatically captured keywords to form the final extended keyword set S_final for in-depth retrieval;

[0105] S4.2. If the user chooses to skip, S_final = S;

[0106] S5. In-depth retrieval, using the final extended set to perform logical or retrieval in the title, author keywords, and extended keywords fields of the database simultaneously, and constructing a search formula;

[0107] In this embodiment, it should be further explained that the implementation process of step S5 is:

[0108] Using S_final = {A1, A2,..., Am}, perform parallel search in TI, AK and KP (or other database's corresponding fields, see the following description) fields of the database, and use logical OR (OR) combination within the field, that is, construct the search formula as (TI = (A1 OR A2 OR... OR Am)) OR (AK = (A1 OR A2 OR... OR Am)) OR (KP = (A1 OR A2 OR... OR Am));

[0109] S6. Result processing, performing merge deduplication, weighted sorting and field adaptation;

[0110] In this embodiment, it should also be noted that the implementation process of step S6 is:

[0111] S6.1. After search execution, collect all documents hit by parallel search in three fields, merge based on the unique identifier (such as DOI, Accession Number, document ID) of each document, remove duplicate documents, and form the deduplicated in-depth search result set M;

[0112] S6.2. Result sorting:

[0113] Calculate the comprehensive score for each document in the result set M, and arrange all documents in descending order of total score and output, see formula (3);

[0114]

[0115] Field weight:

[0116]

[0117] Weight basis: determined based on field semantic importance experiment (TI > KP > AK), and the weight sum is 1;

[0118] AK), the weight sum is 1;

[0119] Adjustability: weight supports user customization (must satisfy 'w TI +w KP +w AK = 1);

[0120] Example 2, to verify the effectiveness of the method in improving the recall rate (recall rate) and precision rate compared with the traditional single search method, the following example experiment is carried out:

[0121] Research topic:

[0122] Application of digital twin in intelligent manufacturing industry;

[0123] Database:

[0124] Web of Science (WOS) Core Collection

[0125] Comparison method:

[0126] Method A (traditional single search):

[0127] Directly search in the title (TI) field:

[0128] TI = ("Digital Twin" AND ("Intelligent Manufacturing" OR "Smart Manufacturing"));

[0129] Record the search result set A;

[0130] Method B (two-step method of the present application):

[0131] Initial word calibration:

[0132] Input "Intelligent Manufacturing" and query the pre-set high-frequency subject keyword library;

[0133] Since "Intelligent Manufacturing" itself is among the top 10% of comprehensive word frequencies in the Author Keywords and Keywords Plus fields of WOS, it is a high-frequency word and does not need to be supplemented with synonyms, and it can be directly used as a seed keyword;

[0134] Preliminary search:

[0135] Search in the Author Keywords (AK) field of WOS:

[0136] AK = "Intelligent Manufacturing";

[0137] Obtain the preliminary search result set N (N = 100 in the example);

[0138] Expansion keyword extraction / filtering / expansion:

[0139] Extract terms from the TI, AK, and KP fields of N; filter through the stop word list (remove general words such as "study" and "analysis"); calculate the TF-IDF value and retain terms with a value ≥ 0.01; based on synonym library matching, word vector semantic similarity calculation (cosine value ≥ 0.7), and substring inclusion relationship, generate a candidate expansion keyword set S (the example includes:

[0140] “Digital Twin”, “DT”, “Virtual Commissioning”, “Cyber-Physical System”, “Smart Factory”, “Industry 4.0”, “Industry 4.0”, “Predictive Maintenance”...);

[0141] User confirmation (optional): system presents S, user can choose to review and manually adjust (delete irrelevant words or add important words) to form the final expansion set S_final;

[0142] In the performance comparison experiment of this embodiment, to isolate and accurately evaluate the core performance of the automatic keyword generation and sorting algorithm, the execution process of method B skips the optional user confirmation link (i.e. user confirmation is not performed), and directly uses the expansion keyword set S generated by the algorithm for in-depth retrieval and sorting, i.e. S_final = S;

[0143] Therefore, the following experimental results reflect the performance of the method of the present application under the automatic core process (without user intervention);

[0144] It can be reasonably expected that when the user confirmation link is enabled to optimize S, a better S_final is formed, and the recall rate and precision rate (especially the precision rate) of the final retrieval results are expected to be further improved or at least remain unchanged;

[0145] In-depth retrieval:

[0146] In the TI, AK, and KP fields of WOS, use S_final to perform logical OR (OR) combined retrieval;

[0147] Result merging and deduplication:

[0148] Collect the retrieval results of the three fields, merge and deduplicate based on the unique identifier (DOI) of the literature;

[0149] Weighted sorting:

[0150] For the deduplicated literature, calculate the comprehensive score according to the preset weight (TI: 0.6, KP: 0.3, AK: 0.1):

[0151] Total score = (TI hit S_final word number * 0.6) + (KP hit S_final word number * 0.3) + (AK hit S_final word number * 0.1), ranked in descending order of total score;

[0152] Record the sorted retrieval result set B;

[0153] Gold standard set (G) construction basis and method:

[0154] Seed literature selection:

[0155] Based on the field consensus literature identification method (including but not limited to journal impact indicators and cited times), 5 authoritative reviews and 10 highly cited research papers are selected as seed literature;

[0156] Note:

[0157] This selection depends on the coverage ability of existing retrieval tools for core literature, and there may be relevant literature that is not effectively retrieved by existing methods;

[0158] Expert-led final decision:

[0159] Three independent experts will conduct double-blind review of candidate literature according to the pre-defined relevance criteria. The prerequisite for literature to be included in the gold standard set G is that at least two experts agree that it is "highly relevant" (Cohen's Kappa = 0.78);

[0160] Expert consensus is the core basis to ensure the authority of G set, and seed literature is only used to expand the scope of candidate pool;

[0161] This selection standard aims to ensure the representativeness and authority of seed literature;

[0162] Literature pool expansion:

[0163] Using the citation network function of WOS, perform forward citation tracing:

[0164] Retrieve all cited literature of these 15 seed literatures; backward citation tracing:

[0165] Retrieve the references of these 15 seed literatures and their important cited literatures (cited times > 50); literature pool merging and deduplication:

[0166] Merge forward and backward tracing results, remove duplicate literature, form an initial pool containing about 600 candidate literatures;

[0167] This process aims to capture as much relevant literature as possible based on the citation network of core literature;

[0168] Independent expert review:

[0169] Three independent experts with more than ten years of research experience in the field of "intelligent manufacturing and digital twin" will conduct double-blind review of candidate literature according to the pre-defined relevance criteria (such as whether it involves modeling methods, application scenarios, integration architecture, etc.);

[0170] The prerequisite for literature to be included in the final gold standard set G is that at least two experts agree that it is "highly relevant";

[0171] In addition, to enhance the objectivity of the review, Cohen's Kappa coefficient was calculated to measure the consistency among experts (in the example, =0.78, indicating high consistency);

[0172] Gold standard set construction results:

[0173] Through this process, the gold standard set G (|G|=152) containing 152 articles was finally constructed.

[0174] Representative Explanation:

[0175] This construction method combines citation networks and expert review to strive to cover highly relevant literature, but it still has limitations.

[0176] Relying on the initial selection of seed literature may lead to a bias towards mature fields and highly cited literature; it may also overlook emerging or niche research that is not covered by the citation network; future evaluations need to explore a more comprehensive gold standard construction method to verify the universality of this approach.

[0177] Evaluation indicators:

[0178] Recall:

[0179] (|A∩G| / |G|)*100% or (|B∩G| / |G|)*100%;

[0180] The recall rate is measured by the proportion of the search result set (A or B) that covers the gold standard set G.

[0181] Precision@K of the top K articles:

[0182] (Number of documents belonging to G in the top K of the search results / K) * 100%;

[0183] The relevance of the top K articles in the ranking results is measured (K = 10, 20 in this experiment), reflecting the precision.

[0184] Mean Average Precision (MAP):

[0185] The comprehensive precision index considers the position of relevant literature in the search results ranking list, with a value range of [0,1]. The higher the value, the higher the relevant literature appears.

[0186]

[0187] Results analysis:

[0188] The recall rate has been significantly improved (89% vs 63%):

[0189] The recall rate of method B in this invention is much higher than that of traditional method A;

[0190] This strongly demonstrates that the method effectively improves the recall rate and captures more relevant documents by dynamic keyword expansion (solving the problem of synonymous / paronymous / variant terms such as "Smart Manufacturing", "Industry 4.0", "Industry 40") and multi-field parallel retrieval (covering TI, AK, KP, solving the problem of missing only in single field retrieval);

[0191] The relevance at the top of the result list is excellent (Precision@10 = 93%, Precision@20 = 87%):

[0192] The relevance (Precision@10 and @20) of method B at the top of the result list is significantly better than that of method A;

[0193] This shows that the dynamic keyword purification mechanism (stop word filtering, TF-IDF threshold filtering) used in the method effectively reduces the interference of noise words, and the ranking model based on field weight (TI weight 0.6 gives higher importance to title matching) significantly improves the probability of high-relevance documents appearing at the top of the result, i.e. improves the precision;

[0194] The overall ranking quality leads (MAP = 0.81 vs 0.52):

[0195] The average precision (MAP) of method B is significantly better than that of method A;

[0196] This comprehensively verifies that the method not only has an advantage in the number of retrievals (recall), but also has a significant advantage in the overall relevance and ranking quality of the retrieval results, ensuring that users see the most relevant documents first;

[0197] Conclusion:

[0198] This embodiment verifies the effectiveness and superiority of the two-step retrieval method proposed in the invention in significantly improving the recall rate (Recall) and precision (Precision) of document retrieval by constructing a rigorous gold standard set for the automatic core process (skipping user confirmation) and evaluating standard information retrieval indicators under the research topic of "digital twin application in smart manufacturing" and the Web of Science (WOS) database environment, using example data;

[0199] In particular, it solves the core challenge of term variation and cross-field retrieval;

[0200] The complete scheme combined with the optional user confirmation link is expected to provide better or at least equivalent retrieval results;

[0201] The above results indicate that the proposed method is effective in the specific scenario described above.

[0202] Future research will further evaluate the universality and robustness of this method across different disciplines, different literature databases (such as Scopus, PubMed, CNKI), and multilingual literature retrieval environments.

[0203] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0204] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A two-step document retrieval method based on keyword expansion, characterized in that, The method comprises the following steps: S1. Initial word calibration, input user keywords, query pre-set high-frequency subject keyword library, and execute calibration rules; S2. Preliminary retrieval, use the calibrated seed keywords to execute retrieval in the AuthorKeywords field of the target database to obtain a preliminary retrieval result set; S3. Expansion keyword extraction and purification, extract candidate words, extract terms in the title, author keywords, and expansion keyword fields from the documents in the result set, and perform two-level noise filtering, keyword expansion, and co-occurrence verification; S4. User confirmation and adjustment, present the expansion set for user modification, the user can delete redundant words or add missing words to generate a final expansion set; S5. In-depth retrieval, use the final expansion set to execute logical or retrieval in the title, author keywords, and expansion keywords fields of the database in parallel, and construct a retrieval formula; S6. Result processing, perform merging, deduplication, weighted sorting, and field adaptation.

2. The two-step document retrieval method based on keyword expansion according to claim 1, wherein, The implementation process of step S1 is as follows: After obtaining the document set N in the preliminary retrieval stage, the three field information of each document, namely the title, author keywords, and expansion keywords, is extracted in turn, and the noise filtering and semantic expansion process is entered.

3. The two-step document retrieval method based on keyword expansion according to claim 2, wherein, The implementation process of step S2 is as follows: S2.

1. Noise filtering: First level: use a pre-defined stop word list to filter out non-informational words; Second level: calculate the TF-IDF value of the remaining candidate words, and set a dynamic TF-IDF threshold, with a default TF-IDF threshold of 0.01; S2.

2. Determine by grid search: Select 100 documents in each of the three fields to form a test set, compare the noise rate and recall rate of the expansion keyword set in the threshold interval [0.001, 0.05], when the threshold is 0.01, the noise rate is ≤15% and the recall rate is ≥88%, the comprehensive F1 value is the highest (0.91±0.03), see formula (1); TF-IDF(t,d,N)=tf(t,d)×idf(t,N) (1); In the formula, tf(t,d): the occurrence frequency of term 't' in document 'd'; |N| : the total number of documents in the preliminary retrieval result set; |d∈N:t∈d| : the number of documents containing term 't'.

4. The two-step document retrieval method based on keyword expansion according to claim 3, wherein, The implementation process of step S3 is as follows: S3.

1. Expansion keyword generation, perform expansion keyword generation operation on the filtered candidate words: Obtain synonyms through a pre-defined synonym library, calculate the cosine similarity of the candidate words and the seed keywords in the vector space using a built-in pre-trained word vector model, and retain words that meet a preset threshold (e.g., ≥0.7), see formula (2); Cosine similarity calculation formula: In the formula, seed keyword and the word vector of the candidate word; S3.

2. Co-occurrence analysis and confirmation: Based on the candidate words and seed keywords in the preliminary retrieval result set N, use the model to calculate their correlation in the semantic space, check and include candidate words that meet the substring inclusion relationship, to overcome the noise that may be introduced by simply relying on semantic similarity or substring relationship, and further analyze the co-occurrence of candidate words with the seed keywords in the preliminary retrieval result set N; The number of documents in the entire set N that meet this co-occurrence condition is counted, and only when the statistical value reaches or exceeds the set empirical threshold, the candidate word is considered as an effective extended keyword. The co-occurrence threshold is a configurable parameter, allowing users to adjust within a reasonable range according to specific needs. The key verification is directly based on the statistical information contained in the preliminary retrieval result set N itself, and the final extended keyword set S = {A1, A2,..., An} is generated. The generation of this set is based on the preliminary retrieval result set N, and semantic resources are used to assist semantic association calculation and preliminary screening. The context co-occurrence verification filters the candidate words based on the statistical information of N, ensuring the actual relevance of the extended words to the theme.

5. The two-step document retrieval method based on keyword expansion according to claim 4, wherein, The implementation process of step S4 is: S4.

1. Present the generated set S to the user, and the user manually removes irrelevant or redundant keywords based on professional knowledge, or adds important but not automatically captured keywords to form the final extended keyword set S_final for in-depth retrieval; S4.

2. If the user chooses to skip, S_final = S.

6. The two-step document retrieval method based on keyword expansion according to claim 5, wherein, The implementation process of step S5 is: using S_final = {A1, A2,..., Am}, performing parallel retrieval in the TI, AK, and KP fields of the database, and using logical OR OR combination within the field, i.e. constructing the retrieval formula as (TI = (A1 OR A2 OR...ORAm)) OR (AK = (A1 OR A2 OR...OR Am)) OR (KP = (A1 OR A2 OR...OR Am)).

7. The two-step document retrieval method based on keyword expansion according to claim 6, wherein, The implementation process of step S6 is: S6.

1. After retrieval, collect all documents hit by parallel retrieval in the three fields, merge based on the unique identifier of each document, remove duplicate documents, and form the de-duplicated in-depth retrieval result set M; S6.

2. Result sorting: calculate the comprehensive score of each document in the result set M, and sort all documents in descending order of total score, see formula (3); Field weight: Weight basis: determined based on field semantic importance experiment (TI > KP > AK), and the weight sum is 1; Including: Adjustability: weights support user customization (must satisfy 'w TI +w KP +w AK = 1).

8. A system for applying the two-step document retrieval method based on keyword expansion according to any one of claims 1 to 7, characterized in that, Word library management module, which is used to store high-frequency theme word library, professional synonym library, and stop word list, and also used to provide word frequency data and synonyms required for calibration for step S1, and provide synonym matching support in step S3 expansion stage, and support periodic automatic / manual update of word library; Retrieval execution module, which is used to generate retrieval formula, call database interface to execute query, and also used to execute step S2 preliminary retrieval and step S5 multi-field parallel retrieval; Keyword processing engine, which is used to realize TF-IDF calculation, word vector similarity analysis, and co-occurrence statistics, and also used to perform candidate word extraction, noise filtering, and semantic expansion in step S3 stage, and rely on pre-trained word vector model; ​ A sorting and adapting module is used for result deduplication, weighted scoring, cross-database field mapping, and processing multi-field search results in step S6, generating a sorting list, and dynamically adjusting weights according to database field structures; A user interaction module is used for receiving user input, presenting an extended keyword set and search results, and receiving initial keywords in step S1, providing a manual adjustment interface in S4, and outputting a final sorting result.

Images