An academic citation intelligent recommendation method and system

Abstract

The application discloses an academic reference intelligent recommendation method and system, wherein the system comprises a literature import and analysis module, a text vectorization module, a literature library priority management module, a selected text receiving and cutting module, a vector semantic retrieval module, an original text evidence extraction module and a reference text generation module. The academic reference intelligent recommendation method and system provided by the application have accurate personalized scheduling capability, realize the rigid retrieval logic of "user will is prior to algorithm similarity", accurately control the priority of the recommendation source, retrieve original text paragraphs from the local literature library as evidence display, completely solve the false reference problem caused by AI "hallucination", and are compatible with the existing workflow without conflict, have the advantages of low cost and high privacy protection, and are also applicable to multi-scene configuration and rapid switching.

Description

Technical Field

[0001] This invention relates to the field of intelligent search engines, and more particularly to an intelligent recommendation method and system for academic citations. Background Technology

[0002] The act of citing in academic papers is essentially the process by which authors establish connections between their current research and existing knowledge units. However, with the exponential growth in the number of papers published across various disciplines, the core dilemma faced by authors has evolved from "the knowledge base not existing" to "failing to recall previously encountered knowledge units while writing." This dilemma stems from the fact that human semantic memory retrieval relies on incomplete contextual cues, while computer systems have significant advantages in storing and semantically matching large-scale texts.

[0003] General academic search engines (such as Google Scholar) employ a passive retrieval model, requiring users to actively input keywords. Search results are highly dependent on keyword selection, and the returned results contain a significant amount of noise. Reference management software (such as EndNote and Zotero) helps users store and organize references, but still relies on manual input and citation insertion, lacking intelligent suggestions based on semantic understanding.

[0004] The publicly available document recommendation systems mainly share the following common limitations:

[0005] Recommendation logic is fixed: the system makes recommendations based on preset similarity algorithms or citation relationships, and users cannot intervene in the recommendation priority according to their personalized needs;

[0006] The triggering method is singular: users need to actively initiate a search request, and it is impossible to automatically complete the full-text review after the paper is completed;

[0007] The search granularity is too coarse: it recommends the entire document as the unit of recommendation, making it impossible to locate specific paragraphs or data in the document;

[0008] Poor output verifiability: It only outputs a list of references without providing original text evidence, making it impossible for users to verify the reliability of the recommendations. Summary of the Invention

[0009] In view of the aforementioned shortcomings of existing technologies, the technical problem to be solved by this invention is that existing literature recommendation systems suffer from problems such as rigid recommendation logic, inability to prioritize recommendations based on users' personalized needs, single triggering methods, excessively coarse-grained retrieval, and poor output verifiability. This invention provides an intelligent academic citation recommendation method and system that can adjust the priority of recommended sources according to users' personalized needs, features a fully automated manuscript review mechanism, verifiable original text evidence, compatibility with existing workflows, high privacy protection for the text, low cost, and applicability to various scenarios.

[0010] To achieve the above objectives, this invention provides an intelligent recommendation method and system for academic citations. The system includes a document import and parsing module, a text vectorization module, a document database priority management module, a selected text receiving and segmentation module, a vector semantic retrieval module, a source text evidence extraction module, and a citation text generation module.

[0011] The document import and parsing module receives PDF documents imported by the user and uses parsing tools to extract the full text content, automatically identifying paragraph boundaries. The text vectorization module segments the parsed full text into natural paragraphs or semantically complete sentences, dividing each segment into independent text blocks. The document library priority management module includes multiple independent document libraries and assigns a search priority number to each library, constructing a search priority queue based on the user-defined numbers. The selected text receiving and segmentation module receives text content selected by the user by dragging and dropping the mouse in the paper document, triggers a search, and automatically segments the selected text. The text is divided into independent sentence units, then converted into vectors, which are passed to the vector semantic retrieval module as query conditions. After receiving the sentence vectors of the selected text, the vector semantic retrieval module performs semantic retrieval layer by layer according to the library sequence number set by the literature library priority management module. The original text evidence extraction module processes the search results, extracts the original text sentences of the literature containing the matching paragraph from the local vector database, i.e., the matched text, and extracts the context of the paragraph and the complete metadata of the literature. The cited text generation module inputs the user's original sentence and the matched original text paragraph into the large language model API and outputs the corresponding results.

[0012] Furthermore, the document import and parsing module also includes parsing the metadata information of the documents, including title, author, journal name, publication year, and DOI.

[0013] Furthermore, the text vectorization module segments the parsed full text of the document into natural paragraphs or semantically complete sentences. After segmenting each text into independent text blocks, it also includes converting each text block into a fixed-dimensional vector representation.

[0014] Furthermore, the system stores the triple mapping relationship between vectors, original text, and document metadata into a local vector database to complete the index construction of the document library.

[0015] Furthermore, multiple independent document repositories are kept logically independent, without overlapping or merging with each other.

[0016] Furthermore, the selected text receiving and segmentation module uses periods, question marks, and exclamation marks as delimiters to automatically segment the selected text into independent sentence units. Then, each sentence is converted into a vector and passed to the semantic retrieval module as a query condition.

[0017] Furthermore, the vector semantic retrieval module sequentially performs vector similarity retrieval in the document database, calculates the cosine similarity between the sentence vector and the paragraph vectors of all documents in the database, and extracts the highest similarity score. If the highest similarity score is greater than the preset similarity threshold, the retrieval process is terminated, and only the matching results in that database are returned to the user. If there are no matching results in the preceding document database, the retrieval is sequentially performed in the subsequent document databases until all active document databases are traversed or a certain layer is hit.

[0018] Furthermore, the citation text generation module inputs the user's original sentence and the matched original paragraph into the large language model API and outputs the corresponding results, specifically including three output results: generating a matching explanation, generating suggested citation statements, and generating reference entries that conform to academic norms.

[0019] In a preferred embodiment of the present invention, an intelligent academic citation recommendation method based on any of the above-mentioned intelligent academic citation recommendation systems is provided, comprising the following steps:

[0020] It receives PDF documents imported by users, uses parsing tools to extract the full text of the documents, and automatically identifies paragraph boundaries.

[0021] Then, the parsed full text of the document is segmented into natural paragraphs or semantically complete sentences, and each paragraph is segmented into independent text blocks;

[0022] Build multiple independent document databases and assign a search priority number to each database. Construct a search priority queue based on the user-defined number.

[0023] Users drag and drop selected text content in the paper document with the mouse, then trigger the search, and the selected text is automatically segmented into independent sentence units, which are then converted into vectors as search conditions;

[0024] Based on the search criteria, semantic retrieval is performed layer by layer according to the database sequence number set by the document database priority management module;

[0025] The search results are processed to extract the original text of the document containing the matching paragraph from the local vector database, i.e. the matched text, and at the same time, the context of the paragraph and the complete metadata of the document in which it is located are extracted.

[0026] The user-selected sentences to be matched are organized side by side with the matched original paragraphs to form a data structure for evidence comparison, which prepares materials for subsequent generation of cited texts;

[0027] The user's original sentence and the matched original text paragraph are input into the large language model API and the corresponding results are output.

[0028] Furthermore, let the set of literature databases be K = {K1, K2, ..., K}. n}, where the subscript is the retrieval priority number.

[0029] The intelligent recommendation method and system for academic citations provided by this invention have the following specific technical effects:

[0030] 1. User-driven priority scheduling mechanism: The "user-defined document database number" is used as the first weight for retrieval, so that the recommendation logic follows the user's will rather than the algorithm's preference. This is a specific application innovation of retrieval priority control methods in the field of information retrieval.

[0031] 2. High-order truncation retrieval strategy: Once the high-order number database is hit, the subsequent retrieval is terminated to ensure the rigid execution of priority rules, which is different from the conventional "fusion sorting" or "weighted summarization" methods.

[0032] 3. Automatic segmentation and sentence-by-sentence matching of selected text: After the user selects the text to be reviewed, the system automatically segments it into sentence units using periods, question marks, and exclamation marks as delimiters and matches them sentence by sentence, which is different from the traditional search method that requires the user to enter keywords or sentences one by one.

[0033] 4. Paragraph-level evidence retrieval and presentation: The retrieval granularity is refined to the paragraph level and presented in the form of original text evidence, enabling users to judge the reliability of recommendations based on real text, which is different from the document-level output of existing systems.

[0034] 5. Layered management of multi-source document repositories: Supports independent overlaying, dynamic scheduling, and scenario-based saving of document repositories, which is different from the single-repository management mode of existing systems;

[0035] 6. It has multiple application scenarios, such as academic journal editors, research institutes and universities, graduate students and researchers for personal use, and developers of academic writing assistance tools.

[0036] The following will further explain the concept, specific structure, and technical effects of the present invention in order to fully understand the purpose, features, and effects of the present invention. Detailed Implementation

[0037] To make the technical problems, solutions, and beneficial effects of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0038] In the following description, specific details, such as particular internal procedures and techniques, are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will appreciate that the invention may be practiced in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of the invention with unnecessary detail.

[0039] This invention provides an intelligent recommendation method and system for academic citations. The system includes a document import and parsing module, a text vectorization module, a document database priority management module, a selected text receiving and segmentation module, a vector semantic retrieval module, a original text evidence extraction module, and a citation text generation module. Details are as follows:

[0040] The document import and parsing module receives PDF document files imported by users, extracts the full-text content of the documents using open-source PDF parsing tools, and automatically identifies paragraph boundaries. This module also parses the document's metadata information, including title, author, journal name, publication year, DOI, etc., providing a complete source of document information for subsequent search result display.

[0041] The text vectorization module segments the parsed full-text document into natural paragraphs or semantically complete sentences, dividing each segment into independent text blocks. Then, this module uses a local embedding model (such as sentence-transformers) to transform each text block into a fixed-dimensional vector. This vector is essentially a "semantic fingerprint" of the text, enabling the computer to compare the semantic relevance between different texts through mathematical calculations (such as cosine similarity) rather than precise keyword matching. The system stores the triple mapping relationship of "vector-original text-document metadata" in a local vector database, completing the index construction of the document library. Furthermore, text vectorization is the process of converting natural language text into a mathematical representation that can be computed by a computer. A text segment is mapped to a fixed-dimensional floating-point array (usually 384-dimensional or 768-dimensional) through an embedding model, which constitutes a point in mathematical space. Semantically similar texts are closer in this space, while semantically unrelated texts are farther apart. Therefore, comparing whether two texts are related essentially involves calculating the cosine distance between the two sets of vectors, rather than performing semantic understanding. The core advantage of this mechanism lies in its ability to identify generalized semantic relationships such as synonyms and hyponyms, rather than being limited to precise keyword matching. This is the technical basis for the system to achieve "semantic retrieval" rather than "keyword retrieval".

[0042] The literature library priority management module supports users importing multiple independent literature libraries simultaneously and assigning a search priority number (1, 2, 3, ...) to each library. Each library remains logically independent, without overwriting or merging others. In this embodiment, multiple independent literature libraries can be moved and adjusted via drag and drop. Users can enable or disable any library at any time, and can also adjust the order of the libraries' numbers through drag-and-drop operations. The system constructs a search priority queue based on the user-defined numbers—the smaller the number, the higher the search priority. Users can also save the current activation status and number sorting of all literature libraries as an independent "scenario configuration file," allowing for one-click switching between different writing tasks without repeated configuration.

[0043] The selected text receiving and segmentation module receives text content that the user selects by dragging and dropping with the mouse in the paper document. After completing the paper, the user selects a specific paragraph or sentence they wish to review and clicks the "Citation Review" button to trigger the search. This module automatically segments the selected text into independent sentence units, using periods, question marks, and exclamation marks as delimiters. Each sentence is then converted into a vector, which is passed to the semantic retrieval module as query conditions. This "select-to-review" trigger mechanism gives the user complete control—the user only initiates the review when needed; the system does not scan the entire text and does not interfere with the user's normal writing process.

[0044] The vector semantic retrieval module receives sentence vectors from selected text and performs semantic retrieval layer by layer strictly according to the library sequence number set by the library priority management module. Specifically, the system first performs vector similarity retrieval in the library with sequence number 1, calculating the cosine similarity between the sentence vector and the paragraph vectors of all documents in the library, and taking the highest similarity score. If this score is greater than the preset similarity threshold (the system default threshold is 0.7, which can be adjusted by the user), the retrieval process is immediately terminated, and only the matching results in that library are returned to the user—this mechanism is called "high-order truncation," and its core effect is that once a high-order library is matched, even if there are documents with higher semantic similarity in lower-order libraries, they will not be presented. If there are no matching results in the library with sequence number 1, the system continues to search in the library with sequence number 2, and so on, until all active library numbers are traversed or a certain layer is matched. The essence of this mechanism is that "the user-defined library sequence number has the highest weight over semantic similarity."

[0045] The system organizes the sentences selected by the user to be matched alongside the matched original paragraphs to form a "data structure for evidence comparison," which prepares materials for subsequent generation of cited text.

[0046] The citation text generation module inputs the user's original sentence and the matched original paragraph into the large language model API. The model then outputs three things: First, it generates a matching explanation—using natural language to explain why the reference paragraph is semantically relevant to the user's original sentence; second, it generates suggested citations—statements that can be directly copied into the paper as citations (e.g., "Zhang et al. (2024) showed that…"); third, it generates academically compliant reference entries—the system formats the literature metadata into standard reference entry text based on the user's preset formatting preferences (supporting multiple mainstream academic formats such as APA, MLA, GB / T 7714, Chicago, etc.). The system only presents this text content to the user through the interactive interface, allowing the user to copy and paste it into the paper; the system does not perform any automatic document modification operations.

[0047] Another preferred embodiment of the present invention provides an intelligent recommendation method for academic citations, implemented based on an intelligent recommendation system for academic citations. The specific implementation process is the same as described above, and the identical parts will not be repeated. It should be noted that in the method of this embodiment, it is assumed that the set of literature databases activated by the user is K = {K1, K2, ..., K}. n}, where the subscript represents the search priority number. For each sentence T obtained after segmenting the user-selected text, the system first sets i=1 and retrieves the highest similarity score S in Kᵢ. max If S max If the value is greater than or equal to the preset threshold θ, then immediately return the matching result of Kᵢ and terminate all searches; if S... max If θ < 1, then let i = i + 1 and repeat the above steps until a certain layer is hit or the entire activation library is traversed. The final recommendation results presented to the user include the following six fields: the user's original sentence, matching evidence (original paragraph and complete literature information), semantic similarity score, explanation of the matching reason, suggested citation statement, and formatted reference entries.

[0048] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.

Claims

1. An intelligent recommendation system for academic citations, characterized in that, It includes modules for document import and parsing, text vectorization, document database priority management, selected text receiving and segmentation, vector semantic retrieval, original text evidence extraction, and cited text generation; among which, The document import and parsing module receives PDF documents imported by the user and extracts the full text content using parsing tools, automatically identifying paragraph boundaries. The text vectorization module segments the parsed full text into natural paragraphs or semantically complete sentences, dividing each paragraph into independent text blocks. The document library priority management module includes multiple independent document libraries and assigns a search priority number to each library, constructing a search priority queue based on the user-defined number. The selected text receiving and segmentation module receives text content selected by the user by dragging and dropping the mouse in the paper document, triggers a search, and automatically segments the selected text. The sentence vectors are then converted into independent sentence units and passed to the vector semantic retrieval module as query conditions. After receiving the sentence vectors of the selected text, the vector semantic retrieval module performs semantic retrieval layer by layer according to the library sequence number set by the literature library priority management module. The original text evidence extraction module processes the search results, extracts the original text sentences of the literature containing the matching paragraph from the local vector database, i.e., the matched text, and extracts the context of the paragraph and the complete metadata of the literature. The cited text generation module inputs the user's original sentence and the matched original text paragraph into the large language model API and outputs the corresponding results.

2. The intelligent recommendation system for academic citations as described in claim 1, characterized in that, The document import and parsing module also includes parsing the metadata information of the documents, including title, author, journal name, publication year, and DOI.

3. The intelligent recommendation system for academic citations as described in claim 1, characterized in that, The text vectorization module segments the parsed full text of the document into natural paragraphs or semantically complete sentences. After segmenting each text into independent text blocks, it also includes converting each text block into a vector representation with fixed dimensions.

4. The intelligent recommendation system for academic citations as described in claim 3, characterized in that, The system stores the triple mapping relationship between vectors, original text, and document metadata into a local vector database to complete the index construction of the document library.

5. The intelligent recommendation system for academic citations as described in claim 1, characterized in that, Multiple independent document repositories are kept logically independent, without overlapping or merging with each other.

6. The intelligent recommendation system for academic citations as described in claim 1, characterized in that, The selected text receiving and segmentation module uses periods, question marks, and exclamation marks as delimiters to automatically segment the selected text into independent sentence units. Then, each sentence is converted into a vector and passed to the semantic retrieval module as a query condition.

7. The intelligent recommendation system for academic citations as described in claim 1, characterized in that, The vector semantic retrieval module sequentially performs vector similarity retrieval in the document database, calculates the cosine similarity between the sentence vector and the paragraph vectors of all documents in the database, and takes out the highest similarity score; if the highest similarity score is greater than the preset similarity threshold, the retrieval process is terminated, and only the matching results in that database are returned to the user; if there are no matching results in the preceding document database, the retrieval is sequentially performed in the subsequent document databases until all active document databases are traversed or a certain layer is hit.

8. The intelligent recommendation system for academic citations as described in claim 1, characterized in that, The cited text generation module inputs the user's original sentence and the matched original paragraph into the large language model API and outputs the corresponding results, specifically including three output results: generating a matching explanation, generating suggested citation statements, and generating reference entries that conform to academic norms.

9. An intelligent academic citation recommendation method based on an intelligent academic citation recommendation system as described in any one of claims 1-8, characterized in that, Includes the following steps: It receives PDF documents imported by users, uses parsing tools to extract the full text of the documents, and automatically identifies paragraph boundaries. Then, the parsed full text of the document is segmented into natural paragraphs or semantically complete sentences, and each paragraph is segmented into independent text blocks; Build multiple independent document databases and assign a search priority number to each database. Construct a search priority queue based on the user-defined number. Users drag and drop selected text content in the paper document with the mouse, then trigger the search, and the selected text is automatically segmented into independent sentence units, which are then converted into vectors as search conditions; Based on the query criteria, semantic retrieval is performed layer by layer according to the library sequence number set by the document library priority management module; The search results are processed to extract the original text of the document containing the matching paragraph from the local vector database, i.e. the matched text, and at the same time, the context of the paragraph and the complete metadata of the document in which it is located are extracted. The user-selected sentences to be matched are organized side by side with the matched original paragraphs to form a data structure for evidence comparison, which prepares materials for subsequent generation of cited texts; The user's original sentence and the matched original text paragraph are input into the large language model API and the corresponding results are output.

10. The method as described in claim 9, characterized in that, Let the set of literature databases be K = {K1, K2, ..., K}. n }, where the subscript is the retrieval priority number.