Manuscript proofreading method and device

Abstract

The application provides a manuscript proofreading method and device, which are applied to the technical field of publishing manuscripts. The method comprises the following steps: analyzing a manuscript to be proofread to obtain a plurality of content units and a plurality of multimodal document object models matched with the plurality of content units; processing the plurality of content units and the plurality of multimodal document object models based on proofreading rules to determine first candidate questions and first segment anchor points; performing entity extraction and external fact extraction on hit content and hit content context through a planner to obtain global fact context; determining proofreading task context and an intelligent agent based on the application scenario of the manuscript to be proofread, the global fact context and a plurality of object nodes; processing the hit content and the hit content context according to the proofreading task context by using the intelligent agent to obtain second candidate questions and second segment anchor points, and combining the first candidate questions and the first segment anchor points to obtain target questions and target anchor points, so as to map the target questions to the manuscript to be proofread and obtain a proofreading result.

Description

Technical Field

[0001] This application relates to the field of published manuscripts, and more specifically to a manuscript review method and apparatus. Background Technology

[0002] Manuscript review is a process of comprehensively reviewing and proofreading the content and form of a manuscript before its formal publication, thereby ensuring that the manuscript meets quality standards.

[0003] In related technologies, before reviewing a manuscript, it is necessary to convert the manuscript to be reviewed into plain text. This will result in the loss of information such as position in the manuscript, making it impossible to accurately map the problems found in the review back to the manuscript, thus affecting the efficiency of subsequent processing of the manuscript. Summary of the Invention

[0004] In view of the above problems, embodiments of this application provide a manuscript review method and apparatus.

[0005] According to a first aspect of this application, a manuscript review method is provided, the method comprising: in response to receiving a manuscript to be reviewed, performing structured parsing on the manuscript to be reviewed to obtain multiple content units and a multimodal document object model matching the multiple content units, the multimodal document object model including multiple object nodes, object nodes representing content units, object nodes corresponding to the position information and structural information of the content units in the manuscript to be reviewed, the content units including at least one of text, title, table, image and formula; processing the multiple content units and the multimodal document object model based on review rules to determine a first candidate question and a first segment anchor point, the first segment anchor point including the hit content matching the first candidate question, the position information corresponding to the hit content and the context of the hit content; and processing the first candidate question through a planner. The process involves selecting a question, the hit content, and the hit content context, and determining the review mode for these elements. If the review mode is expert mode or a hybrid mode, a planner extracts entities and external facts from the hit content and its context to obtain a global fact context. Based on the application scenario of the manuscript to be reviewed, the global fact context, and multiple object nodes, the review task context and the agent are determined. The agent processes the hit content and its context according to the review task context to obtain a second candidate question and a second segment anchor. Based on the first candidate question, the first segment anchor, the second candidate question, and the second segment anchor, the target question and the target anchor are obtained. Based on the location information indicated by the target anchor, the target question is mapped to the manuscript to be reviewed, yielding the review result.

[0006] According to a second aspect of this application, a manuscript review apparatus is provided, comprising: a parsing module, configured to, in response to receiving a manuscript to be reviewed, perform structured parsing on the manuscript to be reviewed to obtain multiple content units and a multimodal document object model matching the multiple content units, the multimodal document object model including multiple object nodes, object nodes representing content units, object nodes corresponding to the position information and structural information of the content units in the manuscript to be reviewed, the content units including at least one of text, titles, tables, images, and formulas; a first review module, configured to process the multiple content units and the multimodal document object model based on review rules, determine a first candidate question and a first segment anchor point, the first segment anchor point including the hit content matching the first candidate question, the position information corresponding to the hit content, and the context of the hit content; and a first planning module, configured to process the first candidate question and the hit content through a planner. The system comprises the following modules: a first module and a second module. The first module defines the review mode based on the content and the context of the hit content. The second module, when the review mode is expert mode or a hybrid mode, uses a planner to extract entities and external facts from the hit content and its context to obtain a global fact context. Based on the application scenario of the manuscript to be reviewed, the global fact context, and multiple object nodes, the system determines the review task context and at least one agent. The second review module uses the agent to process the hit content and its context according to the review task context to obtain a second candidate question and a second segment anchor. The third module, an acquisition module, uses the first candidate question, the first segment anchor, the second candidate question, and the second segment anchor to obtain the target question and the target anchor. The fourth module, a write-back module, uses the location information indicated by the target anchor to map the target question to the manuscript to be reviewed, obtaining the review result.

[0007] According to a third aspect of this application, an electronic device is provided, comprising: one or more processors; and a memory for storing one or more computer programs, wherein the one or more processors execute the one or more computer programs to implement the steps of the method described above.

[0008] According to a fourth aspect of this application, a computer-readable storage medium is also provided, on which a computer program or instructions are stored, wherein the computer program or instructions, when executed by a processor, implement the steps of the above-described method.

[0009] According to a fifth aspect of this application, a computer program product is also provided, including a computer program or instructions that, when executed by a processor, implement the steps of the above-described method. Attached Figure Description

[0010] The above and other objects, features and advantages of this application will become clearer from the following description of embodiments with reference to the accompanying drawings, in which:

[0011] Figure 1 The illustrations depict application scenarios of the manuscript review method, apparatus, device, medium, and program product according to embodiments of this application.

[0012] Figure 2 A flowchart illustrating a manuscript review method according to an embodiment of this application is shown schematically.

[0013] Figure 3 A flowchart illustrating the parsing of a draft document according to an embodiment of this application is shown schematically.

[0014] Figure 4 A flowchart illustrating the use of an intelligent agent to process hit content and hit content context according to an embodiment of this application is shown in the illustration.

[0015] Figure 5 A flowchart illustrating the mapping target problem according to an embodiment of this application is shown schematically;

[0016] Figure 6 A flowchart illustrating the updating review rules according to an embodiment of this application is shown schematically;

[0017] Figure 7 This schematic diagram illustrates the structural block diagram of a manuscript review apparatus according to an embodiment of this application;

[0018] Figure 8 A block diagram schematically illustrates an electronic device suitable for implementing a manuscript review method according to an embodiment of this application. Detailed Implementation

[0019] The embodiments of this application will now be described with reference to the accompanying drawings. However, it should be understood that these descriptions are exemplary only and are not intended to limit the scope of this application. In the following detailed description, numerous specific details are set forth to provide a thorough understanding of the embodiments of this application for ease of explanation. However, it will be apparent that one or more embodiments may be implemented without these specific details. Furthermore, descriptions of well-known structures and technologies are omitted in the following description to avoid unnecessarily obscuring the concepts of this application.

[0020] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The terms “comprising,” “including,” etc., as used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0021] All terms used herein (including technical and scientific terms) have the meanings commonly understood by those skilled in the art, unless otherwise defined. It should be noted that the terms used herein are to be interpreted in a manner consistent with the context of this specification, and not in an idealized or overly rigid way.

[0022] When using expressions such as "at least one of A, B and C", they should generally be interpreted in accordance with the meaning that is commonly understood by those skilled in the art (e.g., "a system having at least one of A, B and C" should include, but is not limited to, a system having A alone, a system having B alone, a system having C alone, a system having A and B, a system having A and C, a system having B and C, and / or a system having A, B and C, etc.).

[0023] Manuscript proofreading is a process of comprehensively reviewing and proofreading the content and format of a manuscript before its formal publication, thereby ensuring that the manuscript meets quality standards. It typically involves simultaneously addressing issues such as typos, punctuation, number formatting, terminology consistency, fact-checking, figures, formulas, references, duplicate content, and layout problems.

[0024] In related technologies, the content of the manuscript to be reviewed is reviewed based on review rules or general large models.

[0025] However, for manuscripts in image or portable document formats awaiting review, conversion tools must first be used to convert them into editable versions in order to read and proofread the content. This results in the destruction of the document structure and the loss of information such as formatting and location, making it difficult to rewrite the final review suggestions back into the manuscript awaiting review.

[0026] Furthermore, when manuscripts awaiting review contain multimodal content such as charts, formulas, and images, the visual information is difficult for the review rules to recognize. This makes it difficult to consistently and accurately detect issues related to tables and images, such as mismatched text and images, and misinterpretation of table content. Additionally, the review rules are not updated with changes to the manuscripts awaiting review, and the accuracy of the review process needs improvement.

[0027] Furthermore, relying solely on general large models for manuscript review can lead to review errors due to the illusion of large models.

[0028] To address at least one of the aforementioned technical problems, embodiments of this application provide a manuscript review method. This method involves structurally parsing the manuscript to be reviewed to obtain multiple content units and a multimodal document object model matching these content units. Based on the multiple content units and the multimodal document object model, a dual review and fusion process is performed using review rules and an intelligent agent to obtain the target problem, thereby improving the review accuracy of the manuscript. Simultaneously, the target problem is mapped to the manuscript using target anchor points, obtaining the review result while avoiding damage to the original manuscript. Furthermore, a planner is used to determine the review task context for the hit content and its context, guiding the intelligent agent to process the aforementioned content, improving the flexibility and security of the agent's calls, ensuring that the task conforms to the characteristics of the aforementioned content, and improving review accuracy.

[0029] Figure 1 The illustration depicts an application scenario of the manuscript review method according to an embodiment of this application. For example... Figure 1 As shown, application scenario 100 according to an embodiment of this application may include a first terminal device 101, a second terminal device 102, a third terminal device 103, a network 104, and a server 105. The network 104 serves as a medium for providing a communication link between the first terminal device 101, the second terminal device 102, the third terminal device 103, and the server 105. The network 104 may include various connection types, such as wired or wireless communication links or fiber optic cables. For example, a user can use the first terminal device 101, the second terminal device 102, and the third terminal device 103 to interact with the server 105 through the network 104 to receive or send information, etc.

[0030] The first terminal device 101, the second terminal device 102, and the third terminal device 103 can be electronic devices such as smartphones, wearable devices, personal computers, intelligent voice interaction devices, smart home appliances, intelligent vehicles, in-vehicle terminals, aircraft, unmanned vending terminals, and extended reality devices. Extended reality devices can include virtual reality devices, augmented reality devices, and mixed reality devices. A client application for the target application can be installed and run on the terminal devices. This target application can include, but is not limited to, financial transaction applications, payment applications, shopping applications, web browser applications, search applications, instant messaging tools, email clients, and social media platform software (these are just examples). Furthermore, this application embodiment does not limit the form of the target application, and it can include, but is not limited to, applications, mini-programs, etc., installed on the terminal devices, and can also be in the form of web pages.

[0031] Server 105 can be a server providing various services, such as a backend management server supporting websites browsed by users using the first terminal device 101, the second terminal device 102, and the third terminal device 103 (this is just an example). The backend management server can analyze and process received user requests and other data, and feed back the processing results (such as web pages, information, or data obtained or generated according to user requests) to the terminal devices. The server can be a standalone physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services such as cloud databases, cloud computing, cloud functions, cloud storage, network services, cloud communication, middleware services, domain name services, security services, content delivery networks, and basic cloud computing services such as big data. The server can be the backend server of the aforementioned target application, used to provide backend services to the clients of the target application.

[0032] It should be noted that the manuscript review method provided in this application embodiment can generally be executed by server 105 and / or terminal devices 101-103. Accordingly, the manuscript review device provided in this application embodiment can generally be set in server 105 and / or terminal devices 101-103.

[0033] It should be understood that Figure 1 The number of terminal devices, networks, and servers shown is merely illustrative. Depending on implementation needs, any number of terminal devices, networks, and servers can be included.

[0034] Figure 2 A flowchart illustrating a manuscript review method according to an embodiment of this application is shown schematically. Figure 2 As shown, the manuscript review method 200 according to the embodiments of this application may include steps S210 to S270.

[0035] In step S210, in response to receiving the manuscript to be reviewed, the manuscript to be reviewed is subjected to structured parsing to obtain multiple content units and a multimodal document object model that matches the multiple content units.

[0036] According to embodiments of this application, a content unit includes at least one of text, title, table, image and formula. A content unit may refer to a block-level unit or paragraph-level unit after content such as text, image, formula, table, etc. are broken down, such as a text block, semantic paragraph, etc.

[0037] The multimodal document object model includes multiple object nodes; each object node represents a content unit, and each object node corresponds to the positional and structural information of the content unit in the manuscript to be reviewed. The node type of the object node matches the modal type of its corresponding content unit. The node type may include at least one of the following: text node, title node, table node, image node, and formula node.

[0038] Structural information may include the formatting information of content units in the manuscript to be reviewed, the reading order, and the multimodal object relationships between content units and adjacent content units. Multimodal object relationships include, for example, images and figure titles, tables and table titles. Positional relationships may include the paragraph numbering of content units in the manuscript to be reviewed, page boundaries, character ranges within paragraphs (including the start and end points of characters within the paragraph), and text segment (run) numbers.

[0039] Furthermore, when the manuscript to be reviewed is in Word document format, the manuscript to be reviewed can first be parsed into multiple content units and the position and format information of each content unit in the manuscript to be reviewed through document fidelity parsing; then, the multiple content units and the position and format information of each content unit in the manuscript to be reviewed are mapped to the corresponding object nodes to obtain MDOM.

[0040] For example, document fidelity parsing can be achieved by parsing the XML file in the corresponding DOCX compressed package of the DOCX document. For instance, the document structure nodes in the XML file can be traversed, and the content such as the body text, tables, headers and footers, and comments can be split into text units or run_text. For each text unit, a structure index can be created that includes positional information such as paragraph index, run number, start and end of characters within the run, and structural information such as style summary.

[0041] In step S220, multiple content units and multimodal document object models are processed based on the review rules to determine the first candidate issue and the first segment anchor point.

[0042] According to an embodiment of this application, the first segment anchor point includes the hit content that matches the first candidate question, the location information corresponding to the hit content, and the context of the hit content.

[0043] For example, the review rules can refer to rules based on regular expressions or fixed dictionary settings, such as deterministic or semi-deterministic rules for number review, punctuation review, space review, obfuscated word review, terminology review, citation review, and format review. These rules can be used to process multiple content units and multimodal document object models to obtain a first candidate question and the matched content and context of the matched content.

[0044] For example, multiple content units can be processed using the above review rules to obtain the first candidate question and the content that matches the first candidate question; based on the content, the position information of the content in the manuscript to be reviewed can be obtained by searching in the multimodal object document model; the content, the position information of the content in the manuscript to be reviewed, and the context of the content are used as the first segment anchor point.

[0045] In step S230, the planner processes the first candidate question, the hit content, and the hit content context to determine the review mode for the hit content and the hit content context.

[0046] According to embodiments of this application, the review mode can characterize the review method for the hit content and the context of the hit content in the subsequent data processing flow. For example, the review mode may include rule confirmation mode, manual review mode, expert scheduling mode and hybrid mode.

[0047] For example, the planner can be a large model, which can perform semantic analysis on the first candidate question, the hit content, and the context of the hit content to determine whether the first candidate question identified by the review rules is accurate, and then determine the review mode for the hit content and the context of the hit content.

[0048] In one or more embodiments described herein, the term "large model" can refer to a deep learning model with a large number of model parameters, which can include hundreds of millions, tens of billions, hundreds of billions, trillions, or even more than ten trillion model parameters. A large model can also be called a foundational model or basic model. It is pre-trained using a large-scale unlabeled corpus to produce a pre-trained model with hundreds of millions of parameters. Such models can adapt to a wide range of downstream tasks and have good generalization ability; examples include large language models and multimodal pre-trained models.

[0049] In step S240, when the review mode is expert mode or hybrid mode, the planner performs entity extraction and external fact extraction on the hit content and the hit content context to obtain the global fact context; based on the application scenario of the manuscript to be reviewed, the global fact context and multiple object nodes, the review task context and intelligent agent are determined.

[0050] According to embodiments of this application, the global fact context may refer to domain knowledge related to the hit content and the context of the hit content. The application scenarios of the manuscript to be reviewed may include exam papers, academic papers, official documents, books, journals, science subjects, etc. The intelligent agent may include expert intelligent agents, general intelligent agents, etc.

[0051] For example, firstly, entity extraction and external fact extraction can be performed on the hit content and hit content context using a large model to obtain a global fact context. Then, based on the hit content and hit content context, multiple object nodes are searched to obtain object nodes that match the hit content and hit content context. Next, the application scenario of the manuscript to be reviewed, the global fact context, and the object nodes matching the hit content and hit content context are concatenated to obtain the review task context. Finally, semantic analysis is performed on the review task context using a large model to determine the agent that matches the review task context. For example, when calling a general agent to process the hit content and hit content context, if the node type of the object node in the review task context involves tables, then a table review agent needs to be called; if the node type of the object node in the review task context involves images, then an image review agent needs to be called.

[0052] In other embodiments, when the review mode is rule confirmation mode, the first candidate question can be directly used as the target question and the first segment anchor point as the target anchor point for write-back.

[0053] In other embodiments, when the review mode is manual review mode, the first candidate issue can be marked as requiring manual processing and await manual processing.

[0054] In other embodiments, a hybrid mode may refer to using review rules and agents for subsequent review.

[0055] In step S250, the agent processes the hit content and the hit content context according to the review task context to obtain the second candidate question and the second segment anchor point.

[0056] In embodiments of this application, the review context may also include the subject area and natural language type of the manuscript to be reviewed.

[0057] For example, the agent processes the hit content and the hit content context based on the review task context and calls the review tool that matches the review task context to obtain the second candidate question and the second segment anchor.

[0058] In step S260, the target problem and the target anchor point are obtained based on the first candidate problem, the first segment anchor point, the second candidate problem, and the second segment anchor point.

[0059] According to embodiments of this application, a target anchor may include target hit content matching the target question, the context of the target hit content, and the location information of the target hit content in the manuscript to be reviewed.

[0060] Further, step S260 can be: arbitrarily fusing the first candidate problem, the first segment anchor point, the second candidate problem, and the second segment anchor point to obtain the target problem and the target anchor point. For example, merging duplicate problems in the first and second candidate problems. For conflicting problems in the first and second candidate problems, an arbitrating agent can be invoked to arbitrate the first and second candidate problems based on domain knowledge related to them, and output the final result.

[0061] For example, merging duplicate questions in the first candidate questions and the second candidate questions may include: first, clustering multiple first candidate questions and multiple second candidate questions to obtain the question types of each first candidate question and each second candidate question; if any two questions in the multiple first candidate questions and multiple second candidate questions have the same question type and the overlap ratio of their corresponding content units exceeds a preset threshold, then the two questions are considered to be duplicates and merged.

[0062] In step S270, based on the location information indicated by the target anchor point, the target issue is mapped to the manuscript to be reviewed, and the review result is obtained.

[0063] For example, step S270 can be: retrieving the corresponding position in the manuscript to be reviewed based on the position information indicated by the target anchor point, mapping the target issue to the manuscript to be reviewed, and obtaining the review result.

[0064] In the embodiments of this application, the manuscript to be reviewed is structurally parsed to obtain multiple content units and a multimodal document object model matching the multiple content units. The multimodal document object model contains the positional and structural information of the content units in the manuscript to be reviewed. Using multiple content units and the multimodal document object model for review avoids only reviewing plain text or other content in the subsequent review process. The multiple content units and the multimodal object model are initially reviewed using review rules to obtain the first candidate question and the first segment anchor point, thereby providing a data foundation for the planner to determine the review mode, avoiding the need to process the entire data, improving processing efficiency and reducing costs. Then, in the case of a hybrid mode or expert mode, the planner determines the review task context based on the hit content and the actual content in the context of the hit content, ensuring that the subsequent processing of the intelligent agent is closely related to the actual content, improving the accuracy of the intelligent agent's review. Based on the positional and structural information indicated by the target anchor point, the target question is mapped to the manuscript to be reviewed, obtaining the review result while avoiding damage to the original format of the manuscript to be reviewed.

[0065] Figure 3 A flowchart illustrating the parsing of a draft document according to an embodiment of this application is shown.

[0066] Different drafts to be reviewed may have different file formats or contain content of different modalities. For example, the document format may be PDF, scanned, or a multi-column formatted document, and it may contain charts, formulas, and other information. In related technologies, processing such drafts typically only extracts the text content, ignoring cross-modal semantic information such as the coordinates, hierarchy, layout area, text-image anchoring relationships, and formula context. Furthermore, the aforementioned cross-modal semantic information depends on the visual layout of the draft, and review rules for pure text are difficult to accurately identify problems, potentially leading to review defects such as misjudgment of table content, retention of duplicate OCR blocks, text-image mismatch, and missing formula context.

[0067] To address the aforementioned issues, embodiments of this application improve the parsing method for the manuscript to be reviewed. By combining the quantity of each modal content in the manuscript, the complexity of each modal content is determined, thereby determining the page level for selecting the parsing path and providing each page with a parsing method that matches its content, ensuring accurate parsing. Simultaneously, the structural and positional information of each text unit is extracted to provide cross-modal semantic information for subsequent review, improving the accuracy of review.

[0068] Next, combined Figure 3 The improved parsing method is further explained below. In some embodiments, the manuscript to be reviewed includes multiple pages; the manuscript to be reviewed is structured and parsed to obtain multiple content items and a multimodal document object model matching the multiple content items, including:

[0069] First, determine the page level of each page based on its complexity.

[0070] According to embodiments of this application, complexity is determined by at least one of the quantity of each modal content on each page and the proportion of noisy text. Complexity is used to indicate the degree of content complexity on a page; the higher the complexity, the more complex the content on the page. Complexity is positively correlated with page level. Each modal content may include images, tables, formulas, text, etc.

[0071] Furthermore, the complexity is obtained using the following method:

[0072] First, based on the document format of each page, the corresponding parsing method is selected for preliminary parsing to obtain the parsing results for each page. Then, the number of image objects, table objects, formulas, vector drawing objects, abnormal character ratio, short text block ratio, and suspected physical line break ratio are extracted from each parsing result. Next, based on the ratio of the number of each image object to the total number of content blocks on each page, the image density of each page is obtained; based on the ratio of the number of each table object to the total number of content blocks on each page, the table density of each page is obtained; based on the ratio of the number of each formula to the total number of content blocks on each page, the formula density of each page is obtained; based on the ratio of the number of each vector drawing object to the total number of content blocks on each page, the drawing density of each page is obtained. Finally, the image density, table density, formula density, drawing density, abnormal character ratio, short text block ratio, and suspected physical line break ratio of each page are weighted and summed to obtain the complexity of each page.

[0073] Here, the percentage of abnormal characters can be obtained by comparing the number of abnormal characters (such as spaces, garbled characters, etc.) on the page to the total number of characters on the page. The percentage of short text blocks can be obtained by comparing the number of short text blocks on the page to the total number of text blocks on the page. Here, a short text block can refer to a text block of no more than three lines of text. The percentage of suspected physical line breaks can be obtained by comparing the number of suspected physical lines broken on the page to the total number of physical lines on the page. Here, a physical line can refer to a continuous and complete line of text on the page, while a suspected physical line break can refer to a physical line that may be cut off by page boundaries, page breaks, or image edges, where only the top or bottom half of the text is displayed, resulting in incomplete content.

[0074] For example, selecting the corresponding parsing method for preliminary parsing based on the document format of each page may include: performing document fidelity parsing on the page when the document format of the page is DOCX; and performing optical character recognition (OCR) or visual model parsing on the page when the document format of the page is a scanned document or image-type PDF.

[0075] For example, complexity score It can be represented as:

[0076] ;

[0077] In the formula, w1, w2, w3, w4, w5, w6, and w7 represent weighting coefficients, respectively. density Represents image density, table density Indicates table density, formula density Represents the formula density, drawing density Indicates plot density, noise ratio Indicates the proportion of abnormal characters, fragmentratio soft_break represents the proportion of short text blocks. ratio This indicates the percentage of suspected physical line cuts.

[0078] For example, for any target page among all pages, if the complexity of the target page is less than a preset first threshold, the target page may be a simple text page, and the page level of the target page is A; if the complexity of the target page is greater than or equal to the preset first threshold and less than the preset second threshold, the target page may be a native PDF page, and the page level of the target page is B; if the complexity of the target page is greater than or equal to the preset second threshold and less than the preset third threshold, the target page may be a mixed text and image page, and the page level of the target page is C; if the complexity of the target page is greater than or equal to the preset third threshold and less than the preset fourth threshold, the target page may be a scanned page or a garbled text page, and the page level of the target page is D; if the complexity of the target page is greater than or equal to the preset fourth threshold and less than the preset fifth threshold, the target page may be a page with dense tables, formulas, and vector graphics, and the page level of the target page is E.

[0079] Then, based on each page level, the target parsing path 302 for each page is determined from multiple candidate parsing paths 301.

[0080] According to embodiments of this application, each candidate parsing path has a corresponding page level. The candidate parsing paths include at least one of document fidelity parsing, structural parsing, hybrid parsing, optical character recognition (OCR), and visual language model parsing. Hybrid parsing may include both OCR and visual language model parsing.

[0081] For example, when the page level is A, the target parsing path is document fidelity parsing; when the page level is B, the target parsing path is structural parsing, such as PDF structural parsing; when the page level is C, the target parsing path is hybrid parsing; when the page level is D, the target parsing path is optical character recognition or visual language model parsing; and when the page level is E, the target parsing path is specialized visual language model parsing or cross-modal layout parsing. In addition, if the confidence level of the parsing result is low, an expert agent can be introduced for verification.

[0082] Then, each page is processed based on the parsing path of each target to obtain the content unit 303 corresponding to each page and the position information and structural information 304 of each content unit in the corresponding page.

[0083] For example, when the page level is E, the target resolution path for this page can be cross-modal layout resolution, and the specific process is as follows:

[0084] First, optical character recognition (OCR) or native PDF text parsing is performed on the page to obtain content blocks and their coordinate frames. Content blocks can include text, images, tables, formulas, etc., and each content block has corresponding formatting information, including font size, line spacing, column spacing, and reading direction. Then, the layout areas on the page are identified. Next, the reading order is reconstructed based on the coordinates indicated by the coordinate frames of the content blocks, font size, line spacing, column spacing, and reading direction, and multimodal object relationships are established between images and figure titles, tables and table titles, formulas and formula numbers, body text and footnotes, and multi-column text.

[0085] Here, the content block refers to the aforementioned content unit. The page layout area may include at least one of the following: body text area, heading area, header and footer area, table area, image area, formula area, and footnote area.

[0086] In other embodiments, when the confidence level of the parsing result is low, the location cannot be reliably determined, or there are conflicts among multiple parsing results, the corresponding page or page area is reviewed.

[0087] Then, each content unit, its corresponding position information and structural information on the page are mapped to each object node 305 to obtain the multimodal document object model.

[0088] According to embodiments of this application, object nodes have corresponding target parsing paths and complexities. That is, the target parsing paths and complexities corresponding to each content object can be synchronously mapped to their corresponding object nodes, thereby providing a basis for subsequent review by the intelligent agent.

[0089] In the embodiments of this application, the page is first parsed using a parsing method that matches the document format of the page. Then, the target parsing path corresponding to the complexity of the content of each page is determined based on the complexity of each page, thereby improving the accuracy of parsing each page. At the same time, data such as structural information, location information, target parsing path and complexity are associated with object nodes to form a multimodal document object model containing various information, so as to provide comprehensive detection data for subsequent review, thereby improving the accuracy of review.

[0090] In other embodiments, a page pre-inspection report can be generated for each page, and each page pre-inspection report can be mapped to the corresponding object node and injected into the multimodal object model.

[0091] For example, a page preflight report can be represented as:

[0092] Page pre-inspection report = {page number, page type, complexity, image density, table density, formula density, drawing density, proportion of short text blocks, proportion of suspected physical line breaks, repetition proportion, target parsing path, alternative parsing path, whether a visual language model is needed, list of reasons}.

[0093] In some embodiments, a planner performs entity extraction and external fact extraction on the hit content and the hit content context to obtain a global fact context; based on the application scenario of the manuscript to be reviewed, the global fact context, and multiple object nodes, the review task context and at least one intelligent agent are determined, including:

[0094] First, the planner extracts entities from the hit content and its context to obtain multiple entities.

[0095] For example, the planner can be a large model, which can extract entities from the hit content and its context according to a preset window length to obtain multiple candidate entities; duplicate entities among the multiple candidate entities are filtered out to obtain multiple entities. Here, multiple entities may include people, organizations, terms, place names, subject concepts, factual nouns, and other entities that require external fact-checking.

[0096] Then, based on multiple entities, a global factual context is obtained by searching the pre-defined domain knowledge base.

[0097] According to embodiments of this application, the global fact context represents domain knowledge related to multiple entities.

[0098] Here, the domain knowledge base may include publishing standards, terminology databases, historical review records, user-defined rules, and authoritative reference format standards.

[0099] In related technologies, if the manuscript content is directly input into an AI agent, which then independently outputs the proofreading results, the agent is prone to generating illusions or false positives due to the lack of constraints from publishing standards, authoritative thesaurus, historical proofreading records, and external knowledge evidence. For example, it may misjudge standard professional terms as incorrect words, misjudge intentional expressions by the author as grammatically incorrect sentences, or generate unfounded revision suggestions.

[0100] Therefore, in the embodiments of this application, in order to improve the reliability of the review results, relevant domain knowledge is retrieved from a preset domain knowledge base as a global factual context for the hit content and the context of the hit content, so as to provide evidentiary constraints for the intelligent agent review and avoid the generation of illusions.

[0101] For example, based on multiple entities, a retrieval enhancement generation tool can be used to retrieve domain knowledge related to multiple entities from a domain knowledge base or external knowledge, and this knowledge can be used as a global fact context.

[0102] Then, based on the application scenario of the manuscript to be reviewed, the global fact context, and multiple object nodes, the review task context and the intelligent agent matching the review task context are obtained.

[0103] For example, firstly, the application scenario, global factual context, and multiple object nodes that match the hit content and the hit content context of the manuscript to be reviewed are concatenated to obtain the review task context; then, agents that match the review task context are selected from multiple candidate agents.

[0104] In other embodiments, the review task context also includes the subject area, natural language type, modality type, etc., to which the hit content and the hit content context belong; the specific process for selecting agents that match the review task context from multiple candidate agents is as follows:

[0105] First, the scheduling weight of each candidate agent is determined based on at least one of the following factors: scene matching factor, domain matching factor, natural language matching factor, historical quality factor, risk factor, and modality matching factor.

[0106] According to embodiments of this application, for any candidate agent among multiple candidate agents, the scenario matching factor is obtained based on the degree of matching between the application scenario of the manuscript to be reviewed and the corresponding candidate agent; the domain matching factor is obtained based on the degree of matching between the subject domain to which the hit content and the context of the hit content belong and the corresponding candidate agent; the natural language matching factor is obtained based on the degree of matching between the natural language type of the hit content and the context of the hit content and the corresponding candidate agent; the historical quality factor is obtained based on the patch adoption rate and human feedback adoption rate of the corresponding candidate agent; the risk factor is obtained based on the degree of risk of the hit content and the context of the hit content; and the modality matching factor is obtained based on the degree of matching between the modality type of the hit content and the context of the hit content and the review capabilities supported by the corresponding candidate agent.

[0107] For example, application scenarios may include proofreading of science and engineering books, proofreading of humanities books, etc.; subject areas may include mathematics, physics, chemistry, biology, history, humanities or general fields; and natural language types may include Chinese, English, Japanese, Russian, Tibetan, Uyghur, Latin script, Cyrillic script, Tibetan alphabet, Arabic script, etc.

[0108] Multiple candidate intelligent agents can include general intelligent agents and expert intelligent agents, such as formula expert intelligent agents, computational expert intelligent agents, science expert intelligent agents, fact-checking expert intelligent agents, knowledge expert intelligent agents, and compliance expert intelligent agents.

[0109] Furthermore, the scheduling weights of each candidate agent are obtained by weighting and summing the scene matching factor, domain matching factor, natural language matching factor, historical quality factor, risk factor, and modality matching factor.

[0110] To further explain the above, in the following discussion, the hit content and the context of the hit content will be collectively referred to as the content to be reviewed.

[0111] Furthermore, when the subject area, natural language type, and modality type of the hit content and its context are the same, the calculation process for the scene matching factor, domain matching factor, natural language matching factor, historical quality factor, risk factor, and modality matching factor of any agent among multiple candidate agents is as follows:

[0112] (1) Based on the similarity between the application scenario of the manuscript to be reviewed and the application scenario supported by the corresponding candidate agent, the scenario matching factor of the corresponding candidate agent is obtained.

[0113] For example, the scene matching factor S(e,u) can be expressed as:

[0114] S(e,u)=S[scenario(u),agent type (e)];

[0115] In the formula, S represents the similarity, scenario(u) represents the application scenario of the content u to be reviewed, and agenttype(e) represents the application scenario supported by the candidate agent e.

[0116] (2) Based on the similarity between the subject area to which the content to be reviewed belongs and the subject area that the corresponding candidate agent is good at, the domain matching factor of the corresponding candidate agent is obtained.

[0117] For example, the domain matching factor D(e,u) can be expressed as:

[0118] D(e,u)=S[domain(u),agent domain (e)];

[0119] In the formula, domain(u) represents the subject area to which the content u to be reviewed belongs, and agent... domain (e) represents the subject area that candidate agent e excels in.

[0120] (3) Based on the similarity between the natural language type of the content to be reviewed and the natural language type supported by the corresponding candidate agent, the natural language matching factor of the corresponding candidate agent is obtained. Here, the natural language type can be a language type or a text type.

[0121] For example, the natural language matching factor L(e,u) can be represented as:

[0122] L(e,u)=S[language(u),agent language (e)];

[0123] In the formula, language(u) represents the natural language type of the content u to be reviewed, and agent... language (e) indicates the type of natural language supported by candidate agent e.

[0124] (4) Based on the number of times the corresponding candidate agent was adopted by human review in history and the number of times the human review had a conclusion in history, the human feedback adoption rate of the corresponding candidate agent is obtained; based on the number of times the patch was actually applied and the total number of times the patch was processed, the patch adoption rate of the corresponding candidate agent is determined; based on the human feedback adoption rate and the patch adoption rate of the corresponding candidate agent, the human feedback score of the corresponding candidate agent is obtained; based on the human feedback score and the arbitration adoption rate of the corresponding candidate agent, the reliability score of the corresponding candidate agent is obtained; based on the reliability score of the corresponding candidate agent, the historical quality factor of the corresponding candidate agent is obtained.

[0125] For example, the historical quality factor Q(e) can be expressed as:

[0126] Q(e)=Clip[0.75+0.5×Rel(e),0.75,1.25];

[0127] The reliability score Rel(e) can be expressed as:

[0128] Rel(e) = 0.7 × F(e) + 0.3 × J(e);

[0129] The human feedback rating F(e) can be expressed as:

[0130] F(e)=α×ReviewAcceptRate(e)+β×PatchAcceptRate(e);

[0131] The review acceptance rate (e) can be expressed as:

[0132] ReviewAcceptRate(e) = the number of times the review was accepted by human reviewers in the past and the number of times the reviewers reached a conclusion in the past.

[0133] The patch acceptance rate, ReviewAcceptRate(e), can be expressed as:

[0134] PatchAcceptRate(e) = Number of times the patch was actually applied / Total number of times the patch was processed;

[0135] In the formula, Clip represents the upper and lower limit clipping function, and α and β are the feedback type weights. For prompting expert agents, α can be greater than β; for expert agents that can directly modify the main text, β can be greater than α.

[0136] (5) Based on the risk level of the content to be reviewed and the degree of matching between the corresponding candidate agent and the risk type of the content to be reviewed, the risk factor of the corresponding candidate agent is obtained.

[0137] For example, the risk factor R(e,u) can be expressed as:

[0138] R(e,u)=1 +λ×RiskLevel(u)×RiskMatch(e,u);

[0139] Here, RiskLevel(u) represents the risk score corresponding to the risk level of the content to be reviewed. The risk score is 0.2 for a low risk level, 0.5 for a high risk level, and 1.0 for a high risk level. RiskMatch(e,u) represents the degree of matching between the candidate agent and the risk type of the content to be reviewed. It is 1 when they match and 0 when they do not match. For example, compliance risk corresponds to compliance expert, factual risk corresponds to fact-checking expert, formula risk corresponds to formula expert, and calculation risk corresponds to calculation expert, which is a match.

[0140] (6) Based on the modal feature indicator value of the content to be reviewed and the capability type indicator value of the corresponding candidate agent, the modal matching factor of the corresponding candidate agent is obtained.

[0141] For example, the modality matching factor M(e,u) can be expressed as:

[0142] M(e,u)=1+μ1×has formula( u)×is formula_agent (e)+μ2×has table (u)×is table_agent (e)+μ3×has image (u)×is visual_agent (e)+μ4×needs vlm (u)×is vlm_agent (e);

[0143] In the formula, has formula( u) represents the formula characteristic indicator value of the content u to be reviewed, used to indicate whether the content u to be reviewed contains a formula; has table (u) represents the table feature indicator value of the content u to be reviewed, used to indicate whether the content u to be reviewed contains a table; has image (u) represents the image feature indicator value of the content u to be reviewed, used to indicate whether the content u to be reviewed contains an image; needs vlm(u) represents the review instruction value for the content u to be reviewed, used to indicate whether the content u to be reviewed needs visual language review; is formula_agent (e) represents the formula review capability indicator value for candidate agent e, used to indicate whether candidate agent e possesses formula review capability; is table_agent (e) represents the table review capability indicator value for candidate agent e, used to indicate whether candidate agent e possesses table review capability; is visual_agent (e) represents the visual content review capability indicator value of candidate agent e, used to indicate whether candidate agent e possesses visual content review capability; is vlm_agent (e) represents the visual language model processing capability indicator value of candidate agent e, used to indicate whether candidate agent e possesses visual language processing capability; μ1, μ2, μ3, and μ4 represent weight parameters. Here, a feature indicator value of 1 indicates that the content to be reviewed u contains the corresponding modal content, a capability indicator value of 1 indicates that the corresponding candidate agent possesses the review capability of the corresponding modal content, and vice versa.

[0144] For example, the scheduling weight W(e,u) can be expressed as:

[0145] W(e,u)=Clip[W0(e)×S(e,u)×D(e,u)×L(e,u)×Q(e)×R(e,u)×M(e,u),Wmin,Wmax];

[0146] In the formula, W0(e) represents the basic weight of the agent, Wmin represents the minimum scheduling weight, and Wmax represents the maximum scheduling weight.

[0147] In the embodiments of this application, the application scenarios supported by the aforementioned candidate intelligent agents, the subject areas they are good at, the types of natural languages ​​they support, the number of times they have been adopted by historical manual review, the number of times they have reached a conclusion in historical manual review, the number of times patches have been actually applied, the total number of times patches have been processed, the arbitration adoption rate, etc., can be obtained through the configuration information corresponding to the candidate intelligent agents.

[0148] In other embodiments, when the subject area, natural language type, or modality of the hit content and the hit content context are different, they can be regarded as multiple contents to be reviewed, and the scheduling weights are calculated separately, and the matching agents are called separately for subsequent review.

[0149] Then, based on each scheduling weight, the agent is determined from multiple candidate agents.

[0150] For example, according to multiple scheduling weights, the corresponding candidate agents are arranged in descending order, and at least one agent ranked at the top is used as the agent for subsequent review.

[0151] In the embodiments of this application, agents for subsequent review are selected from multiple candidate agents based on the review task context, ensuring the matching degree between the review agent's review capabilities and the content to be reviewed, while improving the flexibility of agent invocation during review.

[0152] Figure 4 A flowchart illustrating the use of an agent to process hit content and hit content context according to an embodiment of this application is shown.

[0153] To reduce the risk of intelligent agents accidentally invoking external tools, leaking manuscript content, or incorrectly writing to files, in the embodiments of this application, a tool permission gateway is introduced to control the tool invocation permissions of the intelligent agent during the process of the intelligent agent processing the hit content and the hit content context. Next, combined with... Figure 4 This section further explains the process of using intelligent agents to process hit content and the context of hit content.

[0154] In some embodiments, the agent processes the hit content and the hit content context according to the review task context to obtain the second candidate question and the second segment anchor point, including:

[0155] First, the agent invokes a review tool that matches the review task context. Then, if the review tool invocation is successful, the agent uses the review tool to process the hit content and the hit content context to obtain the second candidate question and the second segment anchor.

[0156] For example, based on the review skills possessed by the intelligent agent, and by using review tools to process the hit content and the context of the hit content, a second candidate question and a second segment anchor point are obtained.

[0157] Here, proofreading skills can include typo proofreading, punctuation proofreading, grammar proofreading, fact-checking, computational proofreading, compliance proofreading, terminology proofreading, and citation proofreading. Proofreading skills can also be referred to as the proofreading capabilities possessed by the agent. Proofreading tools can refer to the external capabilities that the agent can optionally invoke when performing proofreading skills. Proofreading tools include retrieval enhancement generation, sandbox computation, web retrieval, file writing, and visual language models.

[0158] According to embodiments of this application, the invocation result is determined by the tool permission gateway based on the agent's authorized permissions and the tool permission level of the review tool. For example, the permission level for retrieval enhancement generation is level 1, the tool permission level for sandbox computation is level 2, the tool permission level for network retrieval is level 3, and the tool permission level for file writing is level 4.

[0159] Furthermore, the authorization permissions for the intelligent agent can be determined based on the minimum value among the agent's configured permission level, the data sensitivity level of the content to be reviewed, and the tool permission level of the review tool to be invoked. The agent's configured permission level can be obtained from the user's configuration information for the agent. For example, if the user disables network search, the agent's configured permission level can be level 2; if the user disables file writing, the agent's configured permission level can be level 3. The data sensitivity level of the content to be reviewed is determined based on the degree of public access to the content. The data sensitivity level is directly proportional to the degree of public access to the content. For example, if the content to be reviewed is a public document, the data sensitivity level is level 4; if the content to be reviewed is an internal or confidential document, the data sensitivity level is level 2.

[0160] For example, the tool permission gateway can obtain the call result based on the tool permission level of the review tool and the authorization permission of the agent; for example, if the tool permission level is not higher than the authorization permission, it is considered that the verification is passed and the call to the review tool is allowed.

[0161] In the embodiments of this application, the scope of permissions for intelligent agents to call tools is verified through a tool permission gateway to prevent unauthorized calls by intelligent agents, thus meeting the security and confidentiality requirements of the publishing industry. Furthermore, the intelligent agent can be guided to review the hit content and its context based on the global factual context and calculate the confidence level; if the confidence level is lower than a confidence level threshold, a "manual review required" prompt is generated; if the confidence level is greater than or equal to the confidence level threshold, a review conclusion is output.

[0162] For example, the review conclusion can be expressed as:

[0163] Review conclusion = {anchor point identifier, second candidate issue, issue type, original content, suggested revision text, evidence, source, confidence level, action type, whether manual review is required}.

[0164] In other embodiments, if the review conclusion indicates that the confidence level of the second candidate question is lower than a preset confidence threshold or that there is a lack of suggested modification text, the second candidate question is submitted to the arbitration agent for review or marked as requiring manual review.

[0165] In some embodiments, the process of using an agent to process the hit content and the hit content context according to the review task context to obtain the second candidate question and the second segment anchor point further includes: checking the second candidate question and the second segment anchor point based on a plurality of preset checking rules; and discarding the second candidate question if the check result of the original text segment location check fails.

[0166] According to embodiments of this application, the inspection rules include at least one of original text fragment location inspection, structured output format inspection, and field type inspection. The original text fragment location inspection can be exact matching, blank tolerance location inspection, punctuation tolerance location inspection, or fuzzy matching location inspection.

[0167] For example, multiple inspection rules may include: H1: Whether the output is valid JSON; H2: Whether it contains an array of issues; H3: Whether it contains severity, original text fragments, suggested text, and reasoning explanations; H4: Whether high-risk or medium-risk issues contain evidence; H5: Whether the manuscript to be reviewed or the object node can be located through exact matching, blank tolerance location check, punctuation tolerance location check, or fuzzy matching location check; H6: Whether the suggested text is an empty replacement, a complete rewrite, or an out-of-bounds modification; H7: Whether the tool call complies with the authorization permissions of the corresponding agent.

[0168] In the embodiments of this application, the output of the intelligent agent is checked from multiple dimensions to improve the reliability of the review results.

[0169] Figure 5 A flowchart illustrating the mapping target problem according to an embodiment of this application is shown schematically.

[0170] In related technologies, review conclusions may include issues and suggested changes in natural language, such as "it is suggested to replace a certain word with another word." However, the mapping relationship between this suggestion and paragraphs, runs, character offsets, table cells, and comment locations in the original Word document is not preserved. Therefore, when the same issue appears multiple times in the document, or when the issue spans multiple runs, tables, footnotes, or comment areas, it is difficult to accurately write the suggested changes back to the original text, which can easily lead to misaligned changes, lost formatting, or layout corruption.

[0171] To accurately reflect the review findings back to the manuscript awaiting review, in this embodiment, the target issue is mapped to the manuscript using the target anchor point corresponding to the target issue. Next, combined with... Figure 5 The process of mapping target issues to manuscripts awaiting review is further explained.

[0172] In some embodiments, the target anchor point includes target content matching the target question, location information corresponding to the target content, and context of the target content. The location information corresponding to the target content includes the paragraph index and character range within the paragraph of the target content in the manuscript to be reviewed. Based on the location information indicated by the target anchor point, the target question is mapped to the manuscript to be reviewed 501, and the review results include:

[0173] First, based on the target content and the context of the target content in the manuscript to be reviewed 501, the corresponding paragraph index and character range within the paragraph are searched in the manuscript to be reviewed to obtain the target content 502.

[0174] Here, the character range within a segment includes an optional start character and an optional end character.

[0175] For example, target anchor point A can be represented as:

[0176] A = {p, pre, hit, post, c};

[0177] In the formula, p represents the paragraph index, pre represents the context before the target hit content, hit represents the target hit content, post represents the context after the target hit content, and c represents the optional character start point.

[0178] Then, a consistency check is performed on the target hit content 503 and the target content 502. If the consistency check passes, the target anchor point 504 is converted into the patch anchor point 505.

[0179] The following explanation uses the target hit content 503 as an example. For instance, based on the paragraph index p, multiple text fragments corresponding to the paragraph in the manuscript to be reviewed are read and sequentially concatenated into the paragraph plain text. Based on pre, hit, post, and the optional character start point c within the paragraph, the target text is located within the paragraph plain text. Simultaneously, the start and end character positions of the target text within the paragraph plain text are obtained. A consistency check is performed between the target hit text and the target text. If the consistency check passes, based on the text length of each of the multiple runs, the start and end character positions of the target text within the paragraph plain text are mapped to the run sequence number and the start and end offsets within the run, thus obtaining the patch anchor point.

[0180] For example, patch anchor B can be represented as:

[0181] B={p,r,o s ,o e};

[0182] In the formula, r represents the run sequence number, o s Indicates the starting offset within run, o e This indicates the end offset within the run function.

[0183] Then, based on patch anchor 505, target issue 506 is mapped to target content 502 to obtain the review result.

[0184] For example, when the manuscript to be reviewed is a DOCX document, the target content 502 is located in the manuscript to be reviewed 501 based on the patch anchor 505, and w:del and w:ins elements are inserted based on the target issue 506 to form a revision, or comments are generated in the corresponding position, thereby preserving the original style, headers and footers, existing comments, table structure and the ability to accept / reject Word revisions.

[0185] In other embodiments, when the manuscript to be reviewed is a PDF document or a scanned copy, the location information corresponding to the target content in the target anchor point may include page number, coordinate frame, hit text and hit text context; the target content can be located in the page image, preview result or export result of the manuscript to be reviewed based on the target anchor point, and annotations, highlight marks, selection marks or question prompts can be generated at the corresponding positions.

[0186] In the embodiments of this application, it is possible to effectively handle complex versions of documents that are difficult for traditional review tools to handle, and it can also overcome the problem of destroying the original document structure, so that problems and review suggestions can be safely written back to the original text, and increase the editor's disposal and revision permissions.

[0187] Figure 6 A flowchart illustrating the determination of review rules according to an embodiment of this application is shown.

[0188] In related technologies, user feedback on the proofreading results is ignored, making it difficult to adapt to the proofreading preferences of specific publishers, disciplines, or authors. Therefore, in the embodiments of this application, closed-loop learning is performed based on user feedback.

[0189] like Figure 6 As shown, in some embodiments, the method further includes:

[0190] In step S610, in response to receiving user feedback information regarding the review results, the user feedback information is categorized to obtain the feedback type.

[0191] For example, feedback types can include rule false positives, terminology errors, formatting errors, model illusions, and differences in user preferences. Here, user feedback can be categorized using a large model or manually.

[0192] In step S620, if the number of occurrences of user feedback information of the same feedback type is greater than or equal to a preset threshold, a candidate review rule matching the feedback type is generated.

[0193] For example, generating candidate review rules that match the feedback type may include:

[0194] When the feedback type is a rule-based false alarm, cluster analysis is performed on the historical false alarm samples corresponding to the feedback type to extract the common features of compliant but falsely flagged samples. These features are then used as new exclusion conditions for the corresponding review rules to form candidate review rules. For example, if the original review rule is to mark an issue if the text contains "top-level", and the common feature is "top-level industry exhibition", then the candidate review rule is to mark an issue if the text contains "top-level" but does not contain "top-level industry exhibition".

[0195] When the feedback type is terminology error or formatting error, generate a dedicated review rule for the corresponding term or format as a candidate review rule.

[0196] In cases where the feedback type is large model illusion, a new review rule is added as a candidate review rule.

[0197] When the feedback type is user preference difference, the relevant samples and review results are input into the large model, and a rule generation instruction is issued to guide the large model to summarize the common problems and generate candidate review rules.

[0198] In step S630, historical samples are processed based on candidate review rules to obtain the review accuracy of candidate review rules.

[0199] For example, firstly, historical samples are processed through candidate review rules to obtain review results; then, statistical analysis is performed on the review results to obtain their review accuracy in historical samples.

[0200] In step S640, if the review accuracy rate is greater than the preset accuracy rate threshold, the candidate review rule is adopted as the review rule.

[0201] In the embodiments of this application, the review rules are updated based on user feedback information to improve the matching degree between the review method and user preferences, thereby adapting to the review preferences of different users.

[0202] In some embodiments, the method further includes: for any two target content units among a plurality of content units, calculating a global similarity between the two target content units based on at least one of character similarity, cosine similarity, and Hamming distance; obtaining a local coverage rate between the two target content units based on the content overlap between the two target content units; and treating the two content units as similar content units if the global similarity is greater than a preset first threshold or the local coverage rate is greater than a preset second threshold.

[0203] Here, the preset first threshold can be 0.56, and the preset second threshold can be 0.9. Similar content units can be output in the form of a report so that users can decide whether they constitute duplication or a typesetting parsing error.

[0204] Furthermore, the character similarity, cosine similarity, or Hamming distance between any two target content units can be used as the global similarity between them. Alternatively, at least two of the character similarity, cosine similarity, and Hamming distance between any two target content units can be calculated, and then a weighted sum can be taken to obtain the global similarity between them.

[0205] For example, the character similarity between any two target content units is taken as the global similarity between any two target content units. The global similarity J(A,B) is expressed as:

[0206] J(A,B)=|G(A)∩G(B)| / |G(A)∪G(B)|;

[0207] The local coverage C(A,B) is expressed as:

[0208] C(A,B)=|G(A)∩G(B)| / min(|G(A)|,|G(B)|);

[0209] In the formula, G(A) represents the set of character fragments in content unit A, and G(B) represents the set of character fragments in content unit B. Here, J(A,B) ranges from 0 to 1, with a larger value indicating greater overall similarity. C(A,B) ranges from 0 to 1 and is used to detect local repetitions where short text is covered by long text.

[0210] In the embodiments of this application, by performing the above-mentioned global similarity detection on the content units, a quality signal is provided to the user to determine whether there is a duplication or abnormal typesetting parsing.

[0211] In related technologies, regardless of manuscript type, subject area, or content structure, the same set of review rules is used for review. On the one hand, this leads to the frequent triggering of irrelevant review rules, increasing the cost of invocation and generating noise; on the other hand, for manuscripts to be reviewed in specific professional fields, the fixed set of rules is difficult to cover the terminology, format, formulas, references, and publishing standards requirements of specific fields, resulting in missed checks.

[0212] Therefore, in some other embodiments, multiple content units and multimodal document object models are processed based on review rules that match the document to be reviewed. The specific process is as follows:

[0213] First, extract the document features of the manuscript to be reviewed. Then, determine the document category of the manuscript based on the document features. Next, calculate the fit score between multiple candidate review rules and the manuscript based on the document features and document category. Select candidate review rules with a fit score greater than a preset threshold as the review rules. Then, when there are multiple review rules, determine the execution order of the multiple review rules according to risk level and rule cost. For example, high-risk, low-cost rules are prioritized, while low-confidence or high-cost rules are postponed. Finally, supplement the review rules based on the execution results. For example, when a large number of formulas are found, add a formula number consistency check.

[0214] For example, document features may include document type, number of tables, number of formulas, number of images, and language; these can be obtained through statistical analysis of relevant features in the parsing results of the aforementioned pages, such as summing the number of formulas across all pages as the total number of formulas in the manuscript to be reviewed. Document categories can be one of the following: academic papers, textbooks, journal articles, book chapters, popular science articles, or contract texts. Candidate review rules may have corresponding applicable document categories, applicable feature conditions, historical review quality information, and rule execution costs.

[0215] The fit score (r, D) can be represented as:

[0216] ;

[0217] In the formula, type match Indicates category matching degree, feature match Indicates feature matching degree, history precision α represents historical precision, β represents rule cost, and γ, δ represent weights.

[0218] Here, category matching can be obtained by calculating the similarity between the document category and the applicable document category of the candidate review rule, and feature matching can be obtained by calculating the similarity between the document features and the applicable feature conditions of the candidate review rule. Rule cost can refer to the computational resources and execution time required to execute the candidate review rule.

[0219] Based on the above-described manuscript review method, embodiments of this application also provide a manuscript review device. The following will combine... Figure 7 The device is described in detail.

[0220] Figure 7 A schematic block diagram of a manuscript review apparatus according to an embodiment of this application is shown.

[0221] like Figure 7As shown, the manuscript review device 700 of this embodiment includes a parsing module 710, a first review module 720, a first planning module 730, a second planning module 740, a second review module 750, an acquisition module 760, and a write-back module 770.

[0222] The parsing module 710 is used to perform structured parsing on the received manuscript for review, in response to receiving the manuscript to be reviewed, to obtain multiple content units and a multimodal document object model matching the multiple content units. The multimodal document object model includes multiple object nodes, each object node representing a content unit, and each object node corresponds to the positional and structural information of the content unit in the manuscript to be reviewed. The content units include at least one of text, titles, tables, images, and formulas. In one embodiment, the parsing module 710 can be used to perform step S210 described above, which will not be repeated here.

[0223] The first review module 720 is used to process multiple content units and a multimodal document object model based on review rules, and determine a first candidate question and a first segment anchor point. The first segment anchor point includes the matched content that matches the first candidate question, the location information corresponding to the matched content, and the context of the matched content. In one embodiment, the first review module 720 can be used to perform step S220 described above, which will not be repeated here.

[0224] The first planning module 730 is used to process the first candidate question, the hit content, and the hit content context through a planner to determine the review mode for the hit content and the hit content context. In one embodiment, the first planning module 730 can be used to perform the step S230 described above, which will not be repeated here.

[0225] The second planning module 740 is used to extract entities and external facts from the hit content and the hit content context through a planner when the review mode is expert mode or hybrid mode, so as to obtain the global fact context; based on the application scenario of the manuscript to be reviewed, the global fact context and multiple object nodes, the review task context and at least one intelligent agent are determined; in one embodiment, the second planning module 740 can be used to execute the step S240 described above, which will not be repeated here.

[0226] The second review module 750 is used to process the hit content and the hit content context according to the review task context using an intelligent agent to obtain a second candidate question and a second segment anchor point. In one embodiment, the second review module 750 can be used to perform the step S250 described above, which will not be repeated here.

[0227] The acquisition module 760 is used to obtain the target question and the target anchor point based on the first candidate question, the first fragment anchor point, the second candidate question, and the second fragment anchor point. In one embodiment, the acquisition module 760 can be used to perform step S260 described above, which will not be repeated here.

[0228] The write-back module 770 is used to map the target issue to the manuscript to be reviewed based on the location information indicated by the target anchor point, thereby obtaining the review result. In one embodiment, the write-back module 770 can be used to perform step S270 described above, which will not be repeated here.

[0229] According to embodiments of this application, any multiple modules among the parsing module 710, the first verification module 720, the first planning module 730, the second planning module 740, the second verification module 750, the acquisition module 760, and the write-back module 770 can be merged into one module, or any one of these modules can be split into multiple modules. Alternatively, at least some of the functions of one or more of these modules can be combined with at least some of the functions of other modules and implemented in one module. According to embodiments of this application, at least one of the parsing module 710, the first verification module 720, the first planning module 730, the second planning module 740, the second verification module 750, the acquisition module 760, and the write-back module 770 can be at least partially implemented as hardware circuitry, such as a field-programmable gate array, a programmable logic array, a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit, or any other reasonable means of integrating or packaging circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the parsing module 710, the first review module 720, the first planning module 730, the second planning module 740, the second review module 750, the acquisition module 760, and the write-back module 770 can be implemented at least partially as a computer program module, which can perform corresponding functions when the computer program module is run.

[0230] Figure 8 A block diagram schematically illustrates an electronic device suitable for implementing a manuscript review method according to an embodiment of this application.

[0231] like Figure 8As shown, an electronic device 800 according to an embodiment of this application includes a processor 801, which can perform various appropriate actions and processes according to a program stored in a read-only memory 802 or a program loaded from a storage portion 808 into a random access memory 803. The processor 801 may include, for example, a general-purpose microprocessor, an instruction set processor and / or an associated chipset and / or a dedicated microprocessor. The processor 801 may also include onboard memory for caching purposes. The processor 801 may include a single processing unit or multiple processing units for executing different steps of the method flow according to an embodiment of this application.

[0232] Random access memory 803 stores various programs and data required for the operation of electronic device 800. Processor 801, read-only memory 802, and random access memory 803 are interconnected via bus 804. Processor 801 executes various steps of the method flow according to embodiments of this application by executing programs in read-only memory 802 and / or random access memory 803. It should be noted that the programs may also be stored in one or more memories other than read-only memory 802 and random access memory 803. Processor 801 may also execute various steps of the method flow according to embodiments of this application by executing programs stored in said one or more memories.

[0233] According to embodiments of this application, the electronic device 800 may further include an input / output interface 805, which is also connected to a bus 804. The electronic device 800 may also include one or more of the following components connected to the input / output interface 805: an input section 806 including a keyboard, mouse, etc.; an output section 807 including a cathode ray tube, liquid crystal display, etc., and a speaker, etc.; a storage section 808 including a hard disk, etc.; and a communication section 809 including a network interface card, such as a local area network card, modem, etc. The communication section 809 performs communication processing via a network such as the Internet. A drive 810 is also connected to the input / output interface 805 as needed. A removable medium 811, such as a disk, optical disk, magneto-optical disk, semiconductor memory, etc., is installed on the drive 810 as needed so that computer programs read from it can be installed into the storage section 808 as needed.

[0234] Embodiments of this application also provide a computer-readable storage medium, which may be included in the device / apparatus / system described in the above embodiments; or it may exist independently and not assembled into the device / apparatus / system. The computer-readable storage medium carries one or more programs, which, when executed, implement the method according to the embodiments of this application.

[0235] According to embodiments of this application, the computer-readable storage medium can be a non-volatile computer-readable storage medium, such as including but not limited to: portable computer disks, hard disks, random access memory, read-only memory, erasable programmable read-only memory, portable compact disk read-only memory, optical storage devices, magnetic storage devices, or any suitable combination thereof. In embodiments of this application, the computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. For example, according to embodiments of this application, the computer-readable storage medium may include the read-only memory 802 described above, and / or random access memory 803, and / or one or more memories other than read-only memory 802 and random access memory 803.

[0236] Embodiments of this application also include a computer program product comprising a computer program containing program code for performing the methods shown in the flowchart. When the computer program product is run on a computer system, the program code is used to cause the computer system to implement the methods provided in the embodiments of this application.

[0237] In one embodiment, the computer program may rely on a tangible storage medium such as an optical storage device or a magnetic storage device. In another embodiment, the computer program may also be transmitted and distributed in the form of signals over a network medium, and may be downloaded and installed via the communication section 809, and / or installed from a removable medium 811. The program code contained in the computer program can be transmitted using any suitable network medium, including but not limited to: wireless, wired, etc., or any suitable combination thereof.

[0238] In embodiments of this application, the computer program can be downloaded and installed from a network via communication section 809, and / or installed from removable medium 811. When the computer program is executed by processor 801, it performs the functions defined in the system of embodiments of this application. According to embodiments of this application, the systems, devices, apparatuses, modules, units, etc., described above can be implemented by computer program modules.

[0239] According to embodiments of this application, program code for executing the computer programs provided in the embodiments of this application can be written in any combination of one or more programming languages. Specifically, these computational programs can be implemented using high-level procedural and / or object-oriented programming languages, and / or assembly / machine languages. The program code can be executed entirely on the user's computing device, partially on the user's device, partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0240] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this application. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram or flowchart, and combinations of blocks in a block diagram or flowchart, may be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0241] Those skilled in the art will understand that the features described in the various embodiments of this application can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in this application. In particular, the features described in the various embodiments of this application can be combined and / or combined in various ways without departing from the spirit and teachings of this application. All such combinations and / or combinations fall within the scope of this application.

Claims

1. A manuscript review method, characterized in that, The method includes: In response to receiving a manuscript to be reviewed, the manuscript to be reviewed is structured and parsed to obtain multiple content units and a multimodal document object model matching the multiple content units. The multimodal document object model includes multiple object nodes, each object node representing a content unit. Each object node corresponds to the position information and structural information of the content unit in the manuscript to be reviewed. The content unit includes at least one of text, title, table, image and formula. Based on the review rules, the multiple content units and the multimodal document object model are processed to determine the first candidate question and the first segment anchor point. The first segment anchor point includes the hit content that matches the first candidate question, the position information corresponding to the hit content, and the context of the hit content. The planner processes the first candidate question, the hit content, and the hit content context to determine the review mode for the hit content and the hit content context. When the review mode is expert mode or hybrid mode, the planner performs entity extraction and external fact extraction on the hit content and the hit content context to obtain the global fact context; based on the application scenario of the manuscript to be reviewed, the global fact context and multiple object nodes, the review task context and intelligent agent are determined. The agent processes the hit content and the hit content context according to the review task context to obtain the second candidate question and the second segment anchor point; Based on the first candidate question, the first segment anchor, the second candidate question, and the second segment anchor, the target question and the target anchor are obtained; Based on the location information indicated by the target anchor point, the target issue is mapped to the manuscript to be reviewed, and the review result is obtained.

2. The method according to claim 1, characterized in that, The draft to be reviewed includes multiple pages; the structured parsing of the draft to be reviewed to obtain multiple content units and a multimodal document object model matching the multiple content units includes: Based on the complexity of each page, the page level of each page is determined, wherein the complexity is determined by at least one of the quantity of each modal content in each page and the proportion of noisy text; Based on each page level, a target parsing path for each page is determined from multiple candidate parsing paths. Each candidate parsing path has a corresponding page level. The candidate parsing path includes at least one of document fidelity parsing, structure parsing hybrid parsing, optical character recognition, visual language model parsing, and cross-modal layout parsing. Based on the target parsing path, each page is processed to obtain the content unit corresponding to each page and the position information and structural information of each content unit in the corresponding page. The position information includes at least one of paragraph index, character range within paragraph, and page number. The structural information includes format information, reading order, and multimodal object relationship between adjacent content units. Each content unit, its corresponding position information and structural information on the corresponding page are mapped to each object node to obtain a multimodal document object model. The object node has a corresponding target parsing path and complexity.

3. The method according to claim 1, characterized in that, The planner performs entity extraction and external fact extraction on the hit content and the hit content context to obtain the global fact context; Based on the application scenario of the manuscript to be reviewed, the global fact context, and multiple object nodes, determining the review task context and at least one intelligent agent includes: The planner extracts entities from the hit content and its context to obtain multiple entities; Based on the multiple entities, a global fact context is obtained by searching a preset domain knowledge base. The global fact context represents the domain knowledge related to the multiple entities. Based on the application scenario of the manuscript to be reviewed, the global fact context, and the multiple object nodes, a review task context and an intelligent agent matching the review task context are obtained.

4. The method according to claim 3, characterized in that, The review task context also includes at least one of the subject area, natural language type, and modality type to which the hit content and the hit content context belong; the agent matched with the review task context is obtained in the following way: The scheduling weight of each candidate agent is determined based on at least one of the following factors: scene matching factor, domain matching factor, natural language matching factor, historical quality factor, risk factor, and modality matching factor. Based on the scheduling weights, the agent is determined from the plurality of candidate agents; Specifically, for any candidate agent among multiple candidate agents, the scenario matching factor is obtained based on the degree of matching between the application scenario of the manuscript to be reviewed and the corresponding candidate agent; the domain matching factor is obtained based on the degree of matching between the subject domain to which the hit content and the context of the hit content belong and the corresponding candidate agent; the natural language matching factor is obtained based on the degree of matching between the natural language type of the hit content and the context of the hit content and the corresponding candidate agent; the historical quality factor is obtained based on the patch adoption rate and human feedback adoption rate of the corresponding candidate agent; the risk factor is obtained based on the risk level of the hit content and the context of the hit content; and the modality matching factor is obtained based on the degree of matching between the modality type of the hit content and the context of the hit content and the review capabilities supported by the corresponding candidate agent.

5. The method according to claim 1, characterized in that, The step of using the intelligent agent to process the hit content and the hit content context according to the review task context to obtain the second candidate question and the second segment anchor point includes: The agent invokes a review tool that matches the review task context. If the verification tool is successfully invoked, the agent uses the verification tool to process the hit content and the context of the hit content to obtain a second candidate question and a second segment anchor point. The call result is determined by the tool permission gateway based on the authorization permissions of the intelligent agent and the tool permission level of the review tool.

6. The method according to claim 5, characterized in that, The step of using an intelligent agent to process the hit content and the hit content context according to the review task context to obtain the second candidate question and the second segment anchor point further includes: Based on multiple preset inspection rules, the second candidate problem and the second segment anchor point are inspected. The inspection rules include at least one of the following: original text segment location inspection, structured output format inspection, and field type inspection. If the original text fragment location check fails, the second candidate question is discarded.

7. The method according to claim 1, characterized in that, The target anchor point includes the target hit content that matches the target question, the location information corresponding to the target hit content, and the context of the target hit content. The location information corresponding to the target hit content includes the paragraph index and the character range within the paragraph of the target hit content in the manuscript to be reviewed. The step of mapping the target issue to the manuscript to be reviewed based on the location information indicated by the target anchor point, and obtaining the review results, includes: Based on the target content and the paragraph index and character range within the paragraph corresponding to the target content in the manuscript to be reviewed, the target content is retrieved in the manuscript to be reviewed. Perform a consistency check on the target hit content and the target content. If the consistency check passes, convert the target anchor point into a patch anchor point. Based on the patch anchor, the target issue is mapped to the target content to obtain the review result.

8. The method according to claim 1, characterized in that, The method further includes: In response to receiving user feedback information regarding the review results, the user feedback information is categorized to obtain feedback types; If the number of occurrences of user feedback information of the same feedback type is greater than or equal to a preset threshold, a candidate review rule matching the feedback type is generated. Based on the candidate review rules, historical samples are processed to obtain the review accuracy of the candidate review rules; If the review accuracy rate is greater than the preset accuracy rate threshold, the candidate review rule will be used as the review rule.

9. The method according to claim 1, characterized in that, The method further includes: For any two target content units among the plurality of content units, the global similarity between the two target content units is calculated based on at least one of the character similarity, cosine similarity and Hamming distance between the two target content units. Based on the degree of content overlap between any two target content units, the local coverage rate between any two target content units is obtained. If the global similarity is greater than a preset first threshold or the local coverage is greater than a preset second threshold, then any two target content units are considered as similar content units.

10. A manuscript review device, characterized in that, The device includes: The parsing module is used to respond to receiving a manuscript to be reviewed and perform structured parsing on the manuscript to be reviewed and obtained multiple content units and a multimodal document object model matching the multiple content units. The multimodal document object model includes multiple object nodes, each object node representing a content unit. Each object node corresponds to the position information and structural information of the content unit in the manuscript to be reviewed and approved. The content unit includes at least one of text, title, table, image and formula. The first review module is used to process the multiple content units and the multimodal document object model based on review rules, and determine the first candidate question and the first segment anchor point. The first segment anchor point includes the hit content that matches the first candidate question, the position information corresponding to the hit content, and the context of the hit content. The first planning module is used to process the first candidate question, the hit content, and the hit content context through a planner, and determine the review mode for the hit content and the hit content context; The second planning module is used to extract entities and external facts from the hit content and the hit content context through a planner when the review mode is expert mode or hybrid mode, so as to obtain a global fact context; and to determine the review task context and at least one intelligent agent based on the application scenario of the manuscript to be reviewed, the global fact context and multiple object nodes. The second review module is used to process the hit content and the hit content context according to the review task context using the intelligent agent to obtain the second candidate question and the second segment anchor point; The acquisition module is used to obtain the target question and the target anchor point based on the first candidate question, the first segment anchor point, the second candidate question, and the second segment anchor point; The write-back module is used to map the target issue to the manuscript to be reviewed based on the location information indicated by the target anchor point, and obtain the review result.

Images