Entity information association method, device, and storage medium

Abstract

The application discloses an entity information association method and device and a storage medium. The method determines a plurality of semantic regions based on file content information corresponding to a to-be-processed file, extracts entity identifiers and position information, and then determines semantic namespaces to which the entity identifiers belong according to the position information of the entity identifiers and the semantic regions, so that different meanings of the same number in different numbering systems can be effectively distinguished. On this basis, target information entries having an association relationship are determined from a plurality of entity information entries according to an association relationship between the semantic namespaces to which the entity identifiers belong and the entity identifiers, and the plurality of target information entries are associated to the same entity object, so that the risk of incorrectly establishing an association of entity information under different semantic namespaces is reduced, accurate association of multi-source information entries is realized, and the reliability of an association result is significantly improved.

Description

Technical Field

[0001] This application belongs to the field of artificial intelligence technology, and in particular relates to a method, device and storage medium for associating entity information. Background Technology

[0002] In current technical documentation, compound structure diagrams, compound numbers, experimental case numbers, and bioactivity data are typically scattered, and multiple numbering systems often coexist within the same document. Existing methods for associating these scattered information items usually rely solely on identical numbers, page distance, or simple table order to establish correspondences. However, identical numbers in different numbering systems have different meanings, making existing methods prone to information mismatches and compromising the accuracy of the association results.

[0003] Based on this, this application provides a method for associating entity information. Summary of the Invention

[0004] This application provides an entity information association method, device, and storage medium that can accurately associate multi-source information items and significantly improve the reliability of the association results.

[0005] In a first aspect, embodiments of this application provide an entity information association method, the method comprising: Based on the file content information of the file to be processed, multiple semantic regions in the file to be processed are determined, and multiple entity identifiers and the location information corresponding to each entity identifier are extracted from the file content information. The file content information includes at least one text segment and the location information of each text segment in the file to be processed. Extract entity information entries corresponding to each entity object from the semantic region; each entity information entry is associated with a corresponding entity identifier, and multiple entity information entries include structural graph information entries and text information entries; Based on the location information corresponding to each entity identifier and multiple semantic regions, determine the semantic namespace to which each entity identifier belongs; Based on the semantic namespace to which each entity identifier belongs and the association between each entity identifier, determine multiple target information entries with association relationships from multiple entity information entries; Associate multiple target information entries with the same entity object.

[0006] Secondly, embodiments of this application provide an entity information association device, the device comprising: The semantic region processing module is used to determine multiple semantic regions in the file to be processed based on the file content information of the file to be processed, and to extract multiple entity identifiers and the location information corresponding to each entity identifier from the file content information. The file content information includes at least one text segment and the location information of each text segment in the file to be processed. The information entry extraction module is used to extract the entity information entries corresponding to each entity object from the semantic region; among them, the entity information entries are associated with corresponding entity identifiers, and multiple entity information entries include structural diagram information entries and text information entries; The namespace determination module is used to determine the semantic namespace to which each entity identifier belongs based on the location information corresponding to each entity identifier and multiple semantic regions. The information entry determination module is used to determine multiple target information entries with related relationships from multiple entity information entries based on the semantic namespace to which each entity identifier belongs and the association relationship between each entity identifier. The information entry association module is used to associate multiple target information entries with the same entity object.

[0007] Thirdly, embodiments of this application provide an electronic device, which includes: a processor and a memory storing computer program instructions; When the processor executes the computer program instructions, it implements any of the possible implementations of the first aspect described above.

[0008] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer program instructions, which, when executed by a processor, implement the method in any of the possible implementations of the first aspect described above.

[0009] Fifthly, embodiments of this application provide a computer program product in which instructions, when executed by a processor of an electronic device, cause the electronic device to perform a method as described in any of the possible implementations of the first aspect above.

[0010] In this embodiment, multiple semantic regions are determined based on the file content information corresponding to the file to be processed, and entity identifiers and their location information are extracted. Then, the semantic namespace to which each entity identifier belongs is determined based on the location information of the entity identifier and the semantic region. Since entity identifiers in different semantic regions often correspond to different numbering systems, combining location information with semantic regions to divide the semantic namespace can effectively distinguish the different meanings of the same number in different numbering systems. Based on this, according to the semantic namespace to which each entity identifier belongs and the association relationship between entity identifiers, target information entries with association relationships are determined from multiple entity information entries, and multiple target information entries are associated with the same entity object, reducing the risk of incorrectly associating entity information under different semantic namespaces. Therefore, this embodiment can effectively reduce information mismatch problems caused by chaotic numbering systems or spatial dispersion, achieve accurate association of multi-source information entries, and significantly improve the reliability of the association results. Attached Figure Description

[0011] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments of this application will be briefly introduced below. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a diagram illustrating the application environment of the entity information association method in one embodiment; Figure 2 A flowchart illustrating an entity information association method provided in an embodiment of this application; Figure 3 This is a schematic diagram of the overall process of an entity information association method provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an entity information association device provided in an embodiment of this application; Figure 5 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application. Detailed Implementation

[0013] The features and exemplary embodiments of various aspects of this application will be described in detail below. To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only intended to explain this application and not to limit it. For those skilled in the art, this application can be implemented without some of these specific details. The following description of the embodiments is merely to provide a better understanding of this application by illustrating examples.

[0014] It should be noted that, in this document, relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising..." does not exclude the presence of additional identical elements in the process, method, article, or apparatus that includes said element.

[0015] Furthermore, the acquisition, storage, use, and processing of data in this application's technical solution all comply with relevant national laws and regulations.

[0016] As described in the background section, in current technical documents (such as biomedical patent literature), information such as compound structure diagrams, compound numbers, experimental example numbers, and bioactivity data are usually scattered, and multiple numbering systems often coexist in the same document.

[0017] However, existing technologies have the following significant problems when processing such technical documents (especially scanned documents): OCR (Optical Character Recognition) defects: Scanned documents lack a reliable text layer, making OCR prone to issues such as missing number prefixes, truncated digits, confusion of similar characters, misrecognition of table lines, and misrecognition of headers and footers. Spatial heterogeneity and mismatch: Compound structure diagrams and activity data are often scattered across different pages and chapters, easily leading to mismatches between structure diagrams and activity indicators. Conflicts between multiple numbering systems: The same document may contain multiple numbering systems such as compound numbers, experimental example numbers, table row numbers, intermediate numbers, and method numbers, and conflicts can exist between different numbering systems. Interference in structure diagram extraction: Structure diagrams are often located in bordered tables, reaction routes, cross-column layouts, or areas with multiple structures on the same page. Directly taking full-line screenshots can easily mix in adjacent structures, table lines, name text, NMR text, or reaction conditions; excessive cropping may also truncate the chemical structure. Data Table Misinterpretation: Analytical data tables and structure charts are often adjacent on the page and both contain numbered columns, making it easy to mistakenly identify NMR, mass spectrometry, and other analytical data tables as the source of structure charts. Lack of Source Tracing: Automatically extracted results often lack source page numbers, original OCR text, contextual screenshots, structure chart quality status, and relevant evidence, leading to high manual review costs and difficulty in pinpointing the source of errors.

[0018] Because existing technologies have significant problems when processing scanned technical documents, such as OCR recognition defects, spatial heterogeneity and mismatch, conflicts in multiple numbering systems, interference in structural diagram extraction, misjudgment of data tables, and lack of traceability, existing solutions are prone to information mismatch when associating these scattered information items by relying solely on identical numbers, page distance, or simple table order to establish corresponding relationships, making it difficult to guarantee the accuracy of the association results.

[0019] Therefore, in order to solve the problems of the prior art, embodiments of this application provide an entity information association method, device, and storage medium. The entity information association method can be applied to any scenario where there is a need for entity information association.

[0020] For example, the entity information association method provided in the embodiments of this application can be applied to, for example, Figure 1 The application environment is shown. Terminal 110 communicates with server 120 via a network. A data storage system can store the data that server 120 needs to process. The data storage system can be integrated onto server 120 or placed on a cloud or other network server. Server 120 receives the file to be processed uploaded by terminal 110. Based on the file content information, server 120 determines multiple semantic regions in the file and extracts multiple entity identifiers and their corresponding location information from the file content information. The file content information includes at least one text fragment and the location information of each text fragment within the file. It extracts entity information entries corresponding to each entity object from the semantic regions. Each entity information entry is associated with a corresponding entity identifier, and multiple entity information entries include structural diagram information entries and text information entries. Based on the location information corresponding to each entity identifier and multiple semantic regions, it determines the semantic namespace to which each entity identifier belongs. Based on the semantic namespace to which each entity identifier belongs and the association relationship between each entity identifier, it determines multiple target information entries with association relationships from the multiple entity information entries. Finally, it associates the multiple target information entries with the same entity object.

[0021] Terminal 110 can be, but is not limited to, various personal computers, laptops, smartphones, tablets, and other smart terminals. Server 120 can be an independent physical server, a server cluster or distributed system composed of multiple physical servers, or a cloud server providing cloud computing services.

[0022] The entity information association method provided in the embodiments of this application will be introduced first below.

[0023] Figure 2 This illustration shows a flowchart of an entity information association method provided in an embodiment of this application. Figure 2 As shown, the method includes the following steps: S210: Based on the file content information of the file to be processed, determine multiple semantic regions in the file to be processed, and extract multiple entity identifiers and the location information corresponding to each entity identifier from the file content information.

[0024] S220: Extract the entity information entries corresponding to each entity object from the semantic region.

[0025] S230: Determine the semantic namespace to which each entity identifier belongs based on the location information corresponding to each entity identifier and multiple semantic regions.

[0026] S240: Based on the semantic namespace to which each entity identifier belongs and the association between each entity identifier, determine multiple target information entries with association relationships from multiple entity information entries.

[0027] S250: Associate multiple target information entries with the same entity object.

[0028] In this embodiment, multiple semantic regions are determined based on the file content information corresponding to the file to be processed, and entity identifiers and their location information are extracted. Then, the semantic namespace to which each entity identifier belongs is determined based on the location information of the entity identifier and the semantic region. Since entity identifiers in different semantic regions often correspond to different numbering systems, combining location information with semantic regions to divide the semantic namespace can effectively distinguish the different meanings of the same number in different numbering systems. Based on this, according to the semantic namespace to which each entity identifier belongs and the association relationship between entity identifiers, target information entries with association relationships are determined from multiple entity information entries, and multiple target information entries are associated with the same entity object, reducing the risk of incorrectly associating entity information under different semantic namespaces. Therefore, this embodiment can effectively reduce information mismatch problems caused by chaotic numbering systems or spatial dispersion, achieve accurate association of multi-source information entries, and significantly improve the reliability of the association results.

[0029] In S210 above, the "file to be processed" refers to a document carrying entity-related technical data, such as biomedical literature. The file format of the file to be processed can be a native text document (PDF, Portable Document Format), a scanned PDF image document, etc.; among which, native text documents typically contain a reliable text layer, while scanned image documents usually lack a reliable text layer. The document includes, but is not limited to, text content, image content, table content, and page layout information.

[0030] The file content information of the file to be processed is the content data obtained after parsing or recognizing the file to be processed, including but not limited to at least one text fragment and the position information of each text fragment within the file to be processed; the position information is used to characterize the page, coordinates or area range of the text fragment in the file to be processed.

[0031] Semantic regions are layout areas in a file to be processed that have the same or related semantics, such as compound structure regions, supplementary compound regions, preparation regions, analytical data regions, experimental example text regions, bioactivity data regions, or other regions.

[0032] An entity identifier is a symbol or number used to refer to or distinguish an entity object, such as the compound number "Compound 1" or the experimental example number "Experiment A". The entity object is the actual object recorded in the document to be processed, such as a specific chemical compound.

[0033] In some embodiments, the file to be processed is first parsed to obtain file content information consisting of text fragments and corresponding position information. By performing layout analysis on the file content information, the document is divided into multiple semantic regions. Simultaneously, all text fragments in the file content information are traversed. Since entity identifiers are recorded within the text fragments, regular expression matching or a named entity recognition model can be used to identify and extract all entity identifiers within the file to be processed, and the position information corresponding to each entity identifier is recorded. The named entity recognition model (NER) is a machine learning model used to identify specific categories of text from text fragments. In this embodiment, the model is used to identify entity identifiers in various numbered formats.

[0034] For example, page layout analysis can be performed on the document content information to divide the document into multiple semantic regions. Specifically, this can be done in the following ways: Based on the document content information, determine the document layout features, table border features, table titles, and other features of each page in the document to be processed. Based on the document layout features, table border features, and other features of each page, divide each page into one or more semantic regions. Alternatively, the document content information can be input into a trained machine learning model, which can output multiple semantic regions of the document to be processed based on the input document content information, thereby obtaining multiple semantic regions of the document to be processed.

[0035] In S220 above, an entity information entry is an independent information unit used to describe an entity object, and each entity information entry is associated with an entity identifier; entity information entries are divided into structural diagram information entries and text information entries. Structural diagram information entries are information entries containing images of chemical structures, such as images of compound structural formulas. Text information entries are information entries containing textual descriptive information, such as compound names, NMR mass spectrometry data, reaction conditions, and biological activity data.

[0036] In some embodiments, for each semantic region that has been divided, targeted information extraction is performed according to the information type corresponding to that semantic region. For semantic regions carrying image content, the image information contained within the region is read, and a structural diagram information entry is generated based on the image information; for semantic regions carrying text content, one or more text fragments within the semantic region are aggregated, and the contents of the multiple text fragments are combined to generate one or more text information entries.

[0037] One or more entity information entries can be extracted from the same semantic region. After extracting the entity information entries, the pre-identified entity identifiers and their corresponding location information within the current semantic region are obtained; the location range of each entity information entry in the file to be processed is also obtained. The location range of the entity identifiers is compared with the location range of the entity information entries. When the entity identifier and the entity information entry are located in the same semantic region and their page positions are adjacent or belong to the same table row, they are considered to be matched, and each extracted entity information entry is bound to the matching entity identifier.

[0038] In S230 above, the semantic namespace is a set of semantic affiliations of entity identifiers. Different semantic namespaces correspond to different numbering systems and are used to distinguish entity identifiers that have the same number but different meanings.

[0039] In some embodiments, the location information of the entity identifier is used to determine which semantic region the entity identifier falls into. Based on the semantic type represented by that semantic region, a corresponding semantic namespace is assigned to the entity identifier. Entities located in different functional semantic regions will be assigned to different semantic namespaces, thereby isolating different numbering systems.

[0040] For example, the types of semantic namespaces may include compound numbering namespaces, experimental example numbering namespaces, table row numbering namespaces, intermediate numbering namespaces, method numbering namespaces, and activity table numbering namespaces. These types of semantic namespaces are parallel and independent of each other; the same entity identifier belongs to only one type of semantic namespace. For a specific file to be processed, the semantic namespace type actually involved in the entity identifier is one or more of the above types.

[0041] Specifically, the semantic region type where the entity identifier is located can be used as the basis for judgment; or, the semantic region to which the entity identifier initially belongs can be obtained first based on the location information, and then the semantic namespace of some cross-regionally misidentified entity identifiers can be corrected by combining the spatial clustering results of entity identifiers within the same page.

[0042] In S240 above, the association relationship refers to the correspondence or mapping relationship between different entity identifiers. It can be a correspondence formed by the same identifier within the same namespace, or a mapping relationship between different namespaces. Multiple target information entries are multiple entity information entries belonging to the same entity object.

[0043] In some embodiments, the semantic namespace corresponding to each entity identifier is retrieved to define the semantic meaning of the entity identifier. Based on the semantic namespace to which each entity identifier belongs and the association between entity identifiers, multiple target information entries with association relationships are selected from multiple entity information entries. In specific implementation, whether different entity information entries are associated with the same semantic namespace or have a mapping relationship is determined to determine whether they have an association relationship.

[0044] For example, if the entity identifiers associated with multiple entity information entries belong to the same semantic namespace and are identical, then the multiple entity information entries are directly determined to be associated and are used as target information entries; if the entity identifiers associated with multiple entity information entries belong to different semantic namespaces, then a mapping relationship is determined by querying a predefined mapping table or analyzing the contextual semantics. If a mapping relationship exists, then the entity information entries are determined to be associated and are used as target information entries.

[0045] After obtaining multiple target information entries with related relationships, these multiple target information entries are grouped under the same entity object, completing the association and binding between the multiple target information entries and the entity object. This unifies the multiple entity information entries that were originally scattered in different semantic regions of the file to be processed into a single entity object.

[0046] When performing information association, the following methods can be used, including but not limited to: selecting one entity identifier as the index key of the entity object, and attaching all the matched target information entries to the entity object corresponding to the index key to complete the information association; or, creating a data container for the entity object, writing multiple target information entries into the container, and removing completely duplicate information entries during the writing process to complete the association between multiple target information entries and the entity object.

[0047] In some embodiments, the processing flow in S220 above may include the following steps: When the semantic region contains the structure diagram of entity objects, the image sub-region is determined in the semantic region based on the page image where the semantic region is located, and the structure diagram information entries corresponding to each entity object are extracted from the image sub-region. When the semantic region contains text content of entity objects, the text content is extracted from the semantic region to obtain the text information entries corresponding to each entity object.

[0048] The aforementioned structural diagram refers to an image of the chemical structure of a compound, such as a compound structural formula image, a reaction route image, or a final product image; a page image is an image obtained by rendering the file to be processed according to its page layout, used to carry the layout content of each page of the file to be processed; an image sub-region refers to a local area in the semantic region that carries the structural diagram image, such as a cell area in a structure table containing the compound structural formula; and text content refers to the textual descriptive information contained in the semantic region, such as the compound name, NMR mass spectrometry data, bioactivity data, and other textual descriptions.

[0049] In some embodiments, multiple semantic regions in the file to be processed can be divided into various types, such as compound structure region, supplementary compound region, preparation region, analytical data region, experimental example text region, bioactivity data region, or other regions, according to different semantic types.

[0050] Based on the type of semantic region, the type of information contained within that semantic region can be determined. For example: if the semantic region is a compound structure region, it is determined that the semantic region contains a structural diagram of the entity object; if the semantic region is an analytical data region or a bioactivity data region, it is determined that the semantic region contains textual content of the entity object; if the semantic region is a supplementary compound region or an experimental example text region, it is determined that the semantic region can simultaneously contain a structural diagram of the entity object and textual content; if the semantic region is a preparation region, it is determined that the semantic region simultaneously contains a structural diagram of the entity object and textual content; if the semantic region is another type of region, it can be determined whether it contains a structural diagram or textual content based on the actual content contained within the semantic region.

[0051] When a semantic region contains a structure diagram of entity objects, an image sub-region is determined within that semantic region based on the page image where the semantic region is located. Then, the structure diagram information entries corresponding to each entity object are extracted from the image sub-region. The image sub-region can be located within the corresponding local region of the page image based on features such as the table structure and regional layout of the semantic region.

[0052] When a semantic region contains text content related to entity objects, the text content is extracted from that semantic region to obtain text information entries corresponding to each entity object. In specific implementation, one or more text fragments within the semantic region can be aggregated, and the content of multiple text fragments can be combined to generate one or more text information entries; or, the text content within the semantic region can be input into a large language model, and the model can extract and generate text information entries corresponding to each entity object based on semantic understanding.

[0053] The above-mentioned structural diagram information entries and text information entries are parallel branches based on the content contained in the semantic region: the semantic region can contain only structural diagrams, only text content, or both structural diagrams and text content; when both are contained, structural diagram information entries and text information entries can be extracted simultaneously.

[0054] In this embodiment, for the structural diagram, image sub-regions are located and extracted within the semantic region using the page image, preserving the original image information of the chemical structure and avoiding potential structural distortion or information loss when converting the structural diagram into text. For the text content, text fragments are directly extracted to generate text information entries, ensuring the completeness and accuracy of the text description. Thus, the extraction method matches the information type contained in the semantic region, achieving separate extraction of structural diagram information and text information, improving the targeting, accuracy, and completeness of entity information entry extraction.

[0055] In some embodiments, determining the image sub-region within the semantic region based on the page image where the semantic region is located may specifically include the following steps: Based on the page image where the semantic region is located, perform table line detection within the semantic region; In response to the detection of at least one horizontal table line and at least one vertical table line, the column boundaries of each column within the semantic region are determined based on each horizontal table line and each vertical table line. Based on the location information corresponding to each entity identifier within the semantic region, determine the table row where each entity identifier is located; Based on the table row and column boundaries where each entity identifier is located, the target cell containing the structure graph of the entity object is determined within the semantic region, and the target cell is used as the image sub-region.

[0056] The table lines mentioned above refer to the lines used to separate cells in a table, including horizontal and vertical table lines. Horizontal table lines extend horizontally and are used to define the range of table rows. Vertical table lines extend vertically and are used to define the range of table columns.

[0057] Column boundaries refer to the boundaries of each column in a table, and can be used to define numbered columns, structure columns, name columns, analysis data columns, or condition columns, etc. A table row is a row in a table defined by horizontal table lines. A target cell is a cell containing a structure diagram of an entity object, defined by the table row containing the entity identifier and the boundaries of the structure column.

[0058] In some embodiments, firstly, table lines are detected within the semantic region based on the page image where the semantic region is located. Specifically, for framed tables, the page image can be binarized first, and horizontal and vertical table lines can be obtained through morphological long line extraction, projection peak detection, or line detection, and then the cell boundaries can be recovered from the intersection of the horizontal and vertical table lines.

[0059] For tables with broken lines, weak lines, or no borders, a method can be adopted to cluster the text baseline, column center, and row spacing of the text boxes using the semantics of the table header as column anchor points, thereby constructing a virtual row and column grid to form row and column boundaries. If a stable row and column correspondence cannot be formed, the steps of determining the image sub-region within the semantic region and the subsequent extraction process of structural graph information items are not executed, and the semantic region is labeled with relevant items from which structural graph information cannot be extracted.

[0060] In response to the detection of at least one horizontal table line and at least one vertical table line, the column boundaries of each column within the semantic region are determined based on each horizontal and vertical table line. The table row containing each entity identifier is determined based on the location information corresponding to each entity identifier within the semantic region. Since entity identifiers have corresponding location information in the page image, the table row containing each entity identifier can be determined by combining the range of the table rows. Based on the table rows containing each entity identifier and the column boundaries, the target cell containing the structure diagram of the entity object is determined within the semantic region, and the target cell is used as an image sub-region.

[0061] This implementation detects horizontal and vertical table lines within a semantic region, determines the column boundaries of each column within the semantic region based on these lines, and determines the table row containing each entity identifier by combining the position information of each entity identifier. Then, the table row containing the entity identifier and the column boundaries together define the target cell of the structure diagram containing the entity object. This target cell is used as an image sub-region, thereby accurately locating the image sub-region to the target cell. This ensures that the extraction range of subsequent structure diagram information entries is precisely limited to the local area corresponding to the target cell, avoiding the inclusion of content from adjacent cells or irrelevant areas. Furthermore, establishing a row-level correspondence between entity identifiers and the structure diagram based on the table row containing the entity identifier improves the accuracy and specificity of structure diagram location.

[0062] In some embodiments, the extraction of structural diagram information entries corresponding to each entity object from the image sub-region can be performed in the following manner: Extract the initial structure diagrams corresponding to each entity object from the image sub-regions; Remove the long horizontal and vertical lines from each initial structure diagram to obtain the first structure diagram corresponding to each entity object; Calculate the bounding box of the dark pixels corresponding to each first structure diagram based on the remaining dark pixels in each first structure diagram; Based on the dark pixel bounding box and preset border ratio of each first structure diagram, the target boundary of each first structure diagram is determined. Based on the target boundary corresponding to each first structure diagram, the blank area corresponding to each first structure diagram is trimmed to obtain the second structure diagram corresponding to each entity object; If each second structural diagram meets the preset quality conditions, each second structural diagram will be used as a structural diagram information entry corresponding to each entity object.

[0063] In some embodiments, initial structure diagrams corresponding to each entity object are cropped from the image sub-region. Specifically, a coarse cropping region can be formed based on the table boundaries, the column boundaries of the column containing the structure diagram, and the row boundaries of the row containing the structure diagram, and the initial structure diagram is obtained by cropping from the image sub-region.

[0064] Then, table line noise such as long horizontal and vertical lines in each initial structure diagram is removed to obtain the first structure diagram corresponding to each entity object. Specifically, Hough transform or morphological line detection can be used to detect line segments in the initial structure diagram that are approximately horizontal or vertical in direction and exceed a preset length threshold. These segments are removed as table line noise such as long horizontal or vertical lines, thereby preventing table lines from being mixed into the structure diagram.

[0065] Long horizontal lines refer to continuous dark pixel segments extending horizontally and exceeding a preset length threshold. They typically originate from remnants of horizontal table lines in the cropping result. The specific criteria for determining a long horizontal line are: the angle between the line segment's direction and the horizontal direction is less than a preset angle threshold (e.g., less than 10 degrees), and the line segment length is greater than the preset length threshold (e.g., greater than one-third of the width of the structure image). Horizontal line segments whose length does not reach the above thresholds are not considered long horizontal lines; they may belong to structural elements such as chemical bonds.

[0066] Long vertical lines refer to continuous dark pixel segments that extend vertically and exceed a preset length threshold. They are usually remnants of vertical table lines in the cropping result. The specific criteria for judging them can be: the angle between the line segment direction and the vertical direction is less than a preset angle threshold (e.g., less than 10 degrees), and the line segment length is greater than a preset length threshold (e.g., greater than one-third of the height of the structure image).

[0067] The aforementioned dark pixels are pixels whose grayscale values ​​are lower than a preset grayscale threshold (e.g., less than 128) after the image has been grayscaled, or pixels that belong to the foreground region after the image has been binarized. Dark pixels are used to represent ink marks (e.g., drawn atoms, bonds, rings, side chains, etc.) and noise such as table lines in the structural diagram.

[0068] For each first structure graph, traverse all dark pixels in the first structure graph, and count the range of the horizontal and vertical coordinates of all dark pixels. The smallest rectangle enclosed by the minimum horizontal coordinate, the maximum horizontal coordinate, the minimum vertical coordinate, and the maximum vertical coordinate is the bounding box of the dark pixels. This bounding box is used to characterize the distribution range of the effective dark pixels in the structure graph.

[0069] Next, using the dark pixel bounding box as a reference, the target boundary is expanded outward by a preset distance in each of the top, bottom, left, and right directions according to the preset border ratio to obtain the target boundary, thereby preserving an appropriate white border around the structure image; the first structure image is cropped along the target boundary, and the blank area outside the target boundary (i.e., the white border area that does not contain the effective dark pixels of the structure image) is removed, while the structure image content within the target boundary is preserved, thereby automatically shrinking the blank area and preserving an appropriate white border according to the ink distribution of the image to obtain the second structure image after cleaning.

[0070] The preset border ratio can be set according to the structural size ratio, that is, the width and height of the dark pixel outer frame are extended outward by the preset ratio (for example, the width or height of the dark pixel outer frame is extended by 5% to 20% in each direction) to retain appropriate white borders around the structural diagram and avoid cutting off atoms, bonds, rings, side chains or stereochemical markers.

[0071] After obtaining the second structural diagram, it is determined whether the second structural diagram meets the preset quality conditions. If each second structural diagram meets the preset quality conditions, it is used as a structural diagram information entry corresponding to each entity object. If a second structural diagram does not meet the preset quality conditions, it is not used as a structural diagram information entry, but is marked as candidate evidence or pending review, and is not used for subsequent information association, although its source page number and contextual evidence are still retained. Thus, only structural diagrams with high confidence proceed to downstream processing, reducing the risk of erroneous association.

[0072] The aforementioned preset quality conditions are preset conditions used to determine whether the second structural graph can be used as a reliable structural graph. For example, the preset quality conditions may include the structural graph quality score of the second structural graph being greater than a preset quality score threshold.

[0073] In this embodiment, by removing table line noise such as long horizontal and vertical lines from the initial structural diagram, the inclusion of table lines in the chemical structural diagram is effectively avoided, thus improving the purity of the structural diagram. By calculating the bounding box of dark pixels and expanding the target boundary according to a preset border ratio, the blank area can be automatically shrunk based on the ink distribution in the image, while retaining appropriate white borders around the structural diagram to prevent the cutting of atoms, chemical bonds, ring structures, side chains, or stereochemical markers, ensuring the integrity of the structural information. By judging the second structural diagram after cleaning according to preset quality conditions, only structural diagrams with high confidence are used as structural diagram information entries, thereby ensuring the cleanliness, integrity, and reliability of structural diagram information entries and reducing the risk of erroneous associations.

[0074] In some embodiments, the preset quality condition includes the second structural diagram's structural diagram quality score being greater than a preset quality score threshold. The calculation process for the quality score of the structure diagram includes the following steps: Calculate multiple quality indicators for the second structure diagram; The quality score of the second structure diagram is obtained by weighted summation of multiple quality indicators.

[0075] The aforementioned preset quality score threshold is a pre-set critical value used to determine whether the quality score of the structure diagram is qualified; when the quality score of the structure diagram is greater than the preset quality score threshold, the second structure diagram is determined to meet the preset quality conditions.

[0076] The structure graph quality score is a comprehensive score used to quantify the quality of a structure graph. A higher structure graph quality score indicates higher quality. Quality metrics are indicators that quantify the quality of a structure graph from different dimensions, and each quality metric is represented as a sub-score. Multiple quality metrics include at least one of the following: sub-score corresponding to image size, sub-score corresponding to the aspect ratio of effective dark pixels, sub-score corresponding to the proportion of dark pixels, sub-score corresponding to the number of connected regions, and sub-score corresponding to the degree of edge touch.

[0077] The sub-score corresponding to image size is a quality metric calculated based on the width and height dimensions of the structure image, used to measure whether the structure image has reached a recognizable minimum size. For example, when both the width and height of the second structure image are within a preset reasonable size range, the sub-score takes a higher value; otherwise, it decreases according to the degree to which the width or height deviates from the allowable range. The sub-score corresponding to the effective dark pixel aspect ratio is a quality metric calculated based on the aspect ratio of the effective dark pixel distribution in the structure image, used to characterize whether the range of dark pixels reaches the minimum effective size of the structure image. The sub-score corresponding to the dark pixel ratio is a quality metric calculated based on the proportion of dark pixels in the structure image to the total number of pixels in the image (i.e., ink density). A dark pixel ratio that is too low indicates that the structure image may be blank or have too little structure; a ratio that is too high indicates that the structure image may contain text or large areas of noise. This sub-score decreases according to the degree to which it deviates from the reasonable range.

[0078] The sub-score corresponding to the number of connected regions is a quality indicator calculated based on the number of connected regions in the structure graph. Connected regions include effective connected regions and noisy connected regions. When the number and distribution of connected regions conform to the characteristics of a chemical structure graph, the sub-score takes a higher value; when there is only a single long-line region, too few effective connected regions, or too many small noisy regions, the sub-score decreases.

[0079] The sub-score corresponding to the edge touch degree is a quality index calculated based on the degree to which dark pixels in the structure image touch the edge of the structure image. It is used to characterize whether the structure may be truncated. Since the second structure image is obtained by cropping the first structure image according to the target boundary, if the ink marks such as atoms, bonds, side chains, or characters of the chemical structure are located near the target boundary, they may be truncated by the target boundary after cropping and fall on the edge of the second structure image. Therefore, by statistically analyzing the second structure... Figure 4 The proportion of dark pixels in the peripheral edge band can directly characterize the possibility that the structure has been truncated. The higher the proportion of dark pixels at the edge, the greater the possibility that chemical bonds, atomic characters or side chains have been truncated.

[0080] In some embodiments, each quality index can be normalized first to obtain the sub-scores corresponding to each quality index. The sub-scores corresponding to image size, effective dark pixel aspect ratio, dark pixel ratio, and number of connected regions can be weighted according to preset weights (e.g., 0.10, 0.14, 0.12, 0.14, etc.). For quality indicators that represent risk, such as edge touch degree, they can be deducted from the weighted result as deduction items. The final result is limited to a range not lower than a preset lower limit (e.g., a minimum of 0).

[0081] In addition to the quality metrics mentioned above, for example, quality metrics may also include sub-scores corresponding to the proportion of dark pixel bounding boxes, sub-scores corresponding to the concentration of dark pixels in the row and column directions, and sub-scores corresponding to long line segments. The sub-score corresponding to the proportion of dark pixel bounding boxes is calculated based on the ratio of the area of ​​the dark pixel bounding boxes to the image area, and is used to characterize situations where there is too much blank space in the cropped area, the structure is close to the edge, or adjacent content is mixed in; the sub-score corresponding to the concentration of dark pixels in the row and column directions is calculated based on the distribution of dark pixels in each row and column, and is used to characterize situations where dark pixels are concentrated in a single horizontal row or a single vertical column (which may only be table lines); the sub-score corresponding to long line segments is calculated based on the length of continuous horizontal or vertical dark line segments in the second structural image, and is used to characterize whether residual table lines exist.

[0082] Furthermore, quality evaluation can also employ a framework combining normalized sub-score weighted synthesis with hard rejection: positive sub-scores are formed by factors such as image size, effective dark pixel aspect ratio, dark pixel ratio, dark pixel bounding box ratio, connectivity, and chemical structure morphology; risk deductions are formed by factors such as long line residue, text block features, and edge truncation; for blank structural images, severely truncated images, or images where the main ink marks are only table lines, the second structural image can be directly determined not to meet the preset quality conditions without further weighted summation. For example, when the second structural image is blank, or the effective dark pixel width is less than a preset width threshold (e.g., 200 pixels), the effective dark pixel height is less than a preset height threshold (e.g., 100 pixels), or the number of dark pixels is less than a preset number threshold (e.g., 180), the second structural image can be directly determined not to meet the preset quality conditions.

[0083] In addition, quality reasons can be generated based on the abnormality of each quality indicator, such as "edge touch, suspected structure is cut off", "long line segment exists, suspected residual table line" or "ink size and distribution are normal, pass quality access", and the quality reasons and structural diagram quality scores are saved together for subsequent correlation judgment and manual review.

[0084] To facilitate understanding, the calculation process of the structure diagram quality score is illustrated with an example: Let the width and height of the second structure diagram be W and H respectively, the number of dark pixels be N, and the width and height of the outer bounding box of the dark pixels be w and h respectively. The structure diagram quality score Q is calculated according to the following formula: Q = max{0, 100 × (0.1S1 + 0.14S2 + 0.12S3 + 0.12S4 + 0.14S5 + 0.18S6 + 0.2S7) - 20P e}

[0085] Where: S1 is the sub-score corresponding to the image size. When the image width is between 200 and 420 pixels and the height is between 100 and 360 pixels, S1 is 1. When it exceeds this range, it decreases linearly according to the degree to which the width or height deviates from the allowable range, and the minimum value is 0.

[0086] S2 is the sub-score corresponding to the dark pixel ratio. The dark pixel ratio is calculated as dark pixel ratio = N / (W × H). When the dark pixel ratio is within a preset reasonable range, S2 is 1; if the dark pixel ratio is lower than this range, it indicates that the second structure image may be blank or have too little structure, and if the dark pixel ratio is higher than this range, it indicates that the second structure image may contain text or large areas of noise, and S2 is reduced according to the degree of deviation.

[0087] S3 is the sub-score corresponding to the effective aspect ratio of the dark pixels. It is calculated based on the width w and height h of the bounding box of the dark pixels. When both w and h reach the minimum effective size of the structure diagram, S3 is 1; if either size is insufficient, it is reduced according to the ratio of the actual size to the minimum effective size.

[0088] S4 is the sub-score corresponding to the proportion of the dark pixel bounding box. The proportion of the dark pixel bounding box is calculated as dark pixel bounding box proportion = (w×h) / (W×H). When the proportion of the dark pixel bounding box is within a preset reasonable range, S4 is 1; if the proportion is too small, it means there is too much blank space in the cropped area, and if the proportion is too large, it means that the structure may be close to the edge or mixed with adjacent content. S4 is reduced according to the degree of deviation.

[0089] S5 is the sub-score corresponding to the concentration of dark pixels in the row and column directions. The proportion of dark pixels in each row and column of the second structural image is calculated separately. When dark pixels are distributed in multiple row and column regions, S5 takes a higher value; when dark pixels are mainly concentrated in a single horizontal row or a single vertical column, it indicates that the image may only contain table lines, and S5 is reduced accordingly.

[0090] S6 represents the score for long line segments. It detects continuous horizontal or vertical dark-colored line segments in the second structure diagram and calculates the ratio of the longest line segment length to the width or height of the second structure diagram. When the longest line segment's ratio does not exceed 50%, S6 is set to 1; as the ratio exceeds 50%, it gradually decreases; when it reaches or exceeds 72%, S6 is set to 0.

[0091] S7 is the sub-score corresponding to the number of connected regions. Connectivity analysis is performed on dark pixels, and S7 is determined based on the number of effective connected regions, the proportion of the largest connected region to all dark pixels, and the proportion of small noise regions. When the number and distribution of connected regions conform to the characteristics of a chemical structure diagram, S7 takes a higher value; when there is only a single long line region, too few effective connected regions, or too many small noise regions, S7 decreases.

[0092] P e Sub-scores corresponding to the degree of edge contact. Statistical analysis of the second structure. Figure 4 The proportion of dark pixels in the peripheral edge band. When the proportion of dark pixels at the edge does not exceed 2.5%, P eSet to 0; if it exceeds 2.5%, increase according to the degree of excess, with a maximum of 1. The more dark pixels at the edge, the higher the probability that chemical bonds, atomic characters, or side chains have been truncated.

[0093] In a preferred embodiment, if the second structure graph is blank, or if the effective dark pixel width is less than 200 pixels, the effective dark pixel height is less than 100 pixels, or the number of dark pixels is less than 180, it is directly determined that the structure graph admission conditions are not met. For the second structure graph that passes the basic admission judgment, if the structure graph quality score is not lower than 70 points, it is used as a structure graph information entry; if the structure graph quality score is higher than 50 points but lower than 70 points, it is used as a candidate structure graph and marked as pending review; if the structure graph quality score is lower than 50 points, it is not used for subsequent information association. The above weights and thresholds can be adjusted according to the page rendering resolution and actual needs.

[0094] In this embodiment, by calculating corresponding sub-scores from multiple dimensions such as image size, effective dark pixel aspect ratio, dark pixel ratio, number of connected regions, and edge touch degree, and then using these sub-scores as quality indicators for weighted summation, the cleaning quality of the structure graph can be comprehensively and objectively quantified and evaluated. This avoids the one-sidedness of judging by a single threshold, and enables the structure graph quality score to accurately reflect the true usability of the structure graph. Only the second structure graph with a score greater than the preset threshold is considered to meet the preset quality conditions, thereby further ensuring that the structure graph entering downstream processing has a high degree of confidence.

[0095] In some embodiments, the multiple text information entries include bioactivity information entries; The above method of extracting text content from semantic regions to obtain text information entries corresponding to each entity object can be used as follows: Extract the active text content from the semantic region to obtain the activity index data corresponding to each entity object; When the activity index data corresponding to each entity object belongs to the preset classification type, determine the classification symbol of each activity index data. Based on the preset symbol mapping relationship, each grade symbol is mapped to the corresponding activity value range or activity semantic range, thereby obtaining the bioactivity information entries corresponding to each entity object.

[0096] The aforementioned bioactivity information entries are textual information entries containing bioactivity data, such as entries containing activity index data such as median effective concentration (EC50) and maximum effect (Emax). Bioactivity text content refers to text content within the semantic region that contains bioactivity data, such as the cell text recording activity data in a bioactivity table. Bioactivity index data refers to the activity index data extracted from the bioactivity text content that corresponds to the entity object, such as the EC50 value or Emax value corresponding to a specific compound, or data where the above values ​​are presented in a gradation symbol format.

[0097] Preset grading type refers to a data type in which activity index data is presented in the form of grading symbols. That is, the activity data is not a specific numerical value, but a grading symbol. Grading symbols are symbols used to characterize the activity index within a certain activity range, such as letters, plus signs, etc.

[0098] The preset symbol mapping relationship is used to characterize the range of activity values ​​or the range of activity semantics corresponding to different grading symbols. The range of activity values ​​refers to the range of activity values ​​corresponding to the grading symbol. The range of activity semantics is the activity semantic description corresponding to the grading symbol.

[0099] In some embodiments, for the bioactivity table contained in the semantic region, table boundary positioning and cell recognition can be performed to extract the activity number, EC50, Emax, or other activity indicators to obtain the activity indicator data corresponding to each entity object. When the activity indicator data corresponding to each entity object belongs to a preset classification type, the classification symbol can be identified by combining multi-recognition mode optical character recognition with ink morphology features.

[0100] Based on a preset symbol mapping relationship, each tier symbol is mapped to a corresponding activity numerical range or activity semantic range, resulting in bioactivity information entries for each entity object. For example, the preset symbol mapping relationship could be: EC50 tier A represents a value not exceeding 0.05. B indicates a value greater than 0.05. And not higher than 0.5 In the grading system, C indicates a value greater than 0.5. And not higher than 3.0 The grading D indicates a value greater than 3.0. and not higher than 100 Emax can be mapped to different response ranges according to the plus sign level.

[0101] There are multiple ways to determine the above-mentioned preset symbol mapping relationship. The preset symbol mapping relationship is automatically extracted from the table title, table note, footnote and experimental description in the current semantic region. In-text rules take precedence over preset rules. Different tables in the same file to be processed maintain mapping scopes separately. When in-text rules conflict with preset rules, the conversion stops. When there is no basis for either, only the original symbols are retained.

[0102] In some embodiments, determining the classification symbol for each activity index data specifically includes the following steps: For each activity index data, multi-dimensional information recognition is performed on the activity index data to obtain symbol recognition results for each dimension; From multiple symbol recognition results, select the target recognition results that belong to the preset graded symbol set; In the presence of multiple target recognition results, the classification symbols for the activity index data are determined from the multiple target recognition results based on the consistency between the target recognition results, the character confidence of each target recognition result, and the morphological similarity between each target recognition result and other classification symbols in the semantic region.

[0103] The aforementioned recognition methods refer to the specific recognition techniques used when identifying activity index data, with one recognition method corresponding to each dimension. These methods include, but are not limited to, character-constrained OCR, numerical and comparison symbol-constrained OCR, whole-cell OCR with different page segmentation methods, and ink pattern clustering. The symbol recognition result refers to the graded symbol recognition results obtained after recognizing the activity index data across each recognition dimension.

[0104] The preset grading symbol set refers to a pre-defined set of grading symbols. Consistency refers to the degree of consistency between the recognition results of different targets. Character confidence refers to the confidence level of the character corresponding to the target recognition result, i.e., the degree of certainty regarding the recognized character, automatically output by the corresponding recognition method. Morphological similarity refers to the degree of similarity in ink morphology between the recognition result and other grading symbols in the semantic region. Other grading symbols refer to the grading symbols corresponding to other active indicator data in the semantic region.

[0105] In some embodiments, for each activity index data, the above-mentioned multiple identification methods can be used to perform multi-dimensional information identification to obtain symbol identification results for each dimension. From the multiple symbol identification results, the identification results belonging to the preset graded symbol set are selected as the target identification results; that is, the symbol identification results that do not belong to the preset graded symbol set are excluded, and the target identification results that belong to the preset graded symbol set are retained.

[0106] If only one target recognition result exists, that target recognition result is used as the classification symbol for the activity index data. In the case of multiple target recognition results (i.e., different classification symbols are identified in different dimensions), a comprehensive judgment is made: the consistency between the target recognition results, the character confidence of each target recognition result, and the morphological similarity between each target recognition result and other classification symbols in the semantic region are evaluated.

[0107] For example, for each target recognition result, the consistency between the target recognition result and other target recognition results, the character confidence of the target recognition result, and the morphological similarity between the target recognition result and other classification symbols in the semantic region can be weighted according to a preset weight to obtain a comprehensive evidence score; the target recognition result with the highest comprehensive evidence score can be selected from multiple target recognition results as the classification symbol of the activity index data.

[0108] If, among multiple target recognition results, the recognition result corresponding to any of the text recognition methods such as character constraint OCR, numerical and comparison symbol constraint OCR, and whole-cell OCR of different page segmentation methods conflicts with the recognition result corresponding to the morphological recognition method such as ink pattern clustering and cannot form a significant advantage, then the original cell image where the activity index data is located is retained and transferred to the review.

[0109] For example, if character constraint OCR identifies the classification symbol B, but ink pattern clustering shows that the shape of the symbol is highly similar to other classification symbols identified as C in the same semantic region, and the character confidence is low, then the classification symbol of the activity index data can be corrected to C.

[0110] In this embodiment, symbol recognition results for each dimension are obtained through multi-dimensional information recognition, and target recognition results belonging to a preset set of graded symbols are filtered out, thus eliminating illegal symbols from multiple candidates. When multiple target recognition results exist, the consistency, character confidence, and morphological similarity between the results and other graded symbols are comprehensively evaluated, achieving effective fusion of multi-mode recognition results. This fully considers the problems of OCR confusion and difficulty in accurately recognizing graded symbols in scanned files, significantly improving the accuracy and robustness of graded symbol recognition, and providing a reliable foundation for the normalization and association of subsequent activity data.

[0111] In some embodiments, the file to be processed includes multiple file pages, and the file content information includes the page content information of each file page, which includes at least one text fragment and the position information of each text fragment within the file page; based on this, in S210 above, based on the file content information of the file to be processed, multiple semantic regions in the file to be processed are determined, including the following processing steps: For each file page, the file page is divided into multiple candidate regions based on the page content information; For each candidate region, the semantic type of the candidate region is determined based on multiple text fragments, text layout features, table line distribution, and page crossing status of the candidate region. Based on the semantic type of each candidate region and each candidate region, multiple semantic regions in the file to be processed are determined.

[0112] The aforementioned file pages refer to the individual pages contained in the file to be processed, which may include multiple file pages. File content information includes the page content information of each file page. Page content information refers to the content data corresponding to a single file page within the file content information, including at least one text fragment and the position information of each text fragment within the file page. Text layout features are used to characterize the spatial arrangement of text within the candidate region, such as the row distribution, column distribution, and alignment of the text. Table line distribution refers to the distribution of table lines within the candidate region, such as whether there are horizontal table lines, vertical table lines, and their coverage area. Semantic type refers to the semantic category carried by the candidate region or semantic region, such as compound structure region, supplementary compound region, preparation region, analytical data region, experimental example text region, bioactivity data region, or other regions. Page spanning status refers to whether the candidate region spans adjacent file pages and the status of page spanning continuation relationships.

[0113] In some embodiments, for each file page, the file page is divided into one or more candidate regions based on layout features such as the position information of each text fragment contained in the file page, the distribution of table lines, and paragraph boundaries; or the page content information of the file page is input into a trained machine learning model, which can automatically output one or more candidate regions contained in the file page based on the page content information. Each candidate region may correspond to a table, a piece of text, etc.

[0114] For each candidate region, the semantic type of the candidate region is determined based on multiple text fragments within the region, text layout features, table line distribution, and page spanning status. The text layout features are determined based on the positional information of each text fragment within the candidate region on the document page. For example, if a candidate region contains dense table lines, exhibits multi-column alignment in its text layout, and contains the keyword "compound," it may be classified as a compound structure region.

[0115] In one feasible implementation, the cross-page status can be determined based on the following information: whether there is an unclosed area at the end of the previous page, whether there is a continuing area at the beginning of the subsequent page, whether the table name or chapter has changed, whether the column boundaries are consistent, and whether the numbering or data is continuous. When a termination signal such as a new title, a change in column semantics, or a clear restart of numbering occurs, the cross-page status can be determined to have terminated, meaning that the candidate area and the adjacent page area no longer belong to the same continuing content. In addition, for tables in the cross-page status: when the table header is repeated on the continuation page, the table header is re-identified and aligned with the fields of the previous page; when there is no repeated table header, the field definition of the previous page is inherited, and if the evidence of inheritance is insufficient, it needs to be reviewed; the cross-page numbering continues within the same table and the same semantic namespace.

[0116] Finally, each candidate region is used as a semantic region, and its semantic type is used as the semantic type to which the semantic region belongs, thereby determining multiple semantic regions in the file to be processed.

[0117] In this embodiment, for each file page, the file page is divided into multiple candidate regions based on the page content information. The semantic type of each candidate region is determined by combining the text fragments, text layout features, table line distribution, and page crossing status in the candidate regions. This process determines multiple semantic regions in the file to be processed, which can accurately divide different semantic regions in the file page. This provides a reliable layout understanding basis for subsequent extraction of entity information based on semantic regions and effectively avoids extraction errors caused by complex layout or confusion of region types.

[0118] In some embodiments, the determination of the semantic type of a candidate region based on multiple text fragments, text layout features, table line distribution, and page-crossing status of the candidate region can be implemented using the following process: Based on multiple text fragments, text layout features, table line distribution, and page-crossing status of the candidate region, calculate the comprehensive score of the candidate region relative to at least one candidate semantic type. The candidate regions are ranked according to their comprehensive scores relative to each candidate semantic type, and the target score with the highest score is selected from the multiple comprehensive scores. If the target score is greater than the preset comprehensive score threshold, the candidate semantic type corresponding to the target score will be used as the semantic type to which the candidate region belongs.

[0119] The aforementioned candidate semantic types refer to the semantic types to which the candidate region may belong, such as compound structure tables, supplementary compound tables, preparation tables, analytical data tables, experimental example text regions, bioactivity tables, or other regions. The comprehensive score is a combined score of the candidate region relative to a specific candidate semantic type, used to quantify the degree of matching between the candidate region and that candidate semantic type. The comprehensive score can be determined based on at least one of the following: region name evidence score, column structure score, table line density score, keyword score, cross-page continuity score, and exclusion deduction score, with the result limited to a preset range (e.g., 0–1).

[0120] The region name evidence score is a score calculated based on the name-type evidence of a candidate region, used to measure the degree of matching between the name information of the candidate region and the candidate semantic type. For example, region name evidence may include title evidence and header evidence, wherein the title evidence is used to identify the title of a structure table, activity table, or analysis data table, and the header evidence is used to identify number columns, structure columns, activity index columns, or analysis data columns.

[0121] Column structure score is a score calculated based on the column structure features of the candidate region, used to determine whether the candidate region has the column structure corresponding to the candidate semantic type (e.g., structure graph columns or activity index columns). Table line density score is a score calculated based on the distribution density of table lines in the candidate region, used to characterize the degree of tabulation of the candidate region. Keyword score is a score calculated based on the keyword hit rate in the candidate region, used to measure the degree of matching between the domain keywords in the candidate region and the candidate semantic type.

[0122] Cross-page continuity score refers to the score calculated based on the cross-page status of the candidate area, used to measure the continuity between the candidate area and adjacent page areas. Exclusion deduction score refers to the deduction based on the degree of mutual exclusion between the text in the candidate area and the candidate semantic type, used to reduce the impact of noise from headers and footers, ordinary text, supplier information, and analytical data on semantic type determination. Mutual exclusion degree refers to the degree of conflict or repulsion between the text content in the candidate area and the candidate semantic type; the higher the degree of mutual exclusion, the greater the corresponding exclusion deduction score.

[0123] The target score refers to the highest overall score among the comprehensive scores of all candidate semantic types for a candidate region. The preset comprehensive score threshold is a pre-set score threshold used to determine whether a candidate region can adopt the candidate semantic type corresponding to the target score.

[0124] In some embodiments, for each candidate region, based on multiple text fragments within the candidate region, text layout features, table line density, page spanning status, page image of the candidate region, and page order of the file pages containing the candidate region, the region name evidence score, column structure score, table line density score, keyword score, and page spanning continuity score relative to each candidate semantic type are calculated. Exclusion criteria are deducted based on the mutual exclusion between the text and the candidate semantic type within the candidate region. These criteria are then weighted and synthesized according to preset weights, and the result is limited to a preset range to obtain a comprehensive score. Page order is used to determine the consistency of adjacent page states and the continuity of the numbering sequence, while page image is used to assist in the calculation of table line density and OCR clustering.

[0125] The region name evidence score can be composed of a title evidence score and a table header evidence score. For example, the title evidence score is determined based on the degree to which the title matches the candidate semantic type's specialized terms: a higher value (e.g., 1) is taken when the title precisely matches the specialized term corresponding to the semantic type; a middle value (e.g., 0.5) is taken when it matches a synonym or fuzzy match; and a lower value (e.g., 0) is taken when it does not match. The table header evidence score is calculated based on the proportion of the number of identified required target fields to the total number of required fields specified for that semantic type, and the result is limited to 0-1. For example, a structure table containing a structure diagram requires at least an ID column and a structure column, and an activity table containing biological activity data requires at least an ID column and an activity index column.

[0126] The column structure score can be calculated by averaging the proportion of effective target columns, column boundary alignment stability, and cell data type consistency. The table line density score can be calculated by averaging the effective horizontal and vertical line coverage and regular grid integrity; for frameless tables, the row and column clustering stability of OCR text boxes can be used as a substitute. The keyword score can be calculated as the ratio of the actual number of duplicated keywords to the preset number of keywords for that semantic type, with the result limited to 0-1. The cross-page continuity score can be calculated by assigning values ​​(e.g., 0 or 1) to items such as consistency of adjacent page status, consistency of table names, compatibility of normalized column boundaries, and continuity of numbering sequence, and then taking the average.

[0127] The deduction for exclusion items can be determined based on the degree of mutual exclusion between the text in the candidate area and the candidate semantic type. For example, when headers and footers, ordinary text, supplier information, NMR or mass spectrometry analysis data are mutually exclusive with the candidate semantic type, a preset score is deducted for each item according to the degree of conflict (e.g., 0.1 to 0.35 points per item), and the total deduction does not exceed a preset upper limit (e.g., not exceeding 0.6 points).

[0128] The candidate regions are ranked according to their comprehensive scores relative to each candidate semantic type, and the target score with the highest score is selected from multiple comprehensive scores. If the target score is greater than a preset comprehensive score threshold, the candidate semantic type corresponding to the target score is taken as the semantic type to which the candidate region belongs. For example, when the highest score is not lower than a preset threshold (e.g., 0.65) and is at least at a preset interval (e.g., 0.1) higher than the second highest score, the candidate semantic type corresponding to the highest score is determined as the semantic type to which the candidate region belongs; when the highest score is in the middle range (e.g., 0.50 to 0.65) or the difference between the highest score and the second highest score is less than a preset interval, the candidate region is marked as a candidate and enters the review process; when the highest score is lower than a preset lower limit (e.g., 0.5), the candidate region is classified into other regions.

[0129] To facilitate understanding, the calculation process of the comprehensive score is illustrated with an example. The comprehensive score can be calculated according to the following formula: Comprehensive score = clamp [0, 1] (0.22T_k + 0.18H_k + 0.16C_k + 0.12G_k + 0.14K_k + 0.18X_k - P_k).

[0130] Here, `clamp [0,1]` indicates that the final result is limited to 0-1; the sum of the six positive weights is 1, and `P_k` is the deduction for exclusion items. `T_k` is the title evidence score, which is 1 when the title accurately matches the specific term for that state, 0.5 when it matches a synonym or fuzzy match, and 0 when it does not match. `H_k` is the table header evidence score, calculated by dividing the number of identified required target fields by the number of required fields specified for that state, and the result is limited to 0-1; for example, a structure table requires at least an ID column and a structure column, and an active table requires at least an ID column and an active index column. `C_k` is the column structure score, calculated by the average of the effective target column ratio, column boundary alignment stability, and cell data type consistency. `G_k` is the table line density score, calculated by the average of the effective horizontal and vertical line coverage and the regular grid integrity; for frameless tables, the row and column clustering stability of the OCR text boxes is used instead. K_k is the keyword score, calculated as the ratio of the actual number of keywords hitting the desired state after deduplication to the preset number of keywords for that state, with the result limited to 0-1. X_k is the cross-page consecutive score, assigned 0 or 1 to four items: consistent state on adjacent pages, consistent table names, compatibility of normalized column boundaries, and consecutive numbering sequences, and then averaged. P_k is the exclusion deduction, where 0.10-0.35 is deducted for each exclusion when header / footer, regular text, supplier information, NMR or mass spectrometry analysis data are mutually exclusive with the candidate state, with a total deduction not exceeding 0.60, depending on the degree of conflict. The above weights and thresholds are preferred implementation parameters and can be recalibrated according to page resolution, OCR engine, and actual needs.

[0131] This implementation calculates a comprehensive score for each candidate semantic type based on text fragments, text layout features, table line distribution, and page spanning status within the candidate region. It integrates multi-dimensional evidence such as region name, column structure, table line density, keywords, and page spanning continuity, deducting points for exclusion items mutually exclusive with the candidate semantic type. The highest score, exceeding a preset comprehensive score threshold, is then selected as the target score, and the candidate semantic type corresponding to this target score is determined as the semantic type of the candidate region. This achieves quantifiable multi-evidence comprehensive judgment among multiple candidate semantic types, avoiding misjudgment of semantic type based on a single signal, reducing the risk of misclassifying analytical data tables and ordinary text as structural diagram sources, and improving the accuracy and robustness of semantic region classification.

[0132] In some embodiments, the entity information association method further includes the following process: The file to be processed is processed into page images to obtain page images of each page of the file; For each file page, optical character recognition is performed on the page image to obtain at least one text fragment, the text confidence of each text fragment, the position information of each text fragment within the file page, and the word-level bounding box of each text fragment. The text segments are merged into rows based on their word-level bounding boxes to obtain row-level bounding boxes; Based on the multiple line-level bounding boxes, multiple text fragments, the text confidence of each text fragment, the position information of each text fragment within the file page, and the word-level bounding boxes of each text fragment, construct the page content information of the file page.

[0133] For example, the file to be processed is rendered into page images page by page. For example, for a PDF file, a rendering engine is used to convert each page into a high-resolution image, and the correspondence between the file identifier, page number and page image of the file to be processed is recorded.

[0134] For each file page, optical character recognition is performed on each page image. An OCR engine (such as a deep learning-based recognition model) recognizes the image and outputs at least one word-level text fragment, the text confidence score of each text fragment, the page coordinates of each text fragment within the file page, and the word-level bounding box of each text fragment.

[0135] Among them, word-level bounding boxes are the coordinate boxes of text fragments in the page image, usually defined by the coordinates of the top-left and bottom-right corners of the text fragment, representing the precise location of the text fragment; based on the word-level bounding boxes and page coordinates of each text fragment, the positional information of each text fragment can be determined. Text confidence refers to the recognition confidence corresponding to the text fragment, that is, the degree of certainty about the recognition result of the text fragment, which is automatically output by the OCR engine, usually represented by a probability value between 0 and 1, with a higher value indicating more reliable recognition.

[0136] Next, the text fragments are merged into lines based on their word-level bounding boxes. The merging rules can be based on the vertical overlap and horizontal spacing of the word bounding boxes: if the bounding boxes of two words overlap vertically and the horizontal spacing is less than a threshold, they are considered to be on the same line, and their bounding boxes are merged into a single line-level bounding box. This process can also filter out non-body text content such as headers and footers. A line-level bounding box refers to the bounding box of the line-level text within the document page after merging word-level text fragments into lines; it represents the position and extent of the line within the document page.

[0137] Based on multiple line-level bounding boxes, multiple text fragments, text confidence scores of each text fragment, position information of each text fragment within the file page, and word-level bounding boxes of each text fragment, the page content information of the file page is constructed to form a verifiable layer of page evidence.

[0138] This implementation method obtains page images of each file page by performing page image processing on the file to be processed, and then performs optical character recognition on the page images to obtain text fragments, text confidence, location information, and word-level bounding boxes. Then, based on the word-level bounding boxes, the text fragments are merged into lines to obtain line-level bounding boxes. Finally, based on the line-level bounding boxes, text fragments, text confidence, location information, and word-level bounding boxes, the page content information of the file page is constructed, thus forming a page evidence layer that simultaneously retains text content, line-level coordinates, word-level coordinates, and recognition confidence, and laying a solid data foundation for subsequent semantic region segmentation and entity information extraction.

[0139] In some embodiments, the above-mentioned S230 can be implemented in the following manner: Based on the location information corresponding to each entity identifier, the semantic region to which each entity identifier belongs is determined in multiple semantic regions; For each entity identifier, the context information corresponding to the entity identifier is determined based on the location information corresponding to the entity identifier and multiple text fragments within the semantic region to which the entity identifier belongs; For each entity identifier, the semantic identity type of the entity identifier is determined based on multiple text fragments within the semantic region to which the entity identifier belongs, the identifier features of the entity identifier, and the context information corresponding to the entity identifier. Based on the semantic identity type of each entity identifier, determine the semantic namespace to which each entity identifier belongs.

[0140] The aforementioned contextual information refers to information used to assist in determining the semantic identity of an entity identifier, based on its location information and multiple text fragments within the semantic region to which it belongs. Examples include adjacent text, column headers, and table or chapter titles. Identifier features refer to the morphological characteristics inherent in the entity identifier itself, which can be used to determine its semantic identity, such as identifier prefixes and identifier formats. Semantic identity type refers to the type of object the entity identifier actually points to, such as compound number, experimental example number, table row number, intermediate number, and method number.

[0141] In some embodiments, based on the location information of each entity identifier, it is determined which semantic region the location falls within, thereby determining the semantic region to which the entity identifier belongs. For example, if the center coordinates of the bounding box of an entity identifier are located within a semantic region that is determined to be a compound structure region, then the entity identifier belongs to that semantic region.

[0142] For each entity identifier, contextual information is extracted based on its location information and multiple text fragments within its semantic region. Contextual information can be obtained by acquiring text fragments within the row, column, or adjacent region of the entity identifier.

[0143] Subsequently, based on multiple text fragments within the semantic region to which the entity identifier belongs, the identifier features of the entity identifier, and the contextual information corresponding to the entity identifier, the semantic identity type of the entity identifier is comprehensively determined. Among these, the text fragments within the semantic region, the identifier features, and the contextual information provide the basis for judgment from different dimensions.

[0144] Based on the semantic identity type of each entity identifier, entity identifiers with the same semantic identity type are grouped into the same semantic namespace, thus determining the semantic namespace to which each entity identifier belongs. However, objects with the same identifier in different semantic namespaces are not automatically merged simply because they have the same numerical value. For example, all entity identifiers identified as compound numbers are grouped into the compound number namespace.

[0145] In this embodiment, the semantic region to which each entity identifier belongs is determined based on its location information, and the context information of the entity identifier is determined by combining the text fragments in the semantic region. Then, the semantic identity type of the entity identifier is determined based on the text fragments in the semantic region, the identifier features of the entity identifier, and the context information. Finally, the semantic namespace to which each entity identifier belongs is determined based on the semantic identity type. This isolates the numbering from different semantic sources into their respective independent semantic namespaces, providing a clear namespace basis for subsequent entity information association and avoiding confusion between different categories of identifiers.

[0146] In some embodiments, the identifier features include at least one of an identifier prefix and an identifier format; multiple text fragments within the semantic region to which the entity identifier belongs include the region name of the semantic region to which the entity identifier belongs; based on this, the determination of the semantic identity type of the entity identifier according to the multiple text fragments within the semantic region to which the entity identifier belongs, the identifier features of the entity identifier, and the context information corresponding to the entity identifier includes the following steps: Based on the preset evidence priority, the semantic identity type of an entity identifier is determined according to at least one of the following: identifier prefix, identifier format, region name, column position of the entity identifier within the semantic region, and context information corresponding to the entity identifier.

[0147] The aforementioned identifier prefix refers to the prefix portion included in the entity identifier, used to characterize the semantic origin of the number. Identifier format refers to the numbering form of the entity identifier, such as a purely numeric format, a mixed number and letter format, or a format with specific prefixes and suffixes. Region name refers to the name of the semantic region to which the entity identifier belongs, such as the region's title or header field. Column position refers to the column position of the entity identifier within the semantic region, such as whether the column is a numbering column or whether it is separate from the compound object column. Preset evidence priority refers to the priority order of evidence types used to determine the semantic identity type, from highest to lowest: region name or identifier prefix, identifier format, contextual information, and column position.

[0148] In some embodiments, the semantic identity type of an entity identifier is determined according to a preset evidence priority, based on at least one of the following: identifier prefix, identifier format, region name, column position of the entity identifier within the semantic region, and context information corresponding to the entity identifier. For example, the identification can be performed first based on the region name or identifier prefix with the highest priority. If the region name or identifier prefix is ​​insufficient for identification, then the identification can be performed sequentially based on downgraded evidence such as identifier format, context information, and column position.

[0149] Specifically, if a header or prefix such as "Compound", "Compd.", or "化合物" appears, it is determined as a compound number; if a title or prefix such as "Example", "Ex.", or "实施例" appears, it is determined as an experimental example number; if an identifier such as "Intermediate", "Int.", or "中间体" appears, it is determined as an intermediate number; if an identifier such as "Method", "General Procedure", "Procedure", or "方法" appears, it is determined as a method number; if a header such as "序号", "Serial No.", or "Row" appears, and a separate compound number column is additionally provided in the table, it is determined as a table row number. A table row number is only determined as a row number when it is used to represent the physical row order and is separated from the compound object column; purely numeric intermediate numbers and method numbers need to be identified based on trigger words such as headers, titles, or context including "Intermediate" / "Int." and "Method" / "Procedure" respectively. There is no single fixed rule that prioritizes the row number column in all cases. When there is conflicting or insufficient evidence, the candidate semantic identities are retained and a review is initiated, and merging is not performed automatically just because the values are the same.

[0150] In one embodiment, the evidence on which the discrimination is based may also refer to the evidence sequence of clear headers or identifier prefixes > chapter and table roles contained in context information > column position and cell type > sequential numbering rule, which corresponds to the preset evidence priority described above.

[0151] In addition, the activity table number is divided into a source identity indicating that the entity identifier comes from the activity table and an object identity indicating what actual object it points to, so as to avoid hierarchical confusion with compound numbers and experimental example numbers. The activity table number only indicates that the entity identifier originates from the activity table, and is a temporary identity when the specific semantics have not yet been determined, rather than a final semantic identity type parallel to compound numbers, experimental example numbers and the like; based on headers, identifier prefixes, chapter and table roles, number column positions and context evidence, it is further determined as a semantic identity type such as a compound number, an experimental example number, an intermediate number, a method number, or a table row number, and the source identity may be retained continuously to facilitate evidence traceability; when the evidence is insufficient, the activity table source number and the state of undetermined semantic identity are retained.

[0152] Exemplarily, when the header of the activity table number column clearly states "化合物编号", "Compound No." or "Compd.", the header itself can serve as direct evidence for the semantic identity of the number, and the numbers in this column are determined to belong to the compound number namespace. This process belongs to number identity classification, not cross-namespace mapping between two already determined namespaces; if the activity table header only states "No.", and the chapter title can only indicate that the table is a biological activity table, comprehensive determination needs to be made in combination with table footnotes, adjacent descriptions, number formats, naming conventions of other tables, and explicit mapping evidence.

[0153] This implementation determines the semantic identity type of an entity identifier by combining at least one of the following according to a preset evidence priority: region name or identifier prefix, identifier format, context information, and column position. Strong evidence such as region name or identifier prefix takes priority, while weak evidence such as column position is given priority. This allows for rapid determination of semantic identity when high-confidence evidence is sufficient for identification, and the retention of candidate identities for review when evidence is insufficient or conflicting. This improves the accuracy and robustness of semantic identity identification.

[0154] In some embodiments, after determining the semantic namespace to which each entity identifier belongs based on its semantic identity type, the entity information association method further includes the following steps: For multiple entity identifiers within the same semantic namespace, the row and column positions of each entity identifier within the semantic region are determined based on the location information corresponding to each entity identifier. Based on the row and column positions of each entity identifier within the semantic region, candidate identifier sequences corresponding to the row-first reading order and the column-first reading order are constructed respectively. Calculate the sequence model score for each candidate identifier sequence; The candidate identifier sequence with the highest sequence model score that is greater than the preset sequence score threshold is selected as the target identifier sequence. If the target identifier sequence does not trigger the semantic conflict condition corresponding to the semantic namespace, sequence repair is performed on multiple entity identifiers in the semantic namespace based on the target identifier sequence.

[0155] The row and column positions mentioned above refer to the row and column numbers of entity identifiers within the table or page layout of their respective semantic regions, used to determine the layout order of entity identifiers. Row-first reading order refers to the order in which entity identifiers are organized according to the reading order of row first, then column, that is, from left to right, then moving to the next line. Column-first reading order refers to the order in which entity identifiers are organized according to the reading order of column first, then row, that is, from top to bottom, then moving to the next column. A candidate identifier sequence refers to the sequence obtained by sorting multiple entity identifiers within a semantic namespace according to a certain reading order.

[0156] The sequence model score is a score used to evaluate the rationality of candidate identifier sequences. It is determined based on the degree of matching between the expected identifier and the character recognition value of the corresponding entity identifier at each position in the candidate identifier sequence, the degree of deviation between the expected identifier and the character recognition value of the corresponding entity identifier at each position, the gap completion information of the candidate identifier sequence, and the compatibility of the candidate identifier sequence with the semantic namespace.

[0157] Among them, the expected identifier refers to the identifier value that the candidate identifier sequence should correspond to at a certain position according to the consecutive numbering rule; the character recognition value refers to the recognition value obtained after the entity identifier is recognized by character recognition, that is, the original recognition result of the entity identifier, which may contain recognition errors.

[0158] Matching degree refers to the degree of consistency between the expected identifier at each position in the candidate identifier sequence and the character recognition value of the corresponding entity identifier. The higher the matching degree, the higher the corresponding positive sub-score. Deviation degree refers to the degree of difference between the expected identifier at each position in the candidate identifier sequence and the character recognition value of the corresponding entity identifier. The higher the deviation degree, the higher the corresponding negative sub-score. Gap completion information refers to the information or confidence level of the attempt to complete the gap when there is a missing identifier in the candidate identifier sequence (such as a position where the character recognition is not recognized). Compatibility degree refers to the overall consistency between the candidate identifier sequence and the current semantic namespace, such as whether the identifier prefix in the sequence matches the namespace type.

[0159] The preset sequence score threshold refers to the minimum score threshold set in advance to determine whether a candidate identifier sequence can be accepted as a valid sequence. The target identifier sequence is the sequence with the highest sequence model score among the candidate identifier sequences that is greater than the preset sequence score threshold, and is used as the final identifier order.

[0160] Semantic conflict conditions refer to conditions triggered in a target identifier sequence that indicate a contradiction between the sequence and the semantic namespace. For example, semantic conflict conditions may include: a contradiction between a clear character identification value and an inferred value; inference leading to a backward numbering; duplication or unfounded crossing; a numbering format not conforming to the current object namespace; semantic exclusivity between page chapter roles and entity identifiers; the appearance of a new table name after crossing pages; field changes; or explicit restarting of numbering. When any semantic conflict condition occurs, the original identification value is not overwritten by sequence continuity.

[0161] Sequence repair refers to the process of repairing multiple entity identifiers within the same semantic namespace based on the target identifier sequence. Examples include inferring missed numbers, repairing obviously truncated numbers, filling in single sequence gaps, correcting character recognition errors, and adjusting identifier order.

[0162] In some embodiments, for multiple entity identifiers within the same semantic namespace, the row and column positions of each entity identifier within the semantic region are determined based on the location information corresponding to each entity identifier. Based on the row and column positions of each entity identifier within the semantic region, candidate identifier sequences corresponding to row-first reading order and column-first reading order are constructed respectively.

[0163] The process of constructing a candidate identifier sequence may include: for row-first reading order, sorting the row numbers of each entity identifier in ascending order; if the row numbers are the same, sorting them in column numbers from left to right, thus obtaining the candidate identifier sequence corresponding to the row-first reading order; for column-first reading order, sorting the column numbers in ascending order; if the column numbers are the same, sorting them in row numbers from top to bottom, thus obtaining the candidate identifier sequence corresponding to the column-first reading order.

[0164] For example, the sequence model score of each candidate identifier sequence is calculated based on the degree of matching between the expected identifier and the character recognition value of the corresponding entity identifier at each position in the candidate identifier sequence, the degree of deviation between the expected identifier and the character recognition value of the corresponding entity identifier at each position, the gap completion information of the candidate identifier sequence, and the compatibility of the candidate identifier sequence with the semantic namespace.

[0165] Specifically, the scores for complete matches, prefix omission matches, suffix matches, and adjacent consecutive matches of a candidate identifier sequence can be determined based on the degree of matching between the expected identifier at each position and the character recognition value of the corresponding entity identifier. The scores for significant deviations of the candidate identifier sequence can be determined based on the degree of deviation between the expected identifier at each position and the character recognition value of the corresponding entity identifier. The scores for individual gap completion of the candidate identifier sequence can be determined based on the gap completion information of the candidate identifier sequence. The scores for suspected cross-namespace deductions can be determined based on the compatibility of the candidate identifier sequence with the semantic namespace.

[0166] The positive sub-score consists of scores for complete matches, missing prefix matches, suffix matches, adjacent consecutive matches, and single gap completion; negative sub-scores consist of deductions for significant deviations and suspected deductions for cross-namespaces. These sub-items are normalized and then weighted and summed to obtain the sequence model score. Specific weights can be pre-configured according to actual needs.

[0167] For example, the above sub-items can be calculated as follows: Perfect match score: If the character recognition value is exactly the same as the expected identifier, a positive bonus is awarded based on the degree of matching. For example, a perfect match of two or more digits earns 8 points, while a perfect match of one digit earns 3 points.

[0168] Prefix Omission Match Score: If the character recognition value is the last digit after the expected identifier omitting the leading digit, such as the expected identifier being 121 and the recognition value being 21, then a positive bonus is awarded based on the degree of matching. For example, recognizing two or more digits earns 3 points, while recognizing only one digit earns 1.5 points.

[0169] Suffix Matching Score: If the character identification value matches the expected identifier only in the suffix, such as the expected identifier being Cmpd-121 and the character identification value being Cpd-121 (i.e., the suffix numbers are the same but the prefix abbreviations are different), then a positive bonus is awarded based on the degree of matching. Adjacent Continuous Item Score: If the character identification value has the same number of digits as the expected identifier and differs by only one character, such as 120 and 121, then a positive bonus is awarded based on the degree of matching (e.g., 2 points), reflecting the proximity and continuity of the numbering.

[0170] Single Gap Completion Score: When a character identification value is missing at a certain position in the candidate identifier sequence, the system infers the expected identifier at that position based on the consecutive numbering pattern and attempts to complete it. If the completion is successful, a positive bonus is awarded based on the gap completion information.

[0171] Significant deviation penalty: If there is a significant discrepancy between the clear character recognition value and the expected identifier, such as the expected identifier being 121 and the character recognition value being 125 and clearly distinguishable, then negative points will be deducted based on the degree of deviation. For example, a discrepancy of three or more characters will result in a deduction of 8 points, while a discrepancy of two characters will result in a deduction of 1.5 points.

[0172] Cross-namespace suspected deduction: If the expected identifier numbering format or prefix is ​​inconsistent with the type of the current semantic namespace, such as the presence of an experimental example prefix in the compound numbering namespace, negative deduction will be made based on the degree of compatibility.

[0173] When the sequence model score of a candidate identifier sequence is greater than a preset sequence score threshold and higher than the sequence model scores of other candidate identifier sequences, the candidate identifier sequence is adopted as the target identifier sequence. If the target identifier sequence does not trigger the semantic conflict conditions corresponding to the semantic namespace, sequence repair is performed on multiple entity identifiers within the semantic namespace based on the target identifier sequence, such as inferring missed numbers, repairing obviously truncated numbers, or filling in single sequence gaps.

[0174] When the target identifier sequence triggers the semantic conflict condition corresponding to the semantic namespace, such as if the target identifier sequence has a substantial conflict with the original character identification value, semantic region or semantic namespace evidence, it will not be forced to repair, but will be kept in a state pending review, so that the entity identifier repair is subject to the dual constraints of table position and semantic namespace.

[0175] This implementation constructs candidate identifier sequences for multiple entity identifiers within the same semantic namespace, based on row-first and column-first reading orders according to their row and column positions. It then calculates a sequence model score based on the degree of matching, deviation, missing information, and namespace compatibility between the desired identifier and the character recognition value. The target identifier sequence with the highest score, exceeding a preset sequence score threshold, and without triggering semantic conflict conditions is selected for sequence repair. This allows for the recovery of more reliable identifiers based on table position and consecutive numbering patterns when entity identifiers are truncated, missed, or misread. Furthermore, the semantic conflict condition constraint forces repair, preventing erroneous repairs and improving the accuracy and robustness of identifier repair.

[0176] In some embodiments, the specific processing flow of S240 above includes: If multiple first information entries in a set of multiple entity information entries have entity identifiers corresponding to at least two different semantic namespaces and there is a mapping relationship between the entity identifiers corresponding to the multiple first information entries, then the multiple first information entries are determined to be multiple target information entries with an association relationship. If multiple entity information entries contain multiple second information entries whose corresponding entity identifiers belong to the same semantic namespace and whose corresponding entity identifiers are identical, then the multiple second information entries are determined to be multiple target information entries with an association relationship.

[0177] The mapping relationship information mentioned above refers to the information about the correspondence between entity identifiers in different semantic namespaces. For example, the compound number Cmpd-101 and the experimental example number Ex-1 are explicitly or implicitly corresponding in the file to be processed.

[0178] Multiple first-information entries refer to multiple entity information entries belonging to at least two different semantic namespaces, and whose corresponding entity identifiers have a mapping relationship. Multiple second-information entries refer to multiple entity information entries belonging to the same semantic namespace, and whose corresponding entity identifiers are the same.

[0179] In some embodiments, multiple entity information entries are traversed. If a set of entity information entries is retrieved, and the entity identifiers corresponding to the set of entries belong to at least two different semantic namespaces, such as one part belonging to the compound number namespace and the other part belonging to the experimental example number namespace, and there is a mapping relationship between the entity identifiers corresponding to the set of entries, then the set of entity information entries is determined as multiple first information entries, and further determined as multiple target information entries with an association relationship.

[0180] If a set of entity information entries is retrieved, and the entity identifiers corresponding to this set of entries belong to the same semantic namespace (e.g., all belong to the compound number namespace), and the entity identifiers corresponding to this set of entries are the same (e.g., multiple entries all contain the number Cmpd-101), then this set of entity information entries is identified as multiple second information entries, and further identified as multiple target information entries with an association relationship.

[0181] For example, the two determination paths mentioned above are independent of each other. As long as the determination condition corresponding to either path is met, the association of the target information item can be completed.

[0182] This implementation method determines the target information entries with related relationships from multiple entity information entries by using the semantic namespace to which the entity identifier belongs and the mapping or similarity relationships between entity identifiers. This enables the automatic and accurate association of entity information from different or the same sources, avoiding erroneous merging between different numbering systems.

[0183] In some embodiments, the entity information association method further includes the following process: For each text information entry, if the entity identifier corresponding to the text information entry and the entity identifier corresponding to the structure graph information entry belong to different semantic namespaces, at least one candidate structure graph information entry is determined from multiple structure graph information entries based on the file content information of the file to be processed. in, When there are multiple candidate structure graph information entries, obtain the quality score corresponding to each candidate structure graph information entry. The quality score is used to characterize the credibility of the candidate structure graph information entry. The candidate structure diagram information item and text information item with the highest quality score are selected as multiple first information items.

[0184] The aforementioned candidate structure graph information entries refer to structure graph information entries selected from multiple structure graph information entries, where the entity identifiers corresponding to the text information entries and the entity identifiers corresponding to the structure graph information entries belong to different semantic namespaces, and there is a mapping relationship between the entity identifiers corresponding to the text information entries and the entity identifiers corresponding to the text information entries.

[0185] The quality score is used to characterize the credibility of candidate structure graph information entries. The higher the quality score, the higher the credibility of the candidate structure graph information entry. Specifically, the quality score can be obtained by weighted summation of multiple quality indicators of the candidate structure graph information entry. These multiple quality indicators include at least one of the following: image size, effective dark pixel aspect ratio, dark pixel ratio, number of connected regions, and edge touch degree.

[0186] In some embodiments, for each text information entry, it is determined whether the entity identifier corresponding to the text information entry and the entity identifier corresponding to the structure diagram information entry belong to different semantic namespaces. If they belong to different semantic namespaces (e.g., the text information entry belongs to the experimental example number namespace, while the structure diagram information entry belongs to the compound number namespace), then based on the file content information of the file to be processed (e.g., tables, paragraph descriptions, etc. in the file), at least one candidate structure diagram information entry is determined from multiple structure diagram information entries.

[0187] Specifically, for a mapping relationship to exist between entity identifiers corresponding to candidate structure diagram information entries and entity identifiers corresponding to text information entries, the following conditions must be met: A clear mapping is only valid when both types of entity identifiers appear simultaneously in the same sentence, the same table row, or adjacent contexts with a direct referential relationship, and are connected by relationships such as correspondence, i.e., production, or attainment. Relationships derived solely from identical numbers, proximity of pages, or sequence inference can only serve as weak evidence and cannot definitively establish a mapping relationship between the two entities.

[0188] When multiple candidate structure diagram information entries exist, the quality score of each candidate structure diagram information entry is obtained, and the candidate structure diagram information entry with the highest quality score is combined with the aforementioned text information entry to form multiple first information entries. When only one candidate structure diagram information entry exists, it is directly combined with the aforementioned text information entry to form multiple first information entries. When no candidate structure diagram information entry satisfies the mapping relationship information, no subsequent entity information association is performed on the text information entry, but the text information entry, the corresponding entity identifier, and the reason for review are retained for subsequent review.

[0189] To facilitate understanding, the process of associating bioactivity information entries with structural diagram information entries is illustrated below: First, the semantic namespace to which the entity identifier corresponding to the bioactivity information entry belongs is identified. Then, the association path is selected based on the semantic namespace, rather than simply establishing an association based on identical numbers or page distance. For example, when the entity identifier corresponding to the bioactivity information entry belongs to the experimental case number namespace, the file content information of the file to be processed is used to search for evidence of cross-namespace mapping, thereby determining the candidate structural diagram information entry.

[0190] When multiple candidate structural diagram entries exist, the optimal entry is selected primarily based on its quality score. This optimal entry is then combined with bioactivity information entries to form multiple first-order entries. In the auxiliary judgment process, factors such as semantic region matching, entity identifier evidence completeness, and contextual information can be used for comprehensive evaluation. For example, if a candidate structural diagram entry is actually located in the intermediate, starting material, reagent, or analytical data area, or if adjacent text points to another compound, then that candidate structural diagram entry is downgraded in weight; if there are strong conflicts such as namespace incompatibility, then that candidate structural diagram entry is directly excluded.

[0191] In this embodiment, candidate structural diagram information entries are determined by the file content information of the file to be processed under different semantic namespaces, and the candidate with the highest credibility is selected and associated with the text information entry based on the quality score. This improves the accuracy and reliability of cross-namespace mapping, avoids association errors caused by poor structural diagram quality or mapping errors, and further improves the overall accuracy of compound image-text activity association.

[0192] In some embodiments, the multiple semantic regions include at least two of the following: a compound structure region, a supplementary compound region, and an experimental example text region; the multiple semantic namespaces include a compound number namespace and an experimental example number namespace; based on this, the entity information association method further includes the following process: For each text information entry, if the entity identifier corresponding to the text information entry and the entity identifier corresponding to the structure diagram information entry belong to the compound number namespace, at least one first structure diagram information entry belonging to the compound structure region is determined among multiple structure diagram information entries, and the first structure diagram information entry and the text information entry with the same entity identifier as the text information entry are regarded as multiple second information entries. When the entity identifier corresponding to the text information entry and the entity identifier corresponding to the structure diagram information entry belong to the experimental example number namespace, at least one second structure diagram information entry belonging to the experimental example text region and the supplementary compound region is determined from multiple structure diagram information entries. The second structure diagram information entry and the text information entry with the same entity identifier as the text information entry are regarded as multiple second information entries.

[0193] The aforementioned compound structure region refers to the semantic region in the file to be processed that contains compound structure diagrams. The supplementary compound region refers to the semantic region in the file to be processed that contains supplementary compound information, such as the region containing supplementary experimental or intermediate compounds. The experimental example text region refers to the semantic region in the file to be processed that contains experimental example text descriptions, such as experimental example synthesis information, bioactivity tests, experimental example structure tables, etc.

[0194] The compound number namespace refers to the semantic namespace of entities identified as compound numbers, such as the namespace to which entities with the prefix Cmpd- or Compound belong. The experimental example number namespace refers to the semantic namespace of entities identified as experimental example numbers, such as the namespace to which entities with the prefix Ex- or Example belong.

[0195] The first structural diagram information entry refers to the structural diagram information entry that belongs to the compound structural region, determined from multiple structural diagram information entries, when the entity identifier corresponding to the text information entry and the entity identifier corresponding to the structural diagram information entry both belong to the compound number namespace.

[0196] The second structural diagram information entry refers to the structural diagram information entry that belongs to the experimental example text region and the supplementary compound region, determined from multiple structural diagram information entries, when the entity identifier corresponding to the text information entry and the entity identifier corresponding to the structural diagram information entry both belong to the experimental example number namespace.

[0197] In some embodiments, for each text information entry, the semantic namespace to which the entity identifier corresponding to the text information entry and the entity identifier corresponding to the structure graph information entry belong is determined.

[0198] When both the entity identifier corresponding to the text information entry and the entity identifier corresponding to the structure diagram information entry belong to the compound numbering namespace, firstly, based on the semantic region where each structure diagram information entry is located, entries located within the compound structure region are selected from multiple structure diagram information entries and identified as first structure diagram information entries. Subsequently, the entity identifiers corresponding to the first structure diagram information entries are compared with the entity identifiers corresponding to the text information entries (e.g., through exact string matching or matching after number format normalization), and first structure diagram information entries with the same entity identifiers are selected. These first structure diagram information entries are then combined with the text information entry to form multiple second information entries.

[0199] When both the entity identifier corresponding to the text information entry and the entity identifier corresponding to the structure diagram information entry belong to the experimental example number namespace, based on the semantic region where each structure diagram information entry is located, entries located in the experimental example text region or the supplementary compound region are selected from multiple structure diagram information entries and identified as second structure diagram information entries. Subsequently, the entity identifiers corresponding to the second structure diagram information entries are compared with the entity identifiers corresponding to the text information entries, and second structure diagram information entries with the same entity identifiers are selected. These second structure diagram information entries are then combined with the text information entry to form multiple second information entries.

[0200] If there is no first or second structural diagram information entry that meets the above conditions, the text information entry will not be associated with any subsequent entity information entries. The text information entry, the corresponding entity identifier, and the reason for review will be retained for subsequent review.

[0201] For example, when the entity identifier corresponding to a bioactivity information entry belongs to the compound number namespace, the structural diagram information entry with the same entity identifier is searched in the compound structure region; when the entity identifier corresponding to a bioactivity information entry belongs to the experimental example number namespace, the structural diagram information entry with the same entity identifier is searched in the experimental example text region or the supplementary compound region.

[0202] In this embodiment, by targeting the compound number namespace and the experimental example number namespace, and combining different semantic regions such as the compound structure region, the experimental example text region, and the supplementary compound region, structural diagram information entries with the same entity identifier are associated with text information entries as second information entries. This achieves accurate matching of entity information entries and improves the accuracy and execution efficiency of compound diagram-text activity association in specific scenarios.

[0203] In some embodiments, the multiple target information entries associated with an entity object include structural diagram information entries and text information entries; based on this, after S250, the entity information association method further includes the following processing flow: For each entity object, structured association information of the entity object is generated based on the structured attribute information associated with the entity object, the structure diagram information entries associated with the entity object, and the text information entries associated with the entity object. Based on the structural database, the similarity of compound structures is verified for the structural diagram information entries associated with entity objects, and the verification results of the structural diagrams corresponding to the entity objects are obtained. The structured association information of entity objects and the verification results of the corresponding structure diagrams of entity objects are input into the large language model to obtain structured verification information.

[0204] The aforementioned target information items specifically include two categories: structural diagram information items and text information items. Structured attribute information consists of attribute fields related to the entity object and organized in a structured manner, used to characterize the various attribute features of the entity object. The structure database is a pre-built database that includes multiple preset compound structure diagrams for comparative verification of compound structures. The preset compound structure diagrams are standard compound structure diagrams pre-included in the structure database, serving as a comparison benchmark for structural similarity verification.

[0205] In some embodiments, after associating multiple target information entries with the same entity object, the following processing is performed for each entity object: First, the structured attribute information associated with the entity object, the structure diagram information entries associated with the entity object, and the text information entries associated with the entity object are integrated to generate the structured association information of the entity object.

[0206] Subsequently, based on the structure database, similarity verification of compound structures is performed on the structure diagram information entries associated with the entity object. Specifically, the structure diagram is extracted from the structure diagram information entries and compared with multiple preset compound structure diagrams in the structure database to obtain the structure diagram verification result corresponding to the entity object. The structure diagram verification result is used to characterize the degree of matching or verification status between the structure diagram information entry and known structures in the structure database.

[0207] Finally, the structured association information of the entity object and the corresponding structure diagram verification result are input into the large language model. Based on the input structured association information and structure diagram verification results, the large language model performs a comprehensive analysis and outputs structured verification information characterizing the quality status of the entity object association results. The large language model possesses semantic understanding, logical verification, and information integration capabilities, and is used to comprehensively verify the input structured association information and structure diagram verification results.

[0208] In this embodiment, by integrating the structured attribute information, structure diagram information entries, and text information entries of the same entity object into structured association information, and combining structure database similarity verification with comprehensive verification of large language models, multi-dimensional quality control of entity association results is achieved, effectively improving the credibility and completeness of entity object graph-text association results.

[0209] In some embodiments, the entity information association method further includes the following process: Using the entity identifier corresponding to the structural graph information entry associated with the entity object as an index, at least one candidate attribute field of the entity object is extracted from multiple semantic regions; Perform content conflict detection on multiple candidate attribute fields, and filter out multiple target attribute fields that do not have content conflicts from the multiple candidate attribute fields based on the detection results; Based on the confidence level of each target attribute field, the structured attribute information of the entity object is determined from multiple target attribute fields.

[0210] In some embodiments, the structured attribute information of an entity object is generated through the following process: First, using the entity identifier corresponding to the structure graph information entry associated with the entity object as the main index, at least one candidate attribute field related to the entity object is extracted from multiple semantic regions of the file to be processed. The candidate attribute fields include fields such as name, mass spectrometry data, and reaction conditions.

[0211] Next, content conflict detection is performed on the extracted candidate attribute fields to identify any anomalies such as NMR text, supplier information, reaction steps, quantity units, noise markers, and multi-value conflicts. Specifically, for the same attribute type, the values ​​of fields extracted from different semantic regions are compared to see if they are consistent. If the numerical difference exceeds a preset threshold, it is determined that there is a multi-value conflict. At the same time, keyword matching, regular expressions, or format rules are used to identify whether the field content belongs to NMR text (containing features such as chemical shift and coupling constant), supplier information (containing procurement terms such as company name and item number), reaction steps (containing step numbers or condition descriptions), quantity units (containing units of measurement such as mg and mL that do not match the current attribute type), or noise markers (meaningless OCR recognition fragments or garbled text). Among them, major conflicts have a veto effect. For example, when the difference in multi-value conflicts is significant or the field content is confirmed to be completely irrelevant information, the candidate attribute fields with major conflicts are directly excluded. Based on the detection results, candidate attribute fields without content conflicts or with only slight noise are retained as target attribute fields.

[0212] Calculate the confidence score for each target attribute field. During calculation, the evidence for each target attribute field is first normalized. Each target attribute field may come from one or more semantic regions, with each source constituting one piece of evidence. The feature values ​​of each piece of evidence are mapped to a unified dimensional range. An initial confidence score is obtained by weighting and synthesizing the scores based on source reliability, OCR recognition quality, the mapping strength of the entity identifier corresponding to the structural graph information entry associated with the entity object, field format validity, and consistency across multiple sources.

[0213] The aforementioned evidence consists of each original record extracted from the file to be processed for the target attribute field, constituting one piece of evidence for that target attribute field; one target attribute field may correspond to one or more pieces of evidence, and each piece of evidence is associated with its source semantic region, the file page it is located on, and location information, etc.

[0214] The reliability of the source is determined by pre-weighting the semantic region type; for example, the compound structure region is weighted higher than the preparation region, the preparation region is weighted higher than the analysis data region, and the analysis data region is weighted higher than the experimental example text region. OCR recognition quality is determined based on the character recognition confidence score output by the OCR engine. Mapping strength is determined by the spatial distance between the entity identifier corresponding to the structural diagram information item associated with the field and the entity object on the page, whether they are located in the same table row or cell, and whether there is a clear label association; the closer the distance and the more direct the association, the higher the mapping strength. Field format validity is checked using regular expressions or chemical format rules to see if the field content conforms to the expected format of the attribute type; if it does, the score is higher. Multi-source consistency refers to the increased consistency score if the same attribute field is extracted with the same value in multiple semantic regions. Based on the weighted composite score, deductions are made for noisy evidence such as long NMR text, supplier information, and quantity units; the specific weights and thresholds can be determined by training with labeled samples or set by system configuration. The output of the confidence calculation retains the score of each sub-item, the reason for deduction, and the source location to ensure that the confidence is interpretable and traceable. Finally, based on the confidence level of each target attribute field, the field with the highest confidence level is selected from multiple target attribute fields and determined as the structured attribute information of the entity object.

[0215] In this embodiment, candidate attribute fields are extracted by using the entity identifier corresponding to the structure graph information entry associated with the entity object as an index, and content conflict detection and confidence filtering are performed. This effectively eliminates the inconsistency between multi-source attribute information, ensures the accuracy and credibility of structured attribute information, and provides a high-quality data foundation for subsequent structured association and verification.

[0216] In some embodiments, the above-mentioned method of verifying the similarity of compound structures among the structural diagram information entries associated with entity objects based on a structural database to obtain the structural diagram verification result corresponding to the entity object specifically includes the following steps: Chemical structure analysis is performed on the structural diagram information entries associated with entity objects to obtain molecular representation information; Based on molecular representation information, similarity searches or substructure matching are performed in the structural database to obtain similarity scores; The similarity score is compared with a preset similarity threshold to determine the verification status of the structure graph information item; Based on the similarity score and verification status, the verification result of the structure diagram corresponding to the entity object is obtained.

[0217] The aforementioned molecular representation information is used to characterize the digital information of compound structures, including at least one of molecular graph representation, molecular fingerprint, and molecular descriptor. Molecular graph representation is a method of representing molecular structures in a graph structure, where atoms are nodes and chemical bonds are edges. Molecular fingerprint is a method of encoding molecular structural features into fixed-length feature vectors, such as extended connectivity fingerprint (ECFP) and topological fingerprint, used for rapid structural similarity comparison. Molecular descriptor is a quantitative indicator used to describe the physicochemical properties and structural features of molecules, such as molar refractive index, lipid-water partition coefficient (LogP), and polar surface area.

[0218] The similarity score is a quantitative value used to characterize the degree of similarity between the structural diagram information entry to be verified and the preset compound structural diagrams in the structural database. The preset similarity threshold is a preset numerical limit used to judge the degree of matching between the structural diagram information entry and the preset compound structural diagram. When the similarity score is higher than the threshold, the match is considered successful. The specific value can be determined by system configuration or labeled samples.

[0219] The verification status is a status identifier determined based on the comparison between the similarity score and a preset similarity threshold. It includes verified and unverified, and is used to characterize the degree of consistency between the structure diagram and the preset compound structure diagram. The structure diagram verification result is verification result data obtained by combining the similarity score and the verification status, including information such as similarity score, matching compound number, compound name, verification status, and review status.

[0220] In some embodiments, chemical structure analysis is performed on the structural diagram information entries associated with the entity object, converting the image-based structural diagram into a computable digital representation to obtain molecular representation information. Based on the molecular representation information, similarity search or substructure matching is performed in the structure database to obtain a similarity score. The similarity search is the process of calculating similarity indices (such as Tanimoto coefficient, Dice coefficient, or cosine similarity) based on the molecular representation information in the structure database to find preset compound structural diagrams similar to the structural diagram information entry to be verified. Substructure matching is the process of using the molecular representation information as a query structure to retrieve preset compound structural diagrams containing the substructures corresponding to that molecular representation information in the structure database. Furthermore, precise matching can be performed, comparing molecular fingerprints or SMILES representations (simplified linear input specification) with preset compound structural diagrams in the structure database.

[0221] Next, the matching results, similarity scores, matching compound IDs, and compound names are recorded. The similarity scores are compared with a preset similarity threshold to determine the verification status of the structural diagram information entries. When the similarity score is higher than the preset similarity threshold, the verification status is marked as verified; when the similarity score is lower than the preset similarity threshold or no match is found, the verification status is marked as unverified and retained as pending review. Finally, based on the similarity score and verification status, the structural diagram verification result corresponding to the entity object is obtained. This result includes information such as the similarity score, matching compound ID, compound name, verification status, and review status. This verification result can be further input into subsequent consistency checks and review grading processes. For example, if it indicates a possible mismatch between structural diagram information entries and text information entries, it can prevent formal entry into the database or downstream release, thereby reducing the risk of incorrect structural diagrams being used for activity association.

[0222] This implementation method automates the verification of compound structures by parsing structural diagrams into molecular representation information and performing similarity searches or substructure matching in a structural database. It provides objective structural evidence for the association results and effectively improves the rigor and robustness of the image-text activity association method in confirming chemical structures.

[0223] In some embodiments, the structured validation information includes semantic consistency validation results, information contradiction detection results, missing field detection results, review priority corresponding to entity objects, and association confidence corresponding to entity objects. Based on this, the structured association information of entity objects and the structure graph validation results corresponding to entity objects are input into the large language model to obtain structured validation information, specifically including the following steps: The structured association information of entity objects and the verification results of the structure diagram corresponding to the entity objects are input into the large language model. The large language model performs semantic consistency verification, information contradiction detection and missing field detection to obtain the semantic consistency verification results, information contradiction detection results and missing field detection results. Based on the semantic consistency verification results, information contradiction detection results, and missing field detection results, the review priority and association confidence of the entity object are determined.

[0224] The semantic consistency check result described above is the output of the large language model after verifying whether the name, entity identifier, bioactivity data, and other information of the same entity object are consistent across different sources. The information contradiction detection result is the output of the large language model after detecting whether there are contradictions between information from different tables or different file pages. The missing field detection result is the output of the large language model after detecting whether fields such as the name, mass spectrometry, and reaction conditions that an entity object should possess are missing based on context, and can perform contextual inference for missing fields.

[0225] Review priority is an indicator used to characterize the urgency and priority level of the entity object association results requiring review, including three levels: urgent, important, and general. Association confidence is a quantitative rating used to characterize the overall credibility of the entity object association results, including three levels: high, medium, and low.

[0226] In some embodiments, the structured association information of all entity objects, the corresponding structure diagram verification results, and all related information marked as pending review (such as text information entries, structure diagram information entries, etc.) are input into the large language model. At the same time, information such as the association basis, the quality score of the structure diagram information entries, context information, source evidence, preset verification rules, and the page image to which each entity information entry belongs are also input.

[0227] Based on the above information, the large language model performs semantic consistency verification, information contradiction detection, and missing field detection. During these three checks, the large language model generates multiple sub-verification items within each entity object, categorized by fields and relationships. These include entity identifier semantic identity, structure graph, name, mass spectrometry, reaction conditions, various bioactivity fields, structure graph verification results, and the relationships between structure graph information entries and textual information entries (e.g., bioactivity information entries). For each sub-verification item, the model outputs a status of consistency, missing, conflict, or pending confirmation, and saves the corresponding source evidence and the cause of the problem. By summarizing the status of all sub-verification items for each entity object, the model obtains the semantic consistency verification result, information contradiction detection result, and missing field detection result for that entity object.

[0228] For the structure diagram verification results, if the verification status in the structure diagram verification results is unverified, it only means that there is a lack of available external verification evidence in the current structure database. The status of the sub-verification item is recorded as pending confirmation and is not directly judged as a structure error. Only when the matching result is obviously inconsistent with the current structure, entity identifier or name is it judged as a conflict.

[0229] The aforementioned pre-defined verification rules include consistency rules for the name, entity identifier, and bioactivity data of the same compound; rules for determining contradictions across pages and tables; rules for contextual inference of missing fields; rules for calculating association confidence and assigning high, medium, and low ratings; and rules for determining the priority of review based on urgency, importance, and generality. The large language model must execute the above rules based on the source evidence, and its judgment does not independently increase the association confidence rating.

[0230] Then, based on the semantic consistency check results, information contradiction detection results, and missing field detection results, the review priority and association confidence level of the corresponding entity object are determined. The association confidence level is divided into three levels: high, medium, and low, and the review priority is divided into three levels: urgent, important, and general, which are determined separately. The absence of non-critical fields or interpretable inconsistencies only affect the corresponding sub-check items and do not change the review priority of the entity object. When problems such as semantic namespace conflicts, structural graph identity conflicts, or potential mismatches between structural graphs and bioactivity data are involved, the review priority of the entity object is increased accordingly based on the potential impact of the problem on the association relationship between structural graph information entries and bioactivity information entries.

[0231] For example, entity objects that may cause structural mismatches, activity value mismatches, or strong associations in different namespaces are prioritized as urgent and blocked from formal entry into the database and downstream release; entity objects with low association confidence or missing key fields are prioritized as important and placed in the staging area for review; the remaining entity objects are prioritized as general and placed in the sampling review.

[0232] The association confidence level is comprehensively assessed based on the quality score of the information entries in the structure diagram, the semantic identity of the entity identifier, mapping evidence, field consistency, structure diagram verification results, semantic consistency verification results, information contradiction detection results, and missing field detection results: If the above information reflects that the evidence corresponding to the entity object is consistent and there are no major conflicts, the association confidence level of the entity object is rated as high; if the above information reflects that the entity object has interpretable non-critical missing information that does not affect the main mapping, the association confidence level of the entity object is rated as medium; if the above information reflects that the entity object has critical field missing information, namespace conflicts, mapping that relies only on weak inference, or the sub-verification item status corresponding to the structure diagram verification result is conflicting, the association confidence level of the entity object is rated as low. The specific threshold for the above review priority and association confidence level is determined by system configuration or verification samples, and the rating basis and triggering reasons are saved together with the association records.

[0233] The large language model can also generate specific review suggestions for entity objects with low association confidence, contradictions, or missing fields, and summarize the above verification results into a verification report. The verification report includes a verification summary, a list of contradictory records, a list of missing fields, a list of inference results, and a list of review suggestions. It can serve as a guide for review work and as training data for subsequent model optimization. After the review conclusions are written back, the association confidence and association status are recalculated to determine whether to restore them to the database and for further processing.

[0234] This implementation utilizes a large language model to perform multi-dimensional semantic verification of structured association information and structure graph verification results, automatically outputting detection results such as consistency, contradiction, and missing information, and generating review priorities and association confidence based on these results. This not only improves the verification efficiency of association results but also provides clear priority guidance for subsequent reviews, ensuring verification quality while reducing review costs.

[0235] For ease of understanding, let's take... Figure 3 The overall process of the entity information association method provided in this embodiment is described, specifically including the following steps: S310: Perform page image processing on the file to be processed to obtain page images of each file page; for each file page, perform optical character recognition on the page image of the file page, and construct the page content information of each file page based on the recognition results; S320: For each file page, divide the file page into multiple candidate regions based on the page content information and determine the semantic type of each candidate region; determine multiple semantic regions in the file to be processed based on the semantic type of each candidate region and each candidate region; and extract multiple entity identifiers and the location information corresponding to each entity identifier from the page content information of each file page. S330: When the semantic region contains the structure diagram of entity objects, determine the image sub-region in the semantic region according to the page image where the semantic region is located, and extract the structure diagram information entries corresponding to each entity object from the image sub-region; when the semantic region contains the text content of entity objects, extract the text content from the semantic region to obtain the text information entries corresponding to each entity object. S340: Determine the semantic namespace to which each entity identifier belongs based on the location information corresponding to each entity identifier and multiple semantic regions; S350: For multiple entity identifiers within the same semantic namespace, determine the row and column positions of each entity identifier within the semantic region based on the location information corresponding to each entity identifier; construct multiple candidate identifier sequences based on the row and column positions of each entity identifier within the semantic region; and perform sequence repair on multiple entity identifiers within the same semantic namespace based on the target identifier sequence determined from the multiple candidate identifier sequences. S360: If, among multiple entity information entries, there are multiple first information entries whose corresponding entity identifiers belong to at least two different semantic namespaces and there is a mapping relationship between the entity identifiers corresponding to the multiple first information entries, then the multiple first information entries are determined to be multiple target information entries with an association relationship; if, among multiple entity information entries, there are multiple second information entries whose corresponding entity identifiers belong to the same semantic namespace and the entity identifiers corresponding to the multiple second information entries are the same, then the multiple second information entries are determined to be multiple target information entries with an association relationship; and the multiple target information entries are associated with the same entity object; S370: For each entity object, generate structured association information for the entity object based on the structured attribute information associated with the entity object, the structure diagram information entries associated with the entity object, and the text information entries associated with the entity object; based on the structure database, perform compound structure similarity verification on the structure diagram information entries associated with the entity object to obtain the structure diagram verification result corresponding to the entity object. The structure database includes multiple preset compound structure diagrams; input the structured association information of the entity object and the structure diagram verification result corresponding to the entity object into the large language model to obtain structured verification information.

[0236] The entity information association method provided in this application, through page image parsing, semantic region segmentation, namespace isolation, and sequence repair, accurately extracts and associates compound structure diagrams and activity data from multi-source heterogeneous documents, effectively improving the quality of structure diagram cropping and the robustness of number recognition. Combined with external verification of the structure database and semantic verification of the large language model, the accuracy and interpretability of the association results are further guaranteed. Simultaneously, based on rule-based confidence assessment and evidence-based verification workflows, the method significantly reduces manual verification costs while lowering the deployment resource threshold, providing an efficient and reliable end-to-end solution for compound image-text-activity association.

[0237] Based on the entity information association method provided in the above embodiments, this application also provides specific implementation methods of the entity information association device. Please refer to the following embodiments.

[0238] like Figure 4 As shown, the entity information association device 400 provided in this application embodiment includes the following modules: The semantic region processing module 410 is used to determine multiple semantic regions in the file to be processed based on the file content information of the file to be processed, and to extract multiple entity identifiers and the position information corresponding to each entity identifier from the file content information. The file content information includes at least one text segment and the position information of each text segment in the file to be processed. The information entry extraction module 420 is used to extract entity information entries corresponding to each entity object from the semantic region; wherein, the entity information entries are associated with corresponding entity identifiers, and multiple entity information entries include structural diagram information entries and text information entries; The namespace determination module 430 is used to determine the semantic namespace to which each entity identifier belongs based on the location information corresponding to each entity identifier and multiple semantic regions. The information entry determination module 440 is used to determine multiple target information entries with related relationships from multiple entity information entries based on the semantic namespace to which each entity identifier belongs and the association relationship between each entity identifier. The information entry association module 450 is used to associate multiple target information entries with the same entity object.

[0239] The entity information association device 400 described above will be explained in detail below: Optionally, the information entry extraction module 420 includes: The first information entry extraction module is used to determine the image sub-region in the semantic region based on the page image where the semantic region is located, and extract the structure diagram information entries corresponding to each entity object from the image sub-region when the semantic region contains the structure diagram of entity objects. The second information entry extraction module is used to extract text content from the semantic region when the semantic region contains text content of entity objects, and obtain text information entries corresponding to each entity object.

[0240] Optionally, the first information entry extraction module includes: The table line detection module is used to detect table lines within the semantic region based on the page image where the semantic region is located. The column boundary determination module is used to determine the column boundaries of each column within the semantic region based on each horizontal table line and each vertical table line in response to the detection of at least one horizontal table line and at least one vertical table line. The table position determination module is used to determine the table row where each entity identifier is located based on the position information corresponding to each entity identifier within the semantic region. The image region determination module is used to determine the target cell of the structure diagram containing the entity object within the semantic region based on the table row and column boundaries where each entity identifier is located, and to use the target cell as the image sub-region.

[0241] Optionally, the first information entry extraction module includes: The structure diagram cropping module is used to crop out the initial structure diagrams corresponding to each entity object from the image sub-regions; The line removal module is used to remove long horizontal and vertical lines from each initial structure diagram to obtain the first structure diagram corresponding to each entity object. The bounding box calculation module is used to calculate the bounding box of the dark pixels corresponding to each first structure diagram based on the remaining dark pixels in each first structure diagram. The boundary determination module is used to determine the target boundary of each first structure image based on the dark pixel bounding box and the preset border ratio of each first structure image. The region trimming module is used to trim the blank areas corresponding to each first structure diagram according to the target boundary of each first structure diagram, so as to obtain the second structure diagram corresponding to each entity object. The structure diagram determination module is used to determine each second structure diagram as a structure diagram information entry corresponding to each entity object, provided that each second structure diagram meets the preset quality conditions.

[0242] Optionally, the entity information association device 400 further includes: The quality score calculation module is used to calculate multiple quality indicators of the second structural graph. The multiple quality indicators include at least one of the following: sub-score corresponding to image size, sub-score corresponding to effective dark pixel aspect ratio, sub-score corresponding to dark pixel ratio, sub-score corresponding to number of connected regions, and sub-score corresponding to edge touch degree. The quality score of the second structure diagram is obtained by weighted summation of multiple quality indicators.

[0243] Optionally, the second information entry extraction module includes: The content extraction module is used to extract active text content from semantic regions to obtain the activity index data corresponding to each entity object. The symbol determination module is used to determine the classification symbol of each activity index data when the activity index data corresponding to each entity object belongs to a preset classification type. The symbol mapping module is used to map each grade symbol to the corresponding activity value range or activity semantic range based on the preset symbol mapping relationship, so as to obtain the bioactivity information entries corresponding to each entity object; the preset symbol mapping relationship is used to characterize the activity value range or activity semantic range corresponding to different grade symbols.

[0244] Optionally, the symbol determination module includes: The information recognition module is used to perform multi-dimensional information recognition on each activity index data to obtain symbol recognition results for each dimension; each dimension corresponds to a recognition method. The result filtering module is used to filter out target recognition results that belong to a preset graded symbol set from multiple symbol recognition results; The results processing module is used to determine the classification symbols of the activity index data from multiple target recognition results when there are multiple target recognition results, based on the consistency between the target recognition results, the character confidence of each target recognition result, and the morphological similarity between each target recognition result and other classification symbols in the semantic region.

[0245] Optionally, the region processing module 410 includes: The region division module is used to divide each file page into multiple candidate regions based on the page content information. The semantic type determination module is used to determine the semantic type of each candidate region based on multiple text fragments, text layout features, table line distribution, and page crossing status of the candidate region. The text layout features are determined based on the position information of each text fragment in the candidate region within the file page. The semantic region determination module is used to determine multiple semantic regions in the file to be processed based on the semantic type to which each candidate region belongs and each candidate region.

[0246] Optionally, the semantic type determination module includes: The scoring module is used to calculate the comprehensive score of a candidate region relative to at least one candidate semantic type based on multiple text fragments, text layout features, table line distribution, and page-crossing status of the candidate region. The comprehensive score is determined based on at least one of the following: region name evidence score, column structure score, table line density score, keyword score, page-crossing continuity score, and exclusion deduction score. The exclusion deduction score is determined based on the degree of mutual exclusion between the text in the candidate region and the candidate semantic type. The scoring filtering module is used to sort the candidate regions by their comprehensive scores relative to each candidate semantic type, and to filter out the target score with the highest score from multiple comprehensive scores. The semantic type processing module is used to determine the semantic type of the candidate region as the semantic type of the candidate region when the target score is greater than the preset comprehensive score threshold.

[0247] Optionally, the entity information association device 400 further includes: The file processing module is used to perform page image processing on the files to be processed, and obtain page images of each file page; The character recognition module is used to perform optical character recognition on the page image of each file page to obtain at least one text fragment, the text confidence of each text fragment, the position information of each text fragment within the file page, and the word-level bounding box of each text fragment. The text merging module is used to merge text segments into lines based on word-level bounding boxes, thus obtaining line-level bounding boxes; The information construction module is used to construct the page content information of the file page based on multiple line-level bounding boxes, multiple text fragments, the text confidence of each text fragment, the position information of each text fragment within the file page, and the word-level bounding boxes of each text fragment.

[0248] Optionally, the namespace determination module 430 includes: The entity identification processing module is used to determine the semantic region to which each entity identification belongs in multiple semantic regions based on the location information corresponding to each entity identification. The context information determination module is used to determine the context information corresponding to each entity identifier based on the location information corresponding to the entity identifier and multiple text fragments in the semantic region to which the entity identifier belongs. The identity type determination module is used to determine the semantic identity type of each entity identifier based on multiple text fragments within the semantic region to which the entity identifier belongs, the identifier features of the entity identifier, and the context information corresponding to the entity identifier. The namespace processing module is used to determine the semantic namespace to which each entity identifier belongs based on its semantic identity type.

[0249] Optionally, the identity type determination module is specifically used to determine the semantic identity type of an entity identifier according to a preset evidence priority, based on at least one of the following: identifier prefix, identifier format, region name, column position of the entity identifier within the semantic region, and context information corresponding to the entity identifier; wherein, the preset evidence priority from high to low is as follows: region name or identifier prefix, identifier format, context information, and column position.

[0250] Optionally, the entity information association device 400 further includes: The row and column position determination module is used to determine the row and column position of each entity identifier within the semantic region based on the position information corresponding to each entity identifier for multiple entity identifiers within the same semantic namespace. The identifier sequence construction module is used to construct candidate identifier sequences corresponding to the row-first reading order and the column-first reading order, respectively, based on the row and column positions of each entity identifier within the semantic region. The model score calculation module is used to calculate the sequence model score of each candidate identifier sequence. The sequence model score is determined based on the degree of matching between the expected identifier and the character recognition value of the corresponding entity identifier at each position in the candidate identifier sequence, the degree of deviation between the expected identifier and the character recognition value of the corresponding entity identifier at each position, the gap completion information of the candidate identifier sequence, and the compatibility of the candidate identifier sequence with the semantic namespace. The identifier sequence selection module is used to select the candidate identifier sequence with the highest sequence model score that is greater than a preset sequence score threshold as the target identifier sequence. The identifier sequence repair module is used to repair the sequence of multiple entity identifiers in the semantic namespace based on the target identifier sequence when the target identifier sequence does not trigger the semantic conflict condition corresponding to the semantic namespace.

[0251] Optionally, the information entry determination module 440 includes: The first information entry processing module is used to determine multiple first information entries as multiple target information entries with an association relationship when multiple first information entries have entity identifiers corresponding to multiple first information entries belonging to at least two different semantic namespaces and there is mapping relationship information between the entity identifiers corresponding to multiple first information entries. The second information entry processing module is used to determine that multiple second information entries are multiple target information entries with an association relationship when multiple entity information entries have entity identifiers corresponding to multiple second information entries that belong to the same semantic namespace and have the same entity identifiers.

[0252] Optionally, the entity information association device 400 further includes: The candidate entry determination module is used to determine at least one candidate structural graph information entry from multiple structural graph information entries for each text information entry, provided that the entity identifier corresponding to the text information entry and the entity identifier corresponding to the structural graph information entry belong to different semantic namespaces, based on the file content information of the file to be processed; there is a mapping relationship between the entity identifier corresponding to the candidate structural graph information entry and the entity identifier corresponding to the text information entry. in, The quality score acquisition module is used to obtain the quality score corresponding to each candidate structure graph information entry when there are multiple candidate structure graph information entries. The quality score is used to characterize the credibility of the candidate structure graph information entries. The first information entry determination module is used to select the candidate structure diagram information entry with the highest quality score and the text information entry as multiple first information entries.

[0253] Optionally, the entity information association device 400 further includes: The second information entry determination module is used to determine, for each text information entry, at least one first structural diagram information entry belonging to the compound structure region among multiple structural diagram information entries, when the entity identifier corresponding to the text information entry and the entity identifier corresponding to the structural diagram information entry belong to the compound number namespace, and to take the first structural diagram information entry and the text information entry with the same entity identifier as the text information entry as multiple second information entries. The second information entry determination module is further configured to determine at least one second structural diagram information entry belonging to the experimental example text region and the supplementary compound region among multiple structural diagram information entries, when the entity identifier corresponding to the text information entry and the entity identifier corresponding to the structural diagram information entry belong to the experimental example number namespace, and to treat the second structural diagram information entry and the text information entry with the same entity identifier as the text information entry as multiple second information entries.

[0254] Optionally, the entity information association device 400 further includes: The structured information generation module is used to generate structured association information for each entity object based on the structured attribute information associated with the entity object, the structure diagram information entries associated with the entity object, and the text information entries associated with the entity object. The similarity verification module is used to verify the similarity of compound structures of the structural diagram information entries associated with entity objects based on the structural database, and to obtain the structural diagram verification results corresponding to the entity objects. The structural database includes multiple preset compound structural diagrams. The verification information processing module is used to input the structured association information of entity objects and the verification results of the corresponding structure diagrams of entity objects into the large language model to obtain structured verification information.

[0255] Optionally, the entity information association device 400 further includes: The field extraction module is used to extract at least one candidate attribute field of an entity object from multiple semantic regions, using the entity identifier corresponding to the structure diagram information entry associated with the entity object as an index; The field filtering module is used to perform content conflict detection on multiple candidate attribute fields, and filter out multiple target attribute fields that do not have content conflicts from the multiple candidate attribute fields based on the detection results. The attribute information determination module is used to determine the structured attribute information of an entity object from multiple target attribute fields based on the field confidence of each target attribute field.

[0256] Optionally, the similarity verification module includes: The structure parsing module is used to perform chemical structure parsing on the structural graph information entries associated with entity objects to obtain molecular representation information; the molecular representation information includes at least one of molecular graph representation, molecular fingerprint, and molecular descriptor. The structure matching module is used to perform similarity searches or substructure matching in the structure database based on molecular representation information to obtain similarity scores; The verification status determination module is used to compare the similarity score with a preset similarity threshold to determine the verification status corresponding to the structure diagram information item. The verification result determination module is used to obtain the structural diagram verification result corresponding to the entity object based on the similarity score and verification status.

[0257] Optionally, the verification information processing module includes: The information detection module is used to input the structured association information of entity objects and the verification results of the structure diagram corresponding to the entity objects into the large language model. The large language model performs semantic consistency verification, information contradiction detection and missing field detection to obtain the semantic consistency verification results, information contradiction detection results and missing field detection results. The verification information determination module is used to determine the review priority and association confidence of the entity object based on the semantic consistency verification results, information contradiction detection results, and missing field detection results.

[0258] In this embodiment, multiple semantic regions are determined based on the file content information corresponding to the file to be processed, and entity identifiers and their location information are extracted. Then, the semantic namespace to which each entity identifier belongs is determined based on the location information of the entity identifier and the semantic region. Since entity identifiers in different semantic regions often correspond to different numbering systems, combining location information with semantic regions to divide the semantic namespace can effectively distinguish the different meanings of the same number in different numbering systems. Based on this, according to the semantic namespace to which each entity identifier belongs and the association relationship between entity identifiers, target information entries with association relationships are determined from multiple entity information entries, and multiple target information entries are associated with the same entity object, reducing the risk of incorrectly associating entity information under different semantic namespaces. Therefore, this embodiment can effectively reduce information mismatch problems caused by chaotic numbering systems or spatial dispersion, achieve accurate association of multi-source information entries, and significantly improve the reliability of the association results.

[0259] Based on the entity information association method provided in the above embodiments, this application also provides specific implementation methods for electronic devices. Figure 5 A schematic diagram of an electronic device 500 provided in an embodiment of this application is shown.

[0260] Electronic device 500 may include processor 510 and memory 520 storing computer program instructions.

[0261] Specifically, the processor 510 may include a central processing unit (CPU), an application-specific integrated circuit (ASIC), or one or more integrated circuits that can be configured to implement the embodiments of this application.

[0262] Memory 520 may include mass storage for data or instructions. For example, and not limitingly, memory 520 may include a hard disk drive (HDD), floppy disk drive, flash memory, optical disk, magneto-optical disk, magnetic tape, or Universal Serial Bus (USB) drive, or a combination of two or more of these. Where appropriate, memory 520 may include removable or non-removable (or fixed) media. Where appropriate, memory 520 may be internal or external to electronic device 500. In a particular embodiment, memory 520 is a non-volatile solid-state memory.

[0263] In specific embodiments, the memory 520 can be implemented in the form of ROM (Read Only Memory), RAM (Random Access Memory), static storage device, dynamic storage device, etc. The memory 520 can store the operating system and other applications. When the technical solutions provided in the embodiments of this specification are implemented through software or firmware, the relevant program code is stored in the memory 520 and executed by the processor 510. The processor 510 implements any of the entity information association methods in the above embodiments by reading and executing the computer program instructions stored in the memory 520.

[0264] The processor 510 reads and executes computer program instructions stored in the memory 520 to implement any of the entity information association methods in the above embodiments.

[0265] In one example, electronic device 500 may also include communication interface 530 and bus 540. Wherein, as... Figure 5 As shown, the processor 510, memory 520, and communication interface 530 are connected through bus 540 and complete communication with each other.

[0266] The communication interface 530 is mainly used to realize communication between various modules, devices, units and / or equipment in the embodiments of this application.

[0267] Bus 540 includes hardware, software, or both, that couples components of an electronic device together. For example, and not limitingly, the bus may include an Accelerated Graphics Port (AGP) or other graphics bus, an Enhanced Industry Standard Architecture (EISA) bus, a Front Side Bus (FSB), HyperTransport (HT) interconnect, an Industry Standard Architecture (ISA) bus, an Infinite Bandwidth Interconnect, a Low Pin Count (LPC) bus, a memory bus, a Microchannel Architecture (MCA) bus, a Peripheral Component Interconnect (PCI) bus, a PCI-Express (PCI-X) bus, a Serial Advanced Technology Attachment (SATA) bus, a Video Electronics Standards Association Local (VLB) bus, or other suitable buses, or combinations of two or more of these. Where appropriate, bus 540 may include one or more buses. Although specific buses are described and illustrated in embodiments of this application, this application contemplates any suitable bus or interconnect.

[0268] For example, the electronic device 500 can be a mobile phone, tablet computer, laptop computer, handheld computer, in-vehicle electronic device, ultra-mobile personal computer (UMPC), netbook, or personal digital assistant (PDA), etc.

[0269] The electronic device can execute the entity information association method in the embodiments of this application, thereby achieving the combination Figures 2 to 3 The entity information association method described herein, and the beneficial effects of the corresponding method embodiments, will not be elaborated further here.

[0270] Furthermore, in conjunction with the entity information association methods in the above embodiments, this application embodiment can provide a computer-readable storage medium for implementation. This computer-readable storage medium stores computer program instructions; when these computer program instructions are executed by a processor, they implement any of the entity information association methods in the above embodiments.

[0271] The computer-readable medium of this embodiment includes permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information can be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0272] The computer program instructions stored in the storage medium of the above embodiments are used to cause the computer to execute the entity information association method as shown in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0273] Based on the entity information association methods in the above embodiments, this application can provide a computer program product for implementation. When the instructions in this computer program product are executed by the processor of an electronic device, they implement any of the entity information association methods in the above embodiments.

[0274] The computer program products of the above embodiments are used to implement the entity information association method shown in any of the above embodiments, and have the beneficial effects of the corresponding method embodiments, which will not be repeated here.

[0275] It should be clarified that this application is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of this application is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of this application.

[0276] The functional blocks shown in the above-described structural diagram can be implemented as hardware, software, firmware, or a combination thereof. When implemented in hardware, they can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this application are programs or code segments used to perform the required tasks. Programs or code segments can be stored on a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried on a carrier wave. "Machine-readable medium" can include any medium capable of storing or transmitting information. Examples of machine-readable media include electronic circuits, semiconductor memory devices, ROM, flash memory, erasable ROM (EROM), floppy disks, CD-ROMs, optical disks, hard disks, fiber optic media, radio frequency (RF) links, etc. Code segments can be downloaded via computer networks such as the Internet, intranets, etc.

[0277] It should also be noted that the exemplary embodiments mentioned in this application describe methods or systems based on a series of steps or apparatus. However, this application is not limited to the order of the above steps; that is, the steps can be performed in the order mentioned in the embodiments, or in a different order, or several steps can be performed simultaneously.

[0278] The aspects of this application have been described above with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block in the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable entity information association apparatus to produce a machine such that these instructions, executable via the processor of the computer or other programmable entity information association apparatus, enable the implementation of the functions / actions specified in one or more blocks of the flowchart illustrations and / or block diagrams. Such a processor may be, but is not limited to, a general-purpose processor, a special-purpose processor, a special application processor, or a field-programmable logic circuit. It is also understood that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can also be implemented by special-purpose hardware performing the specified functions or actions, or can be implemented by a combination of special-purpose hardware and computer instructions.

[0279] The above description is merely a specific implementation of this application. Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, modules, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here. It should be understood that the protection scope of this application is not limited thereto. Any person skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope disclosed in this application, and these modifications or substitutions should all be covered within the protection scope of this application.

Claims

1. A method for associating entity information, characterized in that, The method includes: Based on the file content information of the file to be processed, multiple semantic regions in the file to be processed are determined, and multiple entity identifiers and the position information corresponding to each entity identifier are extracted from the file content information. The file content information includes at least one text segment and the position information of each text segment in the file to be processed. Extract entity information entries corresponding to each entity object from the semantic region; wherein, each entity information entry is associated with a corresponding entity identifier, and multiple entity information entries include structural graph information entries and text information entries; Based on the location information corresponding to each entity identifier and the multiple semantic regions, determine the semantic namespace to which each entity identifier belongs; Based on the semantic namespace to which each entity identifier belongs and the association relationship between each entity identifier, determine multiple target information entries with association relationships from multiple entity information entries; The multiple target information entries are associated with the same entity object.

2. The method according to claim 1, characterized in that, The step of extracting entity information entries corresponding to each entity object from the semantic region includes: When the semantic region contains a structure diagram of entity objects, an image sub-region is determined in the semantic region based on the page image where the semantic region is located, and the structure diagram information entries corresponding to each entity object are extracted from the image sub-region. When the semantic region contains text content of entity objects, the text content is extracted from the semantic region to obtain text information entries corresponding to each entity object.

3. The method according to claim 2, characterized in that, The step of determining an image sub-region within the semantic region based on the page image where the semantic region is located includes: Based on the page image where the semantic region is located, perform table line detection within the semantic region; In response to detecting at least one horizontal table line and at least one vertical table line, the column boundaries of each column within the semantic region are determined based on each of the horizontal table lines and each of the vertical table lines. Based on the location information corresponding to each entity identifier within the semantic region, determine the table row where each entity identifier is located; Based on the table rows and column boundaries of each entity identifier, a target cell containing the structure diagram of the entity object is determined within the semantic region, and the target cell is used as an image sub-region.

4. The method according to claim 2, characterized in that, The step of extracting the structural diagram information entries corresponding to each entity object from the image sub-region includes: The initial structural diagrams corresponding to each entity object are cropped from the image sub-regions; Remove the long horizontal and vertical lines from each of the initial structure diagrams to obtain the first structure diagrams corresponding to each entity object; Calculate the bounding box of the dark pixels corresponding to each of the first structural diagrams based on the remaining dark pixels in each of the first structural diagrams; Based on the dark pixel bounding box and preset border ratio corresponding to each of the first structural diagrams, the target boundary corresponding to each of the first structural diagrams is determined. Based on the target boundary corresponding to each first structure diagram, the blank area corresponding to each first structure diagram is trimmed to obtain the second structure diagram corresponding to each entity object; If each of the second structural diagrams meets the preset quality conditions, each of the second structural diagrams is used as a structural diagram information entry corresponding to each of the entity objects.

5. The method according to claim 4, characterized in that, The preset quality condition includes that the quality score of the second structural diagram is greater than a preset quality score threshold. The calculation process for the quality score of the structure diagram includes: Calculate multiple quality metrics for the second structural graph, wherein the multiple quality metrics include at least one of the following: sub-score corresponding to image size, sub-score corresponding to effective dark pixel aspect ratio, sub-score corresponding to dark pixel ratio, sub-score corresponding to number of connected regions, and sub-score corresponding to edge touch degree. The quality score of the second structural diagram is obtained by weighted summation of the multiple quality indicators.

6. The method according to claim 2, characterized in that, Multiple of the aforementioned text information entries include bioactivity information entries; Text content is extracted from the semantic region to obtain text information entries corresponding to each entity object, including: Extract the active text content from the semantic region to obtain the activity index data corresponding to each entity object; When the activity index data corresponding to each entity object belongs to a preset classification type, determine the classification symbol of each activity index data. Based on a preset symbol mapping relationship, each of the graded symbols is mapped to a corresponding activity numerical range or activity semantic range, thereby obtaining bioactivity information entries corresponding to each entity object; the preset symbol mapping relationship is used to characterize the activity numerical range or activity semantic range corresponding to different graded symbols.

7. The method according to claim 6, characterized in that, The determination of the grading symbols for each of the activity index data includes: For each of the activity index data, multi-dimensional information recognition is performed on the activity index data to obtain symbol recognition results for each dimension; wherein, each dimension corresponds to a recognition method. From the multiple symbol recognition results, select the target recognition result that belongs to the preset graded symbol set; In the presence of multiple target recognition results, the classification symbol of the activity index data is determined from the multiple target recognition results based on the consistency between the target recognition results, the character confidence of each target recognition result, and the morphological similarity between each target recognition result and other classification symbols in the semantic region.

8. The method according to claim 1, characterized in that, The file to be processed includes multiple file pages, and the file content information includes the page content information of each file page, which includes at least one text fragment and the position information of each text fragment within the file page; The determination of multiple semantic regions in the file to be processed based on the file content information includes: For each file page, the file page is divided into multiple candidate regions based on the page content information of the file page; For each candidate region, the semantic type of the candidate region is determined based on multiple text fragments, text layout features, table line distribution, and page crossing status of the candidate region; the text layout features are determined based on the position information of each text fragment in the candidate region within the file page. Based on the semantic type to which each candidate region belongs and each candidate region, multiple semantic regions in the file to be processed are determined.

9. The method according to claim 1, characterized in that, The step of determining the semantic namespace to which each entity identifier belongs based on the location information corresponding to each entity identifier and the multiple semantic regions includes: Based on the location information corresponding to each entity identifier, the semantic region to which each entity identifier belongs is determined in the plurality of semantic regions; For each entity identifier, the context information corresponding to the entity identifier is determined based on the location information corresponding to the entity identifier and multiple text fragments within the semantic region to which the entity identifier belongs; For each entity identifier, the semantic identity type of the entity identifier is determined based on multiple text fragments within the semantic region to which the entity identifier belongs, the identifier features of the entity identifier, and the context information corresponding to the entity identifier. Based on the semantic identity type of each entity identifier, determine the semantic namespace to which each entity identifier belongs.

10. The method according to claim 1, characterized in that, The step of determining multiple target information entries with related relationships from multiple entity information entries based on the semantic namespace to which each entity identifier belongs and the association relationships between the entity identifiers includes: If, among the multiple entity information entries, the entity identifiers corresponding to multiple first information entries belong to at least two different semantic namespaces and there is a mapping relationship between the entity identifiers corresponding to the multiple first information entries, then the multiple first information entries are determined to be multiple target information entries with an association relationship. If multiple entity information entries contain multiple second information entries whose corresponding entity identifiers belong to the same semantic namespace and whose corresponding entity identifiers are identical, then the multiple second information entries are determined to be multiple target information entries with an association relationship.

11. The method according to any one of claims 1 to 10, characterized in that, The multiple target information entries associated with the entity object include structural diagram information entries and text information entries; After associating the multiple target information entries with the same entity object, the method further includes: For each entity object, structured association information of the entity object is generated based on the structured attribute information associated with the entity object, the structure diagram information entries associated with the entity object, and the text information entries associated with the entity object. Based on the structure database, the similarity of compound structures is verified for the structure diagram information entries associated with the entity object, and the structure diagram verification result corresponding to the entity object is obtained. The structure database includes multiple preset compound structure diagrams. The structured association information of the entity object and the verification result of the structure diagram corresponding to the entity object are input into the large language model to obtain the structured verification information.

12. An electronic device, characterized in that, include: processor; Memory used to store the processor's executable instructions; The processor is configured to execute the entity information association method according to any one of claims 1 to 11.

13. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method according to any one of claims 1 to 11.

Images