Chinese ancient book cross-language standardization translation system based on ontology structure mapping

Abstract

The application relates to the technical field of natural language processing, and discloses a traditional Chinese medical book cross-language standardization translation system based on ontology structure mapping, which comprises the following steps: receiving a book page image and literature metadata, performing layout analysis, character recognition, variant character standardization, text reconstruction and word unit extraction on the page, and generating a source language entity and a source language ontology subgraph; forming a binding edge constraint column grid phase offset fingerprint according to a binding edge, an innermost effective text column, adjacent effective text columns and an effective character frame, and using the fingerprint for candidate mapping sorting and directional review sorting; when a candidate mapping depends on cross-column or cross-page text continuity and lacks same-page independent semantic evidence, the candidate mapping sorting is relatively reduced; a target language preferred term is selected, a target translation is generated, and an audit record is output. The application is used for reducing the influence of binding side column recognition abnormalities on traditional Chinese medical term correspondence and translation consistency.

Description

Technical Field

[0001] This invention belongs to the field of natural language processing technology, specifically relating to a cross-language standardized translation system for ancient Chinese medical texts based on ontology structure mapping. Background Technology

[0002] The digital translation of ancient Chinese medical texts typically consists of image acquisition, OCR, variant character unification, sentence segmentation and word segmentation, entity recognition, ontology mapping, terminology selection, translation generation, and quality control review. In practice, OCR output often includes text block coordinates, character boxes, and confidence scores. Subsequent natural language processing focuses more on words, sentences, conceptual relationships, and translation quality. The source clues in the page binding are often only used for reading order or backtracking.

[0003] When collecting traditional Chinese books bound in thread or in imitation thread binding, issues such as spine shadows, binding holes obscuring the text, compressed inner columns, page cropping, and misordered vertical columns are common. These issues can lead to omissions, mis-splitting, or continuous misjudgments across pages in the innermost valid text column, thus affecting the identification of TCM entity boundaries, ontology correspondence, and the selection of target language terminology. Even if conventional translation systems retain page coordinates, they rarely pass on clues from the binding side columns to the candidate mapping sorting and translation review sorting. Proofreaders typically need to manually switch between text, page images, and glossaries to locate these problems. Summary of the Invention

[0004] The present invention aims to provide a cross-language standardized translation system for ancient Chinese medicine texts based on ontology structure mapping, which solves the technical problems mentioned in the background art.

[0005] To achieve the above-mentioned objectives, the present invention provides the following technical solution.

[0006] A cross-language standardized translation system for ancient Chinese medicine texts based on ontology structure mapping includes a page structure feature generation unit, an ontology structure mapping unit, a target language output unit, and a targeted review unit. The page structure feature generation unit generates a binding edge constraint column grid phase deviation fingerprint based on the existing binding edge, innermost valid text column, adjacent valid text column and valid character box in the page image of ancient Chinese medicine books. This fingerprint is used to characterize the degree of deviation of the binding side column grid of a single page relative to the set of pages in the same volume or the same format. The ontology structure mapping unit forms a candidate mapping ranking based on the source language ontology subgraph and the target language concept and term resources, and identifies whether the candidate mapping depends on cross-column or cross-page text continuity. The independent semantic evidence on the same page is the semantic evidence provided by the source language sentence segment, source language entity or source language ontology subgraph within the same page that supports the establishment of the candidate mapping. When the candidate mapping depends on cross-column or cross-page text continuity and lacks independent semantic evidence on the same page, the candidate mapping is relatively reduced in the candidate mapping ranking, and the original mapping score and the corrected mapping score are retained. The target language output unit selects the preferred terminology of the target language according to the candidate mapping sort and generates the target translation; The targeted review unit obtains semantic conflict values, combines them with binding edge constraint column grid phase deviation fingerprints to generate targeted review risk values, and generates targeted review ranking results based on the targeted review risk values.

[0007] Compared with the prior art, the present invention has the following substantial features and significant progress: This invention preserves the source clues of the binding sidebar column in the processes of ancient book page acquisition, character recognition, ontology mapping, translation generation, and review. It links page sidebar column deviation, candidate mapping sorting, and semantic conflict review to the same processing link, enabling the system to prioritize prompting translation risks caused by omissions, misordering, or cross-column splicing of inner sidebar columns. When page clues are unavailable, the original mapping and review process is still used to avoid misjudgments caused by supplementary values, thereby improving the traceability of TCM terminology correspondence, translation consistency, and proofreading location. Attached Figure Description

[0008] Figure 1 This is a schematic diagram of the system architecture and processing mechanism of the cross-language standardized translation system for ancient Chinese medical texts based on ontology structure mapping, as proposed in this invention. Detailed Implementation

[0009] The following description is provided in conjunction with the accompanying drawings. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the invention; those skilled in the art can make equivalent substitutions or combinations for the specific implementations.

[0010] This implementation method targets page images, electronic photocopies, or already entered text of traditional Chinese medicine books that are bound or imitation bound in thread. It operates in a continuous chain of natural language data processing, knowledge graph processing, ontology structure processing, machine translation, and translation quality control to convert source language text in traditional Chinese medicine books into target translations with terminological consistency and traceable ontology relationships.

[0011] In this embodiment, the system processes not paper restoration, scanning mechanical control, or paper quality evaluation, but rather the natural language text, source language entities, source language ontology subgraphs, candidate mapping ranking, target language preferred terms, target translations, targeted review ranking results, and audit records generated from images of pages from ancient Chinese medical books. Page structure evidence is stored and transmitted as input evidence in the natural language data processing chain, without altering the physical structure of the ancient Chinese medical book pages.

[0012] like Figure 1As shown, the system can be constructed using a processor 100, a memory 200, and a communication interface 300. The communication interface 300 receives page images 410 and metadata 420 from ancient Chinese medicine texts and provides them to the processor 100. The processor 100 calls upon the programs, Chinese medicine ontology, cross-language standard terminology resources, and process configurations stored in the memory 200 to execute the page structure feature generation unit 500, the ontology structure mapping unit 700, the target language output unit 800, and the targeted verification unit 900.

[0013] The page structure feature generation unit 500 generates a binding edge constraint column grid phase deviation fingerprint 520 based on the page image 410 of the ancient Chinese medicine book. The binding edge constraint column grid phase deviation fingerprint 520 represents the degree of deviation of a single page's binding side column grid relative to the set of pages in the same volume or format. This binding edge constraint column grid phase deviation fingerprint 520 is not a general image sharpness indicator, page tilt indicator, or OCR confidence indicator, but a structural scalar determined by the binding edge, the innermost valid text column, adjacent valid text columns, and valid character frames.

[0014] The ontology structure mapping unit 700 is used to form a candidate mapping ranking 720 based on the source language ontology subgraph 631 and the target language concept and term resources 641. The ontology structure mapping unit 700 also identifies whether the candidate mapping depends on cross-column or cross-page text continuity. When a candidate mapping depends on cross-column or cross-page text continuity and lacks independent semantic evidence on the same page, the ontology structure mapping unit 700 relatively lowers the candidate mapping ranking 720, while retaining the original mapping score 721 and the revised mapping score 722.

[0015] The target language output unit 800 selects the preferred target language terms 811 based on the candidate mapping sorting 720 and generates the target translation 830. The targeted review unit 900 obtains the semantic conflict value 911, combines it with the binding edge constraint column phase deviation fingerprint 520 to generate a targeted review risk value 912, and generates a targeted review sorting result 920 based on the targeted review risk value 912. The targeted review sorting result 920 is used to organize review priorities, and the audit record 930 is used to record the mapping, translation, review, and revision processes.

[0016] Figure 1 The solid arrows in the text correspond to the main processing data stream. The main processing data stream enters the communication interface 300 from the page image 410 of the ancient Chinese medicine book and the document metadata 420. The processor 100 calls the page structure feature generation unit 500, the ontology structure mapping unit 700, the target language output unit 800, and the targeted review unit 900 to form the target translation 830, the targeted review and sorting result 920, and the audit record 930. Figure 1The dashed arrow in the middle corresponds to the review feedback flow. The review feedback flow feeds back from the directional review and sorting result 920 to the character recognition and variant character standardization module 600 and the cross-language ontology structure mapping module 710, which is used to trigger local re-recognition, re-execute cross-language ontology structure mapping or record revision chain. Figure 1 The dotted lines in the diagram correspond to the program execution relationship, indicating that the processor 100 executes the program in the memory 200 and calls the functional units in the software function set.

[0017] The page structure feature generation unit 500 receives a page image 410 of a traditional Chinese medicine text. The page image 410 can be a scanned image, a photocopy, or a page image file. The page structure feature generation unit 500 does not require the page image 410 to be obtained from a specific scanning device, nor does it require physical marking, stretching, cutting, or rebinding of the paper pages. As long as the page image 410 contains visible information that can be used to identify page edges, spine shadows, binding holes, blank bands, center lines of text columns, or character boxes, the page structure feature generation unit 500 can execute this embodiment.

[0018] The page structure feature generation unit 500 includes a page structure analysis and fingerprint extraction module 510. The page structure analysis and fingerprint extraction module 510 first performs grayscale conversion, noise reduction, binarization, contrast adjustment, skew correction, edge cropping, text detection, annotation detection, page frame detection, and column line detection.

[0019] The page structure analysis and fingerprint extraction module 510 determines the physical boundary direction of the binding edge in the page image. The physical boundary direction of the binding edge is provided by the binding edge visibility information 511. The binding edge visibility information 511 includes at least one of the page edge, spine shadow, binding holes, and blank band. The page structure analysis and fingerprint extraction module 510 can determine the physical boundary direction of the binding edge based on the width of the blank band between the page edge and the text area, the continuous position of the spine shadow, the arrangement position of the binding holes, or the relationship between the binding side boundary and the page frame.

[0020] When the page edge, spine shadow, binding hole, and blank band cannot form a locatable binding side candidate area, or when multiple binding side candidate areas are formed but the unique physical boundary direction of the binding edge cannot be determined based on the page orientation information, page frame position, and text column arrangement relationship in the document metadata 420, the page structure analysis and fingerprint extraction module 510 determines that the binding edge cannot be determined. When the binding edge cannot be determined, the page structure analysis and fingerprint extraction module 510 does not calculate the distance of the marked binding edge, and the page structure feature generation unit 500 marks the binding edge constraint column grid phase deviation fingerprint 520 as unusable.

[0021] The page structure analysis and fingerprint extraction module 510 determines the center line of the innermost valid text column closest to the binding edge. The innermost valid text column 512 represents the text column closest to the binding edge and usable for text recognition. Adjacent valid text columns 513 represent text columns adjacent to the innermost valid text column 512 and usable for text recognition. Valid character frames 514 represent character frames usable for character position and scale analysis during text recognition. Valid character frames 514 can originate from the character rectangles output by the OCR or from the character regions obtained through connected component analysis.

[0022] The page structure analysis and fingerprint extraction module 510 obtains the marked binding edge distance. (Marked binding edge distance) Page The marked vertical distance from the binding edge to the center line of the innermost valid text column. The absolute value of the marked binding edge distance is derived from the distance measurement in the page image coordinate system, and the sign of the marked binding edge distance is used to preserve the physical boundary direction of the binding edge.

[0023] The page structure analysis and fingerprint extraction module 510 obtains the median of the column center distance in the main text. Page The median distance between adjacent valid text column centers. The median distance between text column centers is calculated using the adjacent distances between the center lines of at least two valid text columns. Using the median distance between text column centers can reduce the impact of single column detection bias, variations in the width of the annotation area, or local damage on column spacing estimation.

[0024] The page structure analysis and fingerprint extraction module 510 obtains the median height of the character boxes. Page The median of the effective character frame height. The median character frame height is used for scale normalization, making the original column raster phase values ​​of the same page comparable under different scanning resolutions.

[0025] The marked binding edge distance, median center distance of the text column, and median height of the character box all use the same page image coordinate unit. The median center distance of the text column must be greater than zero, and the median height of the character box must be greater than zero. When the median center distance of the text column is zero, negative, zero, negative, or any one of the median center distance of the text column or the median height of the character box is missing, the page structure feature generation unit 500 marks the binding edge constraint column grid phase deviation fingerprint 520 as unavailable.

[0026] After obtaining the marked binding margin distance, the median center distance of the text columns, and the median height of the character boxes, the page structure analysis and fingerprint extraction module 510 generates the original column grid phase values ​​of the page. This represents the single-page phase residual constructed based on the distance of the marked binding edge, the median center distance of the text columns, and the median height of the character boxes. The original column grid phase value of the page is not a quantity resulting from deformation of the paper page, but a mathematical construction of the existing structure in the page image 410 of the ancient Chinese medicine book.

[0027] ; in, Indicates the original column grid phase value of the page. Indicates the distance of the marked binding edge. This represents the median of the center distance of the main text column. This indicates the median height of the character box. A sign function representing the distance of the signed binding edge. This indicates taking the nearest integer; when the input is exactly between two integers, the integer furthest from zero is used as the rounding result.

[0028] The page structure analysis and fingerprint extraction module 510 performs median statistics on the original column grid phase values ​​of pages within the same volume or page set of the same layout to form an intra-volume layout baseline. This refers to a collection of pages from the same volume or with the same layout. The median of the original column grid phase value of the page. The volume layout baseline is used to eliminate the influence of the inherent binding side margin design of a certain version, so that the binding edge constraint column grid phase deviation fingerprint 520 mainly reflects the deviation of a single page relative to the set of pages in the same volume or layout.

[0029] When performing median statistics on the intra-volume layout baseline, the page structure analysis and fingerprint extraction module 510 selects only pages for which the original column grid phase values ​​have been obtained as statistical samples. When there are no original column grid phase values ​​available for median statistics in the same volume or the same layout page set, the page structure feature generation unit 500 does not calculate the intra-volume layout baseline and marks the binding edge constraint column grid phase deviation fingerprint 520 as unusable; in this case, the ontology structure mapping unit 700 keeps the candidate mapping sort 720 from being adjusted by the binding edge constraint column grid phase deviation fingerprint 520, and the directional verification unit 900 uses only the semantic conflict value 911 to generate the directional verification sort result 920.

[0030] ; in, Indicates the baseline of the volume layout. This refers to a collection of pages from the same volume or with the same layout. This refers to pages within the same volume or set of pages with the same layout. Page The original column grid phase value of the page.

[0031] The page structure analysis and fingerprint extraction module 510 generates a binding edge constraint column grid phase deviation fingerprint 520 based on the original column grid phase values ​​of the page and the in-volume layout baseline. This indicates the deviation of the original column grid phase value of the page relative to the in-volume layout baseline. The absolute value of the binding edge constraint column grid phase deviation fingerprint 520 represents the page... The innermost valid text column 512 deviates from the normal structure of the same volume in terms of structural deviation. The symbol of the binding edge constraint column phase deviation fingerprint 520 indicates the direction of deviation.

[0032] ; in, This indicates that the binding edge constraint column grid phase deviates from the fingerprint. Page The original column grid phase value of the page. Indicates the baseline of the layout within the volume.

[0033] The page structure feature generation unit 500 associates and stores the binding edge constraint column grid phase deviation fingerprint 520 with the page number, text block coordinates, reading order, marked binding edge distance, median center distance of the main text column, median height of the character box, original column grid phase value of the page, and in-volume layout baseline. This association storage enables subsequent source language segments, source language entities, source language ontology subgraphs 631, candidate mapping sorting 720, and target translations 830 to be traced back to the source page.

[0034] When the binding edge cannot be determined, there are fewer than two valid text columns, the number of valid character boxes is insufficient, the median center distance of the text columns is not greater than zero, the median height of the character boxes is not greater than zero, or there is no original column phase value of the page that can be used for median statistics in the same volume or the same layout page set, the page structure feature generation unit 500 marks the binding edge constraint column phase deviation fingerprint 520 as unusable. Fewer than two valid text columns means that the page structure analysis and fingerprint extraction module 510 cannot obtain at least two valid text column center lines; insufficient number of valid character boxes means that the page structure analysis and fingerprint extraction module 510 cannot obtain valid character boxes 514 for calculating the median height of the character boxes. This marking indicates that the page structure feature generation unit 500 has not obtained sufficient page structure evidence to support the calculation. In this case, the ontology structure mapping unit 700 keeps the candidate mapping sort 720 unadjusted by the binding edge constraint column phase deviation fingerprint 520, and the directional verification unit 900 generates the directional verification sort result 920 using only the semantic conflict value 911.

[0035] The page structure feature generation unit 500 does not replace missing marked binding edge distances, median center distances of text columns, or median character box heights with manually filled values. This constraint prevents the system from fabricating structural evidence out of thin air and ensures that the binding edge constraint column grid phase deviation fingerprint 520 has a clear data source when it is available.

[0036] After the page structure feature generation unit 500 completes the page structure analysis, the character recognition and variant character standardization module 600 performs character recognition on each character block. The character recognition and variant character standardization module 600 outputs characters, candidate characters, confidence scores, and character box coordinates, and concatenates characters according to the reading order given by the page structure analysis and fingerprint extraction module 510.

[0037] The character recognition and variant character standardization module 600 maps verifiable variant characters to standardized characters based on publicly available variant character lists, simplified / traditional character conversion tables, and existing character shape comparison tables in the project. For characters with low confidence or characters not found in the character list, the character recognition and variant character standardization module 600 retains the original recognition results and sets them to a pending manual verification state.

[0038] The text reconstruction and word unit extraction module 610 reconstructs the main text according to page, column number, and block order, and saves the annotations and main text separately. The module 610 uses commonly used function words in Classical Chinese, sentence pattern markers, punctuation recovery rules, or existing sentence segmentation models to generate candidate sentence boundaries, and employs dictionary maximum matching, statistical word segmentation, or existing Classical Chinese word segmentation models to extract drug names, prescription names, disease names, syndrome names, organ names, meridian names, acupoint names, dosage expressions, and common words.

[0039] The source language entity recognition and concept normalization module 620 uses a traditional Chinese medicine ontology and a thesaurus to identify traditional Chinese medicine entities in the source text and associates different expressions with normalized concepts in the ontology. The source language entity recognition and concept normalization module 620 can generate one or more ontology concept candidates based on word form, synonyms, context category constraints, and co-occurrence information, and then select normalized concepts based on word similarity, dictionary priority, category consistency, and context matching.

[0040] The source language ontology subgraph construction module 630 uses standardized concepts as graph nodes and writes the main treatment relationship, composition relationship, compatibility relationship, meridian tropism relationship, causal relationship, or syndrome-treatment relationship into the source language ontology subgraph 631. The source language ontology subgraph 631 stores nodes, relationship edges, source sentences, evidence words, and confidence levels. For relationships that can be directly confirmed by the TCM ontology, the source language ontology subgraph construction module 630 writes definite relationship edges; for relationships that are only inferred from the text and lack sufficient evidence, the source language ontology subgraph construction module 630 stores candidate relationship edges and retains the evidence state.

[0041] The Traditional Chinese Medicine (TCM) ontology and cross-language standard terminology resource loading module 640 loads the TCM ontology, thesaurus, and target language concept and terminology resources 641. The target language concept and terminology resources 641 include target language concepts, a target language terminology list, and related indexes. The TCM ontology and cross-language standard terminology resource loading module 640 can use relational database tables, RDF diagrams, OWL diagrams, JSON files, or key-value indexes to host these resources.

[0042] The aforementioned steps of character recognition, variant character normalization, text reconstruction, word unit extraction, source language entity recognition, concept normalization, and source language ontology subgraph construction all maintain a source association with the binding edge constraint column grid phase deviation fingerprint 520. In other words, the source language entities and the source language ontology subgraph 631 not only contain semantic content but also page numbers, column numbers, block numbers, and character coordinates, enabling the ontology structure mapping unit 700 to identify whether candidate mappings depend on cross-column or cross-page text continuity relationships.

[0043] The ontology structure mapping unit 700 receives the source language ontology subgraph 631, target language concept and term resources 641, and binding edge constraint column grid phase offset fingerprint 520. The cross-language ontology structure mapping module 710 performs target concept candidate query, category constraint, adjacency comparison, and basic mapping scoring processing to obtain the original mapping score 721. This represents the dimensionless normalized ranking score of the candidate mapping before sorting adjustment. The larger the value, the higher the priority of the candidate mapping.

[0044] The cross-language ontology structure mapping module 710 compares the source language entity categories, ontology relation types, hierarchical relations, composition relations, principal relations, compatibility relations, meridian relations, causal relations, or syndrome-treatment relations in the source language ontology subgraph 631 with the target concepts in the target language concepts and terminology resources 641.

[0045] The input to the basic mapping scoring process includes at least the source nodes, source node categories, source node relation edges, source sentence segments, and evidence words in the source language ontology subgraph 631, and the target nodes, target node categories, target node relation edges, and target language preferred terms 811 in the target language concept and terminology resource 641. The output of the basic mapping scoring process includes at least the source nodes, target nodes, original mapping scores 721, the ranking position of candidate mappings in candidate mapping ranking 720, and candidate evidence. The cross-language ontology structure mapping module 710 reads the original mapping scores 721 output from the existing cross-language ontology candidate mapping process and performs deterministic corrections when the structural continuity label 723 condition is met.

[0046] The cross-language ontology structure mapping module 710 sets structural continuity markers 723 based on the cross-column or cross-page text continuity relationship formed between the innermost valid text column 512 and the adjacent valid text column 513 or adjacent pages. Structural continuity markers This indicates whether the candidate mapping depends on cross-column or cross-page text continuity. If the candidate mapping depends on cross-column or cross-page text continuity formed between the innermost valid text column 512 and the adjacent valid text column 513 or adjacent pages, and lacks independent semantic evidence within the same page, then the structural continuity marker 723 is... If the candidate mapping does not depend on cross-column or cross-page textual continuity, or if there is independent semantic evidence within the same page, then structural continuity marker 723 is... .

[0047] Independent semantic evidence on the same page refers to semantic evidence that supports the validity of candidate mappings provided by source language segments, source language entities, or source language ontology subgraph 631 within the same page. The determination of independent semantic evidence on the same page can be achieved through entity co-occurrence, consistent relation edges, consistent concept categories, repeated occurrence of evidence words, or contextual support within the same page.

[0048] When the binding edge constraint column phase deviation fingerprint 520 is available and the structural continuity marker 723 indicates a dependency, the cross-language ontology structure mapping module 710 lowers the position of the corresponding candidate mapping in the candidate mapping sorting 720, while retaining the original mapping score 721, the structural continuity marker 723, and the corrected mapping score 722. (Corrected mapping score) This represents the dimensionless sort score of the candidate mapping after sorting adjustment.

[0049] ; in, This indicates a correction to the mapping score. Represents the original mapping score. This indicates that the binding edge constraint column grid phase deviates from the fingerprint. Indicates structural continuity markers, This represents the source node index in the source language ontology subgraph 631. This represents the target node index in the target language concept and terminology resource 641.

[0050] When the phase deviation of the binding edge constraint column grid from fingerprint 520 is zero, the corrected mapping score 722 equals the original mapping score 721. When the structural continuity marker 723 is... When the binding edge constraint column phase deviation fingerprint 520 is unavailable, the corrected mapping score 722 equals the original mapping score 721. The corrected mapping score 722 is only used for relative ranking in the candidate mapping ranking 720; when the corrected mapping score 722 is less than zero, the cross-language ontology structure mapping module 710 retains the less than zero corrected mapping score 722 as the ranking value and does not interpret it as a probability value.

[0051] The ontology structure mapping unit 700 does not directly delete candidate mappings, nor does it replace semantic judgment with the binding edge constraint column grid phase deviation fingerprint 520. The binding edge constraint column grid phase deviation fingerprint 520 only participates in the sorting adjustment when the candidate mapping depends on cross-column or cross-page text continuity and lacks independent semantic evidence on the same page. This constraint avoids denying the true semantic relationship simply because of page structure deviation, while also placing candidate mappings that may be caused by misordering, missing, or incorrectly continuous innermost valid text column 512 in a lower position.

[0052] The candidate mapping sorting 720 outputs node correspondences, relation correspondences, unmapped items, pre-adjustment sorting information, and post-adjustment sorting information. The candidate mapping sorting 720 also saves the original mapping score 721, the corrected mapping score 722, the structural continuity marker 723, and the binding edge constraint column phase deviation fingerprint 520, for the target language output unit 800 to select the preferred target language term 811 and for the directional review unit 900 to form the directional review sorting result 920.

[0053] The target language output unit 800 includes a target language terminology selection module 810 and a target language sentence generation and word order adjustment module 820. The target language terminology selection module 810 selects preferred target language terms 811 from the target language concept and terminology resources 641 based on the cross-language ontology structure mapping results, and establishes a document-level terminology consistency table. The preferred target language terms 811 represent the selected standard terms from the target language concept and terminology resources 641.

[0054] The target language terminology selection module 810 prioritizes the target language preferred term 811 bound to the target concept identifier in the target language concept and terminology resource 641, and records the terminology source and resource version. When multiple allowed terms exist in the target language concept and terminology resource 641, the target language terminology selection module 810 selects the target language preferred term 811 based on entity category, contextual collocation, target audience, and established project terminology specifications.

[0055] The target language sentence generation and word order adjustment module 820 generates the target translation 830 based on the source language ontology subgraph 631, the cross-language ontology structure mapping results, and the target language preferred terms 811. The target language sentence generation and word order adjustment module 820 can call preset sentence templates based on relation types, or input concept and relation tags into existing statistical machine translation, neural machine translation, or controlled generation models. The target language sentence generation and word order adjustment module 820 adjusts the order of subject, verb, object, modifiers, conditionals, negations, dosages, and coordinate components according to the target language grammar, while maintaining the ontology relation direction.

[0056] The target language sentence generation and word order adjustment module 820 preserves the alignment relationships between the target translation 830 and the source language entities, source language segments, and ontology nodes. These alignment relationships enable the directional verification unit 900 to map terms, segments, and relationships in the target translation 830 back to the source language entities, source language segments, source language ontology subgraph 631, and the source region in the traditional Chinese medicine ancient book page image 410.

[0057] The target language output unit 800 does not automatically translate based on the binding edge constraint column phase deviation fingerprint 520. The binding edge constraint column phase deviation fingerprint 520 affects the candidate mapping sorting 720 and the directional review sorting result 920. The target language output unit 800 still generates the target translation 830 based on the candidate mapping sorting 720, the target language preferred terms 811, and the target language grammar rules.

[0058] The targeted review unit 900 includes a consistency and targeted review module 910. The consistency and targeted review module 910 performs entity coverage, relation preservation, terminology consistency, numerical unit checks, negative information checks, and back-translation checks, resulting in a semantic conflict value 911. Semantic conflict value The value represents the dimensionless normalized semantic conflict result obtained by the consistency and directional review module 910. The larger the value, the stronger the semantic conflict.

[0059] The consistency and targeted review module 910 generates a targeted review risk value 912 based on the semantic conflict value 911 and the binding edge constraint column gate phase deviation fingerprint 520. Targeted review risk value This represents the risk value used to determine review priorities. The targeted review risk value of 912 is used to prioritize items with significant semantic conflicts and large deviations from the page binding structure at the top of the targeted review ranking results 920.

[0060] ; in, Indicates the risk value for targeted review. Indicates semantically conflicting values. This indicates that the binding edge constraint column grid phase deviates from the fingerprint. This represents the source node index in the source language ontology subgraph 631. This represents the target node index in the target language concept and terminology resource 641.

[0061] The consistency and targeted review module 910 generates a targeted review ranking result 920 from the original text page, the innermost valid text column image region, the original recognized text, the candidate mapping, and the target translation 830 according to the targeted review risk value 912, and stores the innermost valid text column image region, the original recognized text, the candidate mapping, and the target translation 830 as a review basis. This associated storage corresponds to... Figure 1 The targeted review sorting results 920 and audit record 930 within the targeted review unit 900.

[0062] The targeted review and sorting result 920 is used to trigger local re-identification, re-execution of cross-language ontology structure mapping, or record revision chain. Local re-identification is performed by the character recognition and variant character normalization module 600, which is used to reprocess the innermost valid text column image region. Re-execution of cross-language ontology structure mapping is performed by the cross-language ontology structure mapping module 710, which is used to re-form the candidate mapping sort 720. The record revision chain is written to the audit record 930 by the consistency and targeted review module 910, which is used to save the mapping before and after review, the target translation 830, the semantic conflict value 911, the binding edge constraint column grid phase deviation fingerprint 520, and the targeted review risk value 912.

[0063] The targeted review unit 900 does not have an automatic deletion threshold or an automatic translation threshold. The targeted review risk value 912 is only used for sorting and interpretation and does not directly replace the semantic check conclusion. Local re-identification, re-execution of cross-language ontology structure mapping, or record revision chain are only triggered when the review item actually has problems with entity coverage, relation preservation, terminology consistency, numerical units, negative information, or back-translation checks.

[0064] When the binding edge constraint column gate phase deviation fingerprint 520 is unavailable, the consistency and directed review module 910 generates a directed review ranking result 920 using only the semantic conflict value 911. In this case, the directed review unit 900 does not generate a directed review risk value 912 based on the binding edge constraint column gate phase deviation fingerprint 520, nor does it fill in the missing binding edge constraint column gate phase deviation fingerprint 520 with a manual value other than zero.

[0065] For example, in a complete workflow embodiment, communication interface 300 first receives a batch of scanned pages of traditional Chinese medicine books bound in thread, along with the book title, volume number, page number, edition, library accession number, and target translation language. Communication interface 300 writes the images 410 of the traditional Chinese medicine book pages and the document metadata 420 into a page input object, and then passes the page input object to processor 100.

[0066] Processor 100 invokes page structure feature generation unit 500. Page structure feature generation unit 500 performs grayscale conversion, noise reduction, binarization, contrast adjustment, tilt correction, edge cropping, text detection, annotation detection, page frame detection, and column line detection on the page image 410 of the ancient Chinese medicine book. Page structure analysis and fingerprint extraction module 510 determines the physical boundary direction of the binding edge based on the visible binding edge information 511, and determines the innermost valid text column 512, adjacent valid text columns 513, and valid character frames 514.

[0067] In this workflow embodiment, the page structure analysis and fingerprint extraction module 510 measures the page Marked binding margin The median of the center distance of the main text column Median height of character box All the above values ​​use the same page image coordinate unit, and the median distance between the center of the text columns and the median height of the character boxes are both greater than zero. The page structure analysis and fingerprint extraction module 510 calculates the original column grid phase values ​​of the page.

[0068] ; in, Indicates the original column grid phase value of the page. This indicates the marked binding edge distance in this workflow embodiment. This represents the median of the center distance of the text column in this workflow embodiment. This indicates the median height of the character box in this workflow embodiment.

[0069] If the intra-volume layout baseline of the same volume or the same set of page layouts Then the page structure analysis and fingerprint extraction module 510 calculates the binding edge constraint column grid phase deviation fingerprint 520.

[0070] ; in, This indicates that the binding edge constraint column grid phase deviates from the fingerprint. This represents the original column grid phase value of the page in this workflow embodiment. This indicates the in-volume layout baseline in this workflow embodiment.

[0071] Subsequently, the character recognition and variant character standardization module 600 recognizes the text on the page block by block, generating raw recognition results with character coordinates and confidence levels, and standardizes variant characters based on the variant character table, the simplified / traditional character conversion table, and the existing character shape comparison table for the project. The text reconstruction and word unit extraction module 610 reconstructs the main text according to the page, column number, and block order, and extracts sentence segments and word units.

[0072] The source language entity recognition and concept normalization module 620 uses a traditional Chinese medicine ontology and a thesaurus to identify drug names, prescription names, disease names, syndromes, symptoms, treatment methods, viscera, meridians, acupoints, and dosage expressions, and associates different expressions with normalized concepts in the ontology. The source language ontology subgraph construction module 630 uses normalized concepts as graph nodes and writes the main treatment relationship, composition relationship, compatibility relationship, meridian tropism relationship, causal relationship, or syndrome-treatment relationship into the source language ontology subgraph 631.

[0073] The cross-language ontology structure mapping module 710 reads the source language ontology subgraph 631 and the target language concept and term resources 641, performs target concept candidate query, category constraint, adjacency comparison, and basic mapping scoring processing to obtain the original mapping score 721. If a candidate mapping depends on the cross-column or cross-page text continuity relationship formed between the innermost valid text column 512 and the adjacent valid text column 513, and lacks independent semantic evidence on the same page, then the cross-language ontology structure mapping module 710 sets the structural continuity marker 723 to... And calculate the corrected mapping score 722 based on the binding edge constraint column grid phase deviation fingerprint 520.

[0074] The target language terminology selection module 810 selects the preferred target language terminology 811 based on the candidate mapping sorting 720. The target language sentence generation and word order adjustment module 820 generates the target translation 830 based on the source language ontology subgraph 631, the cross-language ontology structure mapping results, and the preferred target language terminology 811, and retains the alignment relationship between the target translation 830 and the source language entities, source language segments, and ontology nodes.

[0075] The consistency and targeted review module 910 performs entity coverage, relation preservation, terminology consistency, numerical units, negative information, and back-translation checks to obtain a semantic conflict value 911. If the semantic conflict value 911 corresponding to the same candidate mapping or target translation fragment 830 is high, and the binding edge constraint column phase deviation fingerprint 520 of the source page is large, then the consistency and targeted review module 910 calculates a higher targeted review risk value 912 and places the review item higher in the targeted review ranking results 920.

[0076] After the review is completed, the standardized output and audit archiving process outputs the target translation 830, the targeted review sorting result 920, and the audit record 930. The audit record 930 saves the processing time, resource version, model version, mapping before and after review, target translation 830, semantic conflict value 911, binding edge constraint column phase deviation fingerprint 520, and targeted review risk value 912.

[0077] As an optional embodiment, the page image 410 of the ancient Chinese medicine book has clearly visible binding edge information 511, at least two valid text columns, and sufficient valid character boxes 514. The page structure analysis and fingerprint extraction module 510 can stably obtain the marked binding edge distance, the median center distance of the text columns, and the median height of the character boxes. The original column grid phase value of the page is close to the in-volume layout baseline, and the absolute value of the binding edge constrained column grid phase deviation from the fingerprint 520 is small. In this case, even if the candidate mapping depends on the continuity of text across columns or pages, the change of the corrected mapping score 722 relative to the original mapping score 721 is small, and the candidate mapping ranking 720 is mainly determined by the original mapping score 721.

[0078] As an optional embodiment, the binding side of the page image 410 of the ancient Chinese medicine book has a spine shadow, partial cropping, or inner column compression, causing the innermost valid text column 512 to deviate significantly from the set of pages in the same volume or format. The page structure analysis and fingerprint extraction module 510 can still determine the physical boundary direction of the binding edge, the innermost valid text column 512, the adjacent valid text columns 513, and the valid character frames 514, and calculate the binding edge constraint column grid phase deviation fingerprint 520. If a candidate mapping depends on the cross-column or cross-page text continuity relationship formed between the innermost valid text column 512 and the adjacent valid text columns 513, and lacks independent semantic evidence on the same page, then the structural continuity marker 723 is... The corrected mapping score 722 is lower than the original mapping score 721. This candidate mapping is relatively lower in the candidate mapping ranking 720, thereby reducing cross-language ontology structure mapping errors caused by misjudgment of binding-side text continuity.

[0079] As an optional embodiment, the binding edge of the page image 410 of the ancient Chinese medicine book is occluded, and the page structure analysis and fingerprint extraction module 510 cannot reliably determine the physical boundary direction of the binding edge. The page structure feature generation unit 500 marks the binding edge constraint column grid phase deviation fingerprint 520 as unusable. The ontology structure mapping unit 700 keeps the candidate mapping sort 720 from being adjusted by the binding edge constraint column grid phase deviation fingerprint 520, and corrects the mapping score 722 to be equal to the original mapping score 721. The directional verification unit 900 generates the directional verification sorting result 920 using only the semantic conflict value 911.

[0080] As an optional embodiment, the effective text columns in the page image 410 of the ancient Chinese medicine book are less than two. The page structure analysis and fingerprint extraction module 510 cannot calculate the median of the center distance of the text columns, and therefore does not generate the binding edge constraint column phase deviation fingerprint 520. The system continues to perform character recognition, variant character normalization, text reconstruction, source language entity recognition, source language ontology subgraph construction, cross-language ontology structure mapping, target language output, and consistency and orientation verification, but does not use page structure evidence for candidate mapping ranking 720 and orientation verification ranking results 920.

[0081] As an optional embodiment, the number of valid character frames 514 in the page image 410 of the ancient Chinese medicine book is insufficient. The page structure analysis and fingerprint extraction module 510 cannot calculate the median height of the character frames, and therefore cannot perform scale normalization using the median height of the character frames. The page structure feature generation unit 500 marks the binding edge constraint column grid phase deviation fingerprint 520 as unavailable to avoid substituting the median height of the character frames with manually set values.

[0082] As an optional embodiment, the candidate mapping itself suffers from incomplete entity coverage, inconsistent relationship maintenance, terminology consistency deviation, numerical unit errors, loss of negative information, or inconsistent back-translation results. The consistency and targeted review module 910 obtains a high semantic conflict value 911. If the binding edge constraint column phase deviation fingerprint 520 of the candidate mapping source page is also large, the targeted review risk value 912 is high, and the corresponding review item enters the front of the targeted review sorting results 920. Reviewers or review processes can prioritize viewing the original text page, the innermost valid text column image area, the original recognized text, the candidate mapping, and the target translation 830.

[0083] The processor 100 can be a general-purpose central processing unit, a graphics processing unit, a neural network processor, a server processor, or other processing unit capable of executing computer programs. The memory 200 can be volatile memory, non-volatile memory, a disk, a solid-state drive, or network storage. The communication interface 300 can be a network interface, a file interface, a database interface, or other data interface for receiving images 410 of pages from ancient Chinese medical texts and metadata 420 of the documents.

[0084] The memory 200 may store a computer program for executing this embodiment. When the computer program is executed by the processor 100, it implements the following steps: receiving a page image 410 of a traditional Chinese medicine book and document metadata 420; performing layout structure analysis on the page image and generating a binding edge constraint column grating phase deviation fingerprint 520; performing character recognition, variant character standardization, text reconstruction, and word unit extraction on the page image; generating source language entities and source language ontology subgraphs 631 based on the traditional Chinese medicine ontology; forming a candidate mapping ranking 720 based on the source language ontology subgraph 631 and target language concepts and terminology resources 641; selecting the preferred terminology 811 of the target language based on the candidate mapping ranking 720 and generating the target translation 830; generating a targeted review risk value 912 by combining the binding edge constraint column grating phase deviation fingerprint 520 and semantic conflict value 911; generating a targeted review ranking result 920 based on the targeted review risk value 912 and outputting an audit record 930.

[0085] Computer programs can be deployed as standalone programs, batch programs, server programs, plug-in pipeline programs, or functional components in a collection digitization platform.

[0086] The page images 410 of ancient Chinese medicine books can be saved in common formats such as TIFF, PNG, JPEG, or PDF image pages. Document metadata 420 may include the book title, volume number, page number, edition, library accession number, and target language. The system simply marks missing document metadata 420 as empty without performing speculative completion. The above data formats and metadata management methods are standard practices in digital document management.

[0087] In this embodiment, the OCR model, sentence segmentation model, word segmentation model, named entity recognition model, ontology storage method, and machine translation model can all adopt existing models or existing data structures.

[0088] The embodiments of this example have been described above. However, this example is not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A cross-language standardized translation system for ancient Chinese medical texts based on ontology structure mapping, comprising a page structure feature generation unit, an ontology structure mapping unit, a target language output unit, and a targeted verification unit, characterized in that: The page structure feature generation unit generates a binding edge constraint column grid phase deviation fingerprint based on the existing binding edge, innermost valid text column, adjacent valid text column and valid character box in the page image of ancient Chinese medicine books. This fingerprint is used to characterize the degree of deviation of the binding side column grid of a single page relative to the set of pages in the same volume or the same format. The ontology structure mapping unit forms a candidate mapping ranking based on the source language ontology subgraph and the target language concept and term resources, and identifies whether the candidate mapping depends on cross-column or cross-page text continuity. The independent semantic evidence on the same page is the semantic evidence provided by the source language sentence segment, source language entity or source language ontology subgraph within the same page that supports the establishment of the candidate mapping. When the candidate mapping depends on cross-column or cross-page text continuity and lacks independent semantic evidence on the same page, the candidate mapping is relatively reduced in the candidate mapping ranking, and the original mapping score and the corrected mapping score are retained. The target language output unit selects the preferred terminology of the target language according to the candidate mapping sort and generates the target translation; The targeted review unit obtains semantic conflict values, combines them with binding edge constraint column grid phase deviation fingerprints to generate targeted review risk values, and generates targeted review ranking results based on the targeted review risk values.

2. The cross-language standardized translation system for ancient Chinese medical texts based on ontology structure mapping according to claim 1, characterized in that, The page structure feature generation unit includes a page structure analysis and fingerprint extraction module. The page structure analysis and fingerprint extraction module determines the physical boundary direction of the binding edge in the page image, determines the center line of the innermost effective text column closest to the binding edge, and obtains the distance of the bound edge with symbols, the median distance of the center line of the text column, and the median height of the character frame based on at least one binding edge visibility information among the page edge, spine shadow, binding hole, and blank band, as well as the detection results of the center line of the text column and the character frame.

3. The cross-language standardized translation system for ancient Chinese medicine texts based on ontology structure mapping according to claim 2, characterized in that, The page structure analysis and fingerprint extraction module generates the original column grid phase value of the page based on the distance of the bound edge with symbols, the median distance between the center of the text columns, and the median height of the character box. It also performs median statistics on the original column grid phase values ​​of the pages in the same volume or the same layout page set to form the layout baseline within the volume. The fingerprint characterizes the deviation of the original column grid phase value of the page from the bound edge constraint column grid relative to the layout baseline within the volume.

4. The cross-language standardized translation system for ancient Chinese medical texts based on ontology structure mapping according to claim 1, characterized in that, When the binding edge cannot be determined, there are fewer than two valid text columns, or the number of valid character boxes is insufficient, the page structure feature generation unit marks the binding edge constraint column phase deviation fingerprint as unavailable, and ensures that the ontology structure mapping unit keeps the candidate mapping sorting from being adjusted by the binding edge constraint column phase deviation fingerprint, so that the directional review unit only uses semantic conflict values ​​to generate directional review sorting results.

5. The cross-language standardized translation system for ancient Chinese medical texts based on ontology structure mapping according to claim 1, characterized in that, The ontology structure mapping unit includes a cross-language ontology structure mapping module. This module performs target concept candidate query, category constraint, adjacency comparison, and basic mapping score processing to obtain the original mapping score. It sets a structural continuity marker based on the cross-column or cross-page text continuity relationship formed between the innermost valid text column and adjacent columns or adjacent pages. When the binding edge constraint column phase deviation fingerprint is available and the structural continuity marker indicates a dependency, the position of the corresponding candidate mapping in the candidate mapping ranking is reduced, while retaining the original mapping score, structural continuity marker, and corrected mapping score.

6. The cross-language standardized translation system for ancient Chinese medical texts based on ontology structure mapping according to claim 1, characterized in that, The targeted review unit includes a consistency and targeted review module, which performs entity coverage, relation preservation, terminology consistency, number units, negation information, and back-translation checks to obtain semantic conflict values. Based on semantic conflict values ​​and binding edge constraint column phase deviation fingerprints, a targeted review risk value is generated. This results in a targeted review ranking of the original text page, the innermost valid text column image region, the original recognized text, the candidate mapping, and the target translation according to the targeted review risk value. The innermost valid text column image region, the original recognized text, the candidate mapping, and the target translation are associated and stored as the review basis, which is used to trigger local re-recognition, re-execution of cross-language ontology structure mapping, or record revision chain.

7. The cross-language standardized translation system for ancient Chinese medical texts based on ontology structure mapping according to claim 1, characterized in that, The system also includes a source language entity recognition and concept normalization module and a source language ontology subgraph construction module. The source language entity recognition and concept normalization module uses a traditional Chinese medicine ontology and a thesaurus to identify traditional Chinese medicine entities in the source text and associates different expressions with normalized concepts in the ontology. The source language ontology subgraph construction module uses normalized concepts as graph nodes and writes the main treatment relationship, composition relationship, compatibility relationship, meridian tropism relationship, causal relationship or syndrome-treatment relationship into the source language ontology subgraph.

8. The cross-language standardized translation system for ancient Chinese medical texts based on ontology structure mapping according to claim 1, characterized in that, The target language output unit includes a target language terminology selection module and a target language sentence generation and word order adjustment module. The target language terminology selection module selects preferred target language terms from target language concepts and terminology resources based on cross-language ontology structure mapping results and establishes a document-level terminology consistency table. The target language sentence generation and word order adjustment module generates the target translation based on the source language ontology subgraph, cross-language ontology structure mapping results, and preferred target language terms, and retains the alignment relationship between the target translation and source language entities, source language sentence segments, and ontology nodes.

9. A method for cross-linguistic standardized translation of ancient Chinese medical texts based on ontology structure mapping, characterized in that: include: Receive page images and metadata of ancient Chinese medical books; Perform layout structure analysis on the page image and generate a binding edge constraint column grid phase deviation fingerprint; Perform character recognition, variant character standardization, text reconstruction, and word unit extraction on page images; Source language entities and source language ontology subgraphs are generated based on the TCM ontology; Candidate mappings are ranked based on source language ontology subgraphs and target language concept and term resources. When a candidate mapping depends on the continuity of text across columns or pages and lacks semantic evidence supporting the validity of the candidate mapping provided by source language segments, source language entities or source language ontology subgraphs within the same page, the candidate mapping is relatively reduced in the candidate mapping ranking. Select the preferred terms for the target language based on the candidate mapping sort and generate the target translation; It also combines the binding edge constraint column phase deviation fingerprint and semantic conflict value to generate a targeted review risk value, and generates a targeted review ranking result based on the targeted review risk value and outputs the audit record.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the cross-language standardized translation method for ancient Chinese medicine books based on ontology structure mapping as described in claim 9.

Images