Intelligent document translation correction system

Abstract

The invention relates to the technical field of semantic analysis, in particular to an intelligent document translation correction system which comprises the steps that a source file containing mixed layout content and an initial translation are obtained; constructing a document semantic division model to analyze the source file to obtain source file paragraph data; dividing the initial translation through a document semantic division model to generate translation natural division data; taking the division boundary of the source file paragraph data as a reference, mapping to the initial translation, and generating compulsory division data of the translation; the division structure similarity of the translation natural division data and the translation forced division data is calculated and serves as a macroscopic evaluation parameter; calculating the multi-dimensional feature similarity of corresponding paragraphs in the source file paragraph data and the compulsory translation division data as microscopic verification parameters; according to the method, the initial translation is evaluated based on the macroscopic evaluation parameters and the microscopic verification parameters.

Description

Technical Field

[0001] This invention relates to the field of semantic analysis technology, specifically to an intelligent document translation and correction system. Background Technology

[0002] With the increasing frequency of international trade and technological exchanges, the demand for cross-language conversion of professional technical documents has exploded, especially in high-tech fields such as electromagnetic compatibility and electronic and electrical testing and certification. Various test reports and certification certificates are not only passports for products to enter the international market, but their accuracy also directly constrains product compliance through legal force.

[0003] In existing technologies, document translation mainly relies on neural machine translation models, which achieve automatic conversion from source language to target language through large-scale bilingual corpus training. Although general translation engines have made significant progress in fluency, they still face serious challenges when processing professional technical documents with complex mixed layouts, that is, text and scientific charts coexisting.

[0004] Existing document translation systems primarily rely on plain text context for sequence generation, lacking cross-modal data verification capabilities. Specifically, when processing electromagnetic compatibility (EMC) test reports, conclusions often depend on waveform data from spectrum analysis charts, such as frequency peaks and limit margins; the visual information within the charts is not readily apparent, easily leading to numerical illusions. Furthermore, existing technologies tend to lose implicit logical connections between paragraphs in long document translation, resulting in translated texts that, while grammatically correct, exhibit a loose or even collapsed logical structure, failing to meet the rigorous causal logic requirements of professional reports.

[0005] Therefore, how to establish a cross-modal deep verification mechanism that can accurately monitor key values ​​in the translation by using objective data from charts in the document while ensuring the fluency of the translation, and ensure that the translation strictly preserves the physical structure and logical flow of the source file, is a key technical problem that urgently needs to be solved in the field of intelligent document translation technology.

[0006] To address this, an intelligent document translation and correction system is proposed. Summary of the Invention

[0007] The purpose of this invention is to provide an intelligent document translation correction system that evaluates translations using macroscopic evaluation parameters and microscopic verification parameters.

[0008] To achieve the above objectives, the present invention provides the following technical solution: An intelligent document translation and correction system includes: The data acquisition module acquires the source file containing mixed layout content and the initial translation. The mixed layout content includes text paragraphs and spectrum analysis charts. The source file analysis module constructs a document semantic segmentation model to parse the source file and obtain source file paragraph data. The document semantic segmentation model extracts multi-dimensional feature vectors for each sentence, including spectral consistency features, technical term density features, and logical coherence features. Based on the multi-dimensional feature vectors, semantic clustering is performed to divide the source file into multiple source file paragraphs with independent semantic functions. The translation analysis module divides the initial translation using a document semantic segmentation model, generating naturally segmented translation data; it also maps the segmentation boundaries of the source file paragraph data to the initial translation, generating forced segmentation data for the translation. The anomaly detection module calculates the structural similarity between the naturally segmented data and the forced segmented data of the translated text as a macro-evaluation parameter; it also calculates the multi-dimensional feature similarity between the corresponding paragraphs in the source file paragraph data and the forced segmented data of the translated text as a micro-verification parameter; and evaluates the initial translation based on the macro-evaluation parameter and the micro-verification parameter.

[0009] The specific steps for obtaining the spectrum consistency feature include: Computer vision technology is used to analyze the layout of the source file, identify and locate the area where the spectrum analysis chart is located, and the chart is separated into a background grid layer, a limit curve layer and a measured waveform layer by an image segmentation algorithm. A coordinate mapping transformation is performed on the measured waveform layer to identify the frequency range represented by the horizontal axis and the amplitude range represented by the vertical axis. All key waveform peak points exceeding the limit curve on the waveform curve are extracted by the pixel scanning algorithm to construct an image waveform dataset containing frequency and amplitude values. Named entity recognition technology is used to scan the text of the source file sentences in the context of the spectrum analysis chart and extract all text numerical entities describing the test results. The text numerical entities include frequency values, level values ​​and unit symbols. The text numerical entities are matched with the image waveform dataset using Euclidean distance. If the numerical value in the text can find the corresponding peak point in the image waveform dataset within a preset error range, its confidence score is calculated. The average confidence score of all numerical entities in the sentence is defined as the spectral consistency feature value of the sentence.

[0010] The specific steps for obtaining the density characteristics of the technical terms include: Construct a professional terminology database for electromagnetic compatibility, which covers proper nouns, abbreviations, and units of measurement in the fields of electromagnetic interference, electromagnetic susceptibility, radio frequency common-mode conduction, and electrostatic discharge. Each sentence in the source file is segmented into words, and the segmentation results are matched against the terminology database. The word frequency data of the professional terms in each sentence are counted, and the basic terminology density ratio is calculated by combining the total number of words in the sentence. A syntactic component weighting mechanism is introduced to identify the position of terms in sentence components, assign weights according to their positions, and calculate the weighted values ​​as the professional term density features by combining the basic term density ratio.

[0011] The specific steps in obtaining the logical coherence feature include: Each sentence in the source file is mapped to a high-dimensional semantic vector. The cosine similarity between the current sentence vector and the vectors of the preceding and subsequent sentences is calculated to obtain the semantic smoothness index. Explicit logical connection markers in sentences are identified, and a logical topology network representing the pre-defined relationships between sentences is constructed. The logical topology network defines the theoretical trend of semantic change that adjacent sentences should exhibit, including at least semantic continuity and semantic transition trends. The logical topology network is used as a constraint to verify the consistency of the semantic smoothness index and generate logical coherence features.

[0012] The specific steps for dividing a source file into multiple sections include: Construct a high-dimensional feature space that includes the dimensions of spectral consistency, technical terminology density, and logical coherence. A density-based clustering algorithm is used to cluster sentence vectors in the source file. Spatial proximity constraints are introduced during the clustering process. The distance metric between sentences is defined to include not only the semantic distance in the feature vector space, but also the Manhattan distance of the sentence in the physical layout of the document, so as to achieve semantic coherence and physical continuity of the clustering results. For sentences with high spectral consistency, assign them spectral weights to cluster towards the corresponding chart position, force them to be grouped into an independent technical semantic block with the corresponding chart, and divide the paragraphs based on the independent technical semantic block; The output consists of multiple source file paragraphs, each labeled with a specific attribute tag. These attribute tags include at least the data analysis paragraph, the test method paragraph, the conclusion judgment paragraph, and the administrative information paragraph.

[0013] The steps for obtaining the macroscopic evaluation parameters include: Extract the semantic centroid vector of each paragraph in the naturally segmented data of the translated text to form the first vector sequence; extract the semantic centroid vector of each forced paragraph in the forced segmented data of the translated text to form the second vector sequence; The minimum alignment path distance between the first vector sequence and the second vector sequence is calculated using a dynamic time warping algorithm. This distance represents the degree of deviation between the translation and the physical structure of the source file in the natural semantic flow state. The overlap ratio of two data points at key semantic breakpoints is statistically analyzed; a key semantic breakpoint is defined as the sentence boundary where the paragraph attribute label changes; by combining the minimum alignment path distance and the overlap ratio of key semantic breakpoints, a structural similarity score is calculated and generated through normalized weighting, and this score is used as a macro-evaluation parameter.

[0014] The steps for obtaining the microscopic verification parameters include: For each source file paragraph and its corresponding translation, a forced paragraph division is performed, and data quantification and comparison are conducted. Numerical deviation calculation is performed based on the spectral consistency characteristics; digital entities are extracted from the translated paragraphs and their differences are calculated with the corresponding frequency points and amplitude points in the source file image waveform dataset. The absolute error value is calculated and normalized to obtain the numerical deviation index. For the density characteristics of technical terms, perform term mapping counting; based on the technical term database, count the cardinality of term sets in the source document paragraphs, and retrieve the corresponding target language term matching number in the translation paragraphs, calculate the mapping ratio from the source term set to the translation term set, and obtain term loss rate data; For logical coherence features, perform semantic projection analysis; map the paragraph vectors of the source document to the shared semantic space, and calculate the Euclidean distance between the source document and the forced paragraph vectors of the translation in the same space, as the semantic offset; The numerical deviation index, term loss rate, and semantic offset are constructed into three-dimensional feature coordinates. The spatial distance from these coordinates to the ideal zero error point is calculated, and the reciprocal of the spatial distance is used as a micro-verification parameter to measure the quality of paragraph translation.

[0015] The specific steps for evaluating the initial translation based on macro-level evaluation parameters and micro-level verification parameters include: Construct a tiered evaluation system and set macro-level blocking thresholds and micro-level early warning thresholds; First, determine whether the macro-assessment parameters are below the macro-blocking threshold. If they are below the threshold, it is determined that the initial translation has a structural collapse, the initial translation is marked as unusable, and the entire document is re-translated. If the macro-evaluation parameters are qualified, the micro-verification parameters of each paragraph in the source file paragraph data are traversed; for the initial translation paragraphs whose micro-verification parameters are lower than the micro-warning threshold, correction instructions are generated according to the difference type.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. This invention introduces computer vision and signal processing technologies to construct a cross-modal data verification mechanism for electromagnetic compatibility (EMC) test reports. Through image segmentation and Kalman filtering restoration techniques, it can reconstruct obscured waveform data from unstructured spectrum analysis charts with high fidelity. Combined with a logarithmic regression model to calibrate the coordinate axes, a physical data benchmark is established. By matching the numerical entities extracted from the text with the waveform data reconstructed from the image using Euclidean distance, it achieves supervision through visual verification of the text, ensuring the physical accuracy of the core conclusion data in the test report and improving the credibility of professional technical document translations.

[0017] 2. This invention addresses the common structural collapse and logical breakage issues in long document translation by introducing a dynamic time warping algorithm and logical transition matrix mapping technology. At the macro level, by comparing the topological differences between natural and forced division of the translated text, it quantifies the degree of distortion of the original text's physical structure during the translation process, enabling it to keenly detect major quality issues such as missing chapters or misordered paragraphs. At the micro level, by utilizing a pre-trained logical transition matrix, it can identify common paraphrased sentence structures in Chinese reports and accurately assess whether the translated text has undergone reasonable explicit supplementation. This ensures the rigor of the technical verification process in terms of logical flow, resulting in more accurate machine translation results.

[0018] 3. This invention proposes a multi-dimensional evaluation system that integrates semantic space and physical layout space. By introducing a clustering algorithm with Manhattan distance and spectral weight constraints, it can accurately identify and bind charts and their descriptive text, preventing context loss caused by the separation of charts and captions. At the same time, by combining the weighted term density features of dependency parsing, it can distinguish between core technical parameters and secondary modifiers, thus giving higher attention to core terms in quality scoring. This greatly reduces the workload of manual proofreading and provides a high-precision correction method for document translation in high-precision fields such as electromagnetic compatibility. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the structure of an intelligent document translation and correction system according to the present invention; Figure 2 This is a logical schematic diagram of an intelligent document translation and correction system according to the present invention; Figure 3 This is a schematic diagram of the document semantic segmentation model of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1: This invention proposes an intelligent document translation correction system, the structure of which is as follows: Figure 1 As shown, the logic of the system is as follows: Figure 2 As shown, it includes: The data acquisition module acquires the source file containing mixed layout content and the initial translation. The mixed layout content includes text paragraphs and spectrum analysis charts. The source file analysis module constructs a document semantic segmentation model to parse the source file and obtain source file paragraph data. The document semantic segmentation model extracts multi-dimensional feature vectors for each sentence, including spectral consistency features, technical term density features, and logical coherence features. Based on the multi-dimensional feature vectors, semantic clustering is performed to divide the source file into multiple source file paragraphs with independent semantic functions. The translation analysis module divides the initial translation using a document semantic segmentation model, generating naturally segmented translation data; it also maps the segmentation boundaries of the source file paragraph data to the initial translation, generating forced segmentation data for the translation. The anomaly detection module calculates the structural similarity between the naturally segmented data and the forced segmented data of the translated text as a macro-evaluation parameter; it also calculates the multi-dimensional feature similarity between the corresponding paragraphs in the source file paragraph data and the forced segmented data of the translated text as a micro-verification parameter; and evaluates the initial translation based on the macro-evaluation parameter and the micro-verification parameter.

[0022] The structure of the document semantic segmentation model is as follows: Figure 3 As shown, the steps for obtaining spectral consistency features include: Computer vision technology is used to analyze the layout of the source file, identify and locate the area where the spectrum analysis chart is located, and the chart is separated into a background grid layer, a limit curve layer and a measured waveform layer by an image segmentation algorithm. A coordinate mapping transformation is performed on the measured waveform layer to identify the frequency range represented by the horizontal axis and the amplitude range represented by the vertical axis. All key waveform peak points exceeding the limit curve on the waveform curve are extracted by the pixel scanning algorithm to construct an image waveform dataset containing frequency and amplitude values. Named entity recognition technology is used to scan the text of the source file sentences in the context of the spectrum analysis chart and extract all text numerical entities describing the test results. The text numerical entities include frequency values, level values ​​and unit symbols. The text numerical entities are matched with the image waveform dataset using Euclidean distance. If the numerical value in the text can find the corresponding peak point in the image waveform dataset within a preset error range, its confidence score is calculated. The average confidence score of all numerical entities in the sentence is defined as the spectral consistency feature value of the sentence.

[0023] Page layout analysis and layer separation: First, computer vision algorithms are used to scan the source file to identify and locate the areas where the spectrum analysis charts are located; for the located charts, image segmentation algorithms are used to separate the layers.

[0024] Data reconstruction and peak extraction: For the separated measured waveform layer, coordinate mapping transformation is performed. First, the frequency range represented by the horizontal axis and the amplitude range represented by the vertical axis are identified, and the mapping relationship between pixel coordinates and physical values ​​is established. Then, through a pixel scanning algorithm, the waveform curves are traversed to extract all key waveform peak points that exceed or are close to the limit curves, constructing an image waveform dataset containing frequency and amplitude values.

[0025] The coordinate mapping transformation for the measured waveform layer specifically includes: extracting the scale numerical text on the horizontal and vertical coordinate axes using optical character recognition; constructing a coordinate axis scale regression model to detect the pixel distance distribution of the scale numerical text on the image; if the horizontal coordinate axis is detected to be a logarithmic coordinate distribution, then using a logarithmic regression function to fit the mapping relationship between pixel coordinates and frequency values; if a scale value recognized by OCR is detected to deviate from the regression curve by more than a preset threshold, then the recognition result is determined to be a misrecognition, and the predicted value generated by the regression function is used to correct the scale value, thereby establishing a high-precision coordinate mapping matrix; and the correction content is marked and prompted.

[0026] The horizontal axis of an electromagnetic compatibility spectrum diagram, i.e., the frequency axis, is usually distributed on a logarithmic scale. That is, the physical distance from one megahertz to ten megahertz is equal to the physical distance from ten megahertz to one hundred megahertz. Optical character recognition technology is prone to misidentifying ambiguous numbers, such as misidentifying one hundred megahertz as seven hundred megahertz.

[0027] Tick ​​point extraction and pre-screening; identify all text blocks below the horizontal axis, extract the numerical content and their corresponding geometric center horizontal coordinate pixel values, and assemble these data into the original dataset; The coordinate distribution model makes the following assumptions: First, it is assumed that the coordinate axes follow a linear distribution, and the correlation coefficient is calculated. Then, it is assumed that they follow a logarithmic distribution, and the correlation coefficient is calculated again. In the electromagnetic compatibility spectrum diagram, it is usually found that the pixel distance is proportional to the logarithm of the value, thus confirming the use of a logarithmic regression model for calculation. Regression fitting and outlier removal: Logarithmic function fitting is performed using the random sampling consensus algorithm. This algorithm randomly selects a portion of data points to construct a fitting curve and calculates the distance from the remaining data points to the fitting curve. During this process, it may be found that a certain identification result, such as the pixel corresponding to the value 700, deviates greatly from the fitting curve. This point is marked as an outlier or anomaly and removed from the fitting dataset to ensure the accuracy of the model. Forced correction and remapping: Using the best-fit parameters obtained after removing outliers, the theoretical frequency value corresponding to the position of the outlier pixel is calculated in reverse, ensuring that even if character recognition is wrong, the frequency data extracted from the waveform is still accurate.

[0028] This technical solution addresses the bottleneck of unstable recognition rates in OCR technology on low-quality charts. EMC spectrum diagrams, or electromagnetic compatibility spectrum diagrams, typically use logarithmic coordinate axes, a mathematical principle that represents strong prior knowledge. By utilizing a mathematical regression model to judge OCR recognition results, character recognition errors can be automatically corrected, ensuring that the frequency values ​​(such as peak frequencies) extracted from the waveform are physically accurate and preventing all data from being incorrect due to coordinate axis calibration errors.

[0029] Text numerical extraction: Using named entity recognition technology, the text of the sentences in the source file surrounding the spectrum analysis chart is scanned; the focus is on extracting text numerical entities that describe the test results, specifically including frequency values, level values, and related physical unit symbols.

[0030] Cross-modal matching calculation: Numerical entities extracted from the text are compared with waveform datasets extracted from the image. The Euclidean distance between the two is calculated. If the numerical value in the text can find a corresponding peak point in the image waveform dataset within a preset error range, the match is considered successful and a confidence score is calculated. The average confidence score of all numerical entities in the sentence is defined as the spectral consistency feature value of the sentence.

[0031] In this embodiment, the hybrid layout content is an electromagnetic compatibility test report page. The page adopts a single-column layout or a regular multi-column layout. Each text block and spectrum analysis chart on the page can be marked with a rectangular bounding box and corresponding page coordinates through the layout analysis algorithm. The Manhattan distance is calculated based on the above-mentioned page coordinate system.

[0032] This invention utilizes computer vision technology to reverse engineer unstructured spectrograms, reconstructing the underlying physical data. In the field of detection, report conclusions often rely on key peaks in graphs, and machine translation is prone to numerical illusions, such as missing decimal points or unit errors. This invention constructs a strong verification logic that verifies both text and images, using real data from the images to supervise text translation, ensuring zero tampering and zero mistranslation of core test data.

[0033] The specific steps for obtaining the density characteristics of the technical terms include: Construct a professional terminology database for electromagnetic compatibility, which covers proper nouns, abbreviations, and units of measurement in the fields of electromagnetic interference, electromagnetic susceptibility, radio frequency common-mode conduction, and electrostatic discharge. Each sentence in the source file is segmented into words, and the segmentation results are matched against the terminology database. The word frequency data of the professional terms in each sentence are counted, and the basic terminology density ratio is calculated by combining the total number of words in the sentence. A syntactic component weighting mechanism is introduced to identify the position of terms in sentence components, assign weights according to their positions, and calculate the weighted values ​​as the professional term density features by combining the basic term density ratio.

[0034] Build a professional terminology database: Build a professional terminology database covering the field of electromagnetic compatibility in advance. This database integrates terms from international and national standards, covering proper nouns, abbreviations, and units of measurement in sub-fields such as electromagnetic interference, electromagnetic susceptibility, radio frequency common-mode conduction, and electrostatic discharge.

[0035] Word segmentation and basic density statistics: Each sentence in the source file is segmented into words, and the segmentation results are matched against a professional terminology database. The number of words that match professional terms in each sentence is counted, and the basic terminology density ratio is calculated by combining the total number of words in the sentence.

[0036] Syntactic Component Weighting: To more accurately measure the importance of terms, a syntactic component weighting mechanism is introduced. Sentence structure is analyzed to identify the position of terms within sentence components. Terms in core positions (such as subject or object) are assigned higher weights; terms in modifying positions (such as attributive or adverbial modifiers) are assigned lower weights. The weighted value, calculated by combining the basic term density ratio and positional weights, serves as the final term density feature.

[0037] This invention constructs a weighted term density feature, which not only identifies the presence of terms but also their position in sentences. This enables the system to accurately distinguish between high-density technical paragraphs and irrelevant administrative information. By focusing on terms in core components, the accuracy of core technical concepts can be prioritized in subsequent translation quality checks, preventing the mistranslation of technical terms into general terms and ensuring accurate evaluation of the translation.

[0038] The specific steps in obtaining the logical coherence feature include: Each sentence in the source file is mapped to a high-dimensional semantic vector. The cosine similarity between the current sentence vector and the vectors of the preceding and subsequent sentences is calculated to obtain the semantic smoothness index. Explicit logical connection markers in sentences are identified, and a logical topology network representing the pre-defined relationships between sentences is constructed. The logical topology network defines the theoretical trend of semantic change that adjacent sentences should exhibit, including at least semantic continuity and semantic transition trends. The logical topology network is used as a constraint to verify the consistency of the semantic smoothness index and generate logical coherence features.

[0039] Semantic smoothness index calculation: Map each sentence in the source file to a high-dimensional semantic vector; then, calculate the cosine similarity between the current sentence vector and the vectors of the preceding and following sentences. This similarity reflects the degree of content proximity between adjacent sentences, i.e., the semantic smoothness index. Logical topology network construction: Identify explicit logical connection markers in sentences, such as conjunctions like "but," "therefore," and "and"; based on these markers, construct a logical topology network that represents the pre-defined relationships between sentences. This network defines the theoretical trend of semantic change that adjacent sentences should exhibit; for example, transition words indicate that semantics should change abruptly, while sequential words indicate that semantics should remain smooth.

[0040] Consistency verification and feature generation: The logical topology network is used as a constraint to verify the consistency of the semantic smoothness index; it is determined whether the actual content change conforms to the logical rules specified by the connector; if the two match (for example, a transition word is used and the content does indeed change), a high-weight logical coherence feature is generated; otherwise, a low-weight feature is generated.

[0041] This invention proposes a method based on mutual verification between explicit logical identifiers and implicit semantic vectors. Machine translation often produces sentences that are fluent but illogical. By introducing a logical topology network, it is possible to identify logically flawed sentences that use conjunctions in form but are completely unrelated in content. This is particularly important for the detection report, ensuring that the conclusions in the report maintain a rigorous causal logic with the data analysis.

[0042] The specific steps for dividing a source file into multiple sections include: Construct a high-dimensional feature space that includes the dimensions of spectral consistency, technical terminology density, and logical coherence. A density-based clustering algorithm is used to cluster sentence vectors in the source file. Spatial proximity constraints are introduced during the clustering process. The distance metric between sentences is defined to include not only the semantic distance in the feature vector space, but also the Manhattan distance of the sentence in the physical layout of the document, so as to achieve semantic coherence and physical continuity of the clustering results. For sentences with high spectral consistency, assign them spectral weights to cluster towards the corresponding chart position, force them to be grouped into an independent technical semantic block with the corresponding chart, and divide the paragraphs based on the independent technical semantic block; The output consists of multiple source file paragraphs, each labeled with a specific attribute tag. These attribute tags include at least the data analysis paragraph, the test method paragraph, the conclusion judgment paragraph, and the administrative information paragraph.

[0043] Constructing a high-dimensional feature space: Construct a high-dimensional feature space containing spectral consistency feature dimension, technical term density feature dimension, and logical coherence feature dimension, and map each sentence to a point in the space; Spatiotemporal constrained clustering: A density-based clustering algorithm is used to process sentence vectors; spatial proximity constraints are introduced during the clustering process; when defining the distance metric between sentences, not only the semantic distance in the feature vector space is considered, but also the Manhattan distance of the sentences in the physical layout of the document is taken into account; only sentences that are semantically similar and physically adjacent will be clustered together.

[0044] Spectrum Weight Gravity Model: For sentences with high spectral consistency features (i.e., sentences describing charts), assign them additional spectral weights to cluster towards the corresponding chart positions, forcing these descriptive sentences and their corresponding charts to be grouped into an independent technical semantic block, preventing text-image separation; Paragraph attribute tags: Paragraphs in the multiple output source files are tagged with specific attribute tags, including data analysis paragraphs, test method paragraphs, conclusion judgment paragraphs, and administrative information paragraphs.

[0045] To achieve dual constraints of semantics and physical layout, the specific calculation of the distance metric and the implementation of forced clustering are as follows: Calculation of the comprehensive distance metric: The comprehensive distance between sentences is defined as consisting of two parts: semantic distance and physical distance; Semantic distance: The cosine distance between the semantic vectors of two sentences is used directly, and its value is usually between 0 and 1; Physical distance: Calculate the Manhattan distance between the center points of two sentences on the document page (i.e., the sum of the absolute values ​​of the difference in the horizontal coordinates and the absolute values ​​of the difference in the vertical coordinates); to eliminate the difference in dimensions, divide the physical distance by the length of the diagonal pixels of the document page and normalize it to the range of 0 to 1. Weighted fusion: Set a balancing coefficient (e.g., 0.6), multiply the semantic distance by this coefficient, multiply the normalized physical distance by the remaining weights, and add the two together to get the final composite distance.

[0046] Forced clustering implementation (distance-short-circuit method): For forced clustering between sentences describing charts and their corresponding charts, a distance-short-circuit strategy is used. When constructing the distance matrix required by the clustering algorithm, sentences with spectral consistency feature values ​​exceeding a preset high threshold are first identified, and the nearest chart object to their physical location is found. Subsequently, the distance between the sentence and the chart object is directly set to zero in the distance matrix. When running the density-based clustering algorithm, since the distance between them is zero, they are grouped into the same cluster, thus achieving forced binding without modifying the clustering algorithm kernel.

[0047] Furthermore, the density-based clustering algorithm can employ the DBSCAN algorithm, using the comprehensive distance between sentences as a neighborhood distance metric. Specifically, by statistically analyzing the comprehensive distance distribution between sentences within the same paragraph on a historical corpus, a certain quantile of the comprehensive distance (e.g., the median or upper quartile) is used as the initial reference value for the neighborhood radius. This value is then adjusted based on the actual test results of cluster size and noise point ratio, so that the clustering results conform to manual partitioning habits in both semantics and physical space.

[0048] This invention proposes a clustering method that integrates semantic space and physical layout space. Traditional text clustering often ignores the document's layout structure. This invention, by introducing Manhattan distance and spectral weights, ensures that the generated paragraphs are not only semantically coherent but also visually continuous. In particular, for the typical structure of "image + text description below", forced clustering ensures that the caption and image are never separated, providing an accurate structural foundation for subsequent text-image consistency verification.

[0049] The steps for obtaining the macroscopic evaluation parameters include: Extract the semantic centroid vector of each paragraph in the naturally segmented data of the translated text to form the first vector sequence; extract the semantic centroid vector of each forced paragraph in the forced segmented data of the translated text to form the second vector sequence; The minimum alignment path distance between the first vector sequence and the second vector sequence is calculated using a dynamic time warping algorithm. This distance represents the degree of deviation between the translation and the physical structure of the source file in the natural semantic flow state. The overlap ratio of two data points at key semantic breakpoints is statistically analyzed; a key semantic breakpoint is defined as the sentence boundary where the paragraph attribute label changes; by combining the minimum alignment path distance and the overlap ratio of key semantic breakpoints, a structural similarity score is calculated and generated through normalized weighting, and this score is used as a macro-evaluation parameter.

[0050] Constructing a vector sequence: Extract the semantic centroid vector of each paragraph in the naturally segmented data of the translated text to form the first vector sequence; at the same time, extract the semantic centroid vector of each forced paragraph in the forced segmented data of the translated text to form the second vector sequence.

[0051] Dynamic path alignment: The minimum alignment path distance between the first and second vector sequences is calculated using a dynamic time warping algorithm. This distance characterizes the degree of deviation between the translated text and the physical structure of the source file under natural semantic flow. The larger the distance, the more chaotic the structure of the translated text.

[0052] Key breakpoint overlap statistics: This counts the overlap ratio of two data segments at key semantic breakpoints. A key semantic breakpoint is the location where the paragraph attribute label changes, such as the boundary between "test method" and "test conclusion".

[0053] Score generation: The structural similarity score is generated by combining the minimum alignment path distance and the overlap ratio of key semantic breakpoints through normalized weighted calculation, and this score is used as a macro-evaluation parameter.

[0054] The calculation of the semantic centroid vector of a paragraph includes: calculating the centroid of the paragraph using the average pooling method; Specifically, the semantic vectors of all sentences contained in a paragraph are obtained, these vectors are summed in each dimension, and then divided by the total number of sentences in the paragraph. The average vector obtained is the semantic centroid vector of the paragraph.

[0055] Normalized weighted structural similarity score: The macro-evaluation parameters consist of two parts: minimum alignment path distance and key breakpoint overlap ratio. The specific calculation logic is as follows: Distance normalization: Since the path distance calculated by the dynamic time warping algorithm is an absolute value, it is normalized using an exponential decay function; specifically, the negative exponent of the natural constant is calculated, and the exponent part is the path distance divided by a historical average distance constant, which maps the unbounded distance value to a similarity score between 0 and 1.

[0056] Weighted summation: Set weight coefficients, for example, multiply the normalized distance score by 0.7 and the key breakpoint overlap ratio by 0.3, and add the two together to get the final structural similarity score.

[0057] The weights are obtained using an iterative optimization method based on the validation set. The weight coefficients are set to range from 0 to 1, with a step size of 0.05. Since the sum of the two weight coefficients is 1, only one coefficient needs to be iterated over. Iterative testing is performed on a pre-defined validation set, which contains several typical document samples and their corresponding manually labeled "qualified" or "collapsed". In each iteration, a set of weight values ​​is selected, and a macro-evaluation score is calculated for all documents in the validation set. The matching accuracy between this score and the manually labeled tags is then calculated (e.g., using the F1 score).

[0058] Record the weight values ​​that maximize the matching accuracy and determine them as the optimal weight coefficients for the system.

[0059] This invention introduces a dynamic time warping algorithm to solve the structural alignment problem in translation quality assessment. By calculating the distance between the natural flow and the forced flow, the structural fidelity of the translation can be quantified. Once a structural deviation is detected, such as the omission of a chapter or the disorder of paragraph order, a serious quality incident can be identified immediately, preventing erroneous reports from being released.

[0060] The steps for obtaining the microscopic verification parameters include: For each source file paragraph and its corresponding translation, a forced paragraph division is performed, and data quantification and comparison are conducted. Numerical deviation calculation is performed based on the spectral consistency characteristics; digital entities are extracted from the translated paragraphs and their differences are calculated with the corresponding frequency points and amplitude points in the source file image waveform dataset. The absolute error value is calculated and normalized to obtain the numerical deviation index. For the density characteristics of technical terms, perform term mapping counting; based on the technical term database, count the cardinality of term sets in the source document paragraphs, and retrieve the corresponding target language term matching number in the translation paragraphs, calculate the mapping ratio from the source term set to the translation term set, and obtain term loss rate data; For logical coherence features, perform semantic projection analysis; map the paragraph vectors of the source document to the shared semantic space, and calculate the Euclidean distance between the source document and the forced paragraph vectors of the translation in the same space, as the semantic offset; The numerical deviation index, term loss rate, and semantic offset are constructed into three-dimensional feature coordinates. The spatial distance from these coordinates to the ideal zero error point is calculated, and the reciprocal of the spatial distance is used as a micro-verification parameter to measure the quality of paragraph translation.

[0061] For each source file paragraph and its corresponding translation, the text is forcibly divided into paragraphs, and the following quantitative comparison is performed: Numerical deviation calculation: Based on the spectral consistency feature, extract the digital entities in the translated paragraph, perform difference calculation with the corresponding frequency points and amplitude points in the source file image waveform dataset, calculate the absolute error value and perform normalization to obtain the numerical deviation index; Term mapping count: Based on the density characteristics of professional terms, according to the professional terminology database, the cardinality of the term set in the source document paragraph is counted, and the number of corresponding target language terms matched in the translated paragraph is retrieved. The mapping ratio is calculated to obtain the term loss rate data.

[0062] Semantic projection analysis: Based on the logical coherence feature, the paragraph vectors of the source document are mapped to the shared semantic space, and the Euclidean distance between them and the forced paragraph vectors of the translation in the same space is calculated as the semantic offset.

[0063] The numerical deviation index, term loss rate, and semantic offset mentioned above are used to construct a three-dimensional feature coordinate system. The spatial distance from this three-dimensional coordinate system to the ideal zero error point (i.e., no deviation, no loss, and no offset) is calculated, and the reciprocal of this spatial distance is used as a micro-verification parameter to measure the translation quality of the paragraph. The greater the distance, the lower the score.

[0064] This invention quantifies the abstract translation quality problem into a distance problem in geometric space by constructing a three-dimensional feature coordinate system. This design not only provides a comprehensive score, but also performs attribution analysis through coordinate components. In particular, by treating numerical deviation as an independent dimension, it can accurately locate specific erroneous paragraphs.

[0065] The specific steps for evaluating the initial translation based on macro-level evaluation parameters and micro-level verification parameters include: Construct a tiered evaluation system and set macro-level blocking thresholds and micro-level early warning thresholds; First, determine whether the macro-assessment parameters are below the macro-blocking threshold. If they are below the threshold, it is determined that the initial translation has a structural collapse, the initial translation is marked as unusable, and the entire document is re-translated. If the macro-evaluation parameters are qualified, the micro-verification parameters of each paragraph in the source file paragraph data are traversed; for the initial translation paragraphs whose micro-verification parameters are lower than the micro-warning threshold, correction instructions are generated according to the difference type.

[0066] Establish a tiered evaluation system: set macro-level blocking thresholds and micro-level early warning thresholds as red lines for quality control; Macro-level blocking determination: First, it is determined whether the macro-level evaluation parameters are lower than the macro-level blocking threshold. If they are lower than the threshold, it is determined that the initial translation has a structural collapse, the entire document is directly marked as unusable, and a full document re-translation process is triggered, without further local repairs.

[0067] Microscopic Traversal and Correction: If the macroscopic evaluation parameters are qualified, traverse the microscopic verification parameters of each paragraph in the source file data. For paragraphs with parameters below the microscopic warning threshold, generate correction instructions based on the difference type. If the discrepancy mainly stems from numerical deviation, the system calls upon real data from the image waveform dataset to forcibly overwrite the erroneous values ​​in the translation. If the discrepancy mainly stems from missing terminology, the system calls upon a specialized terminology database to replace and correct common terms in the translation. Specifically, for each paragraph's three features—numerical deviation index, terminology loss rate, and semantic offset—corresponding sub-thresholds are set. When the numerical deviation index exceeds the first sub-threshold, the main problem for the paragraph is determined to be inconsistency between the numbers in the translation and the image waveform data, and a numerical overwriting correction instruction is generated first. When the terminology loss rate exceeds the second sub-threshold, the main problem for the paragraph is determined to be the omission or improper replacement of specialized terms, and a correction instruction for terminology replacement and supplementation is generated first. When the semantic offset exceeds the third sub-threshold, the main problem for the paragraph is determined to be semantic content deviating from the source text, and an instruction prompting manual review or suggesting retranslation is generated first.

[0068] The macroscopic blocking threshold is used to divide the boundary value between "strumentally sound" and "structurally collapsed" within the range of macroscopic evaluation parameter values ​​from 0 to 1. In one implementation, statistics can be performed on a validation set containing manually labeled data to calculate the distribution range of macroscopic evaluation parameters for samples labeled as "structurally normal" and samples labeled as "structurally incorrect". The outer boundary of the overlapping interval of the two distributions is used as a candidate threshold so that samples with macroscopic evaluation parameters below the threshold are statistically likely to correspond to serious structural errors, thereby triggering a full document re-translation process.

[0069] This invention constructs a complete closed loop from global circuit breaking to local repair, firstly ensuring the bottom line of the deliverables, and secondly realizing automated and refined correction; in particular, the mechanism of generating correction instructions based on difference types makes the system no longer a passive inspector, but an active corrector, greatly reducing the workload of manual proofreading and realizing an automated closed loop for the translation of test reports.

[0070] Example 2: This invention proposes an intelligent document translation correction system, comprising: The data acquisition module acquires the source file containing mixed layout content and the initial translation. The mixed layout content includes text paragraphs and spectrum analysis charts. The source file analysis module constructs a document semantic segmentation model to parse the source file and obtain source file paragraph data. The document semantic segmentation model extracts multi-dimensional feature vectors for each sentence, including spectral consistency features, technical term density features, and logical coherence features. Based on the multi-dimensional feature vectors, semantic clustering is performed to divide the source file into multiple source file paragraphs with independent semantic functions. The translation analysis module divides the initial translation using a document semantic segmentation model, generating naturally segmented translation data; it also maps the segmentation boundaries of the source file paragraph data to the initial translation, generating forced segmentation data for the translation. The anomaly detection module calculates the structural similarity between the naturally segmented data and the forced segmented data of the translated text as a macro-evaluation parameter; it also calculates the multi-dimensional feature similarity between the corresponding paragraphs in the source file paragraph data and the forced segmented data of the translated text as a micro-verification parameter; and evaluates the initial translation based on the macro-evaluation parameter and the micro-verification parameter.

[0071] The document semantic segmentation model includes the following steps: Feature extraction steps: For each sentence in the source file, calculate a three-dimensional feature vector, specifically including: Spectral consistency feature: measures the degree of agreement between the numerical description in the sentence and the chart data in the source file; Terminology density feature: measures the density and positional weight of technical terms in the sentence; Logical coherence feature: measures the semantic smoothness and logical connection strength between the sentence and adjacent sentences; Clustering steps: Based on the three-dimensional feature vectors and the physical layout of the sentences, a density-based clustering algorithm (such as DBSCAN) is used to group the sentences in the source file into multiple paragraphs, so that the sentences in the same paragraph are semantically similar and physically adjacent. Attribute tagging step: Assign attribute tags to each generated paragraph. The tag types should include at least data analysis paragraphs, test method paragraphs, conclusion judgment paragraphs, and administrative information paragraphs.

[0072] The output of the document semantic segmentation model is source file paragraph data, which includes multiple source file paragraphs with labeled attributes; wherein, the training set of paragraph attribute labels is obtained based on the annotation of an expert group.

[0073] The specific steps for obtaining the spectrum consistency feature include: Computer vision technology is used to analyze the layout of the source file, identify and locate the area where the spectrum analysis chart is located, and the chart is separated into a background grid layer, a limit curve layer and a measured waveform layer by an image segmentation algorithm. A coordinate mapping transformation is performed on the measured waveform layer to identify the frequency range represented by the horizontal axis and the amplitude range represented by the vertical axis. All key waveform peak points exceeding the limit curve on the waveform curve are extracted by the pixel scanning algorithm to construct an image waveform dataset containing frequency and amplitude values. Named entity recognition technology is used to scan the text of the source file sentences in the context of the spectrum analysis chart and extract all text numerical entities describing the test results. The text numerical entities include frequency values, level values ​​and unit symbols. The text numerical entities are matched with the image waveform dataset using Euclidean distance. If the numerical value in the text can find the corresponding peak point in the image waveform dataset within a preset error range, its confidence score is calculated. The average confidence score of all numerical entities in the sentence is defined as the spectral consistency feature value of the sentence.

[0074] Specifically, this includes: converting the spectral analysis chart to the HSV color space and extracting the initial mask for each layer using an adaptive color threshold; for the background mesh layer, using morphological opening operations to remove noise using preset horizontal and vertical linear structural elements, and extracting the coordinates of mesh intersections to construct the mesh topology; for the limit curve layer and the measured waveform layer, detecting the overlapping area between them in pixel space; for the overlapping area, performing path prediction and separation based on the principle of local slope continuity of the curve, and using the Kalman filter algorithm to interpolate and repair the waveform breaks obscured by the limit curve, thereby obtaining a continuous measured waveform layer.

[0075] In EMC radiated emission test reports, the spectrum graph background is typically a dense green grid, with red limit lines running across the graph. The measured waveform in blue may exceed the limit lines at certain frequencies, causing pixel overlap or adhesion between the waveform and the limit lines. Specific implementation steps include: Step 1: Adaptive mask extraction based on HSV color space; First, convert the RGB format spectrogram to HSV (hue, saturation, brightness) color space. Compared with RGB space, HSV can more effectively separate "color" and "brightness".

[0076] For the background mesh layer: Set the green hue range (e.g., hue value from 60 to 180 degrees, saturation range from 30% to 100%, and brightness range from 50% to 100%), and extract the initial mask.

[0077] For the limit curve layer: set the red hue range (e.g., hue value from 0 to 10 degrees and 350 to 360 degrees, saturation range from 40% to 100%, and brightness range from 40% to 100%), and extract the mask.

[0078] For the measured waveform layer: Set the blue hue range (e.g., hue value 200 to 260 degrees, saturation range 35% to 100%, brightness range 30% to 95%), and extract the mask.

[0079] Because of color cast in the scanned document, the upper and lower limits of the threshold ranges for hue, saturation, and brightness are dynamically adjusted by statistically analyzing the histograms of each channel of the image, instead of using fixed values.

[0080] Step 2: Morphological Mesh Removal; For the background mesh layer, two types of morphological structural elements are constructed: horizontal line elements (e.g., 5 pixels wide and 1 pixel high) and vertical line elements (e.g., 1 pixel wide and 5 pixels high); These two structural elements are used to perform opening operations on the green mask. The opening operation involves erosion followed by dilation, which effectively preserves long straight lines (i.e., mesh lines) while removing isolated green noise points caused by compression noise; The extracted mesh intersection coordinates are used to correct the tilt distortion of the image in subsequent steps.

[0081] Step 3: Kalman filter repair of overlapping areas; When separating the blue measured waveform layer, some pixels were detected to simultaneously meet the thresholds for both red and blue or to exhibit a mixed purple color, causing the waveform to be visually cut off by the red limit line; In order to obtain continuous waveform data, the following repair logic is executed: Breakpoint identification: Scan the blue waveform mask to identify areas where pixels are discontinuous and mark them as occlusion breakpoints; State modeling: Establish a Kalman filter and define the state vector as [current frequency coordinate, current amplitude coordinate, amplitude change rate]; Prediction phase: Using the state of the 5 consecutive pixels before the breakpoint, calculate the amplitude change rate (i.e., local slope) to predict the theoretical trend of the waveform within the occlusion area; Interpolation repair: In the breakpoint area, instead of using the original pixel values, the waveform coordinates are filled with the predicted values ​​of the Kalman filter until the blue pixels reappear in the waveform; in this way, the key peak data that was blocked by the limit line can be reconstructed, ensuring the integrity of subsequent data verification; and overlapping areas and reconstructed content are marked and highlighted.

[0082] This technical solution addresses the limitations of current simple color segmentation methods in handling "curve overlap" and "complex background interference." In EMC reports, waveforms exceeding limit lines (i.e., overlap points) are often the most critical failure points. If segmentation causes waveform breakage, the most important test evidence will be lost. By using slope prediction and interpolation repair, the integrity of key data is ensured, significantly improving the robustness of data verification.

[0083] The specific steps for obtaining the density characteristics of the technical terms include: Construct a professional terminology database for electromagnetic compatibility, which covers proper nouns, abbreviations, and units of measurement in the fields of electromagnetic interference, electromagnetic susceptibility, radio frequency common-mode conduction, and electrostatic discharge. Each sentence in the source file is segmented into words, and the segmentation results are matched against the terminology database. The word frequency data of the professional terms in each sentence are counted, and the basic terminology density ratio is calculated by combining the total number of words in the sentence. A syntactic component weighting mechanism is introduced to identify the position of terms in sentence components, assign weights according to their positions, and calculate the weighted values ​​as the professional term density features by combining the basic term density ratio.

[0084] The introduced syntactic component weighting mechanism specifically includes: using the dependency syntactic analysis tree to parse the structure of the sentence, and determining the core predicate node of the sentence as the root node; calculating the depth of the term node in the dependency syntactic analysis tree, and defining that the basic weight is inversely proportional to the node depth, that is, the closer the node is to the root node, the higher the basic weight; identifying the type of dependency relationship between the term node and the root node; if the dependency relationship is a subject relationship or an object relationship, multiplying the basic weight by a first enhancement coefficient; if the dependency relationship is a modifier relationship or an adverbial relationship, multiplying the basic weight by a second attenuation coefficient; normalizing the finally calculated weight as the position weight of the term in the sentence component.

[0085] The system needs to determine which word in the source file is the technical core of the sentence and which word is only a modifier, so as to give different attentions when calculating the density of professional terms. For example, in the sentence "The test device connected by a linear impedance stabilization network was tested", "test device" is the core subject, while "linear impedance stabilization network" is only a modifier.

[0086] Dependency syntactic tree construction: Invoke the natural language processing parser to parse the structure of the source sentence and generate a tree structure with the core predicate verb as the root node. In the above sentence, "performed" is identified as the root node of the whole sentence.

[0087] Node depth calculation: Traverse the generated syntactic tree and calculate the path length of each word from the root node, which is defined as the node depth. For example, "performed" as the root node has a depth of zero; "tested" as the direct object has a depth of 1; "test device" as the subject has a depth of 1; "connected" as the attributive modifying the test device has a depth of 2; "linear impedance stabilization network" as the prepositional object of the connection has a depth of 3. Set the basic weight to be inversely proportional to the node depth, that is, the smaller the depth value, the more important the word is in the main body of the sentence and the higher the basic weight.

[0088] Relationship-based weight assignment: Preset weight adjustment rules and assign correction coefficients according to the type of dependency relationship between the word and its parent node: For words belonging to the main body components, such as subject relationship and object relationship, multiply by a first enhancement coefficient, which is greater than one, to increase its weight, for example, 1.5. For words belonging to the modifier components, such as attributive relationship, adverbial relationship, and prepositional modification relationship, multiply by a second attenuation coefficient, which is less than one, to reduce its weight, for example, 0.5.

[0089] Final weight calculation: For the term "tested device," a higher final weight is calculated due to its very shallow depth and subject-verb relationship. For the term "linear impedance-stabilized network," a lower final weight is calculated due to its deeper depth and prepositional modification relationship. Through this mechanism, when calculating the term density features of the entire sentence, the focus is on whether the core device name is translated correctly, while the attention to auxiliary devices is relatively reduced.

[0090] This technical solution overcomes the shortcomings of traditional word frequency statistics in distinguishing key points. In EMC reports, translation errors in the core tested equipment and power supply methods are fatal, while errors in auxiliary facilities (such as tables) are relatively minor. Through syntactic weighting, the system concentrates the verification computing power on the core of the sentence, making the scoring results more in line with the review standards of human experts.

[0091] The specific steps in obtaining the logical coherence feature include: Each sentence in the source file is mapped to a high-dimensional semantic vector. The cosine similarity between the current sentence vector and the vectors of the preceding and subsequent sentences is calculated to obtain the semantic smoothness index. Explicit logical connection markers in sentences are identified, and a logical topology network representing the pre-defined relationships between sentences is constructed. The logical topology network defines the theoretical trend of semantic change that adjacent sentences should exhibit, including at least semantic continuity and semantic transition trends. The logical topology network is used as a constraint to verify the consistency of the semantic smoothness index and generate logical coherence features.

[0092] The process of acquiring the logical coherence features may further include: constructing a sliding window contextual attention mechanism to calculate the aggregation similarity between the semantic vector of the current sentence and the topic vector of the paragraph; identifying implicit logical breakpoints within the paragraph; if there is a lack of explicit logical connection markers between adjacent sentences, but there is a strong semantic implication relationship between their semantic vectors, then these are marked as implicit logical breakpoints; the strong semantic implication relationship is determined by calculating the projection of the semantic vector of the preceding sentence into a preset logical transition space, and the distance between the projection and the semantic vector of the following sentence is less than a preset threshold; combining the difference in logical coherence features before and after translation, if the translation adds explicit logical connection markers at the implicit logical breakpoints to improve semantic smoothness, then the translation at that point is judged to have undergone explicit logical correction and is given a positive evaluation weight.

[0093] Chinese reports often employ a paratactic writing style, omitting conjunctions, such as "voltage too high, test stopped," while English translations tend to be hypotactic, using explicit conjunctions, such as "because the voltage was too high, the test stopped." It is necessary to recognize this difference and avoid misjudging it as inconsistency in translation.

[0094] Paragraph topic vectors and context window; Define a sliding window containing several sentences before and after it, use a pre-trained language model to extract the semantic vectors of all sentences within the window, calculate the average to obtain the local topic vector; Calculate the similarity between the current sentence vector and the local topic vector; During the source file analysis process, adjacent sentences are examined to determine whether there is an implicit logical relationship.

[0095] Formal detection: First, scan two adjacent sentences to check for the presence of explicit logical connectors such as "because", "therefore", and "however". If these connectors are not present, proceed to the next step of deep vector analysis.

[0096] Vector space logical mapping detection aims to identify sentences that are closely related in content but disconnected in form; including: Initial screening of semantic relevance: First, calculate the similarity value between the semantic vectors of the preceding and following sentences. If the similarity value is lower than the preset relevance threshold, it indicates that the topics discussed by the two sentences are completely unrelated, and the possibility of their logical connection is directly excluded. Logical transition matrix mapping: A logical transition matrix trained on a large-scale corpus is preloaded. This matrix represents the vector transformation rules from "cause" to "effect" or from "argument" to "evidence" in natural language. The semantic vector of the preceding sentence is input into the logical transition matrix, and a prediction vector is generated after linear transformation operation.

[0097] Projection and Distance Calculation: Subsequently, the spatial distance between the predicted vector and the actual semantic vector of the subsequent sentence is calculated.

[0098] Conclusion: If the calculated spatial distance is less than the preset logical judgment radius, it means that the semantic content of the subsequent sentence falls exactly within the semantic space of the preceding sentence after logical transformation; this indicates that although there is no connecting word between the two sentences, there is a mathematically verifiable strong semantic implication relationship.

[0099] When the above conditions are met, even though the conjunction is missing in form, it is still determined that there is an implicit logical chain between the two sentences, and the boundary position of the two sentences is marked as an implicit logical breakpoint.

[0100] Explicit verification of the translation: First, check the corresponding translated sentences and perform a conjunction search to detect whether any explicit logical words not present in the source text have been added to the translation, such as "therefore" or "so". Second, perform smoothness gain calculation to calculate the semantic smoothness of the source text and the semantic smoothness of the translation. If the semantic smoothness of the translation is higher than that of the source text, it means that the semantic flow becomes smoother after adding conjunctions.

[0101] Positive evaluation was generated; although the translation added a word that did not exist in the source text, resulting in a decrease in the accuracy of the literal correspondence, it was determined that this was an explicit correction based on logical understanding. Not only was no point deducted, but it was also given a positive weight in the logical coherence feature score. It was believed that the translation improved the readability and professionalism of the document and met the characteristics of high-quality human translation.

[0102] The specific construction and training process of the logical transition matrix is ​​as follows: Training data preparation: First, training samples are extracted from a general large-scale corpus. The extraction rule is to identify complex sentences or adjacent sentence pairs containing explicit logical connectors (such as "because...therefore...", "although...but..."). The part before the connector is taken as the "preceding sentence" and the part after the connector is taken as the "following sentence", thus constructing a set of positive samples in pairs.

[0103] Vectorization: Using a pre-trained language model, semantic feature vectors of the preceding and following sentences are extracted respectively.

[0104] Matrix training objective: Define a learnable parameter matrix with the same dimension as the semantic vector, i.e., the logical transition matrix; the training objective is to minimize the difference between the "predicted vector" generated by the linear transformation of the vector of the preceding sentence and the true "subsequent sentence vector".

[0105] Optimization process: The mean squared error is used as the loss function, which is to calculate the sum of squares of the differences between the predicted vector and the true vector; the matrix parameters are iteratively updated using the stochastic gradient descent algorithm until the loss function converges; the final matrix is ​​the logical transition matrix that can capture the logical transformation rules of natural language.

[0106] This invention, by introducing logical transition matrix mapping technology, corrects the shortcomings of traditional methods that rely solely on explicit conjunctions to determine logical relationships. Unlike simple vector statistical properties, this solution calculates the projection of semantic vectors into a specific logical subspace, enabling it to accurately identify semantically related sentences that conform to causality but omit conjunctions in form, such as causal sentences without conjunctions commonly found in Chinese reports. This mechanism can intelligently distinguish between "topic switching" and "implicit logic," thus accurately identifying and rewarding explicit supplementation of implicit logic in the translation during translation revision, such as proactively adding conjunctions to the translation, thereby ensuring the rigor and readability of the technical document's logic.

[0107] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. An intelligent document translation and correction system, characterized in that, include: The data acquisition module acquires the source file containing mixed layout content and the initial translation. The mixed layout content includes text paragraphs and spectrum analysis charts. The source file analysis module constructs a document semantic segmentation model to parse the source file and obtain the source file paragraph data; the document semantic segmentation model extracts multi-dimensional feature vectors for each sentence, including spectral consistency features, technical term density features, and logical coherence features; Semantic clustering is performed based on multidimensional feature vectors to divide the source file into multiple source file segments with independent semantic functions; The translation analysis module divides the initial translation using a document semantic segmentation model, generating natural segmentation data for the translation. Based on the segmentation boundaries of the source file paragraph data, the data is mapped to the initial translation to generate forced segmentation data for the translation. The anomaly detection module calculates the structural similarity between the naturally segmented data and the forced segmented data of the translated text as a macro-evaluation parameter; it also calculates the multi-dimensional feature similarity between the corresponding paragraphs in the source file paragraph data and the forced segmented data of the translated text as a micro-verification parameter; and evaluates the initial translation based on the macro-evaluation parameter and the micro-verification parameter.

2. The intelligent document translation and correction system according to claim 1, characterized in that, The specific steps for obtaining the spectrum consistency feature include: Computer vision technology is used to analyze the layout of the source file, identify and locate the area where the spectrum analysis chart is located, and the chart is separated into a background grid layer, a limit curve layer and a measured waveform layer by an image segmentation algorithm. A coordinate mapping transformation is performed on the measured waveform layer to identify the frequency range represented by the horizontal axis and the amplitude range represented by the vertical axis. All key waveform peak points exceeding the limit curve on the waveform curve are extracted by the pixel scanning algorithm to construct an image waveform dataset containing frequency and amplitude values. Named entity recognition technology is used to scan the text of the source file sentences in the context of the spectrum analysis chart and extract all text numerical entities describing the test results. The text numerical entities include frequency values, level values ​​and unit symbols. The text numerical entities are matched with the image waveform dataset using Euclidean distance. If the numerical value in the text can find the corresponding peak point in the image waveform dataset within a preset error range, its confidence score is calculated. The average confidence score of all numerical entities in the sentence is defined as the spectral consistency feature value of the sentence.

3. The intelligent document translation and correction system according to claim 1, characterized in that, The specific steps for obtaining the density characteristics of the technical terms include: Construct a professional terminology database for electromagnetic compatibility, which covers proper nouns, abbreviations, and units of measurement in the fields of electromagnetic interference, electromagnetic susceptibility, radio frequency common-mode conduction, and electrostatic discharge. Each sentence in the source file is segmented into words, and the segmentation results are matched against the terminology database. The word frequency data of the professional terms in each sentence are counted, and the basic terminology density ratio is calculated by combining the total number of words in the sentence. A syntactic component weighting mechanism is introduced to identify the position of terms in sentence components, assign weights according to their positions, and calculate the weighted values ​​as the professional term density features by combining the basic term density ratio.

4. The intelligent document translation and correction system according to claim 1, characterized in that, The specific steps in obtaining the logical coherence feature include: Each sentence in the source file is mapped to a high-dimensional semantic vector. The cosine similarity between the current sentence vector and the vectors of the preceding and subsequent sentences is calculated to obtain the semantic smoothness index. Explicit logical connection markers in sentences are identified, and a logical topology network representing the pre-defined relationships between sentences is constructed. The logical topology network defines the theoretical trend of semantic change that adjacent sentences should exhibit, including at least semantic continuity and semantic transition trends. The logical topology network is used as a constraint to verify the consistency of the semantic smoothness index and generate logical coherence features.

5. The intelligent document translation and correction system according to claim 1, characterized in that, The specific steps for dividing a source file into multiple sections include: Construct a high-dimensional feature space that includes the dimensions of spectral consistency, technical terminology density, and logical coherence. A density-based clustering algorithm is used to cluster sentence vectors in the source file. Spatial proximity constraints are introduced during the clustering process. The distance metric between sentences is defined to include not only the semantic distance in the feature vector space, but also the Manhattan distance of the sentence in the physical layout of the document, so as to achieve semantic coherence and physical continuity of the clustering results. For sentences with high spectral consistency, assign them spectral weights to cluster towards the corresponding chart position, force them to be grouped into an independent technical semantic block with the corresponding chart, and divide the paragraphs based on the independent technical semantic block; The output consists of multiple source file paragraphs, each labeled with a specific attribute tag. These attribute tags include at least the data analysis paragraph, the test method paragraph, the conclusion judgment paragraph, and the administrative information paragraph.

6. The intelligent document translation and correction system according to claim 1, characterized in that: The steps for obtaining the macroscopic evaluation parameters include: Extract the semantic centroid vector of each paragraph in the naturally segmented data of the translated text to form the first vector sequence; extract the semantic centroid vector of each forced paragraph in the forced segmented data of the translated text to form the second vector sequence; The minimum alignment path distance between the first vector sequence and the second vector sequence is calculated using a dynamic time warping algorithm. The overlap ratio of two data points at key semantic breakpoints is statistically analyzed; a key semantic breakpoint is defined as the sentence boundary where the paragraph attribute label changes; by combining the minimum alignment path distance and the overlap ratio of key semantic breakpoints, a structural similarity score is calculated and generated through normalized weighting, and this score is used as a macro-evaluation parameter.

7. The intelligent document translation and correction system according to claim 1, characterized in that: The steps for obtaining the microscopic verification parameters include: For each source file paragraph and its corresponding translation, a forced paragraph division is performed, and data quantification and comparison are conducted. Numerical deviation calculation is performed based on the spectral consistency characteristics; digital entities are extracted from the translated paragraphs and their differences are calculated with the corresponding frequency points and amplitude points in the source file image waveform dataset. The absolute error value is calculated and normalized to obtain the numerical deviation index. For the density characteristics of technical terms, perform term mapping counting; based on the technical term database, count the cardinality of term sets in the source document paragraphs, and retrieve the corresponding target language term matching number in the translation paragraphs, calculate the mapping ratio from the source term set to the translation term set, and obtain term loss rate data; For logical coherence features, perform semantic projection analysis; map the paragraph vectors of the source document to the shared semantic space, and calculate the Euclidean distance between the source document and the forced paragraph vectors of the translation in the same space, as the semantic offset; The numerical deviation index, term loss rate, and semantic offset are constructed into three-dimensional feature coordinates. The spatial distance from these coordinates to the ideal zero error point is calculated, and the reciprocal of the spatial distance is used as a micro-verification parameter to measure the quality of paragraph translation.

8. The intelligent document translation and correction system according to claim 1, characterized in that: The specific steps for evaluating the initial translation based on macro-level evaluation parameters and micro-level verification parameters include: Construct a tiered evaluation system and set macro-level blocking thresholds and micro-level early warning thresholds; First, determine whether the macro-assessment parameters are below the macro-blocking threshold. If they are below the threshold, it is determined that the initial translation has a structural collapse, the initial translation is marked as unusable, and the entire document is re-translated. If the macro-evaluation parameters are qualified, the micro-verification parameters of each paragraph in the source file paragraph data are traversed; for the initial translation paragraphs whose micro-verification parameters are lower than the micro-warning threshold, correction instructions are generated according to the difference type.

Images