A kind of picture-text translation system based on fusion knowledge graph

Abstract

This invention discloses a text-image translation system based on a fused knowledge graph, relating to the field of text-image translation technology. The invention constructs a fused knowledge graph consisting of an entity terminology knowledge subgraph and an OCR obfuscation knowledge subgraph through a knowledge graph module; a layout analysis module performs layout structure analysis on the input image to be translated, determining each layout region within the image; an OCR recognition and confidence labeling module performs optical character recognition on the text in each layout region, outputting the OCR recognition result of each character and its corresponding confidence score; and a retrieval and reasoning module generates the translation result of the image to be translated based on the fused knowledge graph. This system, through the fused knowledge graph, achieves layout structure perception and source error correction, significantly improving the accuracy and format fidelity of complex document translation.

Description

Technical Field

[0001] This invention relates to the field of image-text translation technology, specifically an image-text translation system based on a fused knowledge graph. Background Technology

[0002] With the acceleration of globalization and the popularization of digital information dissemination, automatic translation of cross-language text and image information has a wide range of application needs in cross-border e-commerce, document localization, educational publishing, and cross-border social networking. Text and image translation aims to convert the source language text embedded in an image into the target language, while preserving the visual layout and typesetting information of the original image as much as possible. To solve this problem, existing technologies have mainly formed two technical routes: one is a cascaded pipeline architecture, which uses an optical character recognition (OCR) module to recognize the text in the image as transcribed text, and then sends it to a plain text machine translation model for translation; the other is an end-to-end multimodal fusion architecture, which directly uses a large visual language model to generate the translation result of the target language from the input image.

[0003] However, the aforementioned existing technologies still have significant shortcomings in practical applications. In cascaded pipeline architectures, OCR recognition errors propagate directly to the translation stage, and the OCR output loses the two-dimensional layout structure of the document, resulting in a chaotic layout and semantic errors in the translation results. Although end-to-end multimodal architectures can alleviate error propagation, they require high computing resources and cannot guarantee the consistency of professional terminology translation within the document or the precise alignment of the translated text with the original layout. Therefore, how to overcome the problems of OCR error propagation, loss of layout structure, inconsistent terminology translation, and poor layout fidelity in existing technologies, and achieve high accuracy, high fidelity, and high consistency in text-image translation, is a problem we need to solve. To this end, we now provide a text-image translation system based on a fused knowledge graph. Summary of the Invention

[0004] The purpose of this invention is to provide a text-image translation system based on a fused knowledge graph.

[0005] The objective of this invention can be achieved through the following technical solution: a text-image translation system based on a fused knowledge graph, comprising:

[0006] The knowledge graph module is used to construct a fused knowledge graph consisting of an entity terminology knowledge subgraph and an OCR obfuscation knowledge subgraph.

[0007] The layout analysis module is used to analyze the layout structure of the input image to be translated and determine the various layout areas within the image.

[0008] The OCR recognition and confidence labeling module is used to perform optical character recognition on the text in each layout area, and output the OCR recognition result of each character and the corresponding confidence score.

[0009] The retrieval and reasoning module is used to generate translation results for the image to be translated based on the fused knowledge graph.

[0010] Furthermore, the process by which the layout analysis module performs layout structure analysis on the input image to be translated includes:

[0011] The input image to be translated is converted to grayscale to obtain a corresponding grayscale image, and the grayscale value of each pixel is extracted;

[0012] The grayscale image is normalized to form several homogeneous connected regions with the same level;

[0013] Extract the minimum bounding rectangle of each homogeneous connected region as the initial pixel segmentation region;

[0014] The initial pixel segmentation region is fitted according to preset rules to determine the final layout region, and the final layout region bounding box coordinates and region type are obtained.

[0015] Each finalized layout area is abstracted into a layout area node. Different edges are established based on the visual layout, semantic hierarchy, and reading logic relationship between layout areas to generate a layout sub-graph.

[0016] Furthermore, the process by which the OCR recognition and confidence labeling module performs optical character recognition on the text in each layout area and outputs the OCR recognition result and corresponding confidence score for each character includes:

[0017] The image within the page area is converted to a grayscale image, and the obtained grayscale image is binarized to obtain a binarized image.

[0018] Projecting the binarized image divides it into individual character image blocks;

[0019] A standard character template library is pre-built, which contains binary image templates of commonly used characters;

[0020] Each character image block to be identified is matched pixel-by-pixel with all binarized image templates in the template library, and the matching degree is calculated.

[0021] The character corresponding to the template with the highest matching degree is taken as the recognition result, and the matching degree is the confidence degree of that character.

[0022] Furthermore, the process by which the knowledge graph module constructs a fused knowledge graph consisting of an entity terminology subgraph and an OCR obfuscation subgraph includes:

[0023] By collecting terminology sets from various professional fields containing multiple terminology entries, each terminology entry corresponds to a professional term;

[0024] Iterate through each term entry and create a term entity node corresponding to each term entry;

[0025] Based on the hierarchical relationship field, cross-language, and related term fields in the term entries, different edges are established between term entity nodes to construct an entity term subgraph;

[0026] By statistically analyzing the OCR recognition and confidence annotation modules on images with ground truth annotations, the confusion probability is calculated for each incorrect sample.

[0027] The error sample refers to the number of times each source character is misidentified as a different obfuscated character;

[0028] Using characters as nodes and confusion probabilities as edge weights, directed edges are established from source characters to confused characters, forming an OCR confusion subgraph.

[0029] Furthermore, the retrieval reasoning module includes:

[0030] The OCR error correction guidance submodule is used to extract the character image block features and contextual semantic features corresponding to characters with confidence scores below a preset threshold, retrieve candidate correct characters in the OCR confusion sub-image, and output the corrected text as the text to be processed.

[0031] The terminology disambiguation submodule is used to retrieve terminology entity nodes that match the corresponding professional terms in the entity terminology subgraph based on the page area type and context semantics of the text to be processed. Based on the retrieval and matching results of the entity nodes, it outputs the standard translation of the professional terms and generates a terminology standard translation mapping table composed of terminology entity nodes and standard translations.

[0032] The reading order restoration submodule is used to reason about the page area nodes in the page layout subgraph and output the linearized text sequence of text content in different page areas and the correspondence between page areas.

[0033] Furthermore, the process by which the retrieval and reasoning module generates translation results for the image to be translated based on the fused knowledge graph includes:

[0034] A text encoder is used to input the linearized text sequence into the text encoder. The bounding box coordinates and region types of each page area in the page layout sub-image are used as page structure features and fused with the output of the text encoder to obtain the translation to be generated.

[0035] When the translation to be generated corresponds to a term entity node in the terminology standard translation mapping table, the standard translation of that term entity node is the preferred output of the translation result of the translation to be generated.

[0036] Furthermore, the retrieval reasoning module also includes a terminology consistency constraint submodule for:

[0037] Create a terminology translation mapping cache table, including keys and corresponding blank values ​​for each key;

[0038] When a technical term is translated for the first time, the standard translation of the technical term is retrieved from the standard translation mapping table, the standard translation is written into a blank value, and the term entity node corresponding to the standard translation is associated with the key corresponding to the written blank value.

[0039] When the same technical term appears again, retrieve the key associated with it from the term translation mapping cache table, and use the standard translation in the blank value corresponding to the key as the translation of the technical term.

[0040] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention constructs a text-image translation system based on a fused knowledge graph. It achieves semantic fusion of layout subgraphs, entity terminology subgraphs, and OCR obfuscation subgraphs within a unified graph structure. Combined with graph neural networks, it performs collaborative reasoning on low-confidence characters, terminology translations, and regional reading order. This integrates document structure understanding, terminology disambiguation, and OCR source error correction from image input to typesetting output. Furthermore, by associating standard terminology translations with layout region types and enforcing consistency in terminology translation during subsequent reasoning, it effectively solves the problem of inconsistent terminology in cross-regional translation of long documents. By unifying error correction, disambiguation, and reading order restoration within the framework of a fused knowledge graph through a retrieval and reasoning module, it helps eliminate the defects of error propagation at each level in traditional pipeline methods, effectively solving the problems of low accuracy and poor layout fidelity in the translation of complex text-image documents. Attached Figure Description

[0041] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in this invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0042] Figure 1 This is a schematic diagram of the present invention. Detailed Implementation

[0043] like Figure 1 As shown, a text-image translation system based on a fused knowledge graph includes:

[0044] The knowledge graph module is used to construct a fused knowledge graph consisting of an entity terminology knowledge subgraph and an OCR obfuscation knowledge subgraph.

[0045] The layout analysis module is used to analyze the layout structure of the input image to be translated and determine the various layout areas within the image.

[0046] The OCR recognition and confidence labeling module is used to perform optical character recognition on the text in each layout area, and output the OCR recognition result of each character and the corresponding confidence score.

[0047] The retrieval and reasoning module is used to generate translation results for the image to be translated based on the fused knowledge graph.

[0048] It should be further explained that, in the specific implementation process, the layout analysis module performs layout structure analysis on the input image to be translated, including:

[0049] The input image to be translated is converted to grayscale to obtain a corresponding grayscale image, and the grayscale value of each pixel is extracted;

[0050] The grayscale image is normalized to form several homogeneous connected regions with the same level; the homogeneous connected regions are specifically continuous blocks with similar pixel values;

[0051] The reclassification process specifically involves setting several grayscale value ranges, matching the grayscale values ​​at different locations in the grayscale image with each set grayscale value range, and assigning the grayscale values ​​at each location to the corresponding pixel level based on the matching results. Each grayscale value range corresponds to one pixel level.

[0052] Construct a UV coordinate system and map the normalized grayscale image into the UV coordinate system;

[0053] Extract the minimum bounding rectangle of each homogeneous connected region as the initial pixel segmentation region;

[0054] The initial pixel segmentation region is fitted according to preset rules to determine the final layout region; and the final layout region bounding box coordinates and region type are obtained.

[0055] It should be noted that the preset rules must simultaneously meet the following conditions:

[0056] The spatial adjacency distance between the bounding boxes of the two regions is less than a preset distance threshold;

[0057] The difference in average grayscale values ​​between the two regions is less than the preset grayscale difference threshold;

[0058] The aspect ratio of the bounding box of the merged new region is not greater than the preset ratio threshold, and there are no continuous edge segments with a length exceeding the preset length in the gap area between the two regions.

[0059] For any two initial pixel segmentation regions, if both of the above conditions are met, they are merged into a larger region.

[0060] After merging, the bounding box of the new region is the union of the bounding boxes of the two regions, and the average gray value of the new region is calculated as a weighted average of the pixel areas of the two regions.

[0061] The new region and the regions that were not merged are entered into the next round of scanning. The above merging process is repeated until no merging operation occurs in a certain round of scanning. The final layout region that fits the initial pixel segmentation region is determined, and the normalized bounding box coordinates, region type, and the layer depth of the region in the document are extracted.

[0062] The area types include body text, headings, tables, images, headers, footers, figure captions, and formula areas; the hierarchy depth specifically refers to the top-level area having a depth of 0, with the depth of nested areas increasing progressively.

[0063] Each finalized layout area is abstracted into a layout area node. Different edges are established based on the visual layout, semantic hierarchy, and reading logic relationships between layout areas to generate a layout subgraph. ;

[0064] in, As a set of nodes, each page area is abstracted as a node. Each node The attribute vector includes the bounding box coordinates, region type, and hierarchy depth mentioned above; As an edge set, each page area integrates the document's visual layout, semantic hierarchy, and reading logic into a unified graph structure through connections between different edges, thereby obtaining a page layout subgraph;

[0065] The visual layout refers to establishing a spatial adjacency edge if the spatial adjacency distance between two page areas is less than a preset threshold; where the spatial adjacency distance is the minimum spacing of the bounding boxes.

[0066] The semantic hierarchy refers to the fact that if two page areas with text content have a semantic relationship, and the first page area is semantically subordinate to the second page area, then the two page areas have a semantic hierarchy, thereby establishing a hierarchical subordinate edge.

[0067] The reading logic refers to establishing reading order edges if there is an order relationship based on document reading habits (such as from left to right or from top to bottom).

[0068] It should be further explained that, in the specific implementation process, the OCR recognition and confidence labeling module performs optical character recognition on the text in each layout area, including the following steps:

[0069] The image within the page area is converted to a grayscale image, and the obtained grayscale image is binarized to obtain a binarized image.

[0070] The binarized image is projected to divide it into individual character image blocks; specifically, the projection involves counting the number of black pixels in each row and dividing the image into several text lines based on the peaks and troughs of the row projection; and counting the number of black pixels in each column and dividing the row image into individual character image blocks based on the trough positions of the column projection.

[0071] A standard character template library is pre-built, which contains binary image templates of commonly used characters;

[0072] Each character image block to be identified is matched pixel by pixel with all binarized image templates in the template library, and the matching degree is calculated. Specifically, each template is a normalized 32×32 pixels in size. For each character image block to be identified, it is first scaled to 32×32 pixels.

[0073] The matching degree is defined as the proportion of corresponding pixels in two binary images that are the same:

[0074]

[0075] in, This represents the pixel value of the image block containing the character to be recognized at position (i,j); This represents the pixel value of the template image at the corresponding position; This is an indicator function; it returns 1 if the match is equal, otherwise it returns 0.

[0076] The character corresponding to the template with the highest matching degree is taken as the recognition result, and the matching degree is the confidence degree of the character; if the highest matching degree is lower than the preset threshold, the character is marked as a low confidence character, and the image block features and context (the recognition results of the two characters before and after) of the character are recorded so that the subsequent retrieval and reasoning module can correct the error.

[0077] By statistically analyzing the OCR recognition and confidence annotation modules on images with ground truth annotations, for each error sample, i.e., each source character... (Real characters) were misidentified as various obfuscated characters. Calculate the probability of confusion based on the number of times (OCR error output) occurs:

[0078] ;

[0079] in Indicates source character Identified as Number of times, express Total number of occurrences;

[0080] Using characters as nodes, with confusion probability Using edge weights, directed edges are established from the source character to the obfuscated character, forming an OCR obfuscation subgraph.

[0081] It should be further explained that, in the specific implementation process, the knowledge graph module constructs a fused knowledge graph consisting of entity terminology subgraphs and OCR obfuscation subgraphs, including:

[0082] Collect terminology sets from various professional fields containing multiple terminology entries, with each terminology entry corresponding to a professional term;

[0083] A term entry includes the following term attributes: term name, domain tag, standard translations of the term in different languages, and hierarchical relationship;

[0084] Iterate through each term entry, create a term entity node corresponding to each term entry, and use the term attributes of each term entry and the semantic embedding vector generated by a pre-trained language model (such as BERT) as the node attributes of each term entity node.

[0085] Based on the hierarchical relationship field in the terminology entries, establish directed hierarchical edges between the corresponding terminology entity nodes; establish equivalent edges based on cross-language alignment (connecting terminology entity nodes of the same term in different languages); establish undirected related edges based on related terminology fields (two terms have co-occurrence or semantic relatedness); complete the construction of the entity terminology subgraph;

[0086] For low-confidence characters output by the OCR recognition and confidence labeling module, a "suspected error" relationship edge is established between the corresponding character node and the source character node in the OCR confusion subgraph that uses the character as the confusion target; thus, the potential error correction candidate set of the character node is explicitly constructed in the fusion graph.

[0087] It should be further explained that, in the specific implementation process, the retrieval and reasoning module, based on the fused knowledge graph, performs graph neural network reasoning, including the following:

[0088] In the specific implementation process, the retrieval reasoning module includes an OCR error correction guidance submodule, a terminology disambiguation submodule, and a reading order restoration submodule;

[0089] OCR error correction guidance submodule: For characters with a confidence level below a preset threshold, this submodule first extracts the character image block features corresponding to the character. and contextual semantic features ;

[0090] The image patch features are specifically defined as the matching degree distribution obtained by template matching as the feature vector; the contextual semantic features are specifically defined as the co-occurrence frequency of adjacent characters before and after a character.

[0091] The source character with the highest similarity to the low-confidence character is retrieved from the OCR confusion subgraph, and the comprehensive score of each candidate source character is calculated using the following formula:

[0092] ;

[0093] in, Indicates cosine similarity; Standard image feature templates for candidate source characters; For character context embedding trained on large corpora; , , These are weighting coefficients, which are set to 0.3, 0.4, and 0.3 in this embodiment.

[0094] Will The highest-ranking source character is used as the correction result and output as the text to be processed. The corrected text replaces the corresponding character in the original OCR recognition result.

[0095] The terminology disambiguation submodule retrieves a set of candidate terminology entity nodes matching the corresponding professional terms from the entity terminology subgraph based on the page area type and contextual semantics of the text to be processed. Each candidate node is labeled with a domain tag. and standard translation and pre-trained semantic embedding vectors and scene prototype embedding Calculate the score for each candidate node:

[0096] ;

[0097] in, The type of the page area where the text to be processed is located; This is the image scene feature vector extracted from the global image using ResNet; This is an indicator function; it is 1 when the layout area type matches the domain label of the candidate term, and 0 otherwise. The cosine similarity between the context word embedding and the candidate node semantic embedding; The similarity between image scene features and candidate node scene prototypes; These are weighting coefficients, which are set to 0.5, 0.3, and 0.2 in this embodiment, respectively.

[0098] choose The standard translation corresponding to the highest candidate node is output as the standard translation of the term, and a term standard translation mapping table consisting of term entity nodes and standard translations is generated.

[0099] Reading order restoration submodule: Inputs the page layout subgraph into a graph attention network (GAT); each node The feature vector is determined by its normalized bounding box coordinates. The aspect ratio is and regional types It is composed of unique hot codes spliced ​​together;

[0100] GAT learns the spatial dependency order between adjacent nodes through a multi-head attention mechanism and outputs a score vector representing the node. The probability of a position within the entire reading sequence;

[0101] A greedy topological sorting algorithm is used to generate linearized text sequences of text content in different page areas and the correspondence between page areas;

[0102] Through the collaborative work of the above three sub-modules, the retrieval reasoning module outputs the corrected text sequence, the terminology standard translation mapping table, and the linearized text sequence and region correspondence that conform to the document logic.

[0103] It should be further explained that, in the specific implementation process, the process by which the retrieval and inference module generates the target language translation results includes:

[0104] Using a Transformer-based multimodal translation architecture, a text encoder is employed. The linearized text sequence is input into the text encoder, and the bounding box coordinates and region types of each page area in the layout sub-graph are used as page structure features. These features are then fused with the output of the text encoder to obtain the translation to be generated.

[0105] When the translation to be generated corresponds to a term entity node in the terminology standard translation mapping table, the standard translation of that term entity node is the preferred output of the translation result of the translation to be generated, and other candidate translations are blocked.

[0106] It should be further explained that, in the specific implementation process, the retrieval reasoning module also includes a terminology consistency constraint submodule, which is used to maintain the consistency of terminology translation within the same document. Its specific workflow is as follows:

[0107] Create a terminology translation mapping cache table, including keys and corresponding blank values ​​for each key;

[0108] When a technical term is translated for the first time, the standard translation of the technical term is retrieved from the standard translation mapping table, the standard translation is written into a blank value, and the term entity node corresponding to the standard translation is associated with the key corresponding to the written blank value.

[0109] When the same technical term appears again, retrieve the key associated with it from the term translation mapping cache table, and use the standard translation in the blank value corresponding to the key as the translation of the technical term;

[0110] By using the terminology consistency constraint submodule, the repeated calls to the terminology disambiguation submodule for graph reasoning calculations are avoided, significantly improving the processing speed of long document translation; at the same time, the forced reuse of the first translation method also ensures the translation consistency of the same term throughout the text.

[0111] Through the above specific implementation methods, the present invention realizes a text-image translation system based on a fused knowledge graph. It preserves the document structure through layout subgraphs, achieves source error correction through OCR obfuscation subgraphs, performs accurate disambiguation by combining entity terminology subgraphs with layout type, and ensures document-level translation quality through terminology consistency constraints. Compared with the prior art, the present invention significantly improves the translation accuracy and layout fidelity of complex text-image documents.

[0112] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any modifications or equivalent substitutions made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A text-image translation system based on a fused knowledge graph, characterized in that, include: The knowledge graph module is used to construct a fused knowledge graph consisting of an entity terminology knowledge subgraph and an OCR obfuscation knowledge subgraph. The layout analysis module is used to analyze the layout structure of the input image to be translated and determine the various layout areas within the image. The OCR recognition and confidence labeling module is used to perform optical character recognition on the text in each layout area, and output the OCR recognition result of each character and the corresponding confidence score. The retrieval and reasoning module is used to generate translation results for the image to be translated based on the fused knowledge graph.

2. The image-text translation system based on a fused knowledge graph according to claim 1, characterized in that, The process by which the layout analysis module performs layout structure analysis on the input image to be translated includes: The input image to be translated is converted to grayscale to obtain a corresponding grayscale image, and the grayscale value of each pixel is extracted; The grayscale image is normalized to form several homogeneous connected regions with the same level; Extract the minimum bounding rectangle of each homogeneous connected region as the initial pixel segmentation region; The initial pixel segmentation region is fitted according to preset rules to determine the final layout region, and the final layout region bounding box coordinates and region type are obtained. Each finalized layout area is abstracted into a layout area node. Different edges are established based on the visual layout, semantic hierarchy, and reading logic relationship between layout areas to generate a layout sub-graph.

3. The image-text translation system based on a fused knowledge graph according to claim 1, characterized in that, The process by which the OCR recognition and confidence labeling module performs optical character recognition on the text in each layout area and outputs the OCR recognition result and corresponding confidence score for each character includes: The image within the page area is converted to a grayscale image, and the obtained grayscale image is binarized to obtain a binarized image. Projecting the binarized image divides it into individual character image blocks; A standard character template library is pre-built, which contains binary image templates of commonly used characters; Each character image block to be identified is matched pixel-by-pixel with all binarized image templates in the template library, and the matching degree is calculated. The character corresponding to the template with the highest matching degree is taken as the recognition result, and the matching degree is the confidence degree of that character.

4. The image-text translation system based on a fused knowledge graph according to claim 1, characterized in that, The process by which the knowledge graph module constructs a fused knowledge graph consisting of an entity terminology subgraph and an OCR obfuscation subgraph includes: By collecting terminology sets from various professional fields containing multiple terminology entries, each terminology entry corresponds to a professional term; Iterate through each term entry and create a term entity node corresponding to each term entry; Based on the hierarchical relationship field, cross-language, and related term fields in the term entries, different edges are established between term entity nodes to construct an entity term subgraph; By statistically analyzing the OCR recognition and confidence annotation modules on images with ground truth annotations, the confusion probability is calculated for each incorrect sample. The error sample refers to the number of times each source character is misidentified as a different obfuscated character; Using characters as nodes and confusion probabilities as edge weights, directed edges are established from source characters to confused characters, forming an OCR confusion subgraph.

5. A text-image translation system based on a fused knowledge graph according to claim 1, characterized in that, The retrieval reasoning module includes: The OCR error correction guidance submodule is used to extract the character image block features and contextual semantic features corresponding to characters with confidence scores below a preset threshold, retrieve candidate correct characters in the OCR confusion sub-image, and output the corrected text as the text to be processed. The terminology disambiguation submodule is used to retrieve terminology entity nodes that match the corresponding professional terms in the entity terminology subgraph based on the page area type and context semantics of the text to be processed. Based on the retrieval and matching results of the entity nodes, it outputs the standard translation of the professional terms and generates a terminology standard translation mapping table composed of terminology entity nodes and standard translations. The reading order restoration submodule is used to reason about the page area nodes in the page layout subgraph and output the linearized text sequence of text content in different page areas and the correspondence between page areas.

6. A text-image translation system based on a fused knowledge graph according to claim 1, characterized in that, The process by which the retrieval and reasoning module generates translation results for the image to be translated based on a fused knowledge graph includes: A text encoder is used to input the linearized text sequence into the text encoder. The bounding box coordinates and region types of each page area in the page layout sub-image are used as page structure features and fused with the output of the text encoder to obtain the translation to be generated. When the translation to be generated corresponds to a term entity node in the terminology standard translation mapping table, the standard translation of that term entity node is the preferred output of the translation result of the translation to be generated.

7. A text-image translation system based on a fused knowledge graph according to claim 1, characterized in that, The retrieval reasoning module also includes a terminology consistency constraint submodule for: Create a terminology translation mapping cache table, including keys and corresponding blank values ​​for each key; When a technical term is translated for the first time, the standard translation of the technical term is retrieved from the standard translation mapping table, the standard translation is written into a blank value, and the term entity node corresponding to the standard translation is associated with the key corresponding to the written blank value. When the same technical term appears again, retrieve the key associated with it from the term translation mapping cache table, and use the standard translation in the blank value corresponding to the key as the translation of the technical term.

Images