Multi-language rapid processing method for ERP system

Abstract

The invention discloses a multi-language rapid processing method for an ERP (Enterprise Resource Planning) system, and relates to the technical field of multi-language processing, and the method comprises the steps: collecting a source language text, inputting the source language text into a constructed feature recognition model, and enabling the feature recognition model to build a bidirectional language mapping matrix based on a knowledge graph; segmenting the source language text through a semantic sectionalizer, extracting a semantic feature group based on the bidirectional language mapping matrix, and generating semantic probability distribution; constructing a fusion transformation network, performing feature matching on the semantic probability distribution and a target language template library, and outputting a multi-language candidate sequence with confidence; and screening the multi-language candidate sequences according to the confidence, performing syntactic recombination through a grammar optimizer, and updating the recombined multi-language candidate sequences to the ERP system. According to the method, the key technical breakthrough of multi-language processing of the ERP system is realized by constructing the two-way language mapping matrix and the fusion transformation network of the knowledge graph.

Description

Technical Field

[0001] The present invention relates to the technical field of multi-language processing, in particular to a multi-language rapid processing method for an ERP system. Background Art

[0002] Multilingual processing technology has become a key technology for connecting different language environments and promoting international business communication. In recent years, with the rapid development of artificial intelligence and natural language processing technology, machine translation and language intelligent processing technology have shown an increasingly broad application prospect in enterprise-level applications. However, traditional multilingual processing methods often have technical bottlenecks such as insufficient semantic accuracy, low processing efficiency, and insufficient depth of context understanding.

[0003] Existing multilingual processing technologies mainly rely on statistical and rule-based translation models, which often encounter significant limitations when dealing with complex semantics and professional terminology. Traditional methods usually use static dictionary mapping or simple neural network translation models, which are difficult to accurately capture the subtle semantic differences and contextual associations between languages. Especially in application scenarios such as ERP systems that are highly sensitive to professional terminology and business context, existing technologies often cannot achieve high-precision, low-latency rapid conversion of multilingual texts. These technical shortcomings not only affect effective communication between enterprises, but may also lead to semantic distortion and reduced translation quality of key business information, thereby increasing the communication costs and potential risks of cross-language collaboration.

[0004] Therefore, a fast multilingual processing method for ERP system is proposed to solve the key problems of insufficient semantic accuracy and processing efficiency in traditional multilingual processing technology. Summary of the invention

[0005] In view of the problems of insufficient semantic accuracy, low processing efficiency and insufficient depth of context understanding in existing multilingual processing technologies, the present invention is proposed.

[0006] Therefore, the problem to be solved by the present invention is how to achieve semantic deep mapping based on knowledge graph, improve the accuracy and processing speed of multilingual text conversion in ERP system, and effectively reduce the cost of semantic understanding in cross-language business communication.

[0007] In order to solve the above technical problems, the present invention provides the following technical solutions: In a first aspect, an embodiment of the present invention provides a multilingual rapid processing method for an ERP system, comprising: collecting source language text, and inputting the source language text into a constructed feature recognition model, wherein the feature recognition model establishes a bidirectional language mapping matrix based on a knowledge graph; segmenting the source language text through a semantic segmenter, extracting semantic feature groups based on the bidirectional language mapping matrix, and generating a semantic probability distribution; constructing a fusion transformation network, performing feature matching between the semantic probability distribution and a target language template library, and outputting a multilingual candidate sequence with confidence; screening the multilingual candidate sequence according to the confidence, and performing syntactic reorganization through a grammar optimizer, and updating the reorganized multilingual candidate sequence to the ERP system.

[0008] As a preferred solution of the multilingual fast processing method for the ERP system of the present invention, wherein: the multilingual candidate sequences are screened according to the confidence, and the syntax is reorganized by the grammar optimizer, and the reorganized multilingual candidate sequences are updated to the ERP system, including: by setting a dynamic threshold function based on historical translation quality data, the multilingual candidate sequences are arranged in descending order according to the confidence, and the minimum confidence requirement is determined; the adjacent candidate sequences are detected for repetition by the sliding window algorithm, and the multilingual candidate sequences with confidence greater than the threshold and repetition less than the set value are retained; a grammar optimizer is established by using a hybrid architecture based on rules and statistics, wherein the grammar optimizer includes a grammar analyzer, a grammar analyzer, and a grammar analyzer. The invention relates to a method for preparing a multilingual candidate sequence of a target language using a grammatical analysis module, a structural reorganization module and a grammar verification module; constructing a syntactic tree for the screened multilingual candidate sequence through the syntactic analysis module; adjusting the sentence pattern based on the grammatical specification and syntactic tree of the target language through the structural reorganization module, and extracting the syntactic dependency of the multilingual candidate sequence; adjusting the positions of sentence components according to the word order rules of the target language, optimizing the hierarchical structure of long sentences, supplementing grammatical components, and verifying the adjusted and reorganized multilingual candidate sequence through the grammar verification module; setting a life cycle and access rights for the verified multilingual candidate sequence, and simultaneously updating the data to a number of cache nodes through an asynchronous write mechanism, and establishing a mapping relationship between the source language text and the target language text.

[0009] As a preferred solution of the multilingual fast processing method for the ERP system described in the present invention, the method for obtaining the multilingual candidate sequence is to establish a fusion transformation network based on the encoder architecture and the decoder architecture, and adopt a multi-layer bidirectional feature extraction module for the encoder architecture to capture the deep semantic features of the source language text, wherein the fusion transformation network includes an input layer and an output layer; integrate the target language knowledge through the decoder architecture, and build a target language template library based on large-scale parallel corpus; based on the feature embedding matrix of the input layer, map the normalized language feature vector probability distribution to a high-dimensional feature space; adopt a multi-head feature extraction module based on the high-dimensional feature space The cross-attention layer performs feature matching, and the similarity between the source language feature and the target language template is calculated through the attention head; based on the similarity, the input feature is projected to different subspaces through the transformation matrix, the attention score is calculated and the feature is reorganized; based on the feature reorganization, a short-circuit connection is established between the layers of the fusion transformation network through a residual connection structure, so that the shallow layer features are transmitted to the deep layer, wherein the residual connection structure includes a feature addition module and a layer normalization module; the current layer output feature and the input feature are added through the feature addition module, and the feature is standardized through the layer normalization module; according to the standardized features, a beam search algorithm is used to generate a multilingual candidate sequence with confidence; the confidence score is calculated according to the multilingual candidate sequence, wherein the confidence score is used to sort and screen the multilingual candidate sequence.

[0010] As a preferred solution of the multi-language fast processing method for the ERP system of the present invention, wherein: the multi-head cross attention layer includes a transformation matrix; the transformation matrix includes a query transformation matrix, a key transformation matrix and a value transformation matrix; the specific formula of the confidence score is as follows: ; in, is the source language feature vector, is the target language feature vector, is the total feature dimension, is the eigenvalue of the i-th dimension of the source language, is the eigenvalue of the target language ith dimension, is the feature weight of the i-th dimension, is the angle between the source language feature vector and the target language feature vector, is the similarity enhancement function, the specific formula is as follows: ; in, is the feature difference vector, is the feature space dimension, is the edit distance, is the maximum allowed edit distance.

[0011] As a preferred solution of the multi-language fast processing method for the ERP system described in the present invention, wherein: the method for generating the language feature vector is to establish a semantic segmenter based on the hierarchical structure of a bidirectional long short-term memory network, identify the natural paragraph boundaries of the source language text through the semantic segmenter, and divide the long paragraphs into text units based on the semantic coherence index, and retain the hierarchical relationship and context information between the natural paragraphs; extract grammatical structure markers through a bidirectional language mapping matrix, and match the extracted grammatical structure markers with the segmented text units, wherein the bidirectional language mapping matrix includes grammatical structure markers, word meaning relationship weights and context association rules; according to the word meaning relationship weights of the matching results, the word meaning relationship weights are matched; Recalculate the semantic association strength between words, and analyze the contextual features of the text unit based on the context association rules; combine the grammatical features, semantic association strength and contextual features to form a semantic feature group; process the semantic feature group in parallel through a self-attention network to capture semantic relationships, and calculate the dot product similarity between the query vector, the key vector and the value vector to obtain the attention weights between the features; perform weighted combination of the feature vectors according to the attention weights to generate a language feature vector; perform quantization calculation on the language feature vector, use a softmax function to convert the quantization result of the language feature vector into a probability value, and map the probability value to the [0,1] interval through a normalization method.

[0012] As a preferred solution of the multi-language rapid processing method for ERP system described in the present invention, wherein: the grammatical structure tag includes part-of-speech tagging, syntactic dependency and grammatical component analysis results; the word meaning relationship weight represents the semantic relevance between words; the context association rule represents the usage rules of words in different contexts; the language feature vector The specific formula is as follows: ; in, is the attention head index, is the weight of the p-th attention head, is the query transformation matrix of the p-th attention head, is the key transformation matrix of the p-th attention head, is the value transformation matrix of the p-th attention head, is the input feature, is the number of attention heads, is the scaling factor, is the activation function, is the attention calculation function.

[0013] As a preferred solution of the multi-language rapid processing method for the ERP system described in the present invention, wherein: the method for constructing the bidirectional language mapping matrix is ​​to collect source language text from the ERP system through a web crawler program, and format the source language text using a text parser to remove special characters and redundant marks; divide the formatted source language text into basic language units according to the syntactic structure, perform semantic annotation, and extract entity relationships and attribute information; construct language knowledge nodes based on the annotated basic language units, and form a mesh structure between the nodes through semantic relationship edges to establish a knowledge graph, wherein the knowledge graph continuously optimizes the correlation strength between nodes through an iterative update mechanism; based on a deep neural network model of the Transformer architecture, a feature recognition model is constructed according to the knowledge graph, wherein the feature recognition model includes a multi-layer self-attention mechanism and an encoding layer; the multi-layer self-attention mechanism is used to capture the long-distance dependency relationship of the source language text; the encoding layer maps the source language text into a high-dimensional feature vector; based on the feature recognition model, language knowledge node pairs in the knowledge graph are extracted, the semantic similarity between nodes is calculated, and mapping weights are set to construct a bidirectional language mapping matrix .

[0014] As a preferred solution of the multi-language rapid processing method for ERP system of the present invention, wherein: the bidirectional language mapping matrix The specific formula is as follows: ; in, is the source language word index, t is the target language word index, is the source language word vector, is the target language word vector, is the grammatical match, is the location distance, is the distance attenuation coefficient, is the context relevance, is the normalized component, n is the vector dimension, and k is the number of knowledge graph relations.

[0015] In a second aspect, an embodiment of the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program instructions are executed by the processor, the steps of the multi-language rapid processing method for an ERP system as described in the first aspect of the present invention are implemented.

[0016] In a third aspect, an embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program instructions are executed by a processor, the steps of the multi-language rapid processing method for an ERP system as described in the first aspect of the present invention are implemented.

[0017] The beneficial effects of the present invention are as follows: the present invention can accurately capture subtle semantic differences across languages ​​through a multi-head cross-attention layer and a residual connection structure; combined with a dynamic threshold function and a grammar optimizer, it can intelligently screen and reorganize multilingual candidate sequences to ensure the grammatical fluency and semantic consistency of the translation results; through the iterative update mechanism of the knowledge graph, continuous learning and adaptive optimization of the multilingual processing model are achieved, providing enterprises with a highly intelligent multilingual text rapid conversion solution, reducing the technical barriers and communication costs of cross-language business communication, and achieving key technical breakthroughs in multilingual processing of ERP systems. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the following briefly introduces the drawings required for describing the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. Among them: Figure 1 This is a flow chart of the multi-language rapid processing method for an ERP system according to Example 1. DETAILED DESCRIPTION

[0019] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0020] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0021] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0022] Example 1 Reference Figure 1 , which is the first embodiment of the present invention, provides a multi-language rapid processing method for an ERP system, comprising: S1: Collect source language text and input the source language text into a constructed feature recognition model, wherein the feature recognition model establishes a bidirectional language mapping matrix based on a knowledge graph.

[0023] Specifically, the method for constructing a bidirectional language mapping matrix is ​​to collect source language texts from the ERP system through a web crawler program, and use a text parser to format the source language texts to remove special characters and redundant tags; the formatted source language texts are divided into basic language units according to the syntactic structure, semantic annotation is performed, and entity relationships and attribute information are extracted.

[0024] Furthermore, language knowledge nodes are constructed based on the annotated basic language units, and the nodes are connected to form a network structure through semantic relationship edges to establish a knowledge graph, in which the knowledge graph continuously optimizes the association strength between nodes through an iterative update mechanism.

[0025] Furthermore, a deep neural network model based on the Transformer architecture builds a feature recognition model according to the knowledge graph, where the feature recognition model includes a multi-layer self-attention mechanism and an encoding layer; the multi-layer self-attention mechanism is used to capture the long-distance dependencies of the source language text; the encoding layer maps the source language text into a high-dimensional feature vector.

[0026] Specifically, based on the feature recognition model, we extract language knowledge node pairs in the knowledge graph, calculate the semantic similarity between nodes, set mapping weights, and construct a bidirectional language mapping matrix. , the specific formula is as follows: ; in, is the source language word index, t is the target language word index, is the source language word vector, is the target language word vector, is the grammatical match, is the location distance, is the distance attenuation coefficient, is the context relevance, is the normalized component, n is the vector dimension, and k is the number of knowledge graph relations.

[0027] S2: segmenting the source language text through a semantic segmenter, extracting semantic feature groups based on the bidirectional language mapping matrix, and generating semantic probability distribution.

[0028] Specifically, the method for generating language feature vectors is to establish a semantic segmenter based on the hierarchical structure of a bidirectional long short-term memory network, identify the natural paragraph boundaries of the source language text through the semantic segmenter, and divide the long paragraphs into text units based on the semantic coherence index, while retaining the hierarchical relationship and contextual information between natural paragraphs.

[0029] Furthermore, grammatical structure markers are extracted through a bidirectional language mapping matrix, and the extracted grammatical structure markers are matched with the segmented text units; the semantic association strength between words is calculated according to the word-meaning relationship weights of the matching results, and the contextual features of the text units are analyzed based on contextual association rules; the grammatical features, semantic association strength and contextual features are combined to form a semantic feature group.

[0030] It should be noted that the bidirectional language mapping matrix includes grammatical structure markers, word meaning relationship weights and context association rules; grammatical structure markers include part-of-speech tagging, syntactic dependency and grammatical component analysis results; word meaning relationship weights represent the degree of semantic relevance between words; context association rules represent the usage rules of words in different contexts.

[0031] Furthermore, the semantic feature groups are processed in parallel through the self-attention network to capture the semantic relationship, and the dot product similarity between the query vector, key vector and value vector is calculated to obtain the attention weight between the features; the feature vectors are weighted and combined according to the attention weights to generate the language feature vector , the specific formula is as follows: ; in, is the attention head index, is the weight of the p-th attention head, is the query transformation matrix of the p-th attention head, is the key transformation matrix of the p-th attention head, is the value transformation matrix of the p-th attention head, is the input feature, is the number of attention heads, is the scaling factor, is the activation function, is the attention calculation function.

[0032] Specifically, the language feature vector is quantized and converted into a probability value by using the softmax function, and the probability value is mapped to the interval [0, 1] by a normalization method.

[0033] Furthermore, the specific formula for the probability distribution of language feature vectors is as follows: ; in, is the word of the current index m, h is the index of the candidate word, is the index of the characteristic function, is the weight coefficient of the f-th characteristic function, is the fth characteristic function, is the total number of candidate words, is the word-gram of candidate word-gram h.

[0034] S3: Construct a fusion transformation network, perform feature matching between the semantic probability distribution and the target language template library, and output a multilingual candidate sequence with confidence.

[0035] Specifically, the method for obtaining multilingual candidate sequences is to establish a fusion transformation network based on the encoder architecture and the decoder architecture, and adopt a multi-layer bidirectional feature extraction module for the encoder architecture to capture the deep semantic features of the source language text, wherein the fusion transformation network includes an input layer and an output layer; integrate the target language knowledge through the decoder architecture, and build a target language template library based on a large-scale parallel corpus; based on the feature embedding matrix of the input layer, map the normalized language feature vector probability distribution to a high-dimensional feature space.

[0036] It should be noted that the target language template library includes a language pair feature mapping table, a sentence structure template and semantic conversion rules; the language pair feature mapping table stores the correspondence between the source language and the target language; the sentence structure template contains the typical expressions of the target language; and the semantic conversion rules define the translation strategy in a specific context.

[0037] Furthermore, a multi-head cross-attention layer is used for feature matching according to the high-dimensional feature space, and the similarity between the source language features and the target language template is calculated through the attention head; based on the similarity, the input features are projected to different subspaces through the transformation matrix, the attention score is calculated and the features are reorganized; based on the feature reorganization, a short-circuit connection is established between the layers of the fusion transformation network through the residual connection structure, so that the shallow features are transmitted to the deep layers; the output features of the current layer and the input features are added through the feature addition module, and the features are standardized through the layer normalization module.

[0038] It should be noted that the multi-head cross attention layer includes a transformation matrix; the transformation matrix includes a query transformation matrix, a key transformation matrix, and a value transformation matrix; the residual connection structure includes a feature summation module and a layer normalization module; Furthermore, based on the standardized features, a beam search algorithm is used to generate a multilingual candidate sequence with confidence; the confidence score is calculated based on the multilingual candidate sequence, wherein the confidence score is used for sorting and screening the multilingual candidate sequence.

[0039] Specifically, the specific formula for the confidence score is as follows: ; in, is the source language feature vector, is the target language feature vector, is the total feature dimension, is the eigenvalue of the i-th dimension of the source language, is the eigenvalue of the target language i-th dimension, is the feature weight of the i-th dimension, is the angle between the source language feature vector and the target language feature vector, is the similarity enhancement function, the specific formula is as follows: ; in, is the feature difference vector, is the feature space dimension, is the edit distance, is the maximum allowed edit distance.

[0040] S4: Screening the multilingual candidate sequences according to the confidence level, performing syntactic reorganization through a grammar optimizer, and updating the reorganized multilingual candidate sequences to the ERP system.

[0041] Specifically, by setting a dynamic threshold function based on historical translation quality data, multilingual candidate sequences are sorted in descending order according to confidence, and the minimum confidence requirement is determined; the sliding window algorithm is used to detect the repetition of adjacent candidate sequences, and the multilingual candidate sequences with confidence greater than the threshold and repetition less than the set value are retained.

[0042] It should be noted that when the system receives a new multilingual candidate sequence, it immediately calculates the confidence score of each sequence and arranges all sequences in descending order of confidence; if the confidence of a multilingual candidate sequence is less than a dynamic threshold set based on the historical translation quality, it will be deleted from the candidate queue; if the confidence of a multilingual candidate sequence is greater than or equal to the dynamic threshold set based on the historical translation quality, a sliding window is started to calculate the repetition of adjacent sequences; if the repetition between adjacent sequences is greater than the set threshold, the multilingual candidate sequence with the highest confidence is retained; if the difference in confidence between repeated sequences is less than a preset difference, the multilingual candidate sequence with a forward position is retained; if the repeated content spans several multilingual candidate sequences, the entire set of sequences is evaluated to retain the optimal sequence combination; if the repetition between adjacent sequences is less than or equal to the set threshold, the screened multilingual candidate sequence is added to the sequence pool; if the number of sequences in the sequence pool is greater than the capacity limit, the sequence with the lowest confidence is removed.

[0043] Furthermore, a grammar optimizer is established by using a hybrid architecture based on rules and statistics, wherein the grammar optimizer includes a syntactic analysis module, a structural reorganization module and a grammar verification module; a syntactic tree is constructed for the screened multilingual candidate sequences through the syntactic analysis module; based on the grammatical norms and syntactic trees of the target language, sentence patterns are adjusted through the structural reorganization module to extract syntactic dependencies of the multilingual candidate sequences; the positions of sentence components are adjusted according to the word order rules of the target language, the hierarchical structure of long sentences is optimized, grammatical components are supplemented, and the adjusted and reorganized multilingual candidate sequences are verified through the grammar verification module.

[0044] It should be noted that when the sentence length of the multilingual candidate sequence is greater than a preset threshold, the sentence is decomposed into several sentences for processing; when the sentence length of the multilingual candidate sequence is less than or equal to the preset threshold, the structure is reorganized based on the grammatical norms of the target language; if a component that does not conform to the word order of the target language is detected, the position is adjusted according to the word order rules; if a component that conforms to the word order of the target language is detected, the grammar verification stage is entered; if a grammatical error is found, the corresponding correction is made according to the error type; if there are multiple possible expressions, the multilingual candidate sequence with the highest confidence is selected for marking, and this content is manually reviewed.

[0045] Furthermore, the verified multilingual candidate sequences are set with lifecycles and access permissions, and the data is simultaneously updated to several cache nodes through an asynchronous write mechanism to establish a mapping relationship between the source language text and the target language text.

[0046] Furthermore, this embodiment also provides a computer device suitable for the multi-language rapid processing method for an ERP system, comprising a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the multi-language rapid processing method for an ERP system as proposed in the above embodiment.

[0047] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected through a system bus. Among them, the processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covered on the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0048] The present embodiment also provides a storage medium having a computer program stored thereon, which, when executed by a processor, implements the following steps: collecting source language text, and inputting the source language text into a constructed feature recognition model, wherein the feature recognition model establishes a bidirectional language mapping matrix based on a knowledge graph; segmenting the source language text through a semantic segmenter, extracting semantic feature groups based on the bidirectional language mapping matrix, and generating a semantic probability distribution; constructing a fusion transformation network, performing feature matching between the semantic probability distribution and a target language template library, and outputting a multilingual candidate sequence with confidence; screening the multilingual candidate sequence according to the confidence, and performing syntactic reorganization through a grammar optimizer, and updating the reorganized multilingual candidate sequence to an ERP system.

[0049] In summary, the present invention can accurately capture subtle cross-language semantic differences through multi-head cross-attention layers and residual connection structures; combined with dynamic threshold functions and grammar optimizers, it can intelligently screen and reorganize multilingual candidate sequences to ensure the grammatical fluency and semantic consistency of the translation results; through the iterative update mechanism of the knowledge graph, it realizes continuous learning and adaptive optimization of the multilingual processing model, providing enterprises with a highly intelligent multilingual text rapid conversion solution, reducing the technical barriers and communication costs of cross-language business communication, and achieving key technical breakthroughs in multilingual processing of ERP systems.

[0050] Example 2 Referring to Table 1, which is a second embodiment of the present invention, this embodiment provides a multi-language rapid processing method for an ERP system. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0051] Specifically, 6,000 business text data from a multinational enterprise's ERP system were selected as test corpus, involving six languages: English, Chinese, Japanese, German, French, and Spanish. To ensure the representativeness of the data, the test corpus includes texts in multiple fields such as product descriptions, technical specifications, business processes, and customer feedback; the experimental environment uses a server configured with an Intel Xeon 8-core CPU, 64GB memory, and NVIDIA Tesla V100 GPU.

[0052] Furthermore, data from the ERP system was collected through a customized web crawler program, with the collection rate set to 200 items per minute, and distributed crawling was used to ensure system load balancing; the Stanford CoreNLP toolkit was used to preprocess the source language text, including word segmentation, part-of-speech tagging, and dependency syntax analysis; in the knowledge graph construction stage, the Neo4j graph database was used to store language knowledge nodes, and Python scripts were used to automatically build and update the knowledge graph; the association strength between nodes was calculated based on co-occurrence frequency and semantic similarity, and cosine similarity was used as the benchmark metric.

[0053] Furthermore, the feature recognition model adopts a 12-layer Transformer encoder structure, each layer contains 8 attention heads, and the hidden layer dimension is set to 512. The model training uses the Adam optimizer, the learning rate is set to 0.0001, and the batch size is 64. Weight decay and dropout mechanisms are introduced during the training process to prevent overfitting. The semantic segmenter is built based on the BiLSTM network, with a hidden layer dimension of 256 and optimized using the cross entropy loss function.

[0054] Furthermore, the fusion transformation network adopts an encoder-decoder architecture, with the encoder and decoder each containing 6 layers, and each layer using 16 attention heads. The target language template library is built based on 10 million pairs of parallel corpora, and the TF-IDF algorithm is used for feature matching. The beam width of the beam search algorithm is set to 4 to generate diverse translation candidates; the grammar optimizer adopts a rule-based syntactic reorganization method combined with a statistical model for grammar verification to ensure the fluency and accuracy of the output text.

[0055] Specifically, as shown in Table 1, in terms of translation quality, the BLEU score of the present invention reaches 42.6 points, which is 36.5% higher than the 31.2 points of the traditional method. This improvement is mainly due to the knowledge graph and bidirectional language mapping matrix introduced by the present invention, which enables the system to more accurately understand and convert the semantic content of the source language text.

[0056] Table 1. Comparison between the present invention and the traditional method Furthermore, in terms of system performance, the processing speed of the present invention reaches 242 words / second, which is 55.1% higher than the 156 words / second of the traditional method. At the same time, through the optimized architecture design, the memory usage of the present invention is only 35GB, which is 22.2% lower than the traditional method, showing higher resource utilization efficiency; the average response time of the present invention is only 55ms, which is 35.3% lower than the 85ms of the traditional method. This fast response capability has important practical significance for application scenarios that require real-time translation processing.

[0057] Furthermore, in terms of translation accuracy, the accuracy rate increased from 82.5% of the traditional method to 91.3%, an increase of 10.7 percentage points; the recall rate increased from 80.3% to 89.7%, an increase of 11.7 percentage points; the F1 score increased from 81.4% to 90.5%, an increase of 11.2 percentage points. These data fully demonstrate that the present invention can not only accurately identify and convert the semantic features of the source language text, but also maintain a high translation coverage.

[0058] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A multi-language rapid processing method for an ERP system, characterized by: include, Collecting source language text and inputting the source language text into a constructed feature recognition model, wherein the feature recognition model establishes a bidirectional language mapping matrix based on a knowledge graph; Segmenting the source language text through a semantic segmenter, extracting semantic feature groups based on the bidirectional language mapping matrix, and generating semantic probability distribution; Constructing a fusion transformation network, performing feature matching between the semantic probability distribution and the target language template library, and outputting a multilingual candidate sequence with confidence; The multilingual candidate sequences are screened according to the confidence level, and syntax is reorganized through a grammar optimizer, and the reorganized multilingual candidate sequences are updated to the ERP system.

2. The multi-language rapid processing method for an ERP system according to claim 1, characterized in that: The multilingual candidate sequences are screened according to the confidence, and syntax is reorganized by a grammar optimizer, and the reorganized multilingual candidate sequences are updated to the ERP system, including: By setting a dynamic threshold function based on historical translation quality data, the multilingual candidate sequences are sorted in descending order according to confidence and the minimum confidence requirement is determined; Use the sliding window algorithm to detect the repetition of adjacent candidate sequences, and retain the multilingual candidate sequences with a confidence level greater than a threshold and a repetition level less than a set value; A grammar optimizer is established by using a hybrid architecture based on rules and statistics, wherein the grammar optimizer includes a syntax analysis module, a structure reorganization module and a grammar checking module; Constructing a syntax tree for the screened multilingual candidate sequences through the syntax analysis module; Based on the grammatical norms and syntax tree of the target language, the sentence pattern is adjusted by the structure reorganization module to extract the syntactic dependency of the multilingual candidate sequences; Adjust the positions of sentence components according to the word order rules of the target language, optimize the hierarchical structure of long sentences, supplement grammatical components, and verify the adjusted and reorganized multilingual candidate sequences through the grammar verification module; The verified multilingual candidate sequences are set with life cycles and access permissions, and the data is simultaneously updated to several cache nodes through an asynchronous write mechanism to establish a mapping relationship between the source language text and the target language text.

3. The multi-language rapid processing method for an ERP system according to claim 2, characterized in that: The method for obtaining the multilingual candidate sequence is: Establishing a fusion transformation network based on an encoder architecture and a decoder architecture, and using a multi-layer bidirectional feature extraction module on the encoder architecture to capture deep semantic features of the source language text, wherein the fusion transformation network includes an input layer and an output layer; Integrate the target language knowledge through the decoder architecture and build a target language template library based on the massive parallel corpus; Based on the feature embedding matrix of the input layer, mapping the normalized language feature vector probability distribution to a high-dimensional feature space; Using a multi-head cross attention layer to perform feature matching according to the high-dimensional feature space, and calculating the similarity between the source language feature and the target language template through the attention head; Based on the similarity, the input features are projected into different subspaces through a transformation matrix, the attention scores are calculated and the features are reorganized; Based on the feature reorganization, a short-circuit connection is established between the layers of the fusion transformation network through a residual connection structure, so that the shallow layer features are transmitted to the deep layer, wherein the residual connection structure includes a feature addition module and a layer normalization module; The output feature of the current layer is added to the input feature by the feature addition module, and the feature is normalized by the layer normalization module; Based on the standardized features, the beam search algorithm is used to generate multilingual candidate sequences with confidence. A confidence score is calculated based on the multilingual candidate sequence, wherein the confidence score is used for sorting and screening the multilingual candidate sequence.

4. The multi-language rapid processing method for an ERP system according to claim 3, characterized in that: The multi-head cross attention layer includes a transformation matrix; the transformation matrix includes a query transformation matrix, a key transformation matrix and a value transformation matrix; the specific formula of the confidence score is as follows: ； in, is the source language feature vector, is the target language feature vector, is the total feature dimension, is the eigenvalue of the i-th dimension of the source language, is the eigenvalue of the target language ith dimension, is the feature weight of the i-th dimension, is the angle between the source language feature vector and the target language feature vector, is the similarity enhancement function, the specific formula is as follows: ； in, is the feature difference vector, is the feature space dimension, is the edit distance, is the maximum allowed edit distance.

5. The multi-language rapid processing method for an ERP system according to claim 3, characterized in that: The method for generating the language feature vector is: A semantic segmenter is established based on the hierarchical structure of the bidirectional long short-term memory network, and the natural paragraph boundaries of the source language text are identified by the semantic segmenter, and the long paragraphs are divided into text units based on the semantic coherence index, and the hierarchical relationship and context information between the natural paragraphs are preserved; Extracting grammatical structure markers through a bidirectional language mapping matrix, and matching the extracted grammatical structure markers with the segmented text units, wherein the bidirectional language mapping matrix includes grammatical structure markers, word meaning relationship weights, and context association rules; Calculating the semantic association strength between words according to the word meaning relationship weights of the matching results, and analyzing the contextual features of the text unit based on the context association rules; Combining grammatical features, semantic association strength, and contextual features to form a semantic feature group; Processing the semantic feature groups in parallel through a self-attention network to capture semantic relationships, and calculating dot product similarities between query vectors, key vectors, and value vectors to obtain attention weights between features; Performing weighted combination on the feature vectors according to the attention weights to generate a language feature vector; The language feature vector is quantized and converted into a probability value by using a softmax function, and the probability value is mapped to the interval [0, 1] by a normalization method.

6. The multi-language rapid processing method for an ERP system according to claim 5, characterized in that: The grammatical structure tag includes part-of-speech tagging, syntactic dependency and grammatical component analysis results; the word meaning relationship weight represents the semantic relevance between words; the context association rule represents the usage rules of words in different contexts; the language feature vector The specific formula is as follows: ； in, is the attention head index, is the weight of the p-th attention head, is the query transformation matrix of the p-th attention head, is the key transformation matrix of the p-th attention head, is the value transformation matrix of the p-th attention head, is the input feature, is the number of attention heads, is the scaling factor, is the activation function, is the attention calculation function.

7. The multi-language rapid processing method for an ERP system according to claim 5, characterized in that: The method for constructing the bidirectional language mapping matrix is: Collecting source language text from the ERP system through a web crawler program, and formatting the source language text using a text parser to remove special characters and redundant tags; The formatted source language text is divided into basic language units according to the syntactic structure, semantic annotation is performed, and entity relationships and attribute information are extracted; Construct language knowledge nodes based on the annotated basic language units, and connect the nodes to form a mesh structure through semantic relationship edges to establish a knowledge graph, wherein the knowledge graph continuously optimizes the association strength between nodes through an iterative update mechanism; A deep neural network model based on the Transformer architecture is used to construct a feature recognition model according to the knowledge graph, wherein the feature recognition model includes a multi-layer self-attention mechanism and an encoding layer; the multi-layer self-attention mechanism is used to capture the long-distance dependency relationship of the source language text; the encoding layer maps the source language text into a high-dimensional feature vector; Based on the feature recognition model, the language knowledge node pairs in the knowledge graph are extracted, the semantic similarity between the nodes is calculated, and the mapping weights are set to construct a bidirectional language mapping matrix. .

8. The multi-language rapid processing method for an ERP system according to claim 7, characterized in that: The bidirectional language mapping matrix The specific formula is as follows: ； in, is the source language word index, t is the target language word index, is the source language word vector, is the target language word vector, is the grammatical match, is the location distance, is the distance attenuation coefficient, is the context relevance, is the normalized component, n is the vector dimension, and k is the number of knowledge graph relations.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the multi-language rapid processing method for an ERP system as described in any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the multi-language rapid processing method for an ERP system as described in any one of claims 1 to 8 are implemented.

Images