Method and device for intelligent semantic error correction and business term optimization of foreign trade letter electricity

Abstract

The invention relates to the technical field of natural language processing, in particular to a foreign trade letter intelligent semantic error correction and business term optimization method and device, and the method comprises the steps: obtaining a target foreign trade letter, and constructing a target corpus; performing Chinese word segmentation and part-of-speech tagging on the target foreign trade letter; performing term optimization based on the word segmentation result and the knowledge graph; identifying the letter title by using a conditional random field model, and converting the letter title into structured data; a Bi-LSTM-CRF model is adopted to carry out risk point detection, including Bi-LSTM coding, feature engineering and Max-pooling technologies, a part-of-speech sequence is obtained through a softmax function and a Viterbi path, and sequence labeling is carried out to obtain a risk point detection result; and finally, performing Chinese error correction based on the word segmentation result after part-of-speech tagging and the knowledge graph. The recognition and correction accuracy of foreign trade terminologies is improved, and communication obstacles caused by nonstandard use of the terminologies are effectively reduced.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing, in particular to an intelligent semantic error correction and commercial term optimization method and device for foreign trade correspondence, which is suitable for the field of business communication of foreign trade enterprises and used for improving the accuracy, professionalism and standardization of foreign trade correspondence. BACKGROUND

[0002] With the rapid development of global cross-border digital trade, foreign trade correspondence as an important means of international business communication directly affects the smooth progress of trade activities. However, due to language differences, different cultural backgrounds and the complexity of professional terms, foreign trade correspondence often faces the following problems in practical application:

[0003] Firstly, foreign trade professional terms are diverse and complex, and non-professionals are prone to use non-standard terms and unprofessional expressions when writing correspondence. Secondly, there are great differences in business customs among different countries and regions, which can easily lead to ambiguity or misunderstanding in correspondence content. In addition, foreign trade correspondence may imply legal risk points, such as inappropriate clause expression which may cause contract disputes. Finally, there is a lack of professional intelligent error correction and optimization tools for foreign trade correspondence in the current market, which cannot meet the growing demand for cross-border digital trade communication.

[0004] In the prior art, general text correction systems cannot recognize foreign trade professional terms and have high error correction rates; traditional rule matching methods cannot adapt to the rapid update of foreign trade terms; statistical-based methods cannot effectively understand complex foreign trade contexts; and single model methods cannot comprehensively handle multi-dimensional problems in foreign trade correspondence. Therefore, it is of great practical value to develop an intelligent semantic error correction and commercial term optimization method for foreign trade correspondence. SUMMARY

[0005] The present application provides an intelligent semantic error correction and commercial term optimization method and device for foreign trade correspondence, aiming to solve the technical problems of insufficient professionalism, weak risk identification ability and low term correction accuracy of existing foreign trade correspondence processing systems.

[0006] The present application provides an intelligent semantic error correction and commercial term optimization method for foreign trade correspondence, which includes:

[0007] Obtaining a target foreign trade correspondence and constructing a target corpus;

[0008] Performing Chinese word segmentation and part-of-speech tagging on the target foreign trade correspondence to obtain a part-of-speech tagged word segmentation result;

[0009] Based on the word segmentation result and the knowledge graph, performing term optimization on the target foreign trade correspondence according to the rules to obtain a first optimization result;

[0010] Based on the word segmentation result, a conditional random field model is used to recognize the foreign trade letter title to obtain a recognition result.

[0011] The recognition result is converted into structured data.

[0012] Based on the structured data, a Bi-LSTM-CRF model is used to detect risk points of the foreign trade letter, including:

[0013] Based on the structured data, Bi-LSTM encoding is performed to extract the dependency relationship between words and obtain a sentence vector.

[0014] Through feature engineering, a convolution kernel in the CNN model is selected, and a second sentence vector is generated.

[0015] Through Max-pooling pooling technology, global maximum pooling is performed on the second sentence vector.

[0016] The sentence vector and the second sentence vector are combined in a splicing manner to obtain a combined sentence vector.

[0017] The combined sentence vector is converted into a part-of-speech label using a softmax function, and a part-of-speech sequence of Chinese words is obtained using a Viterbi path.

[0018] Based on the part-of-speech sequence, word boundary problems in Chinese business text are sequentially labeled to obtain a risk point detection result.

[0019] Based on the word segmentation result after part-of-speech labeling and the knowledge graph, Chinese error correction of the foreign trade letter is performed.

[0020] As a preferred embodiment, the target corpus is constructed, including:

[0021] The open source English business documents in the field of commercial correspondence are format-converted to serve as an English corpus.

[0022] The obtained target foreign trade letter is format-converted to obtain a target business corpus.

[0023] The processed target foreign trade letter and the English business documents in the English corpus are spliced in a text splicing manner to create and collect Chinese business documents in Chinese, thereby generating the target corpus.

[0024] As a preferred embodiment, based on the word segmentation result and the knowledge graph, business terminology optimization of the target foreign trade letter is performed according to rules to obtain the first optimization result, including:

[0025] The word segmentation result and the text information in the knowledge graph are input into a BPE model.

[0026] constructing a dictionary file based on phrase vocabulary in the target business corpus, segmenting the business text to obtain the first optimization result.

[0027] As preferred, the risk point detection further comprises:

[0028] performing multi-layer Bi-LSTM encoding on the structured data of the converted semantic hierarchy, wherein the output of each layer of Bi-LSTM is taken as the input of the next layer of Bi-LSTM;

[0029] In the execution of the feature engineering, the size and number of convolution kernels in the CNN model are dynamically adjusted according to the text characteristics of foreign trade letters and cables;

[0030] In the global maximum value pooling, the maximum value of each feature dimension is retained to form a fixed-length global feature representation.

[0031] As preferred, the Chinese error correction further comprises:

[0032] Based on the word segmentation result after the part-of-speech tagging and the knowledge graph, through an LSTM model based on an attention mechanism, the representation of key business words is learned from the business text in a splicing manner to obtain a second optimization result.

[0033] As preferred, the method of the LSTM model based on the attention mechanism comprises:

[0034] Based on the knowledge of the vocabulary in the target business text, a knowledge embedding vector of the knowledge graph is generated to obtain a word representation of the target business text;

[0035] Taking the word representation as the input of the LSTM network, a representation based on the LSTM network is obtained;

[0036] The knowledge embedding vector and the representation based on the LSTM network are input into the attention mechanism to obtain a representation of the target business text;

[0037] The obtained text representation, word representation of the target business text and representation of the LSTM network are input into a feedforward neural network to obtain the second optimization result.

[0038] As preferred, after the risk point detection, further comprising:

[0039] constructing a compliance module for evaluating whether the foreign trade letter and cable is compliant;

[0040] preprocessing the text data of the foreign trade letter and cable, and converting the text data into a digital vector matrix;

[0041] establishing a mapping relationship between the foreign trade letter and the foreign trade letter compliance;

[0042] analyzing the compliance of the input text according to the mapping relationship, wherein the compliance analysis includes a multi-classification task and a multi-label task.

[0043] As preferred, after the risk point detection, it further includes:

[0044] building a translation module for intelligently translating the foreign trade letter;

[0045] The translation module includes an encoder-decoder model, wherein the encoder adopts a self-attention mechanism to encode the text into a sequence representation;

[0046] After encoding by the encoder, the input sequence is converted into a fixed-dimensional vector;

[0047] The decoding is performed by the decoder to obtain a translation result.

[0048] As preferred, after the Chinese error correction, it further includes:

[0049] building a data error correction module for comparing the risk point detection result with the business terms in the international business terminology library;

[0050] comparing the risk point detection result with the business terms in the international business terminology library, and outputting the corresponding international business terms;

[0051] building an audit and correction module for auditing and correcting the international business terms;

[0052] comparing the international business terms identified by the data error correction module with the actual output data;

[0053] updating the international business terminology library according to the comparison result.

[0054] The foreign trade letter intelligent semantic error correction and business term optimization device comprises:

[0055] A target corpus creation module is configured to obtain target foreign trade letters and build a target corpus.

[0056] A word segmentation module is configured to perform Chinese word segmentation and part-of-speech tagging on the target foreign trade letters to obtain a part-of-speech tagged word segmentation result.

[0057] A business term optimization module is configured to perform term optimization of the target foreign trade letters according to rules based on the word segmentation result and a knowledge graph to obtain a first optimization result.

[0058] The recognition module is configured to recognize the foreign trade letter title based on the word segmentation result by using a conditional random field model to obtain a recognition result.

[0059] The conversion module is configured to convert the recognition result into structured data.

[0060] The detection module is configured to detect risk points of the foreign trade letter based on the structured data by using a Bi-LSTM-CRF model, including: performing Bi-LSTM encoding based on the structured data to extract the dependency relationship between words and obtain a sentence vector; selecting a convolution kernel in a CNN model through feature engineering and generating a second sentence vector; performing global maximum pooling on the second sentence vector through a Max-pooling pooling technique; combining the sentence vector and the second sentence vector in a splicing manner to obtain a combined sentence vector; converting the combined sentence vector into a part-of-speech label by using a softmax function, and obtaining a part-of-speech sequence of Chinese words by using a Viterbi path; and performing sequence labeling on the word boundary problem in the Chinese business text based on the part-of-speech sequence to obtain a risk point detection result.

[0061] The error correction module is configured to perform Chinese error correction of the foreign trade letter based on the part-of-speech labeled word segmentation result and the knowledge graph.

[0062] The present application realizes comprehensive processing and optimization of foreign trade letters by using innovative technologies such as multi-model fusion, knowledge graph application, and intelligent risk point detection, and has the following beneficial effects:

[0063] 1. The recognition and correction accuracy of foreign trade professional terms is improved, and the communication barrier caused by non-standard use of terms is effectively reduced.

[0064] 2. The potential legal risk points in the foreign trade letter are accurately identified, and possible commercial disputes are effectively prevented.

[0065] 3. Intelligent semantic error correction of foreign trade letters is realized, and the professionalism and standardization of the letter are improved.

[0066] 4. The term optimization mechanism based on the knowledge graph enables the system to have continuous learning and updating capabilities.

[0067] 5. Multiple models work together to form a complete processing pipeline, improving the processing efficiency and accuracy of the system.

[0068] 6. The customization design for the characteristics of foreign trade letters solves the problem of poor applicability of general text processing systems in the foreign trade field. BRIEF DESCRIPTION OF DRAWINGS

[0069] Figure 1A flowchart of the method for intelligent semantic error correction and business term optimization of foreign trade correspondence provided by the embodiment of the present application is shown in the figure.

[0070] Figure 2 A structural block diagram of the device for intelligent semantic error correction and business term optimization of foreign trade correspondence provided by the embodiment of the present application is shown in the figure.

[0071] Figure 3 A flowchart of the method for risk point detection based on the Bi-LSTM-CRF model provided by the embodiment of the present application is shown in the figure.

[0072] Figure 4 A structural diagram of the LSTM model based on the attention mechanism provided by the embodiment of the present application is shown in the figure.

[0073] Figure 5 A flowchart of the process of constructing and updating the knowledge graph provided by the embodiment of the present application is shown in the figure.

[0074] Figure 6 A structural diagram of the translation module of foreign trade correspondence provided by the embodiment of the present application is shown in the figure.

[0075] Figure 7 A workflow diagram of the compliance analysis module provided by the embodiment of the present application is shown in the figure. DETAILED DESCRIPTION

[0076] Please refer to the accompanying Figures 1-7 , the present application will be further described in detail below in combination with the drawings and embodiments.

[0077] Reference Figure 1 , Figure 1 A flowchart of the method for intelligent semantic error correction and business term optimization of foreign trade correspondence provided by the embodiment of the present application is shown in the figure. As shown in the figure, Figure 1 the method comprises the following steps:

[0078] Step S1, obtaining a target foreign trade correspondence and constructing a target corpus:

[0079] In an embodiment of the present application, constructing the target corpus comprises the following sub-steps:

[0080] Step S11: converting the open-source English business documents in the field of business correspondence into a format, serving as an English corpus. Preferably, the format conversion comprises converting various formats such as PDF and DOC into a pure text format, facilitating subsequent processing. At the same time, the text is uniformly encoded and stored in UTF-8 format. In practice, about 50,000 English business documents can be collected from the business correspondence documents published by authoritative organizations such as the International Chamber of Commerce (ICC) and the United Nations Commission on International Trade Law (UNCITRAL). For example, for a standard international trade contract template, the system will extract the clause structure, professional terms and expression methods therein, providing a reference basis for subsequent term optimization.

[0081] Step S12: Convert the format of the acquired target foreign trade correspondence to obtain the target business corpus. Specifically, collect different types of foreign trade correspondence from the company's actual business, including inquiry letters, quotations, order confirmations, shipping notices, payment notices, etc., and standardize their formats.

[0082] Step S13: The processed target foreign trade correspondence and English business documents from the English corpus are concatenated to create and collect Chinese business documents, generating the target corpus. Preferably, parallel corpus organization is used to organize the corresponding Chinese and English content, facilitating subsequent terminology matching and translation. In this embodiment, the final target corpus contains approximately 200,000 data entries, covering various sources such as foreign trade dictionaries, professional foreign trade emails, and publicly available internet data.

[0083] Step S2: Perform Chinese word segmentation and part-of-speech tagging on the target foreign trade correspondence to obtain the segmentation results after part-of-speech tagging.

[0084] In one embodiment of the present invention, Chinese word segmentation is performed using a maximum matching algorithm combined with a professional foreign trade dictionary. Specifically, the maximum matching word length is preset to 15 characters, and matching starts from the first word. The text is preprocessed to obtain segmented independent word groups. Each independent word is searched in the dictionary, and words with a length greater than 2 are matched to form sentences. When a single word cannot form a new sentence, the longest matching word is used.

[0085] For example, traditional word segmentation systems may incorrectly segment the compound term "FOB Shenzhen price" commonly used in foreign trade correspondence into "FOB / Shenzhen / price," while this system can correctly identify it as the whole term "FOB Shenzhen price," thus providing an accurate basis for subsequent terminology optimization and risk detection.

[0086] Meanwhile, the ICU algorithm is used to eliminate abnormal symbols in the text from three dimensions: First, phrases longer than 10 characters in the sentence are segmented and Chinese characters in them are identified and processed separately; Second, the strings are sorted and processed according to the UTF-8 encoding standard; Third, for non-Chinese characters, they are judged and replaced in the order of numbers, letters, and other characters. For numbers and letters, prompts are given in Chinese annotation form, and the length of the prompts is controlled within 6 characters.

[0087] The词性标注采用条件随机场(CRF)模型,针对外贸领域的特殊词汇进行专门训练,标注精度达到95%以上。例如,对于"CIF"这样的术语,系统能够正确标注为"外贸术语(FTT)",而不是简单的"名词(N)"或"缩写(ABBR)",为后续的风险识别提供更精准的语法信息。

[0088] Step S3, based on the word segmentation result and the knowledge graph, optimize the terms of the target foreign trade correspondence according to the rules to obtain the first optimization result:

[0089] In the embodiment of the present invention, this step specifically includes:

[0090] Step S31: Input the text information in the word segmentation result and the knowledge graph into the BPE (Byte Pair Encoding) model. The BPE model is a sub-word segmentation algorithm, which is especially suitable for processing compound words and professional terms in the foreign trade field. Preferably, the vocabulary size of the BPE model is set to 32000, and the number of iterations is 1000 times.

[0091] Step S32: Construct a dictionary file based on the phrase vocabulary in the target business corpus, segment the business text, and obtain the first optimization result. Specifically, extract phrases with a frequency of more than 5 times from the target business corpus as candidates, and add them to the dictionary file after manual screening. For terms that are not in the dictionary but appear according to a certain pattern, identify and standardize them through regular expressions.

[0092] In this embodiment, the term optimization mainly aims at the following types of problems: standardization of non-standard abbreviations (such as standardizing "FOB." to "FOB"); normalization of term capitalization (such as normalizing "cif" to "CIF"); correction of term spelling mistakes (such as correcting "proformainvocie" to "proformainvoice"); normalization of Chinese terms (such as normalizing "装箱单" to "箱单"), etc.

[0093] Specific application scenario example: When the system processes a foreign trade email containing "weconfirmtheorderwillbeshippedviacifshenzhen", after term optimization processing, the output result is "Weconfir mtheorderwillbeshippedviaCIFShenzhen", standardizing "cif" to the capitalized standard international trade term "CIF", which improves the professionalism and standardization of the correspondence.

[0094] Step S4, based on the word segmentation result, use the conditional random field model to identify the foreign trade correspondence title and obtain the identification result:

[0095] In the embodiments of the present application, title recognition is a key step in foreign trade letter processing, because the title usually contains the core intention and type information of the letter. Conditional random field model is particularly suitable for sequence labeling tasks and can consider context information to improve the accuracy of title recognition.

[0096] Preferably, the feature templates of the conditional random field model include: word features (current word, previous word, next word); part-of-speech features (current part-of-speech, previous part-of-speech, next part-of-speech); position features (whether at the beginning of the sentence, whether at the beginning of the paragraph); length features (word length, sentence length); format features (capitalization ratio, number ratio), etc.

[0097] The model training uses gradient descent method, and the initial learning rate is set to 0.01, which gradually decreases as the training round increases. Feature weight regularization uses L2 regularization, and the regularization coefficient is 0.1. In this embodiment, the title recognition accuracy is above 97%.

[0098] Actual application example: For the beginning text of a mail "Re: Quotation for 1000 pcs of Model XYZ-123", the system can accurately identify that it is a quotation reply mail and extract key information such as product model "XYZ-123" and quantity "1000 pcs", providing accurate context information for subsequent risk detection and content analysis.

[0099] Step S5, converting the recognition result into structured data:

[0100] In the embodiments of the present application, the recognition result is converted into structured data at the semantic level, providing a basis for subsequent risk point detection. The conversion process includes:

[0101] 1. Title information structuring: extracting key elements in the title, such as letter type, date, number, etc.

[0102] 2. Text segmentation: dividing the text into paragraphs and marking the function of each paragraph (such as opening, content, closing, etc.);

[0103] 3. Term labeling: labeling professional terms in the text and establishing a term index;

[0104] 4. Entity recognition: identifying and labeling entities such as names, company names, locations, dates, amounts, etc. in the text;

[0105] 5. Relationship extraction: extracting relationships between entities in the text, such as supplier-customer, product-price, etc.

[0106] Preferably, the structured data is stored in JSON format for easy processing and analysis. Each piece of email data contains two parts: metadata (e.g. ID, date, source, etc.) and content data (title, body, segments, tags, etc.).

[0107] Application example: For an email containing a product inquiry, the structured data may be as follows:

[0108] json

[0109] {

[0110] "metadata":{

[0111] "id":"E202506050001",

[0112] "date":"2025-06-05",

[0113] "type":"Inquiry"

[0114] },

[0115] "content":{

[0116] "title":"InquiryforProductXYZ-123",

[0117] "entities":[

[0118] {"type":"product","value":"XYZ-123","positions":[12,19]},

[0119] {"type":"company","value":"ABCTradingCo.","positions":[120,135]}

[0120] ],

[0121] "terms":[

[0122] {"term":"FOB","positions":[220,223]},

[0123] {"term":"L / C","positions":[350,353]}

[0124] ],

[0125] "paragraphs":[

[0126] "1":"ThisisaninquiryforproductXYZ-123.",{"type":"greeting","text":"Dear Sir / Madam,","position":0},

[0127] {"type":"inquiry","text":"We are interested in your product XYZ-123...","position":1},

[0128] {"type":"closing","text":"Looking forward to your reply. ","position":5} ]

[0130] }

[0131] }

[0132] Step S6, based on the structured data, using Bi-LSTM-CRF model, foreign trade letter risk point detection is carried out:

[0133] Referring to Figure 3 , Figure 3 The flow chart of the risk point detection method based on the Bi-LSTM-CRF model is shown. As Figure 3 shown, the step specifically comprises:

[0134] Step S61: based on the converted semantic hierarchical structured data, Bi-LSTM encoding is carried out, the dependency relationship between the words is extracted, and the sentence vector A1 is obtained.

[0135] In the embodiment, the Bi-LSTM (Bidirectional Long Short-Term Memory Network) model is composed of a forward LSTM and a backward LSTM, which can capture the forward and backward context information of the text at the same time. The specific implementation is as follows:

[0136] For the input sequence X={x1,x2,...,x n}, wherein x i represents the embedding representation of the i-th word, the calculation process of the forward LSTM is as follows:

[0137]

[0138] The calculation process of the backward LSTM is as follows:

[0139]

[0140] The hidden states of the two directions are combined to obtain:

[0141]

[0142] Finally, the sentence vector A1 = {h1, h2,..., h n} is obtained, where h t represents the bidirectional context representation of the t-th word.

[0143] where: X is an input sequence of length n; x t is the embedding representation of the t-th word in the sequence, with dimension d e (d e = 300 in this embodiment); is the hidden state of the forward LSTM at position t, with dimension d h (d h = 256 in this embodiment); is the hidden state of the backward LSTM at position t, with the same dimension d h ; h t is the merged hidden state, with dimension 2d h ; [;] represents the vector concatenation operation; LSTM(·) represents the calculation function of the LSTM unit.

[0144] Preferably, the hidden layer dimension of Bi-LSTM is set to 256, the number of layers is 2, Dropout is used to prevent overfitting, and the Dropout rate is 0.5. The word embedding dimension is 300, and a pre-trained word vector in the foreign trade field is used for initialization.

[0145] In the processing of foreign trade correspondence, this step can effectively capture the contextual relationship between terms. For example, when processing a payment clause containing "payment terms: 30% deposit, 70% before shipment", Bi-LSTM can understand the relationship between "30% deposit" and "70% before shipment", thereby providing a complete semantic background for subsequent risk point detection.

[0146] Step S62: Through feature engineering, select the convolution kernel in the CNN model, and generate the second sentence vector A2.

[0147] In this embodiment, a convolutional neural network (CNN) is used to extract local semantic features. The specific implementation is as follows:

[0148] Define a multi-size convolution kernel K = {k1, k2,..., k m}, where k i represents a convolution kernel with window size i. For the sentence vector A1, apply each convolution kernel for convolution operation:

[0149] c i = Conv(A1, k i),

[0150] The output results of different convolution kernels are spliced to obtain:

[0151] C = [c1; c2;... ; c m ],

[0152] A second sentence vector is obtained by applying a ReLU activation function:

[0153] A2 = ReLU(C),

[0154] wherein K is a set of convolution kernels, containing m convolution kernels of different window sizes; k i represents a convolution kernel with a window size of i; Conv(·) represents a convolution operation; c i is a feature map obtained using the convolution kernel k i ; C is a spliced feature vector; ReLU(·) represents a rectified linear unit activation function, defined as ReLU(x) = max(0, x); A2 is a generated second sentence vector.

[0155] Preferably, the window size of the convolution kernel is set to 3, 4, or 5 words, and the number of convolution kernels of each size is 128, which can capture local features of different scales. In practice, according to the characteristics of foreign trade terminology, it is found that the convolution kernel with a window size of 3 works best for short terms, and the convolution kernel with a window size of 5 works best for compound terms and habitual expressions.

[0156] In practical applications, this step can effectively identify key phrases in foreign trade correspondence. For example, for the expression "goodswillbeshippedassoonaspossible", CNN can identify that "assoonaspossible" is a vague time expression, which may have a risk of delivery period, and is marked as a potential risk point.

[0157] Step S63: Perform global maximum pooling on the second sentence vector by Max-pooling technology.

[0158] The global maximum pooling operation is as follows:

[0159] For each feature dimension j of the second sentence vector A2, perform:

[0160] pool j = max(A2[:,j)),

[0161] to obtain the pooled feature vector Pool = [pool_{1}, pool_{2}, …, pool_{d}], where d is the feature dimension.

[0162] Where: A2[:,j] represents all values of the jth feature dimension in the second sentence vector A2; max() represents the maximum value operation; pool_{j} is the maximum value of the jth feature dimension; Pool is the pooled feature vector with dimension d (same as the feature dimension).

[0163] Preferably, the pooling operation retains the maximum value of each feature dimension, forming a fixed-length global feature representation that is not affected by the length of the input text. This method is particularly suitable for processing foreign trade correspondence of varying lengths, while significantly reducing computational complexity.

[0164] In foreign trade correspondence processing, global pooling can extract the most important features in the entire document. For example, when processing a long contract email, even if the key risk clause is hidden in the middle of the document, this technique can effectively capture and highlight this information.

[0165] Step S64: Combine the sentence vector A1 and the second sentence vector A2 using concatenation to obtain the combined sentence vector A3.

[0166] The combination operation is:

[0167] A3 = Concat(A1, Pool),

[0168] Where: Concat() represents the vector concatenation operation; A1 is the sentence vector encoded by Bi-LSTM, capturing sequence-level dependencies; Pool is the feature vector after global pooling, capturing global semantic features; A3 is the concatenated combined sentence vector.

[0169] This multi-level feature fusion mechanism is one of the core innovations of the present application, which can simultaneously focus on local details and global semantics, and is particularly suitable for complex context understanding in foreign trade correspondence.

[0170] For example, when processing a foreign trade contract containing multiple clauses, local features can capture the specific content of each clause, while global features can understand the structure and main points of the entire contract, and the combination of the two can more accurately identify potential risk points.

[0171] Step S65: Convert the combined sentence vector to a part-of-speech tag using the softmax function, and use the Viterbi path to obtain the part-of-speech sequence of Chinese words.

[0172] The specific implementation is as follows:

[0173] First, use a linear layer and softmax function to calculate the label probability distribution at each position:

[0174] P(y|A3) = softmax(W·A3 + b),

[0175] Then, the CRF layer models the label transition probability:

[0176] Score(X, y) =∑ i (P(y i |A3) + T(y i-1 , y i )),

[0177] Finally, the Viterbi algorithm is used to solve the optimal label sequence:

[0178]

[0179] where P(y|A3) represents the probability distribution of label y given the combined sentence vector A3; softmax(·) represents the normalized exponential function, defined as W is the weight matrix, b is the bias vector; Score(X, y) represents the score of the input sequence X corresponding to the label sequence y; y i represents the label at position i; T(y i-1 , y i ) represents the transition score from label y i-1 to label y i ;∑ i represents the summation over all positions i; y * represents the label sequence with the highest score; argmax y represents finding the label sequence y that maximizes the score.

[0180] Preferably, the label set includes common part-of-speech tags (such as nouns, verbs, adjectives, etc.) and specific foreign trade terminology tags (such as trade terms, currency units, transportation methods, etc.). In this embodiment, a total of 42 labels are defined, of which 16 are foreign trade professional labels.

[0181] In foreign trade letter processing, this step can accurately identify the part-of-speech and function of professional terms. For example, the system can recognize "30-day L / C" as a complete payment method term, rather than splitting it into the number "30", the hyphen "-", the noun "day", and the abbreviation "L / C", thereby providing accurate grammatical structure information for risk point detection.

[0182] Step S66: Based on the part-of-speech sequence, sequence labeling is performed on the word boundary problem in the Chinese business text to obtain the risk point detection result.

[0183] In embodiments of the present application, risk point detection mainly focuses on the following types of problems:

[0184] 1. Payment clause risk: such as ambiguous payment time, unclear payment method, etc.

[0185] 2. Risks related to delivery terms: such as unclear delivery time, place, or conditions;

[0186] 3. Risks related to quality clauses: such as unclear quality standards or vague acceptance conditions;

[0187] 4. Risks associated with liability clauses: such as unclear liability for breach of contract or ambiguous compensation standards;

[0188] 5. Risks related to force majeure clauses: such as missing clauses or unclear wording.

[0189] For detected risk points, the system will mark their location, type, and risk level (high, medium, low) to facilitate subsequent error correction.

[0190] In a specific example, given the input text "We will ship the goods within 3 days after receiving your payment.", the system detected that "within 3 days" is a delivery time clause, but lacks a clear starting point for calculation (whether it is calendar days or working days), marking it as a medium-risk level and suggesting modification to "within 3 working days" to avoid potential disputes.

[0191] Similarly, the system will mark the statement "Quality should meet standard requirements" as a high-risk quality clause because "standard requirements" is too vague. It is recommended to change it to a specific quality standard such as "Quality should comply with ISO 9001:2015 standards".

[0192] Step S7: Based on the word segmentation results after part-of-speech tagging and the knowledge graph, perform Chinese error correction for foreign trade correspondence:

[0193] In embodiments of the present invention, Chinese character correction further includes:

[0194] Based on the word segmentation results after part-of-speech tagging and the knowledge graph, the representation of key business words in business text is learned from the business text by using an attention-based LSTM model and a concatenation method, thus obtaining the second optimization result.

[0195] See Figure 4 , Figure 4 The diagram illustrates the structure of an LSTM model based on an attention mechanism provided in an embodiment of the present invention. Figure 4 As shown, the method of this model includes:

[0196] First, based on the vocabulary knowledge in the target business text, knowledge embedding vectors of the knowledge graph are generated, resulting in the word representation c of the target business text. Preferably, the knowledge embedding employs the TransE algorithm, mapping entities and relationships in the graph to a low-dimensional continuous vector space. The embedding dimension is set to 100, the training epochs are 1000, and the learning rate is 0.01.

[0197] Then, the word representation c is used as input to the LSTM network to obtain the LSTM-based representation s. The computation process of the LSTM network is as follows:

[0198] f t =σ(W f ·[h t-1 ,x t ]+b f ),

[0199] i t =σ(W i ·[h t-1 ,x t ]+b i ),

[0200]

[0201] o t =σ(W o ·[h t-1 ,x t ]+b o ),

[0202] h t =o t *tanh(C t ),

[0203] Where: f t This represents the forget gate, which controls the proportion of previous state information retained; i t This represents the input gate, controlling the proportion of the current input information; o t This represents the output gate, which controls the proportion of the output in the current state. Indicates the candidate state; C t Indicates the unit state; h t The hidden state represents the output of the LSTM; σ represents the sigmoid function, defined as follows: tanh represents the hyperbolic tangent function; W f W i W C W o b represents the weight matrix of the corresponding gate; f ,b i ,b C ,b oRepresents the corresponding bias vector; [h t-1 ,x t ] indicates that the hidden state h from the previous moment is being displayed. t-1 With the current input x t Concatenation; * indicates element-wise multiplication.

[0204] Next, the knowledge embedding vector c and the LSTM-based representation s are input into the attention mechanism to obtain the representation l of the target business text:

[0205]

[0206] Where: s i This represents the output of the LSTM network, i.e., the context representation of the i-th word; c i This represents the embedding of the knowledge graph for the i-th word; i is the index of the i-th word in the word sequence; e i v represents the attention score for the i-th word; a W a and U a α represents the learnable parameters of the attention mechanism; tanh is the hyperbolic tangent function; exp(·) represents the exponential function; α i represents the attention weight of the i-th word, obtained through softmax normalization; l represents the weighted text representation.

[0207] Finally, the obtained text representation l, the word representation c of the target business text, and the LSTM network representation s are input into a feedforward neural network to obtain the second optimization result o:

[0208] o = W o [l;c;s]+b o ,

[0209] Where: [l; c; s] represents concatenating the text representation l, the word representation c, and the LSTM network representation s; W o This is the weight matrix of the feedforward neural network; b o is the bias vector of the feedforward neural network; o represents the final second optimization result.

[0210] In an embodiment of the present invention, the feedforward neural network adopts a two-layer structure, with a hidden layer dimension of 512 and an activation function of ReLU. The output layer uses the softmax function to generate the final error correction result.

[0211] Specific application scenario example: When the system processes Chinese foreign trade correspondence such as "Our company can produce according to your required product specifications, with a delivery period of 40 days after order confirmation, and payment by sight letter of credit," the attention mechanism will identify "delivery period" and "payment method" as key terms and focus on the accuracy of their wording. The system may detect that the wording "within 40 days" is not precise enough (whether it refers to calendar days or working days), and that "sight letter of credit" should be standardized as "sight irrevocable letter of credit," thus providing corresponding error correction suggestions.

[0212] Using the methods described above, the system can identify and correct errors in Chinese expression in foreign trade correspondence, such as typos, inappropriate word choice, and grammatical errors. Furthermore, the knowledge graph-based approach enables the system to understand specific contexts and expression habits within the foreign trade field, improving the accuracy and professionalism of error correction.

[0213] In another embodiment of the invention, after risk point detection, the method further includes:

[0214] Construct a compliance module for evaluating the compliance of foreign trade correspondence; preprocess the text data of foreign trade correspondence and convert it into a numerical vector matrix; establish a mapping relationship between foreign trade correspondence and its compliance; analyze the compliance of the input text based on the mapping relationship, including multi-classification and multi-label tasks.

[0215] See Figure 7 , Figure 7 A flowchart illustrating the workflow of the compliance analysis module provided in an embodiment of the present invention is shown. Figure 7 As shown, the specific implementation of this module is as follows:

[0216] First, preprocessing includes word segmentation and word vector transformation. Word segmentation uses the same method as in step S2, and word vector transformation uses a pre-trained word embedding model to map each word into a 300-dimensional vector.

[0217] Then, establish a mapping relationship between foreign trade correspondence and its compliance:

[0218] φ:X→Y,

[0219] Where: φ represents the mapping operation on the vector space, mapping the input space to the output space; X represents the vector space of the preprocessed text, where each word is represented as x. i Y represents the compliance evaluation space, which is a subset of the positive integer space, i.e. Each element represents a compliance evaluation category or label.

[0220] Compliance analysis includes multi-class classification tasks and multi-label tasks. Multi-class classification tasks use the cross-entropy loss function:

[0221]

[0222] Where: L CE Represents cross-entropy loss; C represents the number of classes; y i This represents the true label, which is 1 if the sample belongs to class i, and 0 otherwise. denoted as the predicted probability, denoted as the probability that the sample belongs to the i-th class; log represents the natural logarithm.

[0223] Multi-label tasks use the Hamming Loss function:

[0224]

[0225] Where: L HL Represents Hamming Loss; N represents the number of samples; L represents the number of labels; y ij This represents the true value of the j-th label of the i-th sample; This represents the predicted value of the j-th label for the i-th sample; This indicates an indicator function that has a value of 1 when the condition is true and 0 otherwise.

[0226] Preferably, the compliance assessment covers the following dimensions: legal compliance (such as whether it complies with international trade regulations), business ethics compliance (such as whether there are improper promises or misleading statements), cultural sensitivity (such as whether there are expressions of cultural conflict), and information security (such as whether sensitive business information has been leaked).

[0227] Example application scenario: When the system analyzes an email containing "We guarantee 100% product quality and zero defect rate," the compliance module will flag it as "Business Ethical Hazard - Over-promise" because "100% quality" and "zero defects" are impossible absolute promises and could lead to subsequent disputes. The system will suggest changing it to "We strive for high quality standards and minimal defect rate."

[0228] In yet another embodiment of the invention, after risk point detection, the method further includes:

[0229] A translation module for intelligent translation of foreign trade correspondence is constructed. The translation module includes an encoder-decoder model, in which the encoder adopts a self-attention mechanism to encode the text into a sequence representation. After encoding by the encoder, the input sequence is converted into a vector of fixed dimensions. The decoder then decodes the input sequence to obtain the translation result.

[0230] See Figure 6 , Figure 6 A structural diagram of the foreign trade correspondence translation module provided in an embodiment of the present invention is shown.

[0231] like Figure 6 As shown, the specific implementation of this module is as follows:

[0232] The encoder uses a self-attention mechanism to encode text into a sequence representation. The calculation process of the self-attention mechanism is as follows:

[0233] Q = W Q X,K=W K X,V=W V X,

[0234]

[0235] Where: X represents the embedding representation of the input sequence, with dimensions (n, d) model ), where n is the sequence length, d model For model dimensions; W Q W K W V Let represent the learnable parameter matrices for the query, key, and value, respectively, each with dimension (d). model ,d k Q, K, and V represent the query, key, and value matrices, respectively; QK T This indicates that the attention score is calculated by multiplying the query matrix by the transpose of the key matrix; `Q` is a scaling factor to prevent the dot product result from becoming too large; `softmax` indicates softmax normalization performed row by row; `Attention(Q,K,V)` represents the attention calculation result, with the same dimension as `V`.

[0236] After encoding, the input sequence is converted into a fixed-dimensional vector. The overall structure of the encoder-decoder model is as follows:

[0237] Φ(X)=g(f(X)),

[0238] Where: Φ represents the encoder-decoder model; f represents the encoder function, i.e., the nonlinear mapping function, which maps the input sequence X to an intermediate representation; g represents the decoder function, which maps the intermediate representation to the target sequence; and X represents the input sequence.

[0239] Preferably, the translation module is specifically trained for the foreign trade field, supporting multilingual translation between Chinese and English, Chinese and Japanese, Chinese and German, and Chinese and French, with particular attention to the accurate translation of foreign trade terminology. The model adopts a Transformer architecture, with 6 layers each for the encoder and decoder, 8 attention heads, a model dimension of 512, and a feedforward network dimension of 2048.

[0240] Example application scenario: When translating Chinese foreign trade correspondence such as "Our company can provide FOB Shenzhen price, shipping period is 30 working days after order confirmation," a general translation system might incorrectly translate it as "My company can provide FOB Shenzhen price, shipping period is 30 working days after order confirmation," while this system will accurately translate it as "We offer prices on FOB Shenzhen terms. The shipping period is 30 working days after order confirmation," correctly translating the professional terminology and making the expression more in line with English foreign trade habits.

[0241] In another embodiment of the present invention, after Chinese character correction, the method further includes:

[0242] Construct a data correction module to compare risk point detection results with business terms in the international business terminology database; compare risk point detection results with business terms in the international business terminology database and output the corresponding international business terms; construct an audit and correction module for reviewing and correcting international business terms; compare the international business terms identified by the data correction module with the actual output data; update the international business terminology database based on the comparison results.

[0243] See Figure 5 , Figure 5 A flowchart illustrating the knowledge graph construction and updating process provided in an embodiment of the present invention is shown.

[0244] like Figure 5 As shown, the specific implementation of this process is as follows:

[0245] First, the International Business Terminology Database is constructed using a knowledge graph format, containing approximately 50,000 foreign trade professional terms, covering multiple areas such as trade terms, modes of transport, payment methods, and product descriptions. Each node in the knowledge graph represents a term, and the edges between nodes represent the relationships between terms, such as "is a," "contains," and "used for."

[0246] The data error correction module matches the risk point detection results with terms in the knowledge graph, using a fuzzy matching algorithm to handle possible variations and erroneous forms. The matching process considers multiple factors such as edit distance, word form similarity, and semantic similarity, and selects the best matching term after comprehensive scoring.

[0247] The review and correction module manually reviews the matching results, confirming correct matches and marking incorrect matches. Based on the review results, the system automatically updates the knowledge graph, including adding new terms, correcting erroneous relationships, and adjusting term weights. Preferably, the knowledge graph updates adopt an incremental learning approach, updating weekly to maintain the timeliness and accuracy of the terminology database.

[0248] Application scenario example: When the system processes foreign trade correspondence containing the term "payment via T / T," the data correction module compares it with standard terms in the knowledge graph. If it finds that the standard expression should be "payment by T / T" or the more formal "payment by Telegraphic Transfer," the system will automatically correct it. Simultaneously, if new expressions such as "payment through bankwire" are discovered in actual business operations, they will be added to the knowledge graph after review, expanding the terminology database.

[0249] See Figure 2 , Figure 2 This diagram illustrates the structural block diagram of the intelligent semantic error correction and business terminology optimization device for foreign trade correspondence provided in an embodiment of the present invention. Figure 2 As shown, the device includes:

[0250] Target Corpus Creation Module 1 is configured to: acquire target foreign trade correspondence and construct a target corpus;

[0251] Module 2 is configured to perform Chinese word segmentation and part-of-speech tagging on the target foreign trade correspondence, and obtain the segmentation results after part-of-speech tagging.

[0252] The business terminology optimization module 3 is configured to: optimize the terminology of the target foreign trade correspondence based on word segmentation results and knowledge graph, and obtain the first optimization result;

[0253] Recognition module 4 is configured to: based on the word segmentation results, use a conditional random field model to recognize the title of foreign trade correspondence and obtain the recognition result;

[0254] Transformation module 5 is configured to convert the recognition results into structured data;

[0255] Detection module 6 is configured to detect risk points in foreign trade correspondence using a Bi-LSTM-CRF model based on structured data.

[0256] Error correction module 7 is configured to perform Chinese error correction on foreign trade correspondence based on word segmentation results after part-of-speech tagging and knowledge graph.

[0257] Preferably, the device may further include: a compliance module 8 for evaluating whether foreign trade correspondence is compliant; a translation module 9 for intelligent translation of foreign trade correspondence; a data correction module 10 for comparison with an international business terminology database; and an audit and correction module 11 for auditing and updating international business terms.

[0258] The device of this invention can be implemented as a software system deployed on a server to provide API services, or as a desktop application or web application for direct use by foreign trade professionals. Recommended system configuration: 8 or more CPU cores, 16GB or more of memory, 100GB or more of storage space, and support for GPU acceleration.

[0259] In practical applications, the intelligent semantic error correction and business terminology optimization method and apparatus for foreign trade correspondence of the present invention can significantly improve the quality and professionalism of foreign trade correspondence, reduce communication costs and legal risks, and provide strong technical support for foreign trade enterprises. Compared with the prior art, the present invention has higher professionalism, stronger risk identification capabilities, and more accurate terminology optimization effects, meeting the growing demand for intelligentization in the foreign trade field.

[0260] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent semantic error correction and business terminology optimization in foreign trade correspondence, characterized in that: Includes the following steps: Acquire target foreign trade correspondence and construct a target corpus; The target foreign trade correspondence was segmented into Chinese and tagged with parts of speech to obtain the segmented results after part-of-speech tagging. Based on the word segmentation results and knowledge graph, the terminology of the target foreign trade correspondence is optimized according to the rules to obtain the first optimization result; Based on the word segmentation results, the title of the foreign trade correspondence is identified using a conditional random field model to obtain the identification result; The identification results are then converted into structured data. Based on the structured data, a Bi-LSTM-CRF model is used to detect risk points in the foreign trade correspondence, including: Based on the structured data, Bi-LSTM encoding is performed to extract the dependencies between texts and obtain sentence vectors; Feature engineering is used to select convolutional kernels in the CNN model and generate a second sentence vector. The second sentence vector is subjected to global maximum pooling using the Max-pooling technique. The sentence vector and the second sentence vector are combined by concatenation to obtain a combined sentence vector. The combined sentence vectors are converted into part-of-speech tags using the softmax function, and the part-of-speech sequences of Chinese words are obtained using the Viterbi path. Based on the part-of-speech sequence, sequence annotation is performed on word boundary issues in Chinese business texts to obtain risk point detection results; Based on the word segmentation results after part-of-speech tagging and the knowledge graph, Chinese error correction is performed on the foreign trade correspondence.

2. The intelligent semantic error correction and business terminology optimization method for foreign trade correspondence according to claim 1, characterized in that, Constructing the target corpus includes: We will convert open-source English business documents in the field of business communications into their formats to create an English corpus. The acquired target foreign trade correspondence is formatted to obtain the target business corpus; The processed target foreign trade correspondence and the English business documents in the English corpus are combined using text concatenation to create and collect Chinese business documents, thereby generating the target corpus.

3. The method for intelligent semantic error correction and business terminology optimization in foreign trade correspondence according to claim 2, characterized in that, Based on the word segmentation results and the knowledge graph, the business terms of the target foreign trade correspondence are optimized according to rules to obtain the first optimization result, including: Input the word segmentation results and the text information in the knowledge graph into the BPE model; A dictionary file is constructed based on the phrases and words in the target business corpus, and the business text is segmented to obtain the first optimization result.

4. The method for intelligent semantic error correction and business terminology optimization in foreign trade correspondence according to claim 1, characterized in that, The risk point detection also includes: The structured data of the transformed semantic hierarchy is subjected to multi-layer Bi-LSTM encoding, wherein the output of each Bi-LSTM layer is used as the input of the next Bi-LSTM layer. When performing the feature engineering, the size and number of convolutional kernels in the CNN model are dynamically adjusted according to the text characteristics of foreign trade correspondence; During the global max pooling process, the maximum value of each feature dimension is retained to form a fixed-length global feature representation.

5. The method for intelligent semantic error correction and business terminology optimization in foreign trade correspondence according to claim 1, characterized in that, The Chinese error correction also includes: Based on the word segmentation results after part-of-speech tagging and the knowledge graph, the representation of key business words in the business text is learned from the business text by using an attention-based LSTM model and a concatenation method, thus obtaining the second optimization result.

6. The intelligent semantic error correction and business terminology optimization method for foreign trade correspondence according to claim 5, characterized in that, The LSTM model based on the attention mechanism includes: Based on the knowledge of vocabulary in the target business text, a knowledge embedding vector of the knowledge graph is generated to obtain the word representation of the target business text; The word representations are used as input to the LSTM network to obtain LSTM-based representations. The knowledge embedding vector and the LSTM-based representation are input into the attention mechanism to obtain the representation of the target business text; The obtained text representation, the word representation of the target business text, and the representation of the LSTM network are input into a feedforward neural network to obtain the second optimization result.

7. The method for intelligent semantic error correction and business terminology optimization in foreign trade correspondence according to claim 1, characterized in that, Following the risk point detection, the following is also included: Construct a compliance module to evaluate whether the aforementioned foreign trade correspondence is compliant; The text data of the foreign trade correspondence is preprocessed and converted into a digital vector matrix; Establish a mapping relationship between the aforementioned foreign trade correspondence and the compliance of foreign trade correspondence; The compliance of the input text is analyzed based on the mapping relationship, wherein the compliance analysis includes multi-classification tasks and multi-label tasks.

8. The method for intelligent semantic error correction and business terminology optimization in foreign trade correspondence according to claim 1, characterized in that, Following the risk point detection, the following is also included: Construct a translation module for intelligent translation of the aforementioned foreign trade correspondence; The translation module includes an encoder-decoder model, wherein the encoder employs a self-attention mechanism to encode text into a sequence representation; After being encoded by the encoder, the input sequence is converted into a vector of fixed dimensions; The translation result is obtained by decoding using the decoder.

9. The intelligent semantic error correction and business terminology optimization method for foreign trade correspondence according to claim 1, characterized in that, Following the Chinese character correction, the following is also included: Construct a data correction module for comparing the risk point detection results with business terms in the international business terminology database; The risk point detection results are compared with the business terms in the international business terminology database, and the corresponding international business terms are output. Construct an audit and correction module for reviewing and correcting the aforementioned international business terms; The international business terms identified by the data correction module are compared with the actual output data; The international business terminology database will be updated based on the comparison results.

10. A device for intelligent semantic error correction and business terminology optimization in foreign trade correspondence, characterized in that: include: The target corpus creation module is configured to: acquire target foreign trade correspondence and build a target corpus; The word segmentation module is configured to perform Chinese word segmentation and part-of-speech tagging on the target foreign trade correspondence, and obtain the word segmentation result after part-of-speech tagging. The business terminology optimization module is configured to: optimize the terminology of the target foreign trade correspondence according to the rules based on the word segmentation results and knowledge graph, and obtain the first optimization result; The recognition module is configured to: based on the word segmentation results, use a conditional random field model to recognize the title of the foreign trade correspondence, and obtain the recognition result; The conversion module is configured to convert the recognition results into structured data; The detection module is configured to: detect risk points in the foreign trade correspondence using a Bi-LSTM-CRF model based on the structured data, including: performing Bi-LSTM-CRF encoding on the structured data to extract dependencies between texts and obtain sentence vectors; selecting convolutional kernels in a CNN model through feature engineering and generating a second sentence vector; performing global maximum pooling on the second sentence vector using Max-pooling technology; combining the sentence vector and the second sentence vector using concatenation to obtain a combined sentence vector; converting the combined sentence vector into part-of-speech tags using the softmax function and obtaining the part-of-speech sequence of Chinese words using the Viterbi path; and performing sequence labeling on word boundary issues in the Chinese business text based on the part-of-speech sequence to obtain the risk point detection results. The error correction module is configured to perform Chinese error correction on the foreign trade correspondence based on the word segmentation results after part-of-speech tagging and the knowledge graph.

Images