Method and system for identifying industry terms in business English communication and interpreting industry terms in combination with context, electronic equipment and storage medium

Abstract

The invention belongs to the technical field of natural language processing, and provides a method and system for identifying industry terms in business English communication and explaining the industry terms in combination with contexts, electronic equipment and a storage medium. The method comprises the steps of structured text sequence generation, term matching, field matching, dual-path feature extraction, feature fusion, scene classification, conflict degree quantification, scene matching and optimal semantic item determination. According to the method, the hierarchical knowledge graph is combined with the priority rule, so that term positioning is more accurate; through double-path extraction and dynamic weight fusion, domain attributes and semantic information are covered at the same time, and more reliable support is provided for scene classification; through conflict quantification and multi-strategy adaptation, the misjudgment rate of low-confidence and high-conflict scenes is reduced; and by focusing context core information, interpretation fits a specific use scene.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of natural language processing, in particular to a method and system for identifying and explaining industry terms in business English communication, an electronic device and a storage medium. BACKGROUND

[0002] The existing business English automatic translation technology has evolved through several stages. From the early rule-based direct translation system, statistical machine translation probability modeling, to the current intelligent system centered on neural machine translation and large language models. The core relies on the Transformer architecture and the heterogeneous design of heavy encoding and light decoding, integrating deep learning, natural language processing technology, and matching the million-level business term library and industry adapter to accurately adapt to the needs of financial, trade, legal and other vertical fields. The technology integrates ASR, OCR, TTS and other multi-modal capabilities, supports text, voice, document and other multi-input, and solves the problems of term ambiguity and expression habit differences through context awareness and cross-cultural adaptation technology. Main providers such as Microsoft, Google, and Tongyi Qianwen rely on cloud computing to launch dedicated solutions, achieving high scalability and efficient response, and have been deeply applied to cross-border e-commerce, international conferences, contract translation and other scenarios, becoming the core support for enterprise globalization communication.

[0003] However, the existing business English translation method does not establish a hierarchical domain knowledge support, and there is no clear priority rule when multiple terms are matched, which is prone to recognition omission or misjudgment. The existing translation model only relies on a single algorithm to extract text features, and cannot simultaneously consider the field professional nature and the context semantic relevance, resulting in one-sided feature representation. And because the existing method does not quantitatively process the conflict between scene classification results and domain distribution probability, it lacks effective adaptation strategies when facing low confidence or high conflict scenarios. In addition, when facing term polysemy ambiguity, it does not focus on the core associated information around the term, but only relies on the matching of term itself, resulting in reduced translation effect. SUMMARY

[0004] In order to overcome the deficiencies of the prior art, the purpose of the present application is to provide a method and system for identifying and explaining industry terms in business English communication, an electronic device and a storage medium, which solves the problems of existing methods prone to recognition omission or misjudgment, inability to simultaneously consider the field professional nature and the context semantic relevance, not quantitatively processing the conflict between scene classification results and domain distribution probability, and only relying on the matching of term itself.

[0005] To achieve the above purpose, the present application provides the following solutions:

[0006] A method for identifying and explaining industry terms in business English communication, comprising:

[0007] Receive the business English text to be interpreted, and perform word segmentation, part-of-speech tagging, and stop word filtering on the business English text to be interpreted to obtain a structured text sequence;

[0008] The structured text sequence is matched with terms using a business terminology database based on a predefined domain-layered knowledge graph, resulting in term matching results and term position codes.

[0009] Extract the domain dictionary set and domain distribution probability corresponding to the term matching results from the domain hierarchical knowledge graph;

[0010] Based on the term position encoding and the domain dictionary set, dual-path feature extraction is performed on the structured text sequence using TF-IDF and BERT to obtain domain feature vectors and semantic vectors;

[0011] The domain feature vector and the semantic vector are fused according to the domain distribution probability to obtain a fused feature vector;

[0012] A fully connected classifier is used to classify the fused feature vector to obtain detailed scene labels and classification confidence scores.

[0013] The conflict degree between the subdivided scene labels and the domain distribution probability is calculated to obtain the distribution conflict degree;

[0014] Based on a preset scene matching strategy, the classification confidence and the distribution conflict degree are matched to obtain candidate scenes;

[0015] Based on the candidate scenarios, the Siamese network is used to perform candidate sense selection, dynamic context focusing, encoding processing, similarity calculation, and optimal sense determination on the structured text sequence to obtain the optimal sense interpretation.

[0016] Preferably, the structured text sequence is matched using a business terminology database based on a predefined domain-layered knowledge graph to obtain term matching results and term position codes, and the method further includes:

[0017] If no term is matched, output "no business term to be explained";

[0018] When multiple terms are matched, they are sorted according to their length priority and processed according to the sorting order.

[0019] Preferably, based on the term position encoding and the domain dictionary set, dual-path feature extraction is performed on the structured text sequence using TF-IDF and BERT to obtain a domain feature vector and a semantic vector, including:

[0020] The TF-IDF algorithm is used to calculate the domain feature vector of the structured text sequence; the expression for the domain feature vector is:

[0021] ;

[0022] in, The domain feature vector; This represents the TF-IDF algorithm; The structured text sequence; The domain dictionary set; Representing vocabulary; , These are term frequency and inverse document frequency, respectively.

[0023] The term position encoding is set as an attention mask, and the structured text sequence is encoded using the encoder of the BERT to obtain the semantic vector.

[0024] Preferably, the expression for the distribution conflict degree is:

[0025] ;in, For target term and scene The degree of conflict of the distribution; For target term In the scene The probability of belonging to a subdomain.

[0026] Preferably, candidate scenes are obtained by performing scene matching on the classification confidence and the distribution conflict degree according to a preset scene matching strategy, including:

[0027] When the classification confidence is not less than the confidence threshold, if the distribution conflict degree is not greater than the conflict degree threshold, then the current scene is determined as the candidate scene;

[0028] When the classification confidence is not less than the confidence threshold, if the distribution conflict degree is greater than the conflict degree threshold, then the scene corresponding to the second high probability domain of the target term is included in the candidate scene and a secondary classification is performed.

[0029] When the classification confidence is less than the confidence threshold, the scenarios corresponding to the first and second high-probability domains of the target term are determined as the candidate scenarios.

[0030] Preferably, based on the candidate scenarios, a Siamese network is used to perform candidate sense selection, dynamic context focusing, encoding processing, similarity calculation, and optimal sense determination on the structured text sequence to obtain the optimal sense interpretation, including:

[0031] Extract the senses corresponding to the candidate scenarios from the domain hierarchical knowledge graph to obtain a sense set;

[0032] Attention is calculated between each sentence in the structured text sequence and the term position encoding of the target term to obtain the relevance.

[0033] Sentences of a target number are filtered based on the relevance to obtain related sentences, and several words closest to the target term are extracted from the related sentences to obtain candidate core phrases;

[0034] The candidate core phrases are matched with the context feature word set of the sense set to obtain the network input;

[0035] Using Dis lBERT encodes and calculates similarity for the network input to obtain a similarity value;

[0036] The optimal semantic interpretation is obtained by performing optimal semantic matching on the similarity value based on the domain hierarchical knowledge graph.

[0037] Preferably, the domain feature vector and the semantic vector are fused according to the domain distribution probability to obtain a fused feature vector, including:

[0038] The domain feature vector, the semantic vector, and the domain distribution probability are calculated using a preset weight calculation formula to obtain the fusion weight; the expression of the weight calculation formula is:

[0039] ;in, The fusion weights; It is the softmax function; This is the weight matrix; The semantic vector; For bias terms; The probability distribution of the region;

[0040] The domain feature vector and the semantic vector are fused using the fusion weights to obtain the fused feature vector; the expression of the fused feature vector is:

[0041] ;in, The fused feature vector is denoted as .

[0042] Preferably, a system for identifying and interpreting industry-specific terms in business English communication, in conjunction with the context, includes:

[0043] The preprocessing module is used to receive the business English text to be explained, and to perform word segmentation, part-of-speech tagging, and stop word filtering on the business English text to be explained to obtain a structured text sequence.

[0044] The pre-matching module is used to perform term matching on the structured text sequence using a business general terminology library of a preset domain hierarchical knowledge graph, and to obtain term matching results and term position codes.

[0045] The domain matching module is used to extract the domain dictionary set and domain distribution probability corresponding to the term matching results in the domain hierarchical knowledge graph;

[0046] The dual feature extraction module is used to perform dual-path feature extraction on the structured text sequence based on the term position encoding and the domain dictionary set using TF-IDF and BERT to obtain a domain feature vector and a semantic vector;

[0047] The feature fusion module is used to fuse the domain feature vector and the semantic vector according to the domain distribution probability to obtain a fused feature vector;

[0048] The scene classification module is used to classify the fused feature vector using a fully connected classifier to obtain subdivided scene labels and classification confidence scores.

[0049] The conflict quantification module is used to calculate the conflict degree between the subdivided scene labels and the domain distribution probability to obtain the distribution conflict degree;

[0050] The scene matching module is used to perform scene matching on the classification confidence and the distribution conflict degree according to the preset scene matching strategy to obtain candidate scenes;

[0051] The sense selection module is used to perform candidate sense selection, dynamic context focusing, encoding processing, similarity calculation, and optimal sense determination on the structured text sequence based on the candidate scenario, so as to obtain the optimal sense interpretation.

[0052] Preferably, an electronic device includes: at least one processor and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform the aforementioned method for identifying and interpreting industry terms in business English communication in conjunction with context.

[0053] Preferably, a non-transitory computer-readable storage medium stores computer instructions for causing a computer to execute the aforementioned method for identifying and interpreting industry terms in business English communication in conjunction with context.

[0054] The present invention discloses the following technical effects:

[0055] This invention provides a method, system, electronic device, and storage medium for identifying and interpreting industry terms in business English communication within context. By combining hierarchical knowledge graphs with priority rules, it solves the problem of omissions or misjudgments in existing methods, achieving precise terminology localization. Through dual-path extraction and dynamic weight fusion, it addresses the inability of existing methods to simultaneously consider domain specialization and contextual semantic relevance, achieving simultaneous coverage of domain attributes and semantic information. Through conflict quantification and multi-strategy adaptation, it solves the problem of existing methods being unable to adapt to complex business scenarios, reducing the misjudgment rate in low-confidence, high-conflict scenarios. By focusing on core contextual information, it resolves the problem of polysemy in business terms, achieving interpretations that are more closely aligned with specific usage scenarios. Attached Figure Description

[0056] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0057] Figure 1 A schematic diagram illustrating the process of identifying and interpreting industry-specific terms in business English communication in accordance with the context, as provided in this embodiment of the invention.

[0058] Figure 2 This is a schematic diagram of the term matching process provided in an embodiment of the present invention;

[0059] Figure 3 This is a schematic diagram of the dual-path feature extraction process provided in an embodiment of the present invention;

[0060] Figure 4 This is a schematic diagram of the scene matching process provided in an embodiment of the present invention;

[0061] Figure 5 This is a schematic diagram of the optimal definition determination process provided in an embodiment of the present invention. Detailed Implementation

[0062] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0063] The purpose of this invention is to provide a method, system, electronic device, and storage medium for identifying and interpreting industry terms in business English communication in conjunction with context, thereby solving the problems of existing methods that are prone to omissions or misjudgments, inability to simultaneously consider domain professionalism and contextual semantic relevance, failure to quantify the conflict between scenario classification results and domain distribution probability, and reliance solely on matching the meaning of the terms themselves.

[0064] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0065] Figure 1 This is a schematic diagram illustrating the process of identifying and interpreting industry-specific terms in business English communication, as provided in an embodiment of the present invention. Figure 1 As shown, this invention provides a method for identifying and interpreting industry-specific terms in business English communication, including:

[0066] Step 100: Receive the business English text to be explained, and perform word segmentation, part-of-speech tagging, and stop word filtering on the business English text to be explained to obtain a structured text sequence;

[0067] Step 200: Use the business general terminology database of the preset domain hierarchical knowledge graph to perform term matching on the structured text sequence to obtain term matching results and term position codes;

[0068] Step 300: Extract the domain dictionary set and domain distribution probability corresponding to the term matching results in the domain hierarchical knowledge graph;

[0069] Step 400: Based on the term position encoding and the domain dictionary set, perform dual-path feature extraction on the structured text sequence using TF-IDF and BERT to obtain domain feature vectors and semantic vectors;

[0070] Step 500: Fuse the domain feature vector and the semantic vector according to the domain distribution probability to obtain a fused feature vector;

[0071] Step 600: Use a fully connected classifier to classify the fused feature vector to obtain subdivided scene labels and classification confidence scores;

[0072] Step 700: Calculate the conflict degree between the subdivided scene labels and the domain distribution probability to obtain the distribution conflict degree;

[0073] Step 800: Perform scene matching on the classification confidence and the distribution conflict degree according to the preset scene matching strategy to obtain candidate scenes;

[0074] Step 900: Based on the candidate scenarios, use the Siamese network to perform candidate sense selection, dynamic context focusing, encoding processing, similarity calculation, and optimal sense determination on the structured text sequence to obtain the optimal sense interpretation.

[0075] refer to Figure 2 The method further includes: using a pre-defined domain-based hierarchical knowledge graph to perform term matching on the structured text sequence, obtaining term matching results and term position codes; and also includes:

[0076] Step 201: If no term is matched, output "no business industry term to be explained".

[0077] Step 202: When multiple terms are matched, sort them according to the term length priority and process them according to the sorting order.

[0078] refer to Figure 3 Based on the term position encoding and the domain dictionary set, dual-path feature extraction is performed on the structured text sequence using TF-IDF and BERT to obtain domain feature vectors and semantic vectors, including:

[0079] Step 401: Calculate the domain feature vector by using the TF-IDF algorithm on the structured text sequence; the expression for the domain feature vector is:

[0080] ;

[0081] in, The domain feature vector; This represents the TF-IDF algorithm; The structured text sequence; The domain dictionary set; Representing vocabulary; , These are term frequency and inverse document frequency, respectively.

[0082] Step 402: Set the term position encoding as an attention mask and use the encoder of the BERT to encode the structured text sequence to obtain the semantic vector.

[0083] Specifically, the expression for the distribution conflict degree is:

[0084] ;in, For target term and scene The degree of conflict of the distribution; For target term In the scene The probability of belonging to a subdomain.

[0085] refer to Figure 4 Based on a preset scene matching strategy, the classification confidence and the distribution conflict degree are matched to obtain candidate scenes, including:

[0086] Step 801: When the classification confidence is not less than the confidence threshold, if the distribution conflict degree is not greater than the conflict degree threshold, then the current scene is determined as the candidate scene;

[0087] Step 80: When the classification confidence is not less than the confidence threshold, if the distribution conflict degree is greater than the conflict degree threshold, then the scene corresponding to the second high probability domain of the target term is included in the candidate scene and a secondary classification is performed;

[0088] Step 803: When the classification confidence is less than the confidence threshold, the scenarios corresponding to the first and second high-probability domains of the target term are determined as the candidate scenarios.

[0089] refer to Figure 5 Based on the candidate scenarios, the Siamese network is used to perform candidate sense selection, dynamic context focusing, encoding processing, similarity calculation, and optimal sense determination on the structured text sequence to obtain the optimal sense interpretation, including:

[0090] Step 901: Extract the senses corresponding to the candidate scenarios from the domain hierarchical knowledge graph to obtain a sense set;

[0091] Step 902: Perform attention calculation on the term position encoding of each sentence in the structured text sequence and the target term to obtain the relevance;

[0092] Step 903: Filter sentences according to the target number based on the relevance to obtain related sentences, and extract several words closest to the target term from the related sentences to obtain candidate core phrases;

[0093] Step 904: Match the candidate core phrase with the context feature word set of the sense set to obtain the network input;

[0094] Step 905: Using Dis lBERT encodes and calculates similarity for the network input to obtain a similarity value;

[0095] Step 906: Perform optimal semantic matching on the similarity value based on the domain hierarchical knowledge graph to obtain the optimal semantic interpretation.

[0096] Further, the domain feature vector and the semantic vector are fused according to the domain distribution probability to obtain a fused feature vector, including:

[0097] The domain feature vector, the semantic vector, and the domain distribution probability are calculated using a preset weight calculation formula to obtain the fusion weight; the expression of the weight calculation formula is:

[0098] ;in, The fusion weights; It is the softmax function; This is the weight matrix; The semantic vector; For bias terms; The probability distribution of the region;

[0099] The domain feature vector and the semantic vector are fused using the fusion weights to obtain the fused feature vector; the expression of the fused feature vector is:

[0100] ;in, The fused feature vector is denoted as .

[0101] Specifically, input preprocessing (data structuring preparation) provides standardized text data for subsequent domain classification and sense matching. The core is locating terms and extracting key context, connecting to the initial term screening of the knowledge graph. 1) Text Input and Basic Processing: Input is business English text to be explained, with a length ranging from 5 to 500 words; NLTK word segmentation is performed, and spaCy's English part-of-speech tagging model is used for part-of-speech tagging, filtering out meaningless stop words, and generating a structured text sequence. ( (The nth word after word segmentation).

[0102] 2) Initial Terminology Screening and Localization: The business terminology database at the top level of the domain-layered knowledge graph is invoked to perform term matching on the structured text sequence (a combination of exact matching and root word matching); successfully matched terms are marked. and its position in the text ( (The index of the first word of the term in the Text); if no term is matched, output "No business industry terminology needs explanation" and the process terminates; if multiple terms are matched, sort them according to "term length priority", with multi-word terms taking precedence over single-word terms, and process them in sequence (this embodiment uses a single term). For example, in the case of multiple terms, subsequent steps are executed in a loop.

[0103] Furthermore, domain classification (domain-specific filtering to narrow down the scope of meanings) is performed based on preprocessing results and domain data from the knowledge graph to determine the specific scenario to which the text belongs, eliminating irrelevant domain interference for meaning matching. Domain reference data is retrieved by extracting two types of data from the middle layer of the domain-layered knowledge graph:

[0104] 1) Domain dictionary set ( (A dictionary of terms specific to a subdomain)

[0105] 2) Target Terminology Domain distribution probability ( for The probability of occurrence in the m-th sub-domain is from (Read from the node attributes).

[0106] Dual-path feature extraction:

[0107] 1) Domain-specific features: Using the preprocessed Text as input, combined with the domain dictionary set D, the domain feature vector is calculated using the TF-IDF algorithm. :

[0108]

[0109] in, The domain feature vector; This represents the TF-IDF algorithm; The structured text sequence; The domain dictionary set; Representing vocabulary; , These are term frequency and inverse document frequency, respectively.

[0110] 2) General semantic features: Taking Text as input, term_pos is used as an attention mask (to enhance the semantic weights of words surrounding the term), and semantic vectors are generated through encoding by the BERT model. :

[0111]

[0112] Preferably, feature fusion and classification calculation: through dynamic allocation and The weights are used to calculate the fused feature vector. :

[0113]

[0114]

[0115] in, The fusion weights; It is the softmax function; This is the weight matrix; The semantic vector; For bias terms; The probability distribution of the region; This serves as a leading indicator for the terminology field.

[0116] Will Input a fully connected classifier and output the sub-scene label to which the text belongs. And the classification confidence score_c.

[0117] Specifically, terminology bias correction and candidate scenario determination: calculating classification results and... Conflict degree of domain distribution probability :

[0118]

[0119] in, For target term In the scene The probability of belonging to a subdomain;

[0120] If score_c ≥ 0.7 and ≤0.2: The candidate scenario is determined, and only the meanings under this scenario are retained;

[0121] If score_c ≥ 0.7 but >0.2: Triggers secondary classification, forcing the classification to... The scenarios corresponding to the top two probability domains were included in the candidates;

[0122] If score_c < 0.7: include directly. The scenarios corresponding to the top two probability domains are used as the candidate scenario set {C1, C2}.

[0123] Furthermore, based on candidate scenarios and preprocessed context, similarity is calculated using a Siamese network to select the optimal sense from the knowledge graph. Candidate sense selection: The sense set S={s1,s2,...,sm} corresponding to the candidate scenario set C is extracted from the bottom layer of the domain-layered knowledge graph: 1) Determine the association attributes of each sense sj; 2) Selection criteria: sense and terminology. The weight of the associated edge is ≥0.7 (from the edge attributes of the knowledge graph). (Read from the middle), filter low-confidence entries.

[0124] Preferably, dynamic context focusing: 1) Calculate the correlation between each sentence in the preprocessed Text and term_pos. :

[0125]

[0126] in, For sentences Attention weights for term position j in the middle;

[0127] Top three For the corresponding sentence, extract the five words before and after it to form a "candidate core phrase";

[0128] The candidate core phrases are matched with the context feature word set of the sense set S, and finally integrated into the input A (core context text) of the Siamese network.

[0129] Furthermore, lightweight encoding and similarity calculation: Encoder initialization: DistilBERT is used as the shared encoder of the Siamese network, and its pre-trained weights inherit the BERT weights from path 2 above; Dual-input encoding: Input A and reference example sentences Bj for each sense are input into the encoder respectively to generate vectors VA and VBj; Similarity fusion calculation:

[0130]

[0131]

[0132]

[0133] in, Cosine similarity; , These are the values ​​of the q-th dimension of VA and VBj, respectively. Distance to Manhattan; To achieve a comprehensive similarity score, β is the fusion weight, determined by the similarity of the sense sj and... The weight of the associated edges is determined.

[0134] Optimal Sense Determination and Interpretation Output: For all senses sj Sort the results and take the highest score, score_s_max, and its corresponding meaning, sj_max. If score_s_max ≥ 0.8, read the precise definition and scenario description of sj_max from the knowledge graph and generate the final explanation. If score_s_max < 0.8, output the candidate explanations of the first two meanings and prompt the user to supplement the context.

[0135] Optionally, dynamic updates (closed-loop optimization, adapting to new terms / senses) are implemented based on real-time corpus streams and user feedback to update the knowledge graph and model parameters, feeding back into previous steps. Real-time corpus stream preprocessing: Integrating real-time business corpus streams, the aforementioned logic is reused to generate two types of candidate data: new words not matched to the knowledge graph terminology database. and matched terms However, "abnormal contextual terms" whose contextual features do not intersect with existing senses Cj old .

[0136] Unsupervised mining of new terms / senses: Corpus clustering: The BERTopic model is used to cluster real-time corpus streams, and clustering feature words are derived from the domain dictionary D in the knowledge graph (ensuring consistency between clustering and the existing domain system); New term detection: ... Calculate the novelty score :

[0137]

[0138] in, For new terminology In context Frequency of occurrence in; For the term t in context The frequency of occurrence in. If If the value is ≥0.6, it is marked as a candidate for a new term.

[0139] New definition detection: for t old Extract its abnormal context BERT encoded vector V old In knowledge graphs Calculate the distance between the existing sense's reference example vector VBj and the reference vector. If the minimum distance is ≥0.6, mark it as a candidate for a new sense.

[0140] Furthermore, incremental model training and parameter synchronization are performed: using annotated corpora (including scene labels) with new terms / senses, the underlying parameters of BERT are frozen, and only the top-level classification layer is trained. The domain dictionary D and attention weight calculation parameters of TF-IDF are updated. Incremental training of the Siamese network is performed: the input sample library is supplemented with reference example sentences of new senses, the top-level parameters of DistilBERT are fine-tuned, and the constraint rules of similarity fusion weights are updated. The updated model parameters are synchronized to the classifier and encoder of the previous steps, and the updated knowledge graph terminology library and domain dictionary are synchronized to the terminology screening in the first step and the domain reference data retrieval in the second step.

[0141] As an optional implementation, this embodiment also provides a system for identifying and interpreting industry-specific terms in business English communication, including:

[0142] The preprocessing module is used to receive the business English text to be explained, and to perform word segmentation, part-of-speech tagging, and stop word filtering on the business English text to be explained to obtain a structured text sequence.

[0143] The pre-matching module is used to perform term matching on the structured text sequence using a business general terminology library of a preset domain hierarchical knowledge graph, and to obtain term matching results and term position codes.

[0144] The domain matching module is used to extract the domain dictionary set and domain distribution probability corresponding to the term matching results in the domain hierarchical knowledge graph;

[0145] The dual feature extraction module is used to perform dual-path feature extraction on the structured text sequence based on the term position encoding and the domain dictionary set using TF-IDF and BERT to obtain a domain feature vector and a semantic vector;

[0146] The feature fusion module is used to fuse the domain feature vector and the semantic vector according to the domain distribution probability to obtain a fused feature vector;

[0147] The scene classification module is used to classify the fused feature vector using a fully connected classifier to obtain subdivided scene labels and classification confidence scores.

[0148] The conflict quantification module is used to calculate the conflict degree between the subdivided scene labels and the domain distribution probability to obtain the distribution conflict degree;

[0149] The scene matching module is used to perform scene matching on the classification confidence and the distribution conflict degree according to the preset scene matching strategy to obtain candidate scenes;

[0150] The sense selection module is used to perform candidate sense selection, dynamic context focusing, encoding processing, similarity calculation, and optimal sense determination on the structured text sequence based on the candidate scenario, so as to obtain the optimal sense interpretation.

[0151] As an optional implementation, this embodiment also provides an electronic device, including: at least one processor, and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform the aforementioned method for identifying and interpreting industry terms in business English communication in conjunction with context.

[0152] As an optional implementation, this embodiment also provides a non-transitory computer-readable storage medium storing computer instructions for causing a computer to execute the aforementioned method for recognizing and interpreting industry terms in business English communication in conjunction with context.

[0153] The beneficial effects of this invention are as follows:

[0154] This invention combines hierarchical knowledge graphs with priority rules to make terminology localization more accurate; it uses dual-path extraction and dynamic weight fusion to cover both domain attributes and semantic information, providing more reliable support for scene classification; it reduces the misjudgment rate of low-confidence, high-conflict scenarios through conflict quantification and multi-strategy adaptation; and it focuses on core contextual information to make interpretations fit specific usage scenarios.

[0155] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0156] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A method for identifying and interpreting industry-specific terms in business English communication, characterized in that, include: Receive the business English text to be interpreted, and perform word segmentation, part-of-speech tagging, and stop word filtering on the business English text to be interpreted to obtain a structured text sequence; The structured text sequence is matched with terms using a business terminology database based on a predefined domain-layered knowledge graph, resulting in term matching results and term position codes. Extract the domain dictionary set and domain distribution probability corresponding to the term matching results from the domain hierarchical knowledge graph; Based on the term position encoding and the domain dictionary set, dual-path feature extraction is performed on the structured text sequence using TF-IDF and BERT to obtain domain feature vectors and semantic vectors; The domain feature vector and the semantic vector are fused according to the domain distribution probability to obtain a fused feature vector; A fully connected classifier is used to classify the fused feature vector to obtain detailed scene labels and classification confidence scores. The conflict degree between the subdivided scene labels and the domain distribution probability is calculated to obtain the distribution conflict degree; Based on a preset scene matching strategy, the classification confidence and the distribution conflict degree are matched to obtain candidate scenes; Based on the candidate scenarios, the Siamese network is used to perform candidate sense selection, dynamic context focusing, encoding processing, similarity calculation, and optimal sense determination on the structured text sequence to obtain the optimal sense interpretation.

2. The method for identifying and interpreting industry-specific terms in business English communication according to claim 1, characterized in that, The structured text sequence is matched using a business terminology database based on a predefined domain-layered knowledge graph to obtain term matching results and term position codes. The method also includes: If no term is matched, output "no business term to be explained"; When multiple terms are matched, they are sorted according to their length priority and processed according to the sorting order.

3. The method for identifying and interpreting industry-specific terms in business English communication according to claim 1, characterized in that, Based on the term location encoding and the domain dictionary set, dual-path feature extraction is performed on the structured text sequence using TF-IDF and BERT to obtain domain feature vectors and semantic vectors, including: The TF-IDF algorithm is used to calculate the domain feature vector of the structured text sequence; the expression for the domain feature vector is: ； in, The domain feature vector; This represents the TF-IDF algorithm; The structured text sequence; The domain dictionary set; Representing vocabulary; , These are term frequency and inverse document frequency, respectively. The term position encoding is set as an attention mask, and the structured text sequence is encoded using the encoder of the BERT to obtain the semantic vector.

4. The method for identifying and interpreting industry-specific terms in business English communication according to claim 1, wherein the expression for the degree of conflict distribution is: ;in, For target term and scene The degree of conflict of the distribution; For target term In the scene The probability of belonging to a subdomain.

5. The method for identifying and interpreting industry terms in business English communication according to claim 1, wherein candidate scenarios are obtained by matching the classification confidence and the distribution conflict degree according to a preset scenario matching strategy, including: When the classification confidence is not less than the confidence threshold, if the distribution conflict degree is not greater than the conflict degree threshold, then the current scene is determined as the candidate scene; When the classification confidence is not less than the confidence threshold, if the distribution conflict degree is greater than the conflict degree threshold, then the scene corresponding to the second high probability domain of the target term is included in the candidate scene and a secondary classification is performed. When the classification confidence is less than the confidence threshold, the scenarios corresponding to the first and second high-probability domains of the target term are determined as the candidate scenarios.

6. The method for identifying and interpreting industry terms in business English communication according to claim 1, comprising using a Siamese network to perform candidate semantic filtering, dynamic context focusing, encoding processing, similarity calculation, and optimal semantic determination on the structured text sequence based on the candidate scenario, to obtain the optimal semantic interpretation, including: Extract the senses corresponding to the candidate scenarios from the domain hierarchical knowledge graph to obtain a sense set; Attention is calculated between each sentence in the structured text sequence and the term position encoding of the target term to obtain the relevance. Sentences of a target number are filtered based on the relevance to obtain related sentences, and several words closest to the target term are extracted from the related sentences to obtain candidate core phrases; The candidate core phrases are matched with the context feature word set of the sense set to obtain the network input; Using Dis lBERT encodes and calculates similarity for the network input to obtain a similarity value; The optimal semantic interpretation is obtained by performing optimal semantic matching on the similarity value based on the domain hierarchical knowledge graph.

7. The method for identifying and interpreting industry-specific terms in business English communication according to claim 3, characterized in that, The domain feature vector and the semantic vector are fused according to the domain distribution probability to obtain a fused feature vector, including: The domain feature vector, the semantic vector, and the domain distribution probability are calculated using a preset weight calculation formula to obtain the fusion weight; the expression of the weight calculation formula is: ;in, The fusion weights; It is the softmax function; This is the weight matrix; The semantic vector; For bias terms; The probability distribution of the region; The domain feature vector and the semantic vector are fused using the fusion weights to obtain the fused feature vector; the expression of the fused feature vector is: ;in, The fused feature vector is denoted as .

8. A system for identifying and interpreting industry-specific terminology in business English communication, characterized in that, The system is used to implement the method for identifying and interpreting industry terms in business English communication according to claim 1, the system comprising: The preprocessing module is used to receive the business English text to be explained, and to perform word segmentation, part-of-speech tagging, and stop word filtering on the business English text to be explained to obtain a structured text sequence. The pre-matching module is used to perform term matching on the structured text sequence using a business general terminology library of a preset domain hierarchical knowledge graph, and to obtain term matching results and term position codes. The domain matching module is used to extract the domain dictionary set and domain distribution probability corresponding to the term matching results in the domain hierarchical knowledge graph; The dual feature extraction module is used to perform dual-path feature extraction on the structured text sequence based on the term position encoding and the domain dictionary set using TF-IDF and BERT to obtain a domain feature vector and a semantic vector; The feature fusion module is used to fuse the domain feature vector and the semantic vector according to the domain distribution probability to obtain a fused feature vector; The scene classification module is used to classify the fused feature vector using a fully connected classifier to obtain subdivided scene labels and classification confidence scores. The conflict quantification module is used to calculate the conflict degree between the subdivided scene labels and the domain distribution probability to obtain the distribution conflict degree; The scene matching module is used to perform scene matching on the classification confidence and the distribution conflict degree according to the preset scene matching strategy to obtain candidate scenes; The sense selection module is used to perform candidate sense selection, dynamic context focusing, encoding processing, similarity calculation, and optimal sense determination on the structured text sequence based on the candidate scenario, so as to obtain the optimal sense interpretation.

9. An electronic device, characterized in that, include: At least one processor, and a memory communicatively connected to the processor; wherein the memory stores instructions executable by the processor, the instructions being executed by the processor to enable the processor to perform a method for identifying and interpreting industry terms in business English communication in accordance with any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium storing computer instructions, characterized in that, The computer instructions are used to cause the computer to execute a method for identifying and interpreting industry terms in business English communication in accordance with any one of claims 1 to 7.

Images