Intelligent customer service FAQ matching method and system based on deep semantic understanding

Abstract

The application provides an intelligent customer service FAQ matching method and system based on deep semantic understanding, comprising: rewriting the original question submitted by a user into a rewritten question conforming to the format of professional terms in the financial field and a FAQ knowledge base through a large language model; identifying the intention and entity of the original question by using an intention classification model and an entity recognition model, and fusing the identified intention and entity with the original question and the rewritten question respectively to generate original spliced text and rewritten spliced text; encoding the original spliced text and the rewritten spliced text into original vectors and rewritten vectors through a semantic matching model, and calculating the similarity of the original vectors, the rewritten vectors and the semantic vectors of the question items in the preprocessed FAQ knowledge base respectively to screen the optimal answer. The application can significantly improve the FAQ matching accuracy, avoid the illusion risk of the answer generated by the large model, and improve the reliability.

Description

Technical Field

[0001] This invention relates to the field of text recognition technology, and in particular to an intelligent customer service FAQ matching method and system based on deep semantic understanding. Background Technology

[0002] With the widespread application of artificial intelligence technology in the financial industry, intelligent customer service systems have become an important tool for financial institutions to improve service efficiency and customer experience.

[0003] Currently, mainstream financial intelligent customer service systems primarily rely on the FAQ (Frequently Asked Questions) question-and-answer model, using a predefined question-and-answer database to answer user inquiries. However, traditional FAQ matching technology is typically based on keyword or shallow semantic matching, making it difficult to handle the diversity and complexity of user questions. For example, users may use different synonyms, abbreviations, or word orders to express the same question, resulting in a high matching failure rate.

[0004] In recent years, the rise of large pre-trained language models (such as GPT and BERT) has provided new solutions for question-answering systems, improving coverage by directly generating answers. However, these models are prone to the "illusion" problem when generating answers, i.e., generating inaccurate or false information. In the financial sector, such errors can lead to serious compliance risks or customer losses. For example, the model might generate incorrect investment advice or account information, which is a risk that financial users cannot afford. Therefore, relying solely on large models to directly generate answers is insufficient to meet the stringent accuracy and reliability requirements of the financial industry. Summary of the Invention

[0005] In view of this, embodiments of the present invention provide an intelligent customer service FAQ matching method and system based on deep semantic understanding, in order to eliminate or improve one or more defects existing in the prior art, and solve the problems that existing rule-based or small-scale model-based FAQ matching lacks flexibility, while large-scale language models directly generating answers are difficult to guarantee reliability.

[0006] On the one hand, this invention provides an intelligent customer service FAQ matching method based on deep semantic understanding, the method comprising:

[0007] The user-submitted original question is input into a preset large language model for rewriting, generating a rewritten question; wherein the rewritten question has the same semantics as the original question, and the expression of the rewritten question conforms to the preset format of financial professional terminology and FAQ knowledge base;

[0008] The original question is input into a pre-trained intent classification model to generate intent labels for the original question; the original question is input into a pre-trained entity recognition model to identify the entities and their categories in the original question, thereby obtaining an entity sequence.

[0009] The original question is fused with the intent tag and the entity sequence to obtain the original concatenated text; the rewritten question is fused with the intent tag and the entity sequence to obtain the rewritten concatenated text.

[0010] The original concatenated text and the rewritten concatenated text are respectively input into a pre-trained semantic matching model to generate an original vector and a rewritten vector; the text similarity between the original vector and the rewritten vector and the semantic vector of each question item in the pre-processed FAQ knowledge base is calculated and sorted, and the final answer is selected.

[0011] The preprocessing of the FAQ knowledge base includes inputting each question item in the FAQ knowledge base into the intent classification model and the entity recognition model to generate an intent label and entity sequence for each question item; fusing the intent label and entity sequence of each question item with the corresponding question to obtain a concatenated question text; and inputting the concatenated question text into the semantic matching model to generate a semantic vector for each question item and storing it in the FAQ knowledge base.

[0012] In some embodiments of the present invention, before rewriting the original question submitted by the user into a preset large language model, the method further includes:

[0013] The original questions are cleaned and standardized, including at least noise character removal, synonym replacement, and terminology normalization.

[0014] In some embodiments of the present invention, the original question submitted by the user is input into a preset large language model for rewriting to generate a rewritten question, including:

[0015] The original question is rewritten into the rewritten question using a preset prompt template or a sequence-to-sequence generation model.

[0016] In some embodiments of the present invention, the intent classification model uses a bidirectional encoder representation model based on the Transformer architecture for classification; the entity recognition model uses a joint model of the bidirectional encoder representation model and the conditional random field to predict the entity sequence.

[0017] In some embodiments of the present invention, the original question is fused with the intent tag and the entity sequence to obtain the original concatenated text; the rewritten question is fused with the intent tag and the entity sequence to obtain the rewritten concatenated text, including:

[0018] The intent tag and the entity sequence are added before and after the original question with preset markers to generate the original concatenated text; the intent tag and the entity sequence are added before and after the rewritten question with preset markers to generate the rewritten concatenated text.

[0019] In some embodiments of the present invention, the text similarity between the original vector and the rewritten vector and the semantic vector of each question item in the preprocessed FAQ knowledge base is calculated and ranked, including:

[0020] Calculate the cosine similarity between the original vector and the semantic vector of each question item in the preprocessed FAQ knowledge base to obtain a matching score and construct a first matching score list;

[0021] Calculate the cosine similarity between the rewritten vector and the semantic vector of each question item in the preprocessed FAQ knowledge base, obtain the matching score, and construct a second matching score list;

[0022] Merge the same question item in the first matching score list and the second matching score list, and retain the highest matching score; sort the matching scores of the merged question items in descending order, select the first preset number of question items, and use the FAQ answers of the selected question items as the final answer.

[0023] In some embodiments of the present invention, the method further includes:

[0024] Select a preset number of question items as candidates;

[0025] Calculate the consistency score between the intent label of the candidate option and the intent label of the original question. If the consistency score is lower than a preset score, filter the candidate option.

[0026] The proportion of entities in the candidate options that cover the original question is counted. If the proportion is lower than a preset proportion, the candidate options are filtered.

[0027] The FAQ answers for the remaining candidates after filtering will be used as the final answer.

[0028] In some embodiments of the present invention, the concatenated text of the question is input into the semantic matching model to generate semantic vectors for each question item and store them in the FAQ knowledge base, including:

[0029] The semantic matching model employs a sentence bidirectional encoder representation model.

[0030] The sentence bidirectional encoder representation model is used to encode the concatenated text of the question into a semantic vector; the generated semantic vector is associated with the corresponding question item and stored in a structured vector retrieval library to construct the FAQ knowledge base;

[0031] The structured vector retrieval library utilizes an efficient similarity search algorithm to achieve vector indexing and fast retrieval.

[0032] On the other hand, the present invention also provides an intelligent customer service FAQ matching system based on deep semantic understanding, the system comprising:

[0033] The question rewriting module is used to rewrite the original questions mentioned by users into rewritten questions that conform to the professional terminology of the financial field and the preset format of the FAQ knowledge base through a preset large language model;

[0034] The intent classification module is used to identify the intent category of the original question through a pre-trained intent classification model and obtain the intent label;

[0035] The entity recognition module is used to identify the entities and their categories in the original question using a pre-trained entity recognition model, thereby obtaining an entity sequence.

[0036] The structured query representation construction module is used to fuse the original question with the intent tag and the entity sequence to obtain the original concatenated text; and to fuse the rewritten question with the intent tag and the entity sequence to obtain the rewritten concatenated text.

[0037] The FAQ knowledge base preprocessing module is used to generate intent labels and entity sequences for each question item in the FAQ knowledge base through the intent classification model and the entity recognition model; fuse the intent labels and entity sequences of each question item with the corresponding question to obtain the question concatenation text; and encode the question concatenation text through a pre-trained semantic matching model to obtain the semantic vector of each question item.

[0038] The semantic matching module is used to encode the original concatenated text and the rewritten concatenated text respectively through the semantic matching model to obtain the original vector and the rewritten vector; calculate and sort the text similarity between the original vector and the rewritten vector and the semantic vector of each question item in the preprocessed FAQ knowledge base, and select the final answer.

[0039] On the other hand, the present invention also provides a computer-readable storage medium having a computer program / instructions stored thereon, which, when executed by a processor, implement the steps of any of the methods mentioned above.

[0040] This invention provides an intelligent customer service FAQ matching method and system based on deep semantic understanding. Compared with existing technologies, this invention only uses a large language model to rewrite questions instead of directly generating answers, fundamentally avoiding the illusionary risk of the model generating unreliable answers and significantly improving reliability. Simultaneously, rewriting the user's original question makes the expression of the user's question more consistent with the habits of domain-specific FAQs, narrowing the semantic gap; incorporating intent and entity information into the retrieval process enhances the semantic alignment between the query and knowledge base items; and performing parallel matching of the original question, the rewritten question, and question items in the FAQ knowledge base better identifies semantically similar FAQ questions, significantly improving the accuracy of FAQ matching.

[0041] Furthermore, the intelligent customer service FAQ matching system based on deep semantic understanding adopts a modular design, with each module having independent functions and being flexibly replaceable and upgraded according to actual conditions and needs.

[0042] Additional advantages, objects, and features of the invention will be set forth in part in the description which follows, and will also become apparent in part to those skilled in the art upon studying the description, or may be learned by practice of the invention. The objects and other advantages of the invention can be realized and obtained by means of the structures specifically pointed out in the description and drawings.

[0043] Those skilled in the art will understand that the objectives and advantages achievable with the present invention are not limited to those specifically described above, and that the above and other objectives achievable with the present invention will become clearer from the following detailed description. Attached Figure Description

[0044] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this application, are not intended to limit the scope of the invention. In the drawings:

[0045] Figure 1 This is a schematic diagram illustrating the steps of an intelligent customer service FAQ matching method based on deep semantic understanding in one embodiment of the present invention.

[0046] Figure 2 This is a flowchart illustrating an intelligent customer service FAQ matching method based on deep semantic understanding in one embodiment of the present invention. Detailed Implementation

[0047] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments and accompanying drawings. Here, the illustrative embodiments and descriptions of this invention are used to explain the invention, but are not intended to limit the invention.

[0048] It should also be noted that, in order to avoid obscuring the invention with unnecessary details, only the structures and / or processing steps closely related to the solution according to the invention are shown in the accompanying drawings, while other details that are not closely related to the invention are omitted.

[0049] It should be emphasized that the term "including / comprises" as used herein refers to the presence of a feature, element, step, or component, but does not exclude the presence or addition of one or more other features, elements, steps, or components.

[0050] It should also be noted that, unless otherwise specified, the term "connection" in this article can refer not only to a direct connection, but also to an indirect connection involving an intermediary.

[0051] In the following description, embodiments of the invention will be illustrated with reference to the accompanying drawings. In the drawings, the same reference numerals represent the same or similar parts, or the same or similar steps.

[0052] It should be emphasized here that the step markers mentioned below are not a limitation on the order of the steps, but should be understood as meaning that the steps can be executed in the order mentioned in the embodiments, or in a different order than in the embodiments, or several steps can be executed simultaneously.

[0053] To address the shortcomings of existing rule-based or small-scale FAQ matching methods, which lack flexibility, and the reliability issues of directly generating answers using large language models, this invention provides an intelligent customer service FAQ matching method based on deep semantic understanding. Figure 1 As shown, the method includes the following steps S101~S104:

[0054] Step S101: Input the original question submitted by the user into a preset large language model for rewriting, generating a rewritten question. The rewritten question has the same semantics as the original question, and its expression conforms to the preset format of financial terminology and the FAQ knowledge base.

[0055] Step S102: Input the original question into the pre-trained intent classification model to generate intent labels for the original question. Input the original question into the pre-trained entity recognition model to identify the entities and their categories in the original question, thereby obtaining an entity sequence.

[0056] Step S103: Merge the original question with the intent label and entity sequence to obtain the original concatenated text. Merge the rewritten question with the intent label and entity sequence to obtain the rewritten concatenated text.

[0057] Step S104: Input the original concatenated text and the rewritten concatenated text into the pre-trained semantic matching model to generate the original vector and the rewritten vector; calculate and sort the text similarity between the original vector and the rewritten vector and the semantic vector of each question item in the preprocessed FAQ knowledge base, and select the final answer.

[0058] The preprocessing of the FAQ knowledge base includes inputting each question item in the FAQ knowledge base into an intent classification model and an entity recognition model to generate an intent label and entity sequence for each question item; fusing the intent label and entity sequence of each question item with the corresponding question to obtain the question concatenation text; and inputting the question concatenation text into a semantic matching model to generate semantic vectors for each question item and storing them in the FAQ knowledge base.

[0059] like Figure 2 The diagram shows a flowchart of an intelligent customer service FAQ matching method based on deep semantic understanding.

[0060] In step S101, considering that users ask questions in various ways, the same question may have multiple expressions, and these expressions are often colloquial, brief, or ambiguous. Therefore, the original question submitted by the user is input into a preset large language model for rewriting, generating a rewritten question. The generated rewritten question has the same semantics as the original question, but its expression conforms to the preset format of financial terminology and the FAQ knowledge base.

[0061] In some embodiments, the pre-defined large language model is a model fine-tuned on a vertical domain corpus. The vertical domain corpus refers to a collection of text data specifically collected and annotated for a particular industry or application scenario (such as finance, healthcare, law, etc.). The fine-tuned large language model (such as ChatGLM) enables it to generate rewritten questions that better conform to the expression habits and professional requirements of the financial industry, avoiding colloquial or unprofessional expressions generated by general models.

[0062] In some embodiments, before rewriting the user-submitted original question into a preset large language model, the original question is first cleaned and standardized, such as noise character removal, synonym replacement, and terminology standardization, in order to purify the input text, unify semantic expression, adapt domain knowledge, and improve the input data quality for subsequent question rewriting, intent classification, entity recognition, etc.

[0063] In some embodiments, the original question is rewritten into a rewritten question using a preset prompt template or a sequence-to-sequence generation model.

[0064] Specifically, by designing specific prompt templates, the model is guided to generate rewritten questions that meet the requirements. For example, the template is set as follows:

[0065] Please rewrite the following user question into a professional question format that conforms to the financial customer service FAQ, while maintaining the original meaning:

[0066] Original question: [User input]

[0067] Rewrite the question:

[0068] Optionally, add domain constraints, such as:

[0069] "Requirements: Use 'Please ask' at the beginning, include the full name of the product (e.g., 'CSI 300 Index Fund'), and avoid colloquial language."

[0070] Therefore, during the rewriting process, the original question is embedded into the Prompt template and input into the model, and the model outputs the generated rewritten question.

[0071] Alternatively, it can be implemented based on sequence-to-sequence generation, modeling the question rewriting task as an end-to-end generation task of "original question → rewritten question", which can be achieved by fine-tuning the model.

[0072] In step S102, the original question is input into a pre-trained intent classification model to generate intent labels for the original question, such as account opening consultation, financial product inquiry, transaction process, etc. The original question is then input into a pre-trained entity recognition model to identify the entities in the original question (such as bank name, securities code, product name, time, amount, etc.) and their categories, resulting in a structured entity sequence.

[0073] In some embodiments, the intent classification model employs a bidirectional encoder representation model based on the Transformer architecture (BERT model).

[0074] Training methods for intent classification models include:

[0075] Construct an intent classification training set, which includes a large number of user questions in the financial field, and label each user question with a true intent tag.

[0076] The model is constructed using the BERT model, which includes an input layer, a BERT encoding layer, and a classification layer.

[0077] The BERT model is trained using an intent classification training set. User questions with real intent labels are input into the BERT model in batches to generate predicted intent labels for each user question. Loss functions for real intent labels and predicted intent labels are constructed, and the model is optimized by minimizing the loss functions to obtain the intent classification model.

[0078] In some embodiments, the entity recognition model employs a joint model of a bidirectional encoder representation model (BERT model) and a conditional random field (CRF).

[0079] Training methods for entity recognition models include:

[0080] Construct an entity recognition training set, which includes a large corpus of annotated entities in the financial field, covering key entity types.

[0081] The model is constructed using a joint BERT and CRF model, which includes an input layer, a BERT encoding layer, and a CRF correction layer.

[0082] The joint model is trained using an entity recognition training set. Batch input of corpora with real entity annotations into the joint model generates predicted entities and their types. A negative log-likelihood loss is constructed to optimize the difference between the real labels and the predicted values, resulting in the entity recognition model.

[0083] In step S103, the original question is fused with the intent label and entity sequence to obtain the original concatenated text. The rewritten question is then fused with the intent label and entity sequence to obtain the rewritten concatenated text.

[0084] In some embodiments, intent tags and entity sequences are added before and after the original question or the rewritten question with preset markers to obtain the original concatenated text and the rewritten concatenated text.

[0085] For example, the intent label is placed before the original or rewritten question, and the entity sequence is placed after the original or rewritten question:

[0086] [INTENT:Financial Product Inquiry]How to buy Fund A [ENTITY:Fund A].

[0087] In step S104, the original concatenated text and the rewritten concatenated text are input into the pre-trained semantic matching model to generate the original vector and the rewritten vector. The text similarity between the original vector and the rewritten vector and the semantic vector of each question item in the preprocessed FAQ knowledge base is calculated and sorted to obtain the final answer.

[0088] In some embodiments, the semantic matching model employs a bidirectional sentence encoder representation model (Sentence-BERT model).

[0089] Training methods for semantic matching models include:

[0090] Construct a semantic matching training set. This entity recognition training set includes a large number of FAQ question-answer pairs in the financial field. Similar sentence pairs are labeled as 1 (semantic same), and dissimilar sentence pairs are labeled as 0.

[0091] The model is constructed using a BERT-based Siamese network, comprising an input layer, a feature extraction layer, and a similarity calculation layer. For example, the similarity calculation layer can be used for a classification task, where the vectors of two sentences are concatenated and input into a fully connected layer, outputting a similarity probability; or it can be used for a regression task, directly calculating the cosine similarity and optimizing it with mean squared error (MSE).

[0092] The model is trained using a semantic matching training set. Annotated FAQ pairs are input into the model to generate predicted similarity values. For classification tasks, cross-entropy loss is constructed; for regression tasks, mean squared error loss or cosine embedding loss is constructed to optimize the model and obtain the semantic matching model.

[0093] In some embodiments, the cosine similarity between the original vector and the semantic vector of each question item in the preprocessed FAQ knowledge base is calculated to obtain a matching score and a first matching score list is constructed; the cosine similarity between the rewritten vector and the semantic vector of each question item in the preprocessed FAQ knowledge base is calculated to obtain a matching score and a second matching score list is constructed.

[0094] Merge the same question item from the first and second matching score lists, retaining the highest matching score. For example, if question item A has a matching score of 0.8 in the first matching score list and 0.9 in the second, merge the two lists and take the matching score of 0.9. Sort the matching scores of the merged question items in descending order, and select the top K questions. For example, select the top three questions with the highest matching scores. Use the FAQ answers for the selected questions as the final answers.

[0095] In some embodiments, after selecting a preset number of question items, the method further includes: using the selected preset number of question items as candidate options; calculating the consistency score between the intent label of the candidate option and the intent label of the original question; if the consistency score is lower than a preset score, the candidate option is filtered. Simultaneously, calculating the proportion of entities in the candidate options that cover the original question; if the proportion is lower than a preset proportion, the candidate option is filtered. The FAQ answers of the remaining candidate options after filtering are used as the final answers.

[0096] In some embodiments, to improve matching, the FAQ knowledge base is preprocessed, including inputting each question item in the FAQ knowledge base into the intent classification model and entity recognition model trained above to generate an intent label and entity sequence for each question item; fusing the intent label and entity sequence of each question item with the corresponding question to obtain concatenated question text; and inputting the concatenated question text into the semantic matching model trained above to generate semantic vectors for each question item and store them in the FAQ knowledge base.

[0097] In some embodiments, the semantic matching model employs a sentence bidirectional encoder representation model. The sentence bidirectional encoder representation model encodes the concatenated question text into semantic vectors, then associates the generated semantic vectors with the corresponding question items and stores them in a structured vector retrieval library to obtain a preprocessed FAQ knowledge base.

[0098] In some embodiments, structured vector retrieval libraries achieve vector indexing and fast retrieval through efficient similarity search algorithms, such as Facebook AI Similarity Search (Faiss) and Milvus vector database.

[0099] Corresponding to the intelligent customer service FAQ matching method based on deep semantic understanding, this invention also provides an intelligent customer service FAQ matching system based on deep semantic understanding, the system comprising:

[0100] The question rewriting module is used to rewrite the original questions mentioned by users into rewritten questions that conform to the professional terminology of the financial field and the preset format of the FAQ knowledge base through a preset large language model.

[0101] The intent classification module is used to identify the intent category of the original question through a pre-trained intent classification model and obtain the intent label.

[0102] The entity recognition module is used to identify the entities and their categories in the original question using a pre-trained entity recognition model, thereby obtaining an entity sequence.

[0103] The structured query representation construction module is used to fuse the original question with intent labels and entity sequences to obtain the original concatenated text; and to fuse the rewritten question with intent labels and entity sequences to obtain the rewritten concatenated text.

[0104] The FAQ knowledge base preprocessing module generates intent labels and entity sequences for each question item in the FAQ knowledge base using an intent classification model and an entity recognition model; it then fuses the intent labels and entity sequences of each question item with the corresponding question to obtain the question concatenation text; and finally, it encodes the question concatenation text using a pre-trained semantic matching model to obtain the semantic vector of each question item.

[0105] The semantic matching module encodes the original concatenated text and the rewritten concatenated text using a semantic matching model, resulting in original vectors and rewritten vectors. It then calculates and sorts the text similarity between the original vectors and rewritten vectors and the semantic vectors of each question item in the preprocessed FAQ knowledge base, and selects the final answer.

[0106] Corresponding to the above method, the present invention also provides an electronic device including a computer device, the computer device including a processor and a memory, the memory storing computer instructions, the processor executing the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the electronic device performs the steps of the method as described above.

[0107] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, performs the steps of the method as described above. The computer-readable storage medium may be a tangible storage medium, such as random access memory (RAM), main memory, read-only memory (ROM), electrically programmable ROM, electrically erasable programmable ROM, register, floppy disk, hard disk, removable storage disk, CD-ROM, or any other form of storage medium known in the art.

[0108] Those skilled in the art will understand that the exemplary components, systems, and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software, or a combination of both. Whether implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this invention. When implemented in hardware, it can be, for example, electronic circuits, application-specific integrated circuits (ASICs), appropriate firmware, plug-ins, function cards, etc. When implemented in software, the elements of this invention are programs or code segments used to perform the desired tasks. The programs or code segments can be stored in a machine-readable medium or transmitted over a transmission medium or communication link via data signals carried in a carrier wave.

[0109] It should be clarified that the present invention is not limited to the specific configurations and processes described above and shown in the figures. For the sake of brevity, detailed descriptions of known methods are omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown. Those skilled in the art can make various changes, modifications, and additions, or change the order of steps, after understanding the spirit of the present invention.

[0110] In this invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with or in place of features of other embodiments.

[0111] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, various modifications and variations of the embodiments of the present invention are possible. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An intelligent customer service FAQ matching method based on deep semantic understanding, characterized in that, The method comprises: inputting an original question submitted by a user to an intelligent customer service to a preset large language model for rewriting to generate a rewritten question; wherein the rewritten question has the same semantics as the original question, and the expression form of the rewritten question conforms to the professional terms in the financial field and the preset format of the FAQ knowledge base; inputting the original question into a pre-trained intent classification model to generate an intent label of the original question; inputting the original question into a pre-trained entity recognition model to identify the entity and its category of the original question, and obtain an entity sequence; adding the intent label and the entity sequence to the front and back of the original question with preset markers to generate an original spliced text; adding the intent label and the entity sequence to the front and back of the rewritten question with preset markers to generate a rewritten spliced text; inputting the original spliced text and the rewritten spliced text into a pre-trained semantic matching model respectively to generate an original vector and a rewritten vector; calculating the text similarity of the original vector and the rewritten vector with the semantic vector of each question item in the preprocessed FAQ knowledge base respectively and sorting; selecting a preset number of question items with high ranking as candidate items, calculating the consistency score of the intent label of the candidate items with the intent label of the original question, if the consistency score is lower than a preset score, filtering the candidate items; calculating the proportion of entities in the candidate items that cover the original question, if the proportion is lower than a preset proportion, filtering the candidate items; taking the FAQ answer of the filtered remaining candidate items as the final answer; wherein the preprocessing of the FAQ knowledge base comprises inputting each question item in the FAQ knowledge base into the intent classification model and the entity recognition model to generate the intent label and the entity sequence of each question item; fusing the intent label and the entity sequence of each question item with the corresponding question to obtain a question spliced text; inputting the question spliced text into the semantic matching model to generate the semantic vector of each question item and store it in the FAQ knowledge base. 2.The intelligent customer service FAQ matching method based on deep semantic understanding according to claim 1, characterized in that, Before inputting the original question submitted by the user into the preset large language model for rewriting, the method further comprises: cleaning and standardizing the original question, including at least noise character removal, synonym replacement, and term standardization. 3.The intelligent customer service FAQ matching method based on deep semantic understanding of claim 1, characterized in that, inputting the original question submitted by the user into the preset large language model for rewriting to generate a rewritten question, comprising: rewriting the original question into the rewritten question through a preset prompt template or a sequence-to-sequence generation model. 4.The intelligent customer service FAQ matching method based on deep semantic understanding of claim 1, characterized in that, The intent classification model uses a bidirectional encoder representation model based on the Transformer architecture for classification; the entity recognition model uses a joint model of bidirectional encoder representation model and conditional random field for predicting the entity sequence. 5.The intelligent customer service FAQ matching method based on deep semantic understanding of claim 1, characterized in that, calculating the text similarity of the original vector and the rewritten vector with the semantic vector of each question item in the preprocessed FAQ knowledge base respectively and sorting, comprising: calculating the cosine similarity of the original vector with the semantic vector of each question item in the preprocessed FAQ knowledge base to obtain a matching score and construct a first matching score list; Calculate the cosine similarity of the rewriting vector and the semantic vector of each question item in the preprocessed FAQ knowledge base to obtain a matching score and construct a second matching score list; Merge the same question items in the first matching score list and the second matching score list, and keep the highest matching score. Sort the matching scores of the merged question items in descending order. 6.The intelligent customer service FAQ matching method based on deep semantic understanding of claim 1, characterized in that, Input the question splicing text into the semantic matching model to generate semantic vectors of each question item and store them in the FAQ knowledge base, including: The semantic matching model uses a sentence bidirectional encoder representation model; Through the sentence bidirectional encoder representation model, the question splicing text is encoded into a semantic vector. The generated semantic vector is associated with the corresponding question item and stored in a structured vector retrieval library to construct the FAQ knowledge base. The structured vector retrieval library realizes vector indexing and fast retrieval through an efficient similarity search algorithm.

7. An intelligent customer service FAQ matching system based on deep semantic understanding, characterized in that, The system comprises: A question rewriting module configured to rewrite an original question mentioned by a user into a rewritten question conforming to professional terms in the financial field and a pre-set format of an FAQ knowledge base through a pre-set large language model; An intent classification module configured to identify an intent category of the original question through a pre-trained intent classification model to obtain an intent label; An entity recognition module configured to identify an entity and its category of the original question through a pre-trained entity recognition model to obtain an entity sequence; A structured query representation construction module configured to add the intent label and the entity sequence before and after the original question with pre-set markers to generate an original splicing text, and add the intent label and the entity sequence before and after the rewritten question with pre-set markers to generate a rewritten splicing text; An FAQ knowledge base preprocessing module configured to generate an intent label and an entity sequence of each question item in the FAQ knowledge base through the intent classification model and the entity recognition model, fuse the intent label and the entity sequence of each question item with the corresponding question to obtain a question splicing text, and encode the question splicing text through a pre-trained semantic matching model to obtain a semantic vector of each question item; A semantic matching module configured to encode the original splicing text and the rewritten splicing text through the semantic matching model to obtain an original vector and a rewritten vector, respectively, calculate the text similarity of the original vector and the rewritten vector with the semantic vector of each question item in the preprocessed FAQ knowledge base, and sort them; select a pre-set number of question items as candidates, calculate the consistency score of the intent label of the candidates and the intent label of the original question, filter the candidates if the consistency score is lower than a pre-set score, count the proportion of entities in the candidates that cover the original question, and filter the candidates if the proportion is lower than a pre-set proportion; and take the FAQ answer of the remaining candidates after filtering as the final answer.

8. A computer readable storage medium having stored thereon computer programs / instructions, characterized in that, The computer program / instructions are executed by the processor to implement the steps of the method of any one of claims 1 to 6. The computer program / instructions are executed by the processor to implement the steps of the method of any one of claims 1 to 6.

Images