Intelligent finance and tax question answering method based on artificial intelligence question answering system

Abstract

The invention discloses an intelligent finance and tax question answering method based on an artificial intelligence question answering system, and the method comprises the steps: S1, generating a structured semantic representation through a natural language processing technology, and generating an intention tag through a machine learning algorithm; s2, constructing and maintaining a knowledge graph containing finance and tax field related information, storing the information in the knowledge graph in a structured manner, and reasoning by using a logical reasoning algorithm to generate a preliminary answer; s3, designing a document retrieval and indexing mechanism, and verifying and supplementing the preliminary answer; s4, training the system based on the finance and tax related data and question and answer pairs, and continuously updating the model; s5, identifying common finance and taxation question modes by using a mode identification technology, and providing standardized answers; and S6, supporting multiple rounds of dialogues with the user, and generating more detailed and accurate answers by gradually clarifying the questions of the user. The method has the advantages of high accuracy, high response speed and capability of continuously learning and adapting to new problems.

Description

Technical Field

[0001] The present invention relates to the field of intelligent question answering technology, and in particular to an intelligent answering method for financial and taxation problems based on an artificial intelligence question answering system. Background Art

[0002] With the digital transformation of the global economy, the demand for financial and tax management and consulting is growing. Enterprises and individuals need real-time and accurate financial and tax information to cope with complex regulatory environments. However, traditional financial and tax consulting services usually rely on manual operations, which is not efficient and flexible enough in the face of massive data and ever-changing regulations.

[0003] In the existing technology, most of the financial and taxation question-and-answer systems are based on simple keyword matching or static database queries. These systems are usually unable to accurately understand the complex query intentions of users and lack the ability to deeply reason about financial and taxation regulations. Specifically, the existing technology has obvious deficiencies in the following aspects:

[0004] 1. Lack of ability to deeply understand user intent: Existing systems can often only handle simple queries. For questions that contain multiple meanings or require in-depth understanding, these systems cannot provide accurate answers.

[0005] 2. Difficulty in updating knowledge: Financial and tax regulations are constantly changing. The process of updating the knowledge base of traditional systems is cumbersome and time-consuming, making it difficult to keep the information up to date.

[0006] 3. Inefficient information retrieval: Traditional taxation question-and-answer systems often rely on predefined paths and rules for information retrieval, which is inefficient when dealing with complex or non-standard questions.

[0007] 4. Lack of adaptability to complex problems: When faced with complex or rare financial and tax issues, the existing system often cannot provide solutions, or the answers provided lack specificity and accuracy.

[0008] Therefore, how to provide an intelligent solution method for financial and tax issues based on an artificial intelligence question-answering system is an issue that technical personnel in this field urgently need to solve. Summary of the invention

[0009] The purpose of the present invention is to propose an intelligent solution method for financial and taxation problems based on artificial intelligence. The method comprehensively applies natural language processing, knowledge graph construction, efficient document retrieval technology and machine learning, and describes in detail the process of intelligently solving financial and taxation problems. The present invention can accurately understand the user's query intention, quickly retrieve and provide relevant financial and taxation information, and has the advantages of high accuracy, fast response speed and continuous learning to adapt to new problems.

[0010] According to an embodiment of the present invention, a method for intelligently answering financial and taxation questions based on an artificial intelligence question-answering system comprises the following steps:

[0011] S1. Use natural language processing technology to segment, tag and parse the natural language finance and taxation questions input by users to generate structured semantic representations. Use machine learning algorithms to identify the intent of the structured semantic representations and generate intent labels.

[0012] S2. Build and maintain a knowledge graph containing relevant information in the field of finance and taxation, store the information in a structured manner in the knowledge graph, extract relevant knowledge from the knowledge graph based on the intent label, use logical reasoning algorithms to perform reasoning, and generate preliminary answers;

[0013] S3. Design document retrieval and indexing mechanisms to pre-process and index financial and tax regulations, policies, and case documents to support rapid retrieval, and combine database technology to verify and supplement the above preliminary answers;

[0014] S4. Train the system based on a large amount of finance and taxation-related data and question-answer pairs, and continuously update the model;

[0015] S5. Use pattern recognition technology to identify common financial and taxation problem patterns and provide standardized answers. At the same time, detect abnormal problems and provide corresponding solutions;

[0016] S6. Support multiple rounds of dialogue with users, generate more detailed and accurate answers by gradually clarifying users' questions, and use data encryption and permission management technology to securely process and protect the privacy of users' financial and tax information.

[0017] Optionally, the S1 specifically includes:

[0018] S11. Obtain the natural language taxation question input by the user and convert it into text format T:

[0019] T={t1,t2,...,t n};

[0020] Among them, t i The character sequence representing the i-th word;

[0021] S12. Use a word segmentation algorithm to segment the acquired text into several word sequences. The word segmentation algorithm is performed by maximizing the probability of the word sequence:

[0022] W = {w1, w2, ..., w m};

[0023]

[0024] Among them, W represents the word sequence after word segmentation, w i represents the i-th word, satisfying P(W|T) represents the probability of the word sequence W under the text T, P(w i |, T) indicates that the previous word w 1：i-1 Given a text T, the i-th word w i The conditional probability of

[0025] S13, perform part-of-speech tagging on the word sequence after word segmentation to generate a part-of-speech tagging sequence:

[0026] P = {p1, p2, ..., p m};

[0027] Among them, P represents the part-of-speech tag sequence, p i Represents the part-of-speech tag of the i-th word;

[0028] S14, using a semantic parsing algorithm to perform semantic parsing on the vocabulary sequence to generate a structured semantic representation s, where s=f(W, P), f is a semantic parsing function including a structured representation of the semantic relationship between the vocabulary;

[0029] S15. Input the generated structured semantic representation s into the pre-trained machine learning model for intent recognition. The machine learning model is trained based on the training data set D:

[0030] D={(S1, T1), (S2, T2),..., (S k , T k )};

[0031]

[0032] Among them, S i represents the structured semantic representation of the i-th training sample, T i Represents the intent label of the i-th training sample. The machine learning model learns the parameter θ by maximizing the log-likelihood function L(θ);

[0033] S16. The machine learning model generates the intent label T of the user question by learning the mapping relationship between the structured semantic representation and the intent label in the training data set, where T = g (S ; θ), g is the intention recognition function of the machine learning model, θ is the model parameter, and the model parameter θ is optimized by maximizing the log-likelihood function L(θ). The update of the model parameter θ is achieved through the gradient ascent algorithm:

[0034]

[0035]

[0036]

[0037] Among them, P(T i |;θ) represents the given structured semantic representation S i and parameter θ, predict the correct intent label T i The probability of the model for each S in the training data set is evaluated. i Prediction T i is the accuracy, α is the learning rate, is the gradient of L(θ) with respect to θ.

[0038] Optionally, the S2 specifically includes:

[0039] S21. Initialize the finance and taxation knowledge graph and construct a graph structure containing key nodes and edges, where each node represents a specific concept and entity in the finance and taxation field, and the edge represents the relationship between entities;

[0040] S22. Organize the information into a system to form data entries:

[0041] E = {e1, e2, ..., e n};

[0042] where e i Represents an independent information entity;

[0043] S23. For each data entry e i Label key attributes and relationships, map them to corresponding nodes in the knowledge graph, and generate node attributes A i ={a i1 , a i2 , …, a im} and the relation R i = {r i1 , r i2 ,…,r il}, where a ij Represents node e i The jth attribute of ik Represents node e i The kth relationship with other nodes;

[0044] S24. Use graph database technology to store knowledge graphs, use query language for data retrieval and logical reasoning, and realize complex query functions;

[0045] S25. Based on the user's query intention label T, retrieve the nodes and edges related to T from the knowledge graph, perform a graph traversal operation, and form a subgraph SG(T) containing all relevant information, where SG represents the subgraph extracted according to the intention T;

[0046] S26. Use the graph algorithm to perform logical reasoning on the subgraph SG(T) to generate an answer to the user's question, using the reasoning result I(T), where I(T) = f(SG(T), R), f is the logical reasoning function, and R is the set of reasoning rules applied to the subgraph;

[0047] S27. Format the logical reasoning results I(T) into answers that users can understand, and update the knowledge graph to reflect the latest financial and tax information and regulatory changes.

[0048] Optionally, the S3 specifically includes:

[0049] S31. Design and implement a document retrieval and indexing system that specializes in processing financial and tax-related regulations, policies, and case documents;

[0050] S32, the tax document D = {d1, d2, ..., d n}Preprocessing, including text cleaning, standardization and language analysis, where d i Represents a single document;

[0051] S33, using information extraction technology from each preprocessed document d i Extract key information from the tuple to form information:

[0052] I={(k1, v1), (k2, v2),..., (k m , v m )};

[0053] where k j is the keyword, v j is the corresponding value;

[0054] S34, establishing a document index based on the extracted information tuples, each document index consisting of a series of key-value pairs, so that any query for keyword k can quickly locate a document set containing the keyword;

[0055] S35. Use inverted index technology to build index structure:

[0056] IDX={k:{d i1 , d i2 , …, d ik}};

[0057] Where k is the index keyword, d ij is the jth document containing keyword k;

[0058] S36, implement efficient query parsing and execution mechanism, support complex query expressions, and use edit distance algorithm to determine the similarity between query terms and words in documents for approximate matching query, so that relevant documents can be effectively retrieved even if the query terms are slightly misspelled:

[0059] L(s,t)=min(L(s-1,t)+1,L(s,t-1)+1,L(s-1,t-1)+cost);

[0060] Where L(s, t) is the edit distance between strings s and t, and cost is the cost of the replacement operation, which is usually 1 unless the two characters are the same, in which case it is 0;

[0061] S37. Display the retrieved documents and their related information to users, provide document sorting and relevance feedback, and regularly update the indexing system so that the system can reflect the latest regulatory changes and policy updates.

[0062] Optionally, the S4 specifically includes:

[0063] S41. Collect and organize a large amount of tax-related data, including previous question-answer pairs, user feedback, and records of changes in tax laws and regulations, to form a comprehensive data set D = {d1, d2, ..., d n}, where d i Represents a single data instance;

[0064] S42, preprocessing the collected data D;

[0065] S43. Based on the preprocessed data set D, a machine learning model M is trained using a supervised learning method to solve the fiscal and taxation problem. Model training includes selecting appropriate features and algorithms.

[0066] S44, defining a loss function L(θ) to optimize the model M, where θ represents a model parameter;

[0067]

[0068] Among them, y i is the actual label of the data point di, is the model M under parameter θ for data point d i The predicted label of

[0069] S45. Optimize the parameters of the model and use the gradient descent algorithm to adjust the parameter θ to minimize the loss function L(θ):

[0070]

[0071] Among them, α is the learning rate, is the gradient of the loss function L(θ) with respect to the parameter θ;

[0072] S46. Regularly evaluate the performance of model M during the training process. After completing the training of model M, integrate it into the financial and taxation question-and-answer system so that the model can dynamically provide accurate answers based on user queries in actual applications.

[0073] Optionally, the S5 specifically includes:

[0074] S51. Use pattern recognition technology to identify common tax problem patterns. By analyzing the historical question-answering data set Q = {q1, q2, ..., q m}, where q i Represents a historical question-and-answer instance;

[0075] S52. Build a classification model C that uses the random forest algorithm to automatically classify and identify the types of fiscal and tax issues. The random forest consists of multiple decision trees, each of which is trained using a randomly selected feature subset. For each tree T in the random forest i , select m features for training, where m≤M, and the decision rule of each tree is based on maximizing the information gain IG:

[0076]

[0077] Where D is the dataset, f is the feature used to segment the dataset, v is the number of possible values ​​of feature f, and D j is the subset divided by the jth value of feature f, H(D) is the entropy of data set D, |D j | and |D| are the sizes of the subset and the total dataset, respectively;

[0078] S53. Define the performance evaluation criteria of model C, and use precision P, recall R and F1 score F to evaluate model performance:

[0079]

[0080]

[0081]

[0082] Among them, TP, FP, and FN represent the number of true positives, false positives, and false negatives, respectively;

[0083] S54. Conduct in-depth analysis of the identified problem patterns and use the isolation forest algorithm for anomaly detection to identify fiscal and tax issues that deviate from the norm. In the isolation forest, outliers are usually isolated earlier and therefore have a shorter average path length:

[0084]

[0085]

[0086] Where x is an instance, n is the number of samples, E(h(x)) is the average path length of x in the tree, c(n) is the average unisolated path length in the random tree, and H(i) is the Harman number, which can be approximated as ln(i)+γ;

[0087] S55. Develop a response generation module R that dynamically generates answers based on the identified question patterns and anomaly detection results, using natural language generation technology to form text responses that are highly natural and professional;

[0088] S56. Continuously monitor system performance and adjust the model based on user feedback, using online learning mechanisms to continuously optimize the model C and response generation module R.

[0089] Optionally, the S6 specifically includes:

[0090] S61. Support multiple rounds of dialogue with users to gradually clarify the details of user questions and specific needs, and use dialogue state management mechanism to track the progress and context of the dialogue;

[0091] S62. In order to understand the user's natural language input, the system applies natural language processing technology to perform language analysis;

[0092] S63. Based on the user's intent and the context of the question, the system dynamically constructs a query for retrieving relevant financial and tax data and information. The query construction includes converting the user's intent into a query language for a specific database and knowledge base;

[0093] S64, implement a dialogue generation module G, which generates appropriate responses and further inquiries based on the current dialogue state and system available data, and guides the dialogue to develop in the direction of useful information;

[0094] S65. Use deep learning models to optimize the dialogue generation process. Model training uses a sequence-to-sequence architecture:

[0095]

[0096] Where x represents the input sequence, yi represents the i-th word in the response generated by the model after a given input, and N is the total number of words in the response;

[0097] S66. The system uses a real-time feedback mechanism to adjust the dialogue strategy, including user satisfaction evaluation and dialogue effect tracking. The dialogue system continuously learns and adapts to users' specific needs and preferences through user interactions, and applies online learning technology to dynamically update dialogue models and content.

[0098] The beneficial effects of the present invention are:

[0099] (1) The present invention combines advanced natural language processing technology and intent recognition algorithms to accurately parse users' natural language financial and tax questions and provide accurate answers. This enables the system to understand and handle a variety of complex financial and tax issues, especially those that require in-depth understanding of regulations and cases, thereby greatly improving the quality of answers and user satisfaction.

[0100] (2) The present invention utilizes the constructed financial and tax knowledge graph and efficient document retrieval technology to quickly access and extract relevant financial and tax laws, policies and case documents. This not only significantly improves the system's response speed, but also ensures that the information provided to users is the latest and most relevant, meeting the real-time needs in a rapidly changing financial and tax environment.

[0101] (3) Through the training of machine learning models and pattern recognition technology, the present invention enables the system to continuously learn and adapt to new financial and tax problem patterns. This adaptive learning mechanism ensures that the system continues to improve over time, can handle increasingly complex financial and tax problems, and provide standardized answers or corresponding processing methods, thereby enhancing the system's adaptability and ability to handle complex problems. BRIEF DESCRIPTION OF THE DRAWINGS

[0102] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0103] Figure 1 This is a general framework diagram of an intelligent solution system for financial and taxation problems based on artificial intelligence proposed by the present invention;

[0104] Figure 2 A schematic diagram of the natural language processing and intent recognition process of an artificial intelligence-based intelligent solution system for financial and taxation issues proposed by the present invention;

[0105] Figure 3 This is a schematic diagram of the construction and use of the financial and taxation knowledge graph of the artificial intelligence-based intelligent solution system for financial and taxation problems proposed in the present invention. DETAILED DESCRIPTION

[0106] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0107] refer to Figure 1-Figure 3 , a method for intelligently answering financial and taxation questions based on an artificial intelligence question-answering system, characterized in that it includes the following steps:

[0108] S1. Use natural language processing technology to segment, tag and parse the natural language finance and taxation questions input by users to generate structured semantic representations. Use machine learning algorithms to identify the intent of the structured semantic representations and generate intent labels.

[0109] In this implementation, S1 specifically includes:

[0110] S11. Obtain the natural language taxation question input by the user and convert it into text format T:

[0111] T={t1,t2,...,t n};

[0112] Among them, t i The character sequence representing the i-th word;

[0113] S12. Use a word segmentation algorithm to segment the acquired text into several word sequences. The word segmentation algorithm is performed by maximizing the probability of the word sequence:

[0114] W = {w1, w2, ..., w m};

[0115]

[0116] Among them, W represents the word sequence after word segmentation, w i represents the i-th word, satisfying P(W|T) represents the probability of the word sequence W under the text T, P(w i |, T) indicates that the previous word w 1：i-1 Given a text T, the i-th word w i The conditional probability of

[0117] S13, perform part-of-speech tagging on the word sequence after word segmentation to generate a part-of-speech tagging sequence:

[0118] P = {p1, p2, ..., p m};

[0119] Among them, P represents the part-of-speech tag sequence, p i Represents the part-of-speech tag of the i-th word;

[0120] S14, using a semantic parsing algorithm to perform semantic parsing on the vocabulary sequence to generate a structured semantic representation S, where S=f(W, P), f is a semantic parsing function including a structured representation of the semantic relationship between the words;

[0121] S15. Input the generated structured semantic representation S into the pre-trained machine learning model for intent recognition. The machine learning model is trained based on the training data set D:

[0122] D={(S1, T1), (S2, T2),..., (S k , T k )};

[0123]

[0124] Among them, S i represents the structured semantic representation of the i-th training sample, T i Represents the intent label of the i-th training sample. The machine learning model learns the parameter θ by maximizing the log-likelihood function L(θ);

[0125] S16. The machine learning model generates the intent label T of the user question by learning the mapping relationship between the structured semantic representation and the intent label in the training data set, where T = g(S; θ), g is the intent recognition function of the machine learning model, θ is the model parameter, and the model parameter θ is optimized by maximizing the log-likelihood function L(θ). The update of the model parameter θ is achieved by the gradient ascent algorithm:

[0126]

[0127]

[0128]

[0129] Among them, P(T i |;θ) represents the given structured semantic representation S i and parameter θ, predict the correct intent label T i The probability of the model for each S in the training data set is evaluated. i Prediction T i is the accuracy, α is the learning rate, is the gradient of L(θ) with respect to θ.

[0130] S2. Build and maintain a knowledge graph containing relevant information in the field of finance and taxation, store the information in a structured manner in the knowledge graph, extract relevant knowledge from the knowledge graph based on the intent label, use logical reasoning algorithms to perform reasoning, and generate preliminary answers;

[0131] In this implementation, S2 specifically includes:

[0132] S21. Initialize the finance and taxation knowledge graph and construct a graph structure containing key nodes and edges, where each node represents a specific concept and entity in the finance and taxation field, and the edge represents the relationship between entities;

[0133] S22. Organize the information into a system to form data entries:

[0134] E = {e1, e2, ..., e n};

[0135] where e i Represents an independent information entity;

[0136] S23. For each data entry e i Label key attributes and relationships, map them to corresponding nodes in the knowledge graph, and generate node attributes A i ={a i1 , a i2 , …, a im} and the relation R i = {r i1 , r i2 ,…,r il}, where a ij Represents node e i The jth attribute of ik Represents node e i The kth relationship with other nodes;

[0137] S24. Use graph database technology to store knowledge graphs, use query language for data retrieval and logical reasoning, and realize complex query functions;

[0138] S25. Based on the user's query intention label T, retrieve the nodes and edges related to T from the knowledge graph, perform a graph traversal operation, and form a subgraph SG(T) containing all relevant information, where SG represents the subgraph extracted according to the intention T;

[0139] S26. Use the graph algorithm to perform logical reasoning on the subgraph sG(T) to generate an answer to the user's question, using the reasoning result I(T), where I(T) = f(SG(T), R), f is the logical reasoning function, and R is the set of reasoning rules applied to the subgraph;

[0140] S27. Format the logical reasoning results I(T) into answers that users can understand, and update the knowledge graph to reflect the latest financial and tax information and regulatory changes.

[0141] S3. Design document retrieval and indexing mechanisms to pre-process and index financial and tax regulations, policies, and case documents to support rapid retrieval, and combine database technology to verify and supplement the above preliminary answers;

[0142] In this implementation, S3 specifically includes:

[0143] S31. Design and implement a document retrieval and indexing system that specializes in processing financial and tax-related regulations, policies, and case documents;

[0144] S32, the tax document D = {d1, d2, ..., d n}Preprocessing, including text cleaning, standardization and language analysis, where d i Represents a single document;

[0145] S33, using information extraction technology from each preprocessed document d i Extract key information from the tuple to form information:

[0146] I={(k1, v1), (k2, v2),..., (k m , v m )};

[0147] where k j is the keyword, v j is the corresponding value;

[0148] S34, establishing a document index based on the extracted information tuples, each document index consisting of a series of key-value pairs, so that any query for keyword k can quickly locate a document set containing the keyword;

[0149] S35. Use inverted index technology to build index structure:

[0150] IDX={k:{d i1 , d i2 , …, d ik};

[0151] Where k is the index keyword, d ij is the jth document containing keyword k;

[0152] S36, implement efficient query parsing and execution mechanism, support complex query expressions, and use edit distance algorithm to determine the similarity between query terms and words in documents for approximate matching query, so that relevant documents can be effectively retrieved even if the query terms are slightly misspelled:

[0153] L(s,t)=min(L(s-1,t)+1,L(s,t-1)+1,L(s-1,t-1)+cost);

[0154] Where L(s, t) is the edit distance between strings s and t, and cost is the cost of the replacement operation, which is usually 1 unless the two characters are the same, in which case it is 0;

[0155] S37. Display the retrieved documents and their related information to users, provide document sorting and relevance feedback, and regularly update the indexing system so that the system can reflect the latest regulatory changes and policy updates.

[0156] S4. Train the system based on a large amount of finance and taxation-related data and question-answer pairs, and continuously update the model;

[0157] In this implementation, S4 specifically includes:

[0158] S41. Collect and organize a large amount of tax-related data, including previous question-answer pairs, user feedback, and records of changes in tax laws and regulations, to form a comprehensive data set D = {d1, d2, ..., d n}, where d i Represents a single data instance;

[0159] S42, preprocessing the collected data D;

[0160] S43. Based on the preprocessed data set D, a machine learning model M is trained using a supervised learning method to solve the fiscal and taxation problem. Model training includes selecting appropriate features and algorithms.

[0161] S44, defining a loss function L(θ) to optimize the model M, where θ represents a model parameter;

[0162]

[0163] Among them, y i is the data point d i The actual label, is the model M under parameter θ for data point d i The predicted label of

[0164] S45. Optimize the parameters of the model and use the gradient descent algorithm to adjust the parameter θ to minimize the loss function L(θ):

[0165]

[0166] Among them, α is the learning rate, is the gradient of the loss function L(θ) with respect to the parameter θ;

[0167] S46. Regularly evaluate the performance of model M during the training process. After completing the training of model M, integrate it into the financial and taxation question-and-answer system so that the model can dynamically provide accurate answers based on user queries in actual applications.

[0168] S5. Use pattern recognition technology to identify common financial and taxation problem patterns and provide standardized answers. At the same time, detect abnormal problems and provide corresponding solutions;

[0169] In this implementation, S5 specifically includes:

[0170] S51. Use pattern recognition technology to identify common tax problem patterns. By analyzing the historical question-answering data set Q = {q1, q2, ..., q m}, where q i Represents a historical question-and-answer instance;

[0171] S52. Build a classification model C that uses the random forest algorithm to automatically classify and identify the types of fiscal and tax issues. The random forest consists of multiple decision trees, each of which is trained using a randomly selected feature subset. For each tree T in the random forest i , select m features for training, where m≤M, and the decision rule of each tree is based on maximizing the information gain IG:

[0172]

[0173] Where D is the dataset, f is the feature used to segment the dataset, v is the number of possible values ​​of feature f, and D j is the subset divided by the jth value of feature f, H(D) is the entropy of data set D, |D j | and |D| are the sizes of the subset and the total dataset, respectively;

[0174] S53. Define the performance evaluation criteria of model C, and use precision P, recall R and F1 score F to evaluate model performance:

[0175]

[0176]

[0177]

[0178] Among them, TP, FP, and FN represent the number of true positives, false positives, and false negatives, respectively;

[0179] S54. Conduct in-depth analysis of the identified problem patterns and use the isolation forest algorithm for anomaly detection to identify fiscal and tax issues that deviate from the norm. In the isolation forest, outliers are usually isolated earlier and therefore have a shorter average path length:

[0180]

[0181]

[0182] Where x is an instance, n is the number of samples, E(h(x)) is the average path length of x in the tree, c(n) is the average unisolated path length in the random tree, and H(i) is the Harman number, which can be approximated as ln(i)+γ;

[0183] S55. Develop a response generation module R that dynamically generates answers based on the identified question patterns and anomaly detection results, using natural language generation technology to form text responses that are highly natural and professional;

[0184] S56. Continuously monitor system performance and adjust the model based on user feedback, using online learning mechanisms to continuously optimize the model C and response generation module R.

[0185] S6. Support multiple rounds of dialogue with users, generate more detailed and accurate answers by gradually clarifying users' questions, and use data encryption and permission management technology to securely process and protect the privacy of users' financial and tax information.

[0186] In this implementation, S6 specifically includes:

[0187] S61. Support multiple rounds of dialogue with users to gradually clarify the details of user questions and specific needs, and use dialogue state management mechanism to track the progress and context of the dialogue;

[0188] S62. In order to understand the user's natural language input, the system applies natural language processing technology to perform language analysis;

[0189] S63. Based on the user's intent and the context of the question, the system dynamically constructs a query for retrieving relevant financial and tax data and information. The query construction includes converting the user's intent into a query language for a specific database and knowledge base;

[0190] S64, implement a dialogue generation module G, which generates appropriate responses and further inquiries based on the current dialogue state and system available data, and guides the dialogue to develop in the direction of useful information;

[0191] S65. Use deep learning models to optimize the dialogue generation process. Model training uses a sequence-to-sequence architecture:

[0192]

[0193] Where x represents the input sequence, yi represents the i-th word in the response generated by the model after a given input, and N is the total number of words in the response;

[0194] S66. The system uses a real-time feedback mechanism to adjust the dialogue strategy, including user satisfaction evaluation and dialogue effect tracking. The dialogue system continuously learns and adapts to users' specific needs and preferences through user interactions, and applies online learning technology to dynamically update dialogue models and content.

[0195] Embodiment 1:

[0196] In order to verify the effectiveness of the present invention, the present invention is applied in the actual operation of a medium-sized accounting consulting company B. Company B needs to handle a large number of financial and tax consultations every day. Faced with complex and ever-changing financial and tax regulations and customers' diverse consulting needs, the traditional manual answering method is often inefficient and error-prone. After deploying the intelligent answering system for financial and tax issues based on artificial intelligence, the company hopes to improve the efficiency and accuracy of consulting and reduce the work pressure of financial advisors.

[0197] During the implementation process, the system first uses natural language processing technology to understand the customer's query intent, and then quickly retrieves relevant information through the constructed financial and tax knowledge graph. For example, when a customer asks about the specific application of the latest VAT policy, the system can accurately interpret the query intent and provide a detailed interpretation of the latest policy and relevant case analysis.

[0198] Specific data shows that after deploying the present invention, the consultation response time of Company B was reduced from an average of 15 minutes to 5 minutes, and the accuracy of answers increased from 85% to more than 95%. In addition, the system can also self-learn and optimize based on user feedback, gradually improving its ability to handle complex and rare problems. In the operating data of one quarter, the number of consultations handled by the system exceeded 5,000, including tax planning, interpretation of tax preferential policies, cross-border tax issues and other aspects.

[0199] For example, when dealing with a tax consultation issue regarding a cross-border e-commerce company, the system not only provided basic tax rate information, but also guided users to understand the tax incentives and risk prevention measures that may be applicable based on the company's specific operating conditions. In this consultation, the system successfully predicted and warned of potential tax risks by analyzing past cases of similar companies, helping customers avoid possible economic losses.

[0200] By real-time monitoring and analyzing the system's operating results, the management of Company B found that the deployment of the system significantly improved customer service quality and customer satisfaction. In customer feedback, more than 90% of users expressed that they were very satisfied with the answers provided by the intelligent system and believed that the answers were professional and easy to understand. Within a quarter after the deployment of the present invention, Company B recorded the following effective consultation processing cases:

[0201] Table 1 Comparison of data before and after deployment

[0202] Indicator Category describe Before deploying the system After deploying the system Consultation response time Average time to complete a consultation 15 minutes 5 minutes Consultation accuracy Proportion of correct answers to inquiries 85％ More than 95% Customer satisfaction Satisfaction percentage based on customer feedback surveys 80％ More than 90% Number of consultations handled Total number of consultations handled by the system in a quarter 2000 times More than 5000 times

[0203] From these data, we can see that before the deployment of the system, the consultation response time of Company B was 15 minutes, and after the deployment, it was significantly reduced to 5 minutes. In addition, the consultation accuracy rate increased from 85% to more than 95%. Through real-time monitoring and system self-optimization, the dialogue generation module can understand and respond to customer questions more accurately, thereby improving the overall customer satisfaction.

[0204] In specific case analysis, the system successfully dealt with a series of complex financial and tax issues, such as cross-border taxation, corporate income tax declaration, and personal income tax optimization. The system's deep learning model and knowledge graph enable it to provide targeted solutions to different problems, significantly improving the relevance and practicality of the solutions. In the process of dealing with tax issues of cross-border e-commerce companies, the system not only quickly provided basic information on tax rates, but also deeply analyzed preferential measures under specific policy backgrounds, helping customers avoid potential tax risks and thus saving a lot of financial costs.

[0205] In addition, through the abnormal consulting patterns identified by the isolation forest algorithm, the system can warn of possible errors or unusual complex issues, conduct risk assessment and necessary preparation in advance, and ensure that the consulting services provided are more reliable and comprehensive. This function is particularly useful for consulting involving the interpretation of new regulations or changes in tax rates.

[0206] Within a quarter after the implementation of the present invention, the market competitiveness of Company B in the field of financial and tax consulting was significantly improved. The company's internal report showed that the number of new customers increased by 30%, and the return rate of old customers increased by 40%. The improvement in service quality and the acceleration of response speed were clearly felt. These changes not only improved the company's business performance, but also strengthened customers' trust and reliance on the company's services.

[0207] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. An intelligent solution method for financial and taxation problems based on an artificial intelligence question-answering system. It is characterized in that The steps include: S1. Use natural language processing technology to segment, tag and parse the natural language finance and tax issues input by users to generate structured semantic representations. Use machine learning algorithms to identify the intent of the structured semantic representations and generate intent labels. S2. Build and maintain a knowledge graph containing relevant information in the field of finance and taxation, store the information in a structured manner in the knowledge graph, extract relevant knowledge from the knowledge graph based on the intent label, use logical reasoning algorithms to perform reasoning, and generate preliminary answers; S3. Design a document retrieval and indexing mechanism to pre-process and index financial and tax regulations, policies and case documents to support rapid retrieval, and combine database technology to verify and supplement the above preliminary answers; S4. Train the system based on a large amount of finance and taxation-related data and question-answer pairs, and continuously update the model; S5. Use pattern recognition technology to identify common financial and taxation problem patterns and provide standardized answers. At the same time, detect abnormal problems and provide corresponding solutions; S6. Support multiple rounds of dialogue with users, generate more detailed and accurate answers by gradually clarifying users' questions, and use data encryption and permission management technology to securely process and protect the privacy of users' financial and tax information.

2. According to claim 1, a method for intelligently answering financial and taxation questions based on an artificial intelligence question-answering system, It is characterized in that The S1 specifically includes: S11. Obtain the natural language taxation question input by the user and convert it into text format T: T={t 1 ,t 2 ,…,t n }; Among them, t i The character sequence representing the i-th word; S12. Use a word segmentation algorithm to segment the acquired text into several word sequences. The word segmentation algorithm is performed by maximizing the probability of the word sequence: In={in 1 ,In 2 ,…,In m }; Among them, W represents the word sequence after word segmentation, w i represents the i-th word, satisfying P(W|T) represents the probability of the word sequence W under the text T, O(w i |,T) indicates that the previous word w 1:i-1 Given a text T, the i-th word w i The conditional probability of S13, perform part-of-speech tagging on the word sequence after word segmentation to generate a part-of-speech tagging sequence: O={p 1 ,p 2 ,…,p m }; Among them, P represents the part-of-speech tag sequence, p i Represents the part-of-speech tag of the i-th word; S14, using a semantic parsing algorithm to perform semantic parsing on the vocabulary sequence to generate a structured semantic representation S, where S = f(W, P), f is a semantic parsing function, including a structured representation of the semantic relationship between the words; S15. Input the generated structured semantic representation S into the pre-trained machine learning model for intent recognition. The machine learning model is trained based on the training data set D: D={(S 1 ,T 1 ),(S 2 ,T 2 ),…,(S k ,T k )}; Among them, S i represents the structured semantic representation of the i-th training sample, T i Represents the intent label of the i-th training sample. The machine learning model learns the parameter θ by maximizing the log-likelihood function L(θ); S16. The machine learning model generates the intent label T of the user question by learning the mapping relationship between the structured semantic representation and the intent label in the training data set, where T = g(S; θ), g is the intent recognition function of the machine learning model, θ is the model parameter, and the model parameter θ is optimized by maximizing the log-likelihood function L(θ). The update of the model parameter θ is achieved by the gradient ascent algorithm: Among them, P(T i |;θ) represents the given structured semantic representation S i and parameter θ, predict the correct intent label T i The probability of the model for each S in the training data set is evaluated. i Prediction T i is the accuracy, α is the learning rate, is the gradient of L(θ) with respect to θ.

3. According to claim 2, a method for intelligently answering financial and taxation questions based on an artificial intelligence question-answering system, It is characterized in that The S2 specifically includes: S21. Initialize the finance and taxation knowledge graph and construct a graph structure containing key nodes and edges, where each node represents a specific concept and entity in the finance and taxation field, and the edge represents the relationship between entities; S22. Organize the information into a system to form data entries: And={and 1 ,And 2 ,…,And n }; where e i Represents an independent information entity; S23. For each data entry e i Label key attributes and relationships, map them to corresponding nodes in the knowledge graph, and generate node attributes A i ={a i1 ,a i2 ,…,a im } and the relation R i = {r i1 ,r i2 ,…,r il }, where a ij Represents node e i The jth attribute of ik Represents node e i The kth relationship with other nodes; S24. Use graph database technology to store knowledge graphs, use query language for data retrieval and logical reasoning, and realize complex query functions; S25. Based on the user's query intention label T, retrieve the nodes and edges related to T from the knowledge graph, perform a graph traversal operation, and form a subgraph SG(T) containing all relevant information, where SG represents the subgraph extracted according to the intention T; S26, using the graph algorithm to perform logical reasoning on the subgraph SG(T), generate an answer to the user's question, and use the reasoning result I(T), where I(T) = f(SG(T), R), f is the logical reasoning function, and R is the set of reasoning rules applied to the subgraph; S27. Format the logical reasoning results I(T) into answers that users can understand, and update the knowledge graph to reflect the latest financial and tax information and regulatory changes.

4. According to claim 3, a method for intelligently answering financial and taxation questions based on an artificial intelligence question-answering system, It is characterized in that The S3 specifically includes: S31. Design and implement a document retrieval and indexing system that specializes in processing financial and tax-related regulations, policies, and case documents; S32, the tax document D = {d 1 ,d 2 ,…,d n }Preprocessing, including text cleaning, standardization and language analysis, where d i Represents a single document; S33, using information extraction technology from each preprocessed document d i Extract key information from the tuple to form information: I={(k 1 ,v 1 ),(k 2 ,v 2 ),…,(k m ,v m )}; where k j is the keyword, v j is the corresponding value; S34, establishing a document index based on the extracted information tuples, each document index consisting of a series of key-value pairs, so that any query for keyword k can quickly locate a document set containing the keyword; S35. Use inverted index technology to build index structure: IDX={k:{d i1 ,d i2 ,…,d ik }}; Where k is the index keyword, d ij is the jth document containing keyword k; S36, implements efficient query parsing and execution mechanism, supports complex query expressions, and uses edit distance algorithm to determine the similarity between query terms and words in documents, so that even if the query terms are slightly misspelled, relevant documents can be effectively retrieved: L(s,t)=min(L(s-1,t)+1,L(s,t-1)+1,L(s-1,t-1)+cost); Where L(s,t) is the edit distance between strings s and t, and cost is the cost of the replacement operation, which is usually 1 unless the two characters are the same, in which case it is 0; S37. Display the retrieved documents and their related information to users, provide document sorting and relevance feedback, and regularly update the indexing system so that the system can reflect the latest regulatory changes and policy updates.

5. According to claim 4, a method for intelligently answering financial and taxation questions based on an artificial intelligence question-answering system, It is characterized in that The S4 specifically includes: S41. Collect and organize a large amount of tax-related data, including previous Q&A sessions, user feedback, and tax law and regulation change records, to form a comprehensive data set D = {d 1 ,d 2 ,…,d n }, where d i Represents a single data instance; S42, preprocessing the collected data D; S43. Based on the preprocessed data set D, a machine learning model M is trained using a supervised learning method to solve the fiscal and taxation problem. Model training includes selecting appropriate features and algorithms. S44, defining a loss function L(θ) to optimize the model M, where θ represents a model parameter; Among them, y i is the data point d i The actual label, is the model M under parameter θ for data point d i The predicted label of S45. Optimize the parameters of the model and use the gradient descent algorithm to adjust the parameter θ to minimize the loss function L(θ): Among them, α is the learning rate, is the gradient of the loss function L(θ) with respect to the parameter θ; S46. Regularly evaluate the performance of model M during the training process. After completing the training of model M, integrate it into the financial and taxation question-and-answer system so that the model can dynamically provide accurate answers based on user queries in actual applications.

6. According to claim 5, a method for intelligently answering financial and taxation questions based on an artificial intelligence question-answering system, It is characterized in that The S5 specifically includes: S51. Use pattern recognition technology to identify common tax problem patterns. By analyzing the historical question-answering data set Q = {q 1 ,q 2 ,…,q m }, where q i Represents a historical question-and-answer instance; S52. Build a classification model C that uses the random forest algorithm to automatically classify and identify the types of fiscal and tax issues. The random forest consists of multiple decision trees, each of which is trained using a randomly selected feature subset. For each tree T in the random forest i , select m features for training, where m≤M, and the decision rule of each tree is based on maximizing the information gain IG: Where D is the dataset, f is the feature used to segment the dataset, v is the number of possible values ​​of feature f, and D j is the subset divided by the jth value of feature f, H(D) is the entropy of data set D, |D j | and |D| are the sizes of the subset and the total dataset, respectively; S53. Define the performance evaluation criteria of model C, and use precision P, recall R and F1 score F to evaluate model performance: Among them, TP, FP, and FN represent the number of true positives, false positives, and false negatives, respectively; S54. Conduct in-depth analysis of the identified problem patterns and use the isolation forest algorithm for anomaly detection to identify fiscal and tax issues that deviate from the norm. In the isolation forest, outliers are usually isolated earlier and therefore have a shorter average path length: Where x is an instance, n is the number of samples, E(h(x)) is the average path length of x in the tree, c(n) is the average unisolated path length in the random tree, and H(i) is the Harman number, which can be approximated as ln(i)+γ; S55. Develop a response generation module R that dynamically generates answers based on the identified question patterns and anomaly detection results, using natural language generation technology to form text responses that are highly natural and professional; S56. Continuously monitor system performance and adjust the model based on user feedback, using online learning mechanisms to continuously optimize the model C and response generation module R.

7. According to claim 6, a method for intelligently answering financial and taxation questions based on an artificial intelligence question-answering system, It is characterized in that The S6 specifically includes: S61. Support multiple rounds of dialogue with users to gradually clarify the details of user questions and specific needs, and use dialogue state management mechanism to track the progress and context of the dialogue; S62. In order to understand the user's natural language input, the system applies natural language processing technology to perform language analysis; S63. Based on the user's intent and the context of the question, the system dynamically constructs a query for retrieving relevant financial and tax data and information. The query construction includes converting the user's intent into a query language for a specific database and knowledge base; S64, implement a dialogue generation module G, which generates appropriate responses and further inquiries based on the current dialogue state and system available data, and guides the dialogue to develop in the direction of useful information; S65. Use deep learning models to optimize the dialogue generation process. Model training uses a sequence-to-sequence architecture: Among them, x represents the input sequence, y i represents the i-th word in the response generated by the model after a given input, and N is the total number of words in the response; S66. The system uses a real-time feedback mechanism to adjust the dialogue strategy, including user satisfaction evaluation and dialogue effect tracking. The dialogue system continuously learns and adapts to users' specific needs and preferences through user interactions, and applies online learning technology to dynamically update dialogue models and content.

Images