Tax hotline consultation agent method and device

Abstract

The invention relates to the technical field of artificial intelligence, and particularly provides a tax hotline consultation agent method and device, and the method comprises the following steps: S1, directly obtaining a session audio stream from a telephone end through a hardware telephone line universal demultiplexer RJ9, and carrying out the analysis; s2, converting the voice dialogue between the seat personnel and the taxpayer into a text form in real time, and performing voice transcription in real time; s3, performing deep analysis on the text content generated by real-time voice transcription; s4, constructing a tax knowledge graph; s5, accurate and detailed answers are automatically generated for the tax consultation questions of the clients; s6, extracting key information from the dialogue content, and performing structured processing for subsequent summarization and extraction; and S7, feeding back the generated dialogue summary and question answers to the seat personnel, and evaluating and modifying the summary content and the question answers by the seat personnel. Compared with the prior art, accurate and comprehensive information support can be provided for seat personnel, and the customer service quality and the working efficiency are improved.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and specifically provides a tax hotline consultation agent method and device. Background Art

[0002] Agents handle a large number of inquiries and questions from taxpayers every day. These inquiries and questions are diverse and cover a wide range of fields and topics. Agents need to spend a considerable amount of time and energy to categorize, summarize, and refine them in order to quickly and accurately provide solutions to customers. However, with the increase in business volume, traditional manual processing methods are inefficient, prone to errors and omissions, and require a high level of professional knowledge and experience from agents. Agents face tremendous workload pressure and struggle to accurately summarize and refine the large volume of conversations in a short period of time. This can lead to problems such as missing information and the inability to preserve high-quality consultation data.

[0003] To address these issues, some existing solutions use simple keyword extraction and template matching techniques. However, these methods often fail to accurately understand the semantics and context of conversations, resulting in inaccurate and incomplete summarization and distillation results. Therefore, a tool is needed that can automatically classify and summarize conversations between agents and customers, providing agents with accurate and comprehensive information support, improving service quality and work efficiency. Summary of the Invention

[0004] The present invention aims to address the deficiencies of the above-mentioned prior art and provides a highly practical tax hotline consultation seat method.

[0005] A further technical task of the present invention is to provide a tax hotline consultation seat device that is rationally designed, safe and applicable.

[0006] The technical solution adopted by the present invention to solve its technical problem is:

[0007] A tax hotline consultation agent method includes the following steps:

[0008] S1, directly obtain the conversation audio stream from the phone end through the hardware telephone line universal tap RJ9 and analyze it;

[0009] S2. The voice conversation between the agent and the taxpayer is converted into text in real time, and the voice is transcribed in real time;

[0010] S3. Conduct in-depth analysis of the text content generated by real-time speech transcription;

[0011] S4. Build a tax knowledge graph;

[0012] S5. Automatically generate accurate and detailed answers to clients’ tax consultation questions;

[0013] S6. Extract key information from the conversation and structure it for subsequent summary and refinement;

[0014] S7. Feedback the generated conversation summary and answers to the questions to the agent, who will evaluate and modify the summary content and answers to the questions.

[0015] Furthermore, in step S2, the convolutional neural network (CNN) technology is used to map the input audio signal to a corresponding phoneme sequence, and the N-gram model is used to predict the next word or phrase that is most likely to appear given the previous context.

[0016] Furthermore, in step S3, natural language processing (NLP) technology is used to conduct in-depth analysis of the text content generated by real-time speech transcription to capture taxpayer needs, emotional states, and key information points in the conversation, and provide comprehensive consulting advice in combination with an external tax knowledge base.

[0017] Furthermore, in step S4, entities, attributes and relationships are extracted from tax documents, laws and regulations, and case analysis to construct a tax knowledge graph. The knowledge graph is stored and managed using the graph database Neo4j, and knowledge mining and association recommendations are achieved through graph algorithms and logical reasoning techniques.

[0018] Furthermore, in step S5, the information provided by the conversational semantic understanding and the tax knowledge graph is combined to automatically generate accurate and detailed answers to the client’s tax consultation questions;

[0019] Based on intent recognition, knowledge graph query and natural language generation technology, the generated answers are ensured to comply with the latest tax regulations.

[0020] Furthermore, in step S6, the Qwen-7B model is used as the base model and trained on a large-scale corpus to acquire language comprehension capabilities. Then, the encoder-decoder architecture is combined with the attention mechanism to understand and generate summaries.

[0021] Throughout the entire process, the ROUGE evaluation metric is used to quantify the accuracy of the summary, ensuring that the generated content is both faithful to the original conversation and highly generalizable. Appropriate fine-tuning training is performed on the large model to enable the model to have good conversation understanding and summarization capabilities.

[0022] Furthermore, in step S7, the generated conversation summary and the answers to the questions are fed back to the agent, who then evaluates and modifies the summary and the answers to the questions.

[0023] At the same time, based on the feedback from the agents, the large model is optimized and improved to improve the accuracy and practicality of the summary and refinement.

[0024] A tax hotline consultation agent device, comprising: at least one memory and at least one processor;

[0025] The at least one memory is configured to store a machine-readable program;

[0026] The at least one processor is configured to call the machine-readable program to execute a tax hotline consultation agent method.

[0027] Compared with the prior art, the tax hotline consultation agent method and device of the present invention have the following outstanding beneficial effects:

[0028] This invention automatically categorizes and summarizes conversation content, providing agents with accurate and comprehensive information support, better meeting customer needs and improving taxpayer satisfaction. It also reduces the time and energy agents spend searching and organizing information, improving work efficiency and shortening taxpayer waiting times.

[0029] With the help of knowledge graphs and summary and refinement functions, agents can continuously learn and accumulate tax knowledge and experience, improving their professional skills. By analyzing and mining consulting data, they can identify the hot spots and difficult issues that taxpayers need, providing a basis for optimizing the tax consulting service process. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0031] Figure 1 It is a flowchart of a tax hotline consultation agent method. DETAILED DESCRIPTION

[0032] In order to enable those skilled in the art to better understand the solutions of the present invention, the present invention will be further described in detail below in conjunction with specific embodiments. Obviously, the embodiments described are only some embodiments of the present invention, rather than all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0033] A best embodiment is given below:

[0034] like Figure 1 As shown, a tax hotline consultation agent method in this embodiment has the following steps:

[0035] S1. The present invention does not need to be connected to any telephone system and is compatible with all types of telephones. It directly obtains the conversation audio stream from the telephone end through the hardware telephone line universal tap RJ9 and performs analysis.

[0036] S2. Convert the voice conversation between the agent and the taxpayer into text in real time. Using convolutional neural network (CNN) technology, the input audio signal is mapped to the corresponding phoneme sequence. The N-gram model is used to predict the next word or phrase that is most likely to appear given the previous context. This reduces ambiguity in the transcription process and improves the fluency and accuracy of the output text.

[0037] The module supports multiple languages ​​and dialects, and has powerful noise suppression capabilities and personalized customization options, ensuring low-latency, high-precision transcription effects in various complex environments, thereby providing reliable data support for subsequent business classification, information extraction and other links.

[0038] S3. Use natural language processing (NLP) technology, including deep learning methods based on pre-trained language models such as BERT and RoBERTa, to conduct in-depth analysis of the text content generated by real-time speech transcription.

[0039] Through core functions such as intent recognition, sentiment analysis, and named entity recognition, this module accurately captures taxpayer needs, emotional states, and key information points in conversations, and provides comprehensive consulting advice in combination with external tax knowledge bases. It supports multi-round conversation management and context awareness, ensuring that it can provide efficient and accurate support to agents in various complex conversation scenarios.

[0040] S4. Extract entities, attributes, and relationships from a large amount of tax documents, laws, regulations, and case studies to build a comprehensive and dynamic tax knowledge graph. Leveraging a graph database (Neo4j) to store and manage the knowledge graph, we support efficient keyword and semantic search, and leverage graph algorithms and logical reasoning techniques to achieve deep knowledge mining and associated recommendations.

[0041] S5. Combine conversational semantic understanding with information from the tax knowledge graph to automatically generate accurate and detailed answers to customers' tax consultation questions. By leveraging intent recognition, knowledge graph query, and natural language generation technologies, it supports multi-round conversation management, ensuring that generated answers comply with the latest tax regulations while remaining highly coherent and personalized, significantly improving agent efficiency and customer satisfaction.

[0042] S6. Based on the results of conversation understanding, key information is extracted from the conversation content, such as the customer's questions, needs, and opinions, and structured for subsequent summary and refinement. Leveraging the language generation capabilities of the large model, the extracted key information is classified, summarized, and refined to produce a concise and clear conversation summary.

[0043] Using the Qwen-7B as the base model, the model is trained on a large-scale corpus to acquire language comprehension capabilities. Then, an encoder-decoder architecture combined with an attention mechanism is used to understand and generate summaries. Throughout this process, the ROUGE evaluation metric is used to quantify the accuracy of the summaries, ensuring that the generated content is both faithful to the original conversation and highly generalizable. By fine-tuning the large model and combining the various techniques mentioned above, the model can achieve excellent conversation comprehension and summarization capabilities.

[0044] S7. The generated conversation summary and answers are fed back to the agent, who can evaluate and modify them. Simultaneously, the system optimizes and improves the model based on agent feedback to enhance the accuracy and practicality of the summary.

[0045] For example:

[0046] Collect a large amount of conversation data between tax hotline agents and taxpayers, including voice and text data. Preprocess the collected data, including voice-to-text conversion, text cleaning, word segmentation, and part-of-speech tagging, to facilitate subsequent model training and application.

[0047] High-quality hotline consultation data in the tax field is used as training data, and the preprocessed data is used to train the dialogue understanding model, information extraction model, and summary and refinement model.

[0048] Integrating the trained model into the intelligent agent assistant enables automatic understanding, information extraction, and summarization of conversations. The intelligent agent assistant can be integrated with the 12366 tax consultation hotline system to capture real-time conversations between agents and customers and present the summarized and refined results to the agent. Agents can evaluate and modify the summarized and refined results based on their actual needs, and the system will optimize and improve based on this feedback.

[0049] Assume that the agent has the following conversation with the taxpayer:

[0050] Agent: Hello, consultant No. 716 is here to serve you. Please go ahead.

[0051] Taxpayer: Hello, I have a question. When I filed my tax return, I got an invoice, but the amount on the invoice was wrong. The amount was written as 0, so the income tax was also 0. If I correct it later, do I still need to pay late payment fees?

[0052] Agent: Ma’am, if it involves paying additional taxes, there will be a late payment fee.

[0053] Taxpayer: So there is a late payment fee, right? If the correction involves paying back taxes, there will be a late payment fee.

[0054] Agent: Yes, if you have underpaid taxes before, there will be late payment fees.

[0055] Taxpayer: Okay, thank you.

[0056] Agent: You’re welcome. Thank you for calling. Please rate our service later. Goodbye.

[0057] After analyzing and summarizing the above conversation, the intelligent agent assistant generates the following summary:

[0058] The conversation primarily revolved around the taxpayer's inquiry into how to correct erroneous tax returns and the potential late payment fees that might be incurred during the correction process. The tax customer service representative provided a clear response.

[0059] **Correction**: The taxpayer mentioned that they made an error in their previous return and used an invoice with a leading 0, resulting in an income tax return of 0. This indicates that they need to make a correction to correct the incorrect data.

[0060] **Involving Tax Payment**: When taxpayers inquired about whether the correction would result in late payment fees, the tax customer service representative clearly stated that if the correction involved tax payment (i.e., additional tax payment required after the correction), then late payment fees would apply. "Tax payment" here means that the corrected declared amount is greater than the previously declared amount, and the difference in tax must be paid.

[0061] **Late Payment Fees**: Late payment fees are typically imposed on taxpayers for failing to pay their taxes within the prescribed timeframe. In this case, since a corrected return may result in an outstanding tax payment, the taxpayer may be required to pay the applicable late payment fee, as required by tax law.

[0062] To summarize, this conversation can be categorized as "tax declaration / declaration correction and cancellation / declaration correction and cancellation" because it covers the taxpayer's correction of erroneous declarations, the tax liability (tax payment) that may be caused by the correction, and its potential consequences (late payment fees).

[0063] Based on the above method, a tax hotline consultation agent device in this embodiment includes: at least one memory and at least one processor;

[0064] The at least one memory is configured to store a machine-readable program;

[0065] The at least one processor is configured to call the machine-readable program to execute a tax hotline consultation agent method.

[0066] The above-mentioned specific implementation methods are only specific cases of the present invention. The patent protection scope of the present invention includes but is not limited to the above-mentioned specific implementation methods. Any technical solutions that conform to the above-mentioned specific implementation methods of the present invention and any appropriate changes or substitutions made thereto by ordinary technicians in the relevant technical field shall fall within the patent protection scope of the present invention.

[0067] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A tax hotline consultation agent method, characterized in that: The steps are as follows: S1, directly obtain the conversation audio stream from the phone end through the hardware telephone line universal tap RJ9 and analyze it; S2. The voice conversation between the agent and the taxpayer is converted into text in real time, and the voice is transcribed in real time; S3. Conduct in-depth analysis of the text content generated by real-time speech transcription; S4. Build a tax knowledge graph; S5. Automatically generate accurate and detailed answers to clients’ tax consultation questions; S6. Extract key information from the conversation and structure it for subsequent summary and refinement; S7. Feedback the generated conversation summary and answers to the questions to the agent, who will evaluate and modify the summary content and answers to the questions.

2. A tax hotline consultation seat method according to claim 1, characterized in that: In step S2, the convolutional neural network (CNN) technology is used to map the input audio signal to the corresponding phoneme sequence, and the N-gram model is used to predict the next word or phrase that is most likely to appear given the previous context.

3. A tax hotline consultation seat method according to claim 2, characterized in that: In step S3, natural language processing (NLP) technology is used to conduct in-depth analysis of the text content generated by real-time speech transcription to capture taxpayer needs, emotional states, and key information points in the conversation, and provide comprehensive consulting advice in combination with an external tax knowledge base.

4. A tax hotline consultation seat method according to claim 3, characterized in that: In step S4, entities, attributes, and relationships are extracted from tax documents, laws and regulations, and case analysis to construct a tax knowledge graph. The knowledge graph is stored and managed using the graph database Neo4j, and knowledge mining and association recommendations are achieved through graph algorithms and logical reasoning techniques.

5. A tax hotline consultation agent method according to claim 4, characterized in that: In step S5, the system automatically generates accurate and detailed answers to customers’ tax consultation questions by combining conversational semantic understanding with information provided by the tax knowledge graph. Based on intent recognition, knowledge graph query and natural language generation technology, the generated answers are ensured to comply with the latest tax regulations.

6. A tax hotline consultation seat method according to claim 5, characterized in that: In step S6, the Qwen-7B model is used as the base model and trained on a large-scale corpus to acquire language comprehension capabilities. Then, the encoder-decoder architecture is combined with the attention mechanism to understand and generate summaries. Throughout the entire process, the ROUGE evaluation metric is used to quantify the accuracy of the summary, ensuring that the generated content is both faithful to the original conversation and highly generalizable. Appropriate fine-tuning training is performed on the large model to enable the model to have good conversation understanding and summarization capabilities.

7. A tax hotline consultation seat method according to claim 6, characterized in that: In step S7, the generated conversation summary and answers to the questions are fed back to the agent, who then evaluates and modifies the summary and answers to the questions. At the same time, based on the feedback from the agents, the large model is optimized and improved to improve the accuracy and practicality of the summary and refinement.

8. A tax hotline consultation seat device, characterized in that: include: at least one memory and at least one processor; The at least one memory is configured to store a machine-readable program; The at least one processor is configured to call the machine-readable program to execute the method according to any one of claims 1 to 7.

Images