Method and device for optimizing reply content of a dialogue robot, computer device

Abstract

The application discloses a method and device for optimizing the reply content of a dialogue robot, a computer device and a storage medium. Specifically, the method for optimizing the reply content of a dialogue robot comprises the following steps: analyzing a knowledge document to generate atomic paragraphs and table data; generating question and answer pairs from the atomic paragraphs and table data according to a hierarchical extraction strategy and known prompts; and inputting the question and answer pairs into a large language model of the dialogue robot to optimize the reply content of the dialogue robot. In this way, the large language model of the dialogue robot can be trained, and the large language model of the dialogue robot can provide more accurate, unified and correct answers when replying to questions.

Description

Technical Field

[0001] This application relates to the field of large language models, and more specifically, to a method, apparatus, computer device, and storage medium for optimizing the response content of a chatbot. Background Technology

[0002] In modern industrial and commercial operations, rapid and effective troubleshooting, maintenance guidance, technical support, and customer service play a crucial role. For users, timely problem identification and finding appropriate solutions are essential for improving work efficiency and productivity. These services are also vital for many businesses and organizations because they quickly and accurately meet customer needs, thereby increasing customer satisfaction and loyalty. Therefore, exploring an efficient, accurate, and cost-effective solution is imperative. Summary of the Invention

[0003] This summary section is provided to introduce some selected concepts in a simplified form, which will be further described in the detailed description section below. This summary section is not intended to identify any key or essential features of the claimed subject matter, nor is it intended to help determine the scope of the claimed subject matter.

[0004] Based on this, this application provides a method for optimizing the response content of a chatbot, comprising:

[0005] The knowledge document is parsed to generate atomic paragraphs and table data;

[0006] Based on a hierarchical extraction strategy and known hints, question-answer pairs are generated from the atomic paragraphs and table data;

[0007] The question-and-answer pairs are input into the large language model of the chatbot to optimize the chatbot's responses.

[0008] The above method can be used to train the large language model of the chatbot, enabling it to provide more accurate, consistent, and correct answers when responding to questions.

[0009] Furthermore, the step of generating question-answer pairs from the atomic paragraphs and table data based on a hierarchical extraction strategy and known hints includes:

[0010] Based on the data about sentences, paragraphs, documents, and tables, generate question-and-answer pairs from the atomic paragraph and table data.

[0011] The above methods allow for more refined and multi-dimensional analysis of atomic paragraphs and table data, generating question-and-answer pairs at different levels or across different knowledge spans.

[0012] Furthermore, based on a hierarchical extraction strategy and known hints, question-answer pairs are generated from the atomic paragraphs and table data, including:

[0013] Based on the known hints regarding yes / no questions, relationship questions, maximum / minimum questions, opinion questions, and causal questions, generate question-and-answer pairs from the atomic paragraphs and table data.

[0014] In this way, different types of questions can be generated, allowing the rich large language model to produce different types of question-answer pairs.

[0015] Furthermore, generating question-and-answer pairs from the atomic paragraphs and table data includes:

[0016] Generate question-and-answer pairs containing macroscopic or microscopic knowledge from the atomic paragraphs and table data;

[0017] The atomic paragraphs and table data are used to generate question-and-answer pairs containing explicit or tacit knowledge.

[0018] By using the methods described above, a large language model can be trained to generate direct or macroscopic answers to different questions, thus fully addressing various questioning scenarios.

[0019] Furthermore, it also includes generating question-answer pairs from the atomic paragraphs and table data based on multiple large language models.

[0020] By using multiple different large language models, various question-answer pairs can be generated, which can improve the diversity of question-answer pairs and make the large language model of the chatbot learn more richly.

[0021] Furthermore, the question-and-answer pairs are input into the large language model of the chatbot to optimize the chatbot's responses, including:

[0022] Based on the question-answer pairs, the large language model of the chatbot is subjected to supervised fine-tuning.

[0023] By using the methods described above, the answers generated by the large language model of the chatbot can be made more accurate.

[0024] This application also discloses an apparatus for optimizing the response content of a chatbot, comprising:

[0025] The document parsing module is used to parse knowledge documents and generate atomic paragraphs and table data;

[0026] The hierarchical strategy module is used to generate question-answer pairs from the atomic paragraphs and table data based on a hierarchical extraction strategy and known hints.

[0027] An input optimization module is used to input the question-answer pair into the large language model of the chatbot to optimize the chatbot's response content.

[0028] This application also provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the above-described method.

[0029] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the above-described method.

[0030] This application also provides a computer program product tangibly stored on a computer-readable medium and comprising computer-executable instructions that, when executed, cause at least one processor to perform the methods described above. Attached Figure Description

[0031] Implementations of this disclosure are illustrated in the accompanying drawings by way of example rather than limitation, and similar reference numerals in the drawings denote the same or similar parts.

[0032] Figure 1 This is a schematic diagram of a method for optimizing the response content of a chatbot according to one embodiment of this application.

[0033] Figure 2 This is a schematic diagram of an apparatus for optimizing the response content of a chatbot according to one embodiment of this application.

[0034] Figure 3 This is a schematic diagram of a computer device for optimizing the response content of a chatbot according to one embodiment of this application.

[0035] Figure 4 This is a schematic diagram of a multi-level knowledge extraction strategy according to one embodiment of this application.

[0036] The accompanying figure is labeled as follows:

[0037] Steps S101-S103

[0038] 200: Device

[0039] 201: Module

[0040] 202: Module

[0041] 203: Module

[0042] 300: Computer equipment

[0043] 302: Processor

[0044] 304: Memory

[0045] 401: Document

[0046] 402: Paragraph Collection

[0047] 403: Table Set

[0048] 404: Question and Answer Correct

[0049] 4041: Sentence Level

[0050] 4042: Table level

[0051] 4043: Paragraph level

[0052] 4044: Document Level

[0053] 405: Fine-tuning a large language model

[0054] 406: A Common Large Language Model Detailed Implementation

[0055] In the following description, numerous specific details are set forth for illustrative purposes. However, it will be understood that the invention can be implemented without these specific details. In other examples, well-known circuits, structures, and techniques have not been shown in detail so as not to affect the understanding of the description.

[0056] Large Language Models (LLMs) are powerful tools that learn from vast corpora, understand user queries, and generate corresponding answers. Enterprises typically possess large amounts of documentation, including user manuals, installation guides, product samples, standard documents, and maintenance records, providing valuable data for various applications. Therefore, this application proposes a multi-level question-answer pair extraction strategy to encompass comprehensive document knowledge. Through supervised fine-tuning, this knowledge is infused into the large language model, thereby enhancing its generative capabilities.

[0057] In existing technologies, to meet the needs of the aforementioned scenarios, in addition to traditional methods relying on manual search and document reading, two main approaches are employed. The first approach utilizes intelligent retrieval technology to process documents. When a user queries, the system quickly retrieves relevant paragraphs, shifting from manual to intelligent retrieval, thus improving efficiency and accuracy. The second approach combines Retrieval-Augmented Generation (RAG) with a large language model. The search results are used as reference documents and input along with the query statement into the large language model, relying on its capabilities to answer the question.

[0058] However, the first method has the drawback of only providing the retrieved paragraphs and not generating a final answer. Furthermore, the retrieved paragraphs may be incomplete. For example, in device malfunctions, symptoms often appear simultaneously. If only one symptom is searched, the retrieved text may not provide sufficient and appropriate support. The second method often suffers from problems such as excessive contextual input and long response times. Currently, lengthy contextual input is usually truncated, leading to lost information and inaccurate results.

[0059] Therefore, the multi-level question-answer pair extraction strategy proposed in this application, combined with LLM fine-tuning, can effectively alleviate or avoid the aforementioned drawbacks.

[0060] Specifically, this application discloses a method for optimizing the response content of a chatbot, comprising:

[0061] S101 parses the knowledge document to generate atomic paragraphs and table data.

[0062] Document Parsing: In the industrial sector, text data exists in various formats such as PDF, Word, Excel, and PowerPoint. Data, especially operation manuals, installation guides, and standard documents, forms the core of the work. To effectively utilize these documents, they must be converted to plain text. Furthermore, specific strategies should be employed to segment documents into logically coherent paragraphs, providing more accurate reference segments for subsequent LLM and RAG integration. Based on the above approach, a large number of atomic paragraphs and tabular data can be generated, where atomic paragraphs are represented by (P_set) and tabular data by (T_set).

[0063] S102, Based on the hierarchical extraction strategy and known hints, generate question-answer pairs from the atomic paragraphs and table data.

[0064] To capture the different categories of knowledge mentioned above, the hierarchical extraction strategy proposed in this application effectively covers all these types of knowledge. Specifically, based on the atomic paragraphs (P_set) and table data (T_set) obtained after document parsing, the question-answer pair extraction strategy includes sentence-level, paragraph-level, document-level, and table-level extraction.

[0065] Specifically, the hierarchical question-answer pair extraction strategy includes:

[0066] At the sentence level, the focus is on key information within the sentence, such as entities and events. Through keyword matching and semantic analysis, sentences that may contain the answer can be identified.

[0067] At the paragraph level, consider the paragraph's topic and contextual information to filter out paragraphs relevant to the question.

[0068] At the document level, analyze the overall structure and content of the document to identify the paragraphs or chapters most likely to contain the answers.

[0069] For the tabular level, relevant data items are extracted by analyzing and processing the tabular data.

[0070] Therefore, the step of generating question-answer pairs from the atomic paragraphs and table data based on a hierarchical extraction strategy and known hints includes:

[0071] Based on the data about sentences, paragraphs, documents, and tables, generate question-and-answer pairs from the atomic paragraph and table data.

[0072] The following is a specific example with a known hint:

[0073] Please generate five questions and their corresponding answers based on the provided "text to be extracted". The text types mainly include sentences, paragraphs, documents, and tables.

[0074] At the sentence level, the focus is on key information within sentences, such as entities and events. Keyword matching and semantic analysis are used to identify sentences that may contain the answer. At the paragraph level, the topic and context of paragraphs are considered to filter out those relevant to the question. At the document level, the overall structure and content of the text are further analyzed to determine the paragraphs or sections most likely to contain the answer. At the table level, relevant data items are extracted by analyzing and processing tabular data.

[0075] When designing questions, keep the following points in mind: Question diversity is crucial. Questions can include, but are not limited to, yes / no questions, relationship questions, extremum questions (e.g., maximum, best, highest), opinion questions (involving subjective evaluation), and causal questions (e.g., why is science and technology the primary productive force?). Furthermore, maintaining a balance in question difficulty is also essential. Some questions may be obvious, while others may require deduction, induction, reasoning, or summarization to answer.

[0076] ---------------------------------------------------------------------

[0077] Text type to extract: {sentence|paragraph|document|table}

[0078] Text to be extracted: {Sentence text|Paragraph text|Document text|Table text}.

[0079] ---------------------------------------------------------------------

[0080] Please continue answering.

[0081] Using the methods described above, and based on a hierarchical extraction strategy and known hints, question-answer pairs are generated from the atomic paragraphs and table data, including:

[0082] Based on the known hints regarding yes / no questions, relationship questions, maximum / minimum questions, opinion questions, and causal questions, generate question-and-answer pairs from the atomic paragraphs and table data.

[0083] Furthermore, to enable large-scale language models to learn the knowledge contained in documents, it is expected that different question-answer pairs (QA pairs) will be extracted based on the data obtained from the document parsing. However, since articles may contain different categories of knowledge, they are classified into macro-knowledge and micro-knowledge, as well as explicit knowledge and implicit knowledge.

[0084] Macro-level knowledge refers to information that requires reading multiple paragraphs or the entire text to answer related questions. Conversely, micro-level knowledge can be easily grasped from just a sentence or a single line of data in a table. Explicit knowledge is information that can be understood directly from its literal meaning. Tacit knowledge, on the other hand, requires a certain level of deductive, inductive, summarizing, and reasoning abilities to acquire.

[0085] Therefore, generating question-and-answer pairs from the atomic paragraphs and table data includes:

[0086] Generate question-and-answer pairs containing macroscopic or microscopic knowledge from the atomic paragraphs and table data;

[0087] The atomic paragraphs and table data are used to generate question-and-answer pairs containing explicit or tacit knowledge.

[0088] Example 1: Based on the given hints and the input sentence, the following question-answer pairs were extracted:

[0089] Tripping: When the closing electromagnet loses power, the compressed trip spring and contact pressure spring begin to act, driving the actuator to separate the contacts in the vacuum arc chamber. At the same time, the trip spring provides sufficient force to balance the self-closing force of the vacuum arc chamber, keeping the vacuum contactor in the tripped state.

[0090] Based on the aforementioned power outage, the potential question-answer pairs output by the large language model might be:

[0091] Question 1: What is a circuit breaker tripping?

[0092] A: Tripping refers to the operation of disconnecting a circuit in a power system. Specifically, when the closing electromagnet is de-energized, the compressed trip spring and contact pressure spring are released, which drives the actuator to separate the contacts in the vacuum arc chamber, thereby disconnecting the circuit.

[0093] Question 2: What are the main components involved in the tripping process?

[0094] A: The main components involved in the tripping process include the closing electromagnet, the tripping spring, the contact pressure spring, and the actuator. These components work together to ensure a safe and effective disconnection of the circuit.

[0095] Question 3: Why is the trip spring so crucial during operation?

[0096] A: The trip spring is crucial during operation because it provides the necessary force to balance the self-closing force of the vacuum arc chamber, ensuring the vacuum contactor remains in the tripped state. If the trip spring does not have sufficient force, the vacuum arc chamber may not be able to remain in the tripped state, thus affecting the safety and efficiency of the equipment.

[0097] The above Q&A effectively captures knowledge at both the explicit and micro levels for questions 1, 2, and 3.

[0098] Example 2, when the following paragraph is input into a large language model:

[0099] Introduction and Structure of Vacuum Contactors

[0100] Closing: When the closing electromagnet is energized, it generates electromagnetic force, which compresses the trip spring and contact pressure spring via the transmission rod. This force overcomes the electrostatic force generated during closing, causing the contacts of the vacuum arc chamber to close, and the vacuum contactor to be in the closed state. After the vacuum contactor closes, the current in the closing electromagnet decreases, and the generated electromagnetic force is only used to compress the trip spring and contact pressure spring to maintain the closed state, thereby reducing energy loss.

[0101] Tripping: When the closing electromagnet is de-energized, the compressed trip spring and contact pressure spring activate, driving the drive rod to separate the contacts of the vacuum interrupter. Simultaneously, the trip spring provides sufficient force to balance the self-closing force of the vacuum interrupter, ensuring the vacuum contactor remains in the tripped state.

[0102] Under different large-scale language models, based on the above descriptions of closure and tripping, the generated question-answer pairs can include:

[0103] Question 1 (Is this a question): During the closing process, is it necessary to keep the closing electromagnet continuously powered to keep the vacuum contactor in the closed state?

[0104] A: No, it's not necessary. After the closing electromagnet completes the closing operation, its current decreases, and the generated electromagnetic force is only used to compress the trip spring and contact pressure spring to maintain the closed state.

[0105] Question 2 (Extreme / Minimum Problem): During the tripping process, which component reacts first to the de-energization of the closing electromagnet?

[0106] A: During the tripping process, the compression tripping spring and the contact pressure spring first respond to the de-energization of the closing electromagnet.

[0107] Question 3 (Super Question): What force must a vacuum contactor overcome during the closing process?

[0108] A: The force that a vacuum contactor must overcome is the electrostatic charge generated during the closing process.

[0109] Question 4 (Opinion-based question): Why does the current in the closing electromagnet decrease after the vacuum contactor is closed?

[0110] A: When the vacuum contactor is closed, the current in the closing electromagnet will decrease to reduce energy loss, while ensuring that the generated electromagnetic force is sufficient to maintain the closed state.

[0111] Question 5 (Cause and Effect): Why does the trip spring play a key role in the tripping process of a vacuum contactor?

[0112] A: The trip spring plays a crucial role in the tripping process of a vacuum contactor because it drives the transmission rod to separate the contacts and provides sufficient force to balance the self-closing force of the vacuum arc chamber, ensuring that the vacuum contactor remains stably in the tripped state.

[0113] The above Q&A effectively captures macro-level knowledge and tacit knowledge for questions 1, 2, 3, 4, and 5.

[0114] In summary, based on multiple large language models, question-answer pairs are generated from the atomic paragraphs and table data.

[0115] Furthermore, table-level input is essentially similar to sentence-level input, while document-level input is similar to paragraph-level input. However, it's important to note that if the document-level input is too long, the contextual support of many large language models will be very limited. In such cases, it's necessary to summarize each part to some extent, connect them to form an abstract summary of the entire document, and then extract question-answer pairs as a paragraph.

[0116] By employing this four-level extraction strategy, both explicit and implicit knowledge in the text can be comprehensively covered, effectively extracting both macroscopic and microscopic information. This hierarchical extraction method not only improves the accuracy and efficiency of the question-answering system but also provides users with richer and more precise answers.

[0117] Furthermore, different large language models and prompts can be used to extract question-answer pairs for different sentences, paragraphs, documents, and tables. This approach ensures diversity in extraction and allows the complementary capabilities of multiple models to cover a more comprehensive range of knowledge.

[0118] S103, input the question-answer pair into the large language model of the chatbot to optimize the chatbot's response content.

[0119] Specifically, the question-answer pairs are input into the large language model of the chatbot, and the large language model is optimized. This is achieved through supervised fine-tuning of large language models, where the model is fine-tuned in a supervised manner based on the extracted question-answer pair data. Furthermore, to ensure data security, open-source models can be selected to inject knowledge into the large language model while protecting the confidentiality of domain-specific data. Finally, when a user asks a question, the large language model can directly provide the final answer, thereby facilitating end-to-end question answering from query to response. This method saves users time and improves the quality and consistency of responses.

[0120] In summary, inputting the question-answer pairs into the large language model of the chatbot to optimize the chatbot's responses includes: performing supervised fine-tuning of the chatbot's large language model based on the question-answer pairs.

[0121] like Figure 4 As described above, 401 represents a document, which is parsed and split into a set of atomic paragraphs 402 and a set of data tables 403. These two sets are then processed by ChatGPT or other large language models to obtain question-answer pairs 404, where 4041 represents the sentence level, 4042 the table level, 4043 the paragraph level, and 4044 the document level. These are then input into a common large language model 406 for further fine-tuning, resulting in a larger language model 405.

[0122] The technical effects of this application are as follows: To ensure completeness, the method described above includes parts of domain-specific text document parsing, question-expansion techniques, and fine-tuning of large language models. It comprehensively proposes a question-answering chatbot construction scheme based on a large language model suitable for vertical domains, and introduces a concept and method of multi-level knowledge extraction, which can cover different categories of article knowledge.

[0123] To determine whether the method used in this application has been employed, the following steps can be taken: First, check for keywords such as large language model, question-answer pair extraction, multi-level, and hierarchical strategy. Then, compare the similarities and differences between the solutions in detail to confirm whether the method has been used.

[0124] Throughout the specification, references to "an implementation," "implementation," "exemplary implementation," "some implementations," "various implementations," etc., indicate that the implementation of the invention described may include specific features, structures, or characteristics. However, it is not implied that every implementation must include these specific features, structures, or characteristics. Furthermore, some implementations may have some, all, or none of the features described for other implementations.

[0125] It should be understood that, although Figure 1 The steps in the flowchart are shown sequentially as indicated by the arrows, but these steps are not necessarily executed in the order indicated by the arrows. Unless otherwise specified herein, there is no strict order in which these steps are executed, and they can be performed in other orders. Figure 1At least some of the steps in the process may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but may be executed at different times. The execution order of these steps or stages is not necessarily sequential, but may be executed in turn or alternately with other steps or at least some of the steps or stages in other steps.

[0126] Figure 2 An optimization device 200 for the response content of a chatbot is provided, comprising:

[0127] Document parsing module 201 is used to parse knowledge documents and generate atomic paragraphs and table data;

[0128] The hierarchical strategy module 202 is used to generate question-answer pairs from the atomic paragraphs and table data according to the hierarchical extraction strategy and known hints;

[0129] The input optimization module 203 is used to input the question-answer pair into the large language model of the chatbot to optimize the chatbot's response content.

[0130] It should be noted that the device can contain more or fewer modules to achieve the described functions. For example, Figure 2 At least one module can be further divided into a plurality of distinct sub-modules, each sub-module being used to perform at least a portion of the operations described herein in conjunction with the corresponding module. Furthermore, in some examples, device 200 may also include additional modules for performing other operations already described in the specification. Moreover, those skilled in the art will understand that exemplary device 200 can be implemented using software, hardware, firmware, or any combination thereof.

[0131] Figure 3 A computer device is provided. According to one embodiment, the computer device 300 may include a processor 302 that executes a computer program stored in a memory 304. When executed by the processor, the computer program implements the method described above.

[0132] Those skilled in the art will understand that Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. A specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0133] Those skilled in the art will understand that all or part of the processes in the methods described above can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, or optical storage, etc. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM can be in various forms, such as static random access memory (SRAM) or dynamic random access memory (DRAM), etc.

[0134] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, performs the above steps.

[0135] This application also provides a computer program product tangibly stored on a computer-readable medium and including computer-executable instructions that, when executed, cause at least one processor to perform the methods described above.

[0136] Furthermore, the computer program can be stored and run in the cloud to execute the method. Furthermore, the components of the program can be deployed on multiple devices or in the cloud; for example, corresponding steps can be deployed and run on a local computer, or run on different cloud devices, transmitting signals via communication connections, or they can also be deployed and run on a local computer. This application does not limit the described approach or method; corresponding technologies can be flexibly deployed and fully utilized to execute and complete the method using cloud computing, big data, supercomputing capabilities, and other equipment and technologies.

[0137] Some implementations of this disclosure may include an article of writing. The article of writing may include a storage medium for storing logic. Examples of storage media may include one or more types of computer-readable storage media capable of storing electronic data, including volatile or non-volatile memory, removable or non-removable memory, erasable or non-erasable memory, writable or rewritable memory, and so on. Examples of logic may include various software units, such as software components, programs, applications, computer programs, application programs, system programs, machine programs, operating system software, middleware, firmware, software modules, routines, subroutines, functions, methods, procedures, software interfaces, application programming interfaces (APIs), instruction sets, computational code, computer code, code segments, computer code segments, words, values, symbols, or any combination thereof. In some implementations, for example, the article of writing may store executable computer program instructions that, when executed by a processor, cause the processor to perform the methods and / or operations described herein. Executable computer program instructions may include any suitable type of code, such as source code, compiled code, interpreted code, executable code, static code, dynamic code, and so on. Executable computer program instructions can be implemented according to a predefined computer language, method, or syntax used to command the computer to perform specific functions. These instructions can be implemented using any suitable high-level, low-level, object-oriented, visual, compiled, and / or interpreted programming language.

[0138] The examples described above include those of the disclosed architecture. It is certainly impossible to describe every conceivable combination of components and / or methods, but those skilled in the art will understand that many other combinations and arrangements are also possible. Therefore, this novel architecture is intended to cover all such alternatives, modifications, and variations that fall within the spirit and scope of the appended claims.

Claims

1. A method for optimizing the response content of a chatbot, wherein, include: The knowledge document is parsed to generate atomic paragraphs and table data; Based on the hierarchical extraction strategy and known hints, question-answer pairs are generated from the atomic paragraphs and table data; The question-and-answer pairs are input into the large language model of the chatbot to optimize the chatbot's responses.

2. The method according to claim 1, wherein, The hierarchical extraction strategy and known hints generate question-answer pairs from the atomic paragraphs and table data, including: Based on the hierarchical extraction strategy for sentences, paragraphs, documents, and tables, question-answer pairs are generated from the atomic paragraph and table data.

3. The method according to claim 1, wherein, Based on a hierarchical extraction strategy and known hints, question-answer pairs are generated from the atomic paragraphs and table data, including: Based on the known hints regarding yes / no questions, relationship questions, maximum / minimum questions, opinion questions, and causal questions, generate question-and-answer pairs from the atomic paragraphs and table data.

4. The method according to claim 1, wherein, Generate question-and-answer pairs from the atomic paragraphs and table data, including: Generate question-and-answer pairs containing macroscopic or microscopic knowledge from the atomic paragraphs and table data; The atomic paragraphs and table data are used to generate question-and-answer pairs containing explicit or tacit knowledge.

5. The method according to any one of claims 1 to 4, further comprising: Based on multiple large language models, the atomic paragraphs and table data are used to generate various question-answer pairs.

6. The method according to claim 1, wherein, The question-and-answer pairs are input into the large language model of the chatbot to optimize the chatbot's responses, including: Based on the question-answer pairs, the large language model of the chatbot is subjected to supervised fine-tuning.

7. An apparatus for optimizing the response content of a chatbot, wherein, include: The document parsing module is used to parse knowledge documents and generate atomic paragraphs and table data; The hierarchical strategy module is used to generate question-answer pairs from the atomic paragraphs and table data based on a hierarchical extraction strategy and known hints. An input optimization module is used to input the question-answer pair into the large language model of the chatbot to optimize the chatbot's response content.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein... When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, wherein, When the computer program is executed by a processor, it implements the steps of the method according to any one of claims 1 to 6.

10. A computer program product tangibly stored on a computer-readable medium and comprising computer-executable instructions that, when executed, cause at least one processor to perform the method according to any one of claims 1 to 6.

Images