Information processing method and device, electronic equipment and storage medium

Abstract

The invention relates to the field of artificial intelligence, in particular to the technical field of natural language processing, provides an information processing method and device, electronic equipment and a storage medium, and aims to improve the accuracy of a large language model question and answer result. The method comprises the steps of receiving a target question input by an object; inputting the target question into a trained classification model, and obtaining a question intention to which the target question belongs; determining a processing flow corresponding to the target problem according to the problem intention; wherein if the problem intention is the first type of intention, the processing flow is that after intermediate processing is carried out on the target problem, an intermediate processing result is input into the finance and taxation large model; if the problem intention is a second type intention, the processing flow is to directly input the target problem into the finance and taxation large model; and taking an output result of the large finance and taxation model as an answer to the target question. The target questions are classified through the classification model, so that question intention recognition can be simply and efficiently realized, and the accuracy of a big language model question and answer result is improved.

Description

Background Art

[0002] As a strategic technology for the new round of technological and industrial revolution, the big language model has triggered changes and reshapes in various fields, giving rise to various vertical industry big models. Among them, the question-answering scenario is one of the important application scenarios of the big language model. For example, in the field of finance and taxation, the big language model can be applied to intelligent question-answering systems, such as intelligent customer service systems, to provide more convenient and effective finance and taxation consulting services to the target.

[0003] In related technologies, the conversation of the object (such as the questions to be answered) is generally directly input into a large language model, and the model outputs the answer. However, due to the complexity and diversity of the subdivision scenarios in the field of finance and taxation, inputting the object questions into the large language model indiscriminately will cause problems such as generating hallucinations, high reasoning costs, and difficulty in optimization.

[0004] Therefore, how to improve the accuracy of question and answer results of large language models in question and answer scenarios is an urgent problem to be solved. Summary of the invention

[0005] Embodiments of the present application provide an information processing method, device, electronic device, and storage medium for improving the accuracy of question and answer results of a large language model.

[0006] An information processing method provided in an embodiment of the present application includes:

[0007] Receive target questions input by subjects;

[0008] Input the target question into a trained classification model to obtain the question intent to which the target question belongs; the classification model is obtained after fine-tuning based on the labeled financial and taxation question and answer data;

[0009] According to the problem intent, determine the processing flow corresponding to the target problem; wherein, if the problem intent is the first type of intent, the processing flow is to perform intermediate processing on the target problem and then input the intermediate processing result into the finance and taxation model; if the problem intent is the second type of intent, the processing flow is to directly input the target problem into the finance and taxation model;

[0010] The output result of the financial and taxation model is used as the answer to the target question.

[0011] An information processing device provided in an embodiment of the present application includes:

[0012] A receiving unit, used for receiving a target question input by a subject;

[0013] An acquisition unit, used for inputting the target question into a trained classification model to acquire the question intent to which the target question belongs; the classification model is obtained after fine-tuning based on the labeled financial and taxation question and answer data;

[0014] A determination unit is used to determine a processing flow corresponding to the target problem according to the problem intention; wherein, if the problem intention is the first type of intention, the processing flow is to perform intermediate processing on the target problem and then input the intermediate processing result into the finance and taxation model; if the problem intention is the second type of intention, the processing flow is to directly input the target problem into the finance and taxation model;

[0015] An output unit is used to use the output result of the financial and tax model as the answer to the target question.

[0016] Optionally, the first type of intent includes a first sub-intent and a second sub-intent, and the determining unit is specifically configured to:

[0017] If the question intent is the first sub-intent, after rewriting the target question and searching the knowledge base, the rewritten question and search results are input into the trained tax model;

[0018] If the question intent is the second sub-intent, the target question is input into the trained computational agent, and the output result of the computational agent is input into the financial and taxation model.

[0019] Optionally, the determining unit is specifically configured to:

[0020] Rewrite the target question, and input the rewritten question into a knowledge base for retrieval to obtain knowledge fragments;

[0021] The rewritten questions, knowledge fragments and preset first task instructions are spliced ​​together and input into the trained finance and taxation model.

[0022] Optionally, if the question intent is the first sub-intent, the determining unit is specifically configured to:

[0023] According to the preset mapping relationship for the time statement, the fuzzy time text in the target question is converted into a specific time text through regular matching; the specific time text is used to limit the relatively accurate time involved in the target question;

[0024] According to the pre-built unique dictionary, the industry customary text in the target question is converted into industry standard text.

[0025] Optionally, if the target question is in a multi-round dialogue, the determining unit is further configured to:

[0026] If the number of rounds of the multi-round dialogue does not exceed a preset value, rewriting the target question according to dialogue information of a preceding round in the multi-round dialogue;

[0027] If the number of rounds of the multiple conversations exceeds the preset value, a preceding conversation of the specified round is selected, and based on the relevance of other conversations in the multiple conversations except the target question and the preceding conversation to the target question, a preset number of other conversations are selected, and the target question is rewritten based on the other conversations and the preceding conversation.

[0028] Optionally, the determining unit is specifically configured to:

[0029] Inputting the target problem into the computational agent;

[0030] Based on the computing agent, decompose the target problem into complex tasks to obtain at least one sub-problem and a decomposition process;

[0031] Based on the computing agent, determine a processing solution corresponding to each of the sub-problems, and execute the corresponding processing solution on each of the sub-problems to obtain a corresponding processing result; the sub-problems correspond to the processing solutions one by one;

[0032] At least one of the processing solutions, at least one of the processing results, the disassembly process and the second task instruction are input into the financial and tax model.

[0033] Optionally, the classification model is a general text classification model; and the financial and taxation question and answer data is mixed data of various question intentions.

[0034] An electronic device provided in an embodiment of the present application includes a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of any one of the above-mentioned information processing methods.

[0035] An embodiment of the present application provides a computer-readable storage medium, which includes a computer program. When the computer program is run on an electronic device, the computer program is used to enable the electronic device to execute the steps of any one of the above-mentioned information processing methods.

[0036] An embodiment of the present application provides a computer program product, which includes a computer program, and the computer program is stored in a computer-readable storage medium; when a processor of an electronic device reads the computer program from the computer-readable storage medium, the processor executes the computer program, so that the electronic device performs the steps of any one of the above-mentioned information processing methods.

[0037] The beneficial effects of this application are as follows:

[0038] The embodiments of the present application provide an information processing method, device, electronic device and storage medium. Since the present application first classifies the target problem through a classification model, it realizes the classification of target problems with different problem intentions through one model. While ensuring the classification effect, it avoids the problem of building multiple classification models for different classification tasks in traditional technical solutions, training the models separately and not sharing data and knowledge. It effectively solves the problems of little professional data in the field, high difficulty in obtaining data, and high threshold for annotation, and can realize problem intention identification simply and efficiently. Then, according to the determined problem intention, after implementing the corresponding processing flow for the target problem, the processing result is input into the large fiscal and taxation model, which solves the problem that the large fiscal and taxation model is prone to calculation errors and improves the reasoning ability and interpretability of the large model on complex tasks.

[0039] Other features and advantages of the present application will be described in the following description, and partly become apparent from the description, or be understood by practicing the present application. The purpose and other advantages of the present application can be realized and obtained by the structures specifically pointed out in the written description, claims, and drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0040] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0041] Figure 1 A schematic diagram of an application scenario of an information processing method provided in an embodiment of the present application;

[0042] Figure 2 An overall flow chart of an information processing method provided in an embodiment of the present application;

[0043] Figure 3 A schematic diagram of an information processing method provided in an embodiment of the present application;

[0044] Figure 4 A schematic diagram of the structure of an information processing device provided in an embodiment of the present application;

[0045] Figure 5 A schematic diagram of a hardware structure of an electronic device provided in an embodiment of the present application;

[0046] Figure 6 A schematic diagram of the hardware structure of another electronic device provided for application of an embodiment of the present application. DETAILED DESCRIPTION

[0047] In order to make the purpose, technical solution and advantages of the embodiments of the present application clearer, the technical solution of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are part of the embodiments of the technical solution of the present application, rather than all of the embodiments. Based on the embodiments recorded in the application documents, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the technical solution of the present application.

[0048] The following is an introduction to some concepts involved in the embodiments of the present application.

[0049] The first type of intent refers to the type of target problem that requires intermediate processing. This type of intent includes the first sub-intention and the second sub-intention. Among them, the target problem corresponding to the first sub-intention is an intensive problem that requires continuous knowledge updating. If it is directly answered by the financial and taxation big model, it is easy to lack domain knowledge, inaccurate information, and generate false content. The first sub-intention can be policy consultation, digital invoice Q&A, and exam Q&A, etc. The target problem corresponding to the second sub-intention can be a financial and taxation calculation problem, which is a multi-step complex problem, such as financial and taxation calculation, etc.

[0050] The second type of intention refers to the type of target problem that is directly input into the financial and taxation model after task classification without intermediate processing, such as accounting entries, etc.

[0051] Classification model: In the embodiment of the present application, it refers to the Universal Text Classification (UTC) model, which has excellent small sample learning ability, can complete task adaptation in the financial and taxation fields based on a small number of labeled samples, and greatly improve the classification effect.

[0052] Fuzzy time text: refers to fuzzy expressions that can refer to time information. For example, in daily communication, fuzzy expressions such as "soon after", "soon", and "a few days ago" are often used to describe time. These expressions are not precise, but convey a general concept of time.

[0053] Specific time text: refers to the pre-set text corresponding to the fuzzy time text, which can express relatively accurate time. For example, if the fuzzy time text is "these two days", the corresponding specific time text can be pre-set as "this week". Then, the current absolute time can be obtained through the specific time text. For example, if the conversation takes place on October 30, 2024, after the fuzzy time text "these two days" in the target question is converted into the specific time text "this week", it can be determined that the absolute time involved in this conversation is from October 28, 2024 to November 3, 2024.

[0054] Computational agent: It can decompose complex problems, decompose complex tasks into multi-step sub-problems, and determine and execute the solution for each sub-problem. In the embodiment of the present application, the computational agent can be implemented by some functions in the fiscal and taxation model, or it can be a model independent of the fiscal and taxation model.

[0055] The preferred embodiments of the present application are described below in conjunction with the drawings in the specification. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application. In addition, the embodiments and features in the embodiments of the present application may be combined with each other if there is no conflict.

[0056] like Figure 1 As shown, it is a schematic diagram of an application scenario of an embodiment of the present application. The application scenario diagram includes two terminal devices 110 and a server 120.

[0057] It should be noted that the information processing method in each embodiment of the present application can be executed by an electronic device, which can be a terminal device 110 or a server 120, that is, the method can be executed by the terminal device 110 or the server 120 alone, or by the terminal device 110 and the server 120 together. For example, when the terminal device 110 and the server 120 are executed together, the terminal device 110 receives the target question input by the object and sends it to the server 120, and the server 120 inputs the target question into the trained classification to obtain the problem intent to which the target question belongs; the server 120 determines the processing flow corresponding to the target question according to the problem intent; wherein, if the server 120 determines that the problem intent to which the target question belongs is the first type of intent, the server 120 performs intermediate processing on the target question and inputs the intermediate processing result into the finance and taxation model; if the server 120 determines that the problem intent to which the target question belongs is the second type of intent, the server 120 directly inputs the target question into the finance and taxation model; the server 120 sends the output of the finance and taxation model to the terminal device 110, so that the object can view the answer to the target question on the terminal device 110.

[0058] In an optional implementation, the terminal device 110 and the server 120 may communicate with each other via a communication network.

[0059] In an optional implementation, the communication network is a wired network or a wireless network.

[0060] It should be noted that Figure 1 The figure is only an example. In fact, the number of terminal devices and servers is not limited and is not specifically limited in the embodiments of the present application.

[0061] The following describes the information processing method provided by the exemplary embodiment of the present application in combination with the application scenarios described above and with reference to the accompanying drawings. It should be noted that the above application scenarios are only shown to facilitate understanding of the spirit and principles of the present application, and the implementation methods of the present application are not limited in this regard.

[0062] See also Figure 2 As shown, it is an implementation flow chart of an information processing method provided in an embodiment of the present application. The specific implementation process of the method is as follows S201-S204:

[0063] S201: receiving a target question input by a subject;

[0064] The information processing method provided in the embodiment of the present application can be applied to question-and-answer scenarios implemented through large models. The information processing method will be introduced later using the finance and taxation field as an example.

[0065] S202: Input the target question into the trained classification model to obtain the question intent to which the target question belongs.

[0066] Among them, the classification model is obtained after fine-tuning based on the labeled financial and taxation question and answer data. The financial and taxation question and answer data is a mixed data of various question intentions, and the financial and taxation question and answer data has a small sample size.

[0067] The embodiment of the present application takes into account the timeliness, regionality and professionalism of fiscal and taxation policies in the fiscal and taxation field, which leads to extremely high data annotation thresholds in the fiscal and taxation field, strong reliance on business personnel with rich fiscal and taxation knowledge and business experience, and slow data accumulation. The classification model in the embodiment of the present application adopts UTC technology. UTC has excellent small sample learning capabilities and can complete task adaptation in the fiscal and taxation field based on a small amount of labeled fiscal and taxation question and answer data, greatly improving the classification effect. In addition, UTC can uniformly model classification tasks with different question intentions, without the need to separately model and train fiscal and taxation question and answer data with different question intentions.

[0068] Specifically, when training the classification model, we take the problem intent including policy consultation, digital invoice Q&A, taxation examination Q&A, accounting entries and taxation calculation as examples. We annotate a small amount of data for the taxation Q&A data of each type of problem intent. Then, we convert each type of data into data that conforms to the UTC input format and mix them. The candidate labels of the data are [policy consultation, digital invoice Q&A, examination Q&A, accounting entries, taxation calculation, and others]. After dividing the training set, validation set and test set in a ratio of 8:1:1, we fine-tune the UTC model. We continuously adjust the data by observing the training effect on the validation set, and try to ensure that the targets under each category are as rich and diverse as possible, and the boundaries between target problems of different categories are clear, until the classification effect requirements are met and the model is saved for inference deployment.

[0069] In the above implementation, UTC universal text classification technology is used to uniformly model classification tasks with different labels. Based on a small number of labeled samples, multi-label classification of question intent is achieved while ensuring the classification effect. This avoids the problem of building multiple classification models for different classification tasks in traditional technical solutions, where the models are trained separately and data and knowledge are not shared. This effectively solves the problems of little domain-specific data, difficulty in obtaining data, and high annotation thresholds, and can simply and efficiently realize question intent identification.

[0070] After determining the problem intent of the target problem through the above method, the target problem can be processed through the following steps:

[0071] S203: Determine the processing flow corresponding to the target problem according to the problem intention;

[0072] Among them, if the intention of the problem is the first type of intention, the processing flow is to perform intermediate processing on the target problem and then input the intermediate processing results into the financial and taxation model; if the intention of the problem is the second type of intention, the processing flow is to directly input the target problem into the financial and taxation model.

[0073] Specifically, the first type of intent refers to the type of target problem that requires intermediate processing. This type of intent includes the first sub-intention and the second sub-intention. Among them, the target problem corresponding to the first sub-intention is an intensive problem that requires continuous knowledge updating. If it is directly answered by the financial and taxation big model, it is easy to lack domain knowledge, inaccurate information, and generate false content. The first sub-intention can be policy consultation, digital invoice Q&A, and exam Q&A, etc. The target problem corresponding to the second sub-intention can be a financial and taxation calculation problem, which is a multi-step complex problem, such as financial and taxation calculation, etc.

[0074] The second type of intention refers to the type of target problem that is directly input into the financial and taxation model after task classification without intermediate processing, such as accounting entries, etc.

[0075] 1. The following is an introduction to the intermediate processing of the target problem of the first type of intention:

[0076] If the question intent is the first sub-intent, the target question is rewritten and the knowledge base is searched, and the rewritten question and search results are input into the trained finance and taxation model.

[0077] In the embodiment of the present application, when the target question of the first type of intention is intermediately processed, an optional implementation method is as follows:

[0078] Rewrite the target question and input the rewritten question into the knowledge base for retrieval to obtain knowledge fragments; splice the rewritten question, knowledge fragments and the preset first task instructions, and input them into the trained finance and taxation model.

[0079] Among them, the knowledge base is a knowledge base in the field of finance and taxation, which contains sub-knowledge bases in multiple sub-fields under the field of finance and taxation. If the first sub-intention includes intentions such as policy consultation, digital invoice Q&A, and exam Q&A in the field of finance and taxation, then the policy sub-knowledge base, digital invoice sub-knowledge base, and exam sub-knowledge base in the knowledge base are searched to obtain the corresponding knowledge fragments. The first task instruction refers to the instruction that instructs the finance and taxation big model to generate an answer based on the input knowledge fragment and the target question after the question is rewritten.

[0080] Specifically, the target question corresponding to the first sub-intent is an intensive question that requires continuous knowledge updating. If it is answered directly by the finance and taxation model, it is easy to lack domain knowledge, inaccurate information, and generate false content. Therefore, taking the policy consultation, digital invoice Q&A, and exam Q&A in the finance and taxation field as an example, the embodiment of the present application introduces additional information sources such as external knowledge bases through the retrieval enhancement generation (RAG) technology to provide an effective mitigation strategy for these problems. According to the target question, relevant knowledge fragments are retrieved in the knowledge base, and then these knowledge fragments are combined with the first task instruction and handed over to the finance and taxation model to generate the final answer. However, since the target question often involves relative time or fuzzy time descriptions, such as "recently", "last month", and industry idioms, such as general enterprises (especially general taxpayers), small and micro (especially small and micro-profit enterprises), sometimes it involves multiple rounds of dialogue. If the latest question of the object is directly used, it will affect the accuracy and recall of the retrieval. Therefore, rewriting the question before retrieval is a necessary pre-step to improve the retrieval effect.

[0081] Therefore, the embodiment of the present application combines the characteristics of the target problems in the finance and taxation field such as the first sub-intention, and designs a combination of technologies such as time rewriting, industry jargon replacement, and semantic rewriting based on historical conversations in multi-round conversation scenarios to complete the rewriting of the target problems.

[0082] The following is an introduction to the steps for rewriting the question for the first sub-intention:

[0083] (1) Time rewriting.

[0084] In the embodiment of the present application, when the target question of the first sub-intention is rewritten in time, an optional implementation method is as follows:

[0085] According to the preset mapping relationship for time statements, the fuzzy time text in the target question is converted into specific time text through regular matching.

[0086] Among them, fuzzy time text refers to fuzzy expressions that can refer to time information. For example, in daily communication, fuzzy expressions such as "soon after", "soon", and "a few days ago" are often used to describe time. These expressions are not precise, but convey a general concept of time.

[0087] The specific time text is used to limit the relatively precise time involved in the target question. It refers to the pre-set text corresponding to the fuzzy time text that can express relatively precise time. For example, if the fuzzy time text is "these two days", the corresponding specific time text can be pre-set as "this week". Then, the current absolute time can be obtained through the specific time text. For example, if the conversation takes place on October 30, 2024, after the fuzzy time text "these two days" in the target question is converted to the specific time text "this week", it can be determined that the absolute time involved in this conversation is from October 28, 2024 to November 3, 2024.

[0088] The embodiment of the present application can customize the mapping relationship between fuzzy time and specific time, identify the fuzzy time text in the target question through regular matching, convert it into specific time text according to the mapping relationship, calculate the time according to the current system time, and replace the fuzzy time in the question with the obtained absolute time, thereby completing the time rewriting.

[0089] For example, the mapping relationship between the fuzzy time regular expression and the specific time can be as follows:

[0090] The fuzzy time "this year, latest, recent, current, now, this year" can be mapped to the specific time "this year"; the fuzzy time "this week, this week, this week, these two days" can be mapped to the specific time "this week"; the fuzzy time "last week, last week, these few days, recent days, a few days ago" can be mapped to the specific time "last week"; the fuzzy time "this month, this month, this month, this month, this period of time" can be mapped to the specific time "this month"; the fuzzy time "today, today, today, today" can be mapped to the specific time "today"; the fuzzy time "last year, the year before last, last year" can be mapped to the specific time "last year"; the fuzzy time "last month, last month|, last month" can be mapped to the specific time "last month", and so on.

[0091] For example, if the target question is "When did the tax collection end last month?", "last month" is identified through regular matching and mapped to "last month" according to the above mapping relationship. Assuming that the current question time is "July 23, 2024", the current month is obtained as "July 2024", and the previous month is calculated as "June 2024", therefore, the target question is changed to "When does the tax collection end in June 2024?". While the precise time expression avoids retrieving outdated information, it also alleviates the uncertainty of retrieval using fuzzy time.

[0092] (2) Replacement of industry customary text.

[0093] In the embodiment of the present application, when the target question of the first sub-intent is replaced with the industry-customized text, an optional implementation method is as follows:

[0094] According to the pre-built unique dictionary, the industry idiomatic text in the target question is converted into industry standard text.

[0095] Among them, taking the field of finance and taxation as an example, the industry customary text is the abbreviation of proper nouns, idioms and other "jargon" that has formed a consensus in the industry in the field of finance and taxation, and the industry standard text is the professional and authoritative expression in the field of finance and taxation. The embodiment of this application collects the industry customary text and the corresponding industry standard text to form a special dictionary, and constructs a mapping relationship between the industry jargon such as the abbreviation of proper nouns, idioms and professional and authoritative expressions in the field of finance and taxation. For example, "small scale" is an industry customary text, and its mapped industry standard text is "small-scale taxpayer", then its mapping relationship can be recorded in the special dictionary as: {'small scale':'small-scale taxpayer'}; "general invoice" is an industry customary text, and its mapped industry standard text is "value-added tax general invoice", then its mapping relationship can be recorded in the special dictionary as: {'general invoice':'value-added tax general invoice'}, etc. When the question intent is determined to be the first sub-intent, the industry jargon in the target question is first identified through regular matching, and then the industry jargon is replaced with professional expressions according to the mapping relationship. After rewriting the target question, the relevant knowledge fragments are retrieved from the knowledge base and sent to the finance and taxation big model after being spliced ​​with the context and instruction template and the question.

[0096] (3) Rewriting of multiple rounds of dialogue.

[0097] When the target question involves multiple rounds of dialogue, the target question, that is, the last question of the multiple rounds of dialogue, is semantically rewritten in combination with historical dialogue information to form a question with reference disambiguation and semantic completeness, that is, a question that can be understood without the help of previous information.

[0098] In an embodiment of the present application, if the target question of the first sub-intention is in a multi-round dialogue, an optional implementation method is as follows when rewriting the target question in a multi-round dialogue:

[0099] ① If the number of rounds of multi-round dialogue does not exceed the preset value, the target question is rewritten according to the dialogue information of the previous round in the multi-round dialogue;

[0100] Specifically, the preset value is a user-defined smaller value, such as 3, 4, etc.

[0101] Taking the preset value of 3 as an example, when the dialogue does not exceed (inclusive of) 3 rounds, all dialogue information in the current multi-round dialogue is used to rewrite the target question.

[0102] ② If the number of rounds of multi-round dialogues exceeds a preset value, the previous dialogue of the specified round is selected, and based on the relevance of other dialogues in the multi-round dialogues except the target question and the previous dialogue with the target question, a preset number of other dialogues are selected, and the target question is rewritten based on the other dialogues and the previous dialogue.

[0103] The designated rounds are also a user-defined smaller value, such as 2, 3, etc. Generally, the number of designated rounds can be set to the preset value minus one.

[0104] When selecting other conversations based on relevance, there may be a situation where the number of other conversation rounds does not meet the preset value. In this case, all other conversations in the multiple conversation rounds are selected. If the number of other conversation rounds meets or exceeds the preset value, they are sorted from high to low in terms of relevance, and the conversations with the highest relevance and the preset value are selected.

[0105] Considering that adjacent rounds of dialogue are more relevant to the target question, the preceding round can be the first two rounds of dialogue adjacent to the target question, or any specified round of preceding dialogue before the target question.

[0106] Still taking the preset value of 3 as an example, when the specified rounds are the first two rounds adjacent to the target question, when the dialogue exceeds 3 rounds, select the 3 rounds of dialogue including the target question and record them as dialogue_1. After removing dialogue_1 from the historical dialogues of the multi-round dialogue, calculate the correlation between each round of dialogue and the target question, select the 3 rounds of dialogue with the highest correlation ranking (if there are less than 3 rounds, select all; if there are more than 3 rounds, only select 3 rounds), keep the dialogue order unchanged, splice them with dialogue_1, integrate the multi-round rewriting instructions and the spliced ​​multi-round dialogues, and complete the multi-round dialogue rewriting.

[0107] It should be noted that each round of dialogue in the multi-round dialogue has gone through the aforementioned time rewriting, industry customary text replacement and multi-round dialogue problem rewriting steps.

[0108] The multi-round dialogue rewriting can be implemented by other models independent of the finance and taxation big model, or can be implemented directly by the finance and taxation big model. When conducting multi-round dialogues, question rewriting task instructions can be input to the big model to instruct the big model to rewrite the questions.

[0109] For example, the question rewriting task instructions could be:

[0110] You are tasked with rewriting the questions according to the given dialogue:

[0111] Your input is the conversation between user (object) and bot (machine);

[0112] You rewrite the user's last statement based on the conversation content, rewriting it into a sentence that can be understood without the previous information;

[0113] Your output is strictly in JSON format;

[0114] ## Output:

[0115] JSON format: {{"user": rewritten statement that can be understood without the above context}};

[0116] ##Require:

[0117] If the user's last statement is not related to the previous text, there is no need to rewrite it;

[0118] If the user's last statement is related to the previous one, but can be understood without context, there is no need to rewrite it;

[0119] If the user's last statement is related to the previous one and requires context to understand, rewrite it;

[0120] ##enter:

[0121] {history(historical dialogue of multiple rounds of dialogue)}.

[0122] The following is an example of rewriting the target question based on historical dialogue information in multiple rounds of dialogue:

[0123] ##enter:

[0124] {"user":"Is it okay for the company to charge the purchase of safety helmets to labor insurance fees?"};

[0125] {"bot":"If the purchase is for use in the production process, it should be recorded as labor insurance fee"};

[0126] {"user":"It is sometimes used in construction projects, how to record it?"};

[0127] ## Output:

[0128] {"user": How should the cost of purchasing hard hats used on construction projects be accounted for?}

[0129] In the above implementation, for the target questions of the first type of intent, such as fiscal and tax policy consultation, digital invoice Q&A, and exam Q&A, a combined question rewriting strategy is designed to rewrite the target questions through time rewriting, industry jargon replacement, and multi-round dialogue rewriting based on historical dialogues, thereby eliminating ambiguity and improving semantics. In order to facilitate the subsequent steps based on the rewritten questions, the accuracy of the model's answers can be improved by using retrieval enhancement technology.

[0130] 2. The following is an introduction to the intermediate processing of the target problem of the second sub-intention:

[0131] If the question intent is the second sub-intention, the target question is input into the trained computational agent, and the output result of the computational agent is input into the finance and taxation model.

[0132] In the embodiment of the present application, when the target question of the second sub-intention is intermediately processed, an optional implementation method is as follows:

[0133] Input the target problem into the computational agent; based on the computational agent, decompose the target problem into complex tasks to obtain at least one sub-problem and the decomposition process; based on the computational agent, determine the corresponding processing plan for each sub-problem, and execute the corresponding processing plan for each sub-problem to obtain the corresponding processing result; the sub-problems correspond to the processing plans one by one; input at least one processing plan, at least one processing result, the decomposition process and the second task instruction into the finance and taxation big model.

[0134] Among them, the computational agent can be part of the function of the fiscal and taxation big model, or it can be another model independent of the fiscal and taxation big model. It fine-tunes the model by constructing unique complex task training data, realizes the complex problem decomposition function of the big model, decomposes complex tasks into multi-step sub-problems, and determines and executes the corresponding processing plan for each sub-problem to obtain the corresponding processing result. The sub-problem corresponds to the processing plan one by one, and the sub-problem also corresponds to the processing result one by one.

[0135] The processing scheme may include searching the target problem in the knowledge base, calling computing tools to perform calculations, and so on.

[0136] Among them, the knowledge base is a knowledge base in the field of finance and taxation, which contains sub-knowledge bases in multiple sub-fields under the field of finance and taxation. If the second sub-intention includes intentions such as financial and taxation calculations in the field of finance and taxation, then a search is performed in the financial and taxation calculation sub-knowledge base in the knowledge base to obtain the corresponding knowledge fragment.

[0137] The second task instruction refers to the instruction that instructs the financial and taxation model to generate answers based on the input processing plan, processing results and disassembly process.

[0138] Specifically, the target problem with the second sub-intention of problem intention identification such as financial and tax calculations is first input into the computing agent for complex problem decomposition, and the target problem is decomposed into at least one sub-problem, and the processing plan for each sub-problem is determined. After the processing plan is executed for each sub-problem, the decomposition process of at least one sub-problem, the processing plan of each sub-problem, the processing result of each sub-problem and the second task instruction are input into the financial and tax model.

[0139] For example, if the computational agent breaks down the target problem into two sub-problems, namely sub-problem 1 and sub-problem 2, and determines that sub-problem 1 needs to be searched in the knowledge base, and sub-problem 2 needs to call computing tool A for calculation, then the computational agent inputs sub-problem 1 into the knowledge base for retrieval, obtains the retrieved knowledge fragments, and calls computing tool A to calculate sub-problem 2 to obtain the calculation results. Then, the computational agent inputs the decomposition process, the processing plan for each sub-problem, and the retrieval results and calculation results into the taxation model.

[0140] In the above implementation, by building a computing agent, the target problems of the second sub-intention such as financial and tax calculations are decomposed into tasks, the execution order of tasks is planned, and with the help of a knowledge base, tools such as calculators or code executors are called to complete the reasoning calculations, thereby enhancing the performance of the financial and tax model in reasoning tasks of complex computational problems and improving the accuracy and explainability of answers.

[0141] S204: Use the output results of the financial and tax model as the answer to the target question.

[0142] After processing the target problem through the above steps, input it into the financial and tax model to get the answer to the target problem.

[0143] like Figure 3 As shown, it is a schematic diagram of information processing provided in an embodiment of the present application. Figure 3 After receiving the target question of the object, the target question is first classified into problem intents; if the problem intent of the target question is the first sub-intention such as policy consultation, digital invoice Q&A and examination Q&A, then the target question is rewritten and searched in the corresponding knowledge bases such as policy knowledge base, digital invoice knowledge base and examination knowledge base, and then the retrieved knowledge fragments, the rewritten target question and the first task instruction are spliced ​​and input into the finance and taxation model; if the problem intent of the target question is the second sub-intention such as finance and taxation calculation, then the target question is input into the computational agent for complex problem decomposition, and the decomposition process, processing plan and processing result output by the computational agent are input into the finance and taxation model; wherein, Figure 3 The processing scheme of the computing agent in the above example includes searching the sub-problems obtained by disassembly. If the question intent of the target question is a second-class intent such as accounting entries, the target question is directly input into the taxation model. The answer generated by the taxation model is fed back to the object.

[0144] In the embodiment of the present application, the technologies such as multi-classification of intent, question rewriting, and computational agents are used to jointly realize the refined recognition and differential processing of the object intent. For the target questions identified as the first sub-intent such as policy consultation, digital invoice Q&A, and exam Q&A, the rewritten questions are obtained through time rewriting, industry jargon replacement, and complex semantic rewriting based on historical dialogues. The rewritten questions are used for retrieval and different external vector knowledge base retrieval, which significantly improves the retrieval accuracy and recall rate compared to the retrieval with the original questions. The design of the instruction template integrates the retrieved knowledge fragments with the rewritten questions and then sends them to the large language model, which can improve the accuracy and reliability of the answers generated by the model and avoid the problems of large model hallucinations and the generation of outdated answers. For the target questions identified as the second sub-intent such as fiscal and tax calculations, the task decomposition, tool calling, and model reasoning are completed by building a computational agent, which solves the problem of calculation errors of the fiscal and taxation large model and improves the reasoning ability and interpretability of the large model on complex tasks.

[0145] Based on the same inventive concept, the embodiment of the present application also provides an information processing device 400. Figure 4 As shown, it is a schematic diagram of the structure of the information processing device 400, which may include:

[0146] A receiving unit 401 is used to receive a target question input by a subject;

[0147] The acquisition unit 402 is used to input the target question into the trained classification model to obtain the question intention to which the target question belongs; the classification model is obtained after fine-tuning based on the labeled financial and taxation question and answer data;

[0148] The determination unit 403 is used to determine the processing flow corresponding to the target problem according to the problem intention; wherein, if the problem intention is the first type of intention, the processing flow is to perform intermediate processing on the target problem and input the intermediate processing result into the finance and taxation big model; if the problem intention is the second type of intention, the processing flow is to directly input the target problem into the finance and taxation big model;

[0149] The output unit 404 is used to use the output result of the financial and tax model as the answer to the target question.

[0150] Optionally, the first type of intent includes a first sub-intent and a second sub-intent, and the determining unit 403 is specifically configured to:

[0151] If the question intent is the first sub-intent, rewrite the target question and search the knowledge base, and input the rewritten and search results into the trained tax model;

[0152] If the question intent is the second sub-intent, the target question is input into the trained computational agent, and the output result of the computational agent is input into the financial and taxation model.

[0153] Optionally, the determining unit 403 is specifically configured to:

[0154] Rewrite the target question, and input the rewritten question into the knowledge base for retrieval to obtain knowledge fragments;

[0155] The rewritten questions, knowledge fragments and preset first task instructions are spliced ​​together and input into the trained finance and taxation model.

[0156] Optionally, if the question intent is the first sub-intent, the determining unit 403 is specifically configured to:

[0157] According to the preset mapping relationship for the time statement, the fuzzy time text in the target question is converted into a specific time text through regular matching; the specific time text is used to limit the relatively accurate time involved in the target question;

[0158] According to the pre-built unique dictionary, the industry customary text in the target question is converted into industry standard text.

[0159] Optionally, if the target question is in a multi-round dialogue, the determining unit 403 is further configured to:

[0160] If the number of rounds of the multi-round dialogue does not exceed a preset value, rewriting the target question according to dialogue information of a preceding round in the multi-round dialogue;

[0161] If the number of rounds of the multiple conversations exceeds the preset value, a preceding conversation of the specified round is selected, and based on the relevance of other conversations in the multiple conversations except the target question and the preceding conversation to the target question, a preset number of other conversations are selected, and the target question is rewritten based on the other conversations and the preceding conversation.

[0162] Optionally, the determining unit 403 is specifically configured to:

[0163] Input the target problem into the computational agent;

[0164] Based on computational agents, the target problem is decomposed into complex tasks to obtain at least one sub-problem and decomposition process;

[0165] Based on the computing agent, determine the corresponding processing solution for each sub-problem, and execute the corresponding processing solution for each sub-problem to obtain the corresponding processing result; the sub-problem and the processing solution correspond one to one;

[0166] At least one processing plan, at least one processing result, a disassembly process and a second task instruction are input into the finance and taxation big model.

[0167] Optionally, the classification model is a general text classification model; and the financial and taxation question and answer data is mixed data of various question intentions.

[0168] After introducing the information processing method and apparatus according to the exemplary embodiment of the present application, next, an electronic device according to another exemplary embodiment of the present application is introduced.

[0169] Those skilled in the art will appreciate that various aspects of the present application may be implemented as a system, method or program product. Therefore, various aspects of the present application may be specifically implemented in the following forms, namely: a complete hardware implementation, a complete software implementation (including firmware, microcode, etc.), or a combination of hardware and software, which may be collectively referred to as "circuit", "module" or "system" herein.

[0170] Based on the same inventive concept as the above method embodiment, an electronic device is also provided in the embodiment of the present application. In one embodiment, the electronic device may be a server, such as Figure 1 In this embodiment, the structure of the electronic device can be as follows: Figure 5 As shown, it includes a memory 501 , a communication module 503 and one or more processors 502 .

[0171] The memory 501 is used to store computer programs executed by the processor 502. The memory 501 may mainly include a program storage area and a data storage area, wherein the program storage area may store an operating system and programs required for running the instant messaging function, etc.; the data storage area may store various instant messaging information and operation instruction sets, etc.

[0172] The memory 501 may be a volatile memory, such as a random-access memory (RAM); the memory 501 may also be a non-volatile memory, such as a read-only memory, a flash memory, a hard disk drive (HDD) or a solid-state drive (SSD); or the memory 501 may be any other medium that can be used to carry or store a desired computer program in the form of instructions or data structures and can be accessed by a computer, but is not limited thereto. The memory 501 may be a combination of the above memories.

[0173] The processor 502 may include one or more central processing units (CPU) or a digital processing unit, etc. The processor 502 is used to implement the above information processing method when calling the computer program stored in the memory 501 .

[0174] The communication module 503 is used to communicate with terminal devices and other servers.

[0175] The specific connection medium between the memory 501, the communication module 503 and the processor 502 is not limited in the embodiment of the present application. Figure 5 In the embodiment, the memory 501 and the processor 502 are connected via a bus 504. The bus 504 is connected to the processor 502 via a bus 504. Figure 5 The connections between the other components are described with bold lines, which are only for illustrative purposes and are not intended to be limiting. The bus 504 can be divided into an address bus, a data bus, a control bus, etc. For ease of description, Figure 5 The diagram shows that only one thick line is used, but this does not mean that there is only one bus or only one type of bus.

[0176] The memory 501 stores a computer storage medium, and the computer storage medium stores computer executable instructions, and the computer executable instructions are used to implement the information processing method of the embodiment of the present application. The processor 502 is used to execute the above-mentioned information processing method, such as Figure 2 shown.

[0177] In another embodiment, the electronic device may also be other electronic devices, such as Figure 1 The terminal device 110 shown in FIG. 1 is a terminal device 110 shown in FIG. 1 . In this embodiment, the structure of the electronic device can be as follows: Figure 6 As shown, it includes: a communication component 610, a memory 620, a display unit 630, a camera 640, a sensor 650, an audio circuit 660, a Bluetooth module 670, a processor 680 and other components.

[0178] The communication component 610 is used to communicate with the server. In some embodiments, a wireless fidelity (WiFi) module may be included. The WiFi module belongs to a short-range wireless transmission technology. The electronic device can help the object to send and receive information through the WiFi module.

[0179] The memory 620 can be used to store software programs and data. The processor 680 executes various functions and data processing of the terminal device 110 by running the software programs or data stored in the memory 620. The memory 620 may include a high-speed random access memory, and may also include a non-volatile memory, such as at least one disk storage device, a flash memory device, or other volatile solid-state storage device. The memory 620 stores an operating system that enables the terminal device 110 to run. In the present application, the memory 620 can store an operating system and various application programs, and can also store a computer program for executing the information processing method of the embodiment of the present application.

[0180] The display unit 630 may also be used to display information input by the object or information provided to the object and a graphical user interface (GUI) of various menus of the terminal device 110. Specifically, the display unit 630 may include a display screen 632 disposed on the front of the terminal device 110. The display screen 632 may be configured in the form of a liquid crystal display, a light emitting diode, etc.

[0181] The display unit 630 can also be used to receive input digital or character information and generate signal input related to user settings and function control of the terminal device 110. Specifically, the display unit 630 may include a touch screen 631 arranged on the front of the terminal device 110, which can collect touch operations of objects on or near it, such as clicking a button, dragging a scroll box, etc.

[0182] The touch screen 631 can be covered on the display screen 632, or the touch screen 631 and the display screen 632 can be integrated to realize the input and output functions of the terminal device 110, and the integrated display screen can be referred to as a touch display screen. In this application, the display unit 630 can display the application and the corresponding operation steps.

[0183] The camera 640 can be used to capture static images, and the object can publish the images taken by the camera 640 through the application. The camera 640 can be one or more. The object generates an optical image through the lens and projects it onto the photosensitive element. The photosensitive element can be a charge coupled device (CCD) or a complementary metal-oxide-semiconductor (CMOS) phototransistor. The photosensitive element converts the optical signal into an electrical signal, and then transmits the electrical signal to the processor 680 to convert it into a digital image signal.

[0184] The terminal device may further include at least one sensor 650, such as an acceleration sensor 651, a distance sensor 652, a fingerprint sensor 653, and a temperature sensor 654. The terminal device may also be configured with other sensors such as a gyroscope, a barometer, a hygrometer, a thermometer, an infrared sensor, a light sensor, and a motion sensor.

[0185] The audio circuit 660, the speaker 661, and the microphone 662 can provide an audio interface between the object and the terminal device 110. The audio circuit 660 can transmit the electrical signal converted from the received audio data to the speaker 661, which is converted into a sound signal for output. The terminal device 110 can also be equipped with a volume button for adjusting the volume of the sound signal. On the other hand, the microphone 662 converts the collected sound signal into an electrical signal, which is received by the audio circuit 660 and converted into audio data, and then the audio data is output to the communication component 610 to be sent to, for example, another terminal device 110, or the audio data is output to the memory 620 for further processing.

[0186] The Bluetooth module 670 is used to exchange information with other Bluetooth devices having Bluetooth modules through the Bluetooth protocol. For example, the terminal device can establish a Bluetooth connection with a wearable electronic device (such as a smart watch) that also has a Bluetooth module through the Bluetooth module 670 to exchange data.

[0187] The processor 680 is the control center of the terminal device. It uses various interfaces and lines to connect various parts of the entire terminal. It executes various functions of the terminal device and processes data by running or executing software programs stored in the memory 620 and calling data stored in the memory 620. In some embodiments, the processor 680 may include one or more processing units; the processor 680 may also integrate an application processor and a baseband processor, wherein the application processor mainly processes the operating system, user interface, and application programs, and the baseband processor mainly processes wireless communications. It is understandable that the above-mentioned baseband processor may not be integrated into the processor 680. In the present application, the processor 680 can run the operating system, application programs, user interface display and touch response, as well as the information processing method of the embodiment of the present application. In addition, the processor 680 is coupled to the display unit 630.

[0188] In some possible implementations, various aspects of the information processing method provided by the present application may also be implemented in the form of a program product, which includes a computer program. When the program product is run on an electronic device, the computer program is used to enable the electronic device to execute the steps of the information processing method according to various exemplary implementations of the present application described above in this specification. For example, the electronic device may execute the following steps: Figure 2 Follow the steps shown in .

[0189] The program product may use any combination of one or more readable media. The readable medium may be a readable signal medium or a readable storage medium. The readable storage medium may be, for example, but not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or device, or any combination of the above. More specific examples of readable storage media (a non-exhaustive list) include: an electrical connection with one or more wires, a portable disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the above.

[0190] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may adopt the form of a complete hardware embodiment, a complete software embodiment, or an embodiment in combination with software and hardware. Moreover, the present application may adopt the form of a computer program product implemented in one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) that include computer-usable program code.

[0191] The present application is described with reference to the flowcharts and / or block diagrams of the methods, devices (systems), and computer program products according to the present application. It should be understood that each process and / or box in the flowchart and / or block diagram, as well as the combination of the processes and / or boxes in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing device to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing device generate instructions for implementing the processes in the flowchart and / or block diagram. Figure 1 A process or multiple processes and / or boxes Figure 1 A device that provides the functions specified in a block or multiple blocks.

[0192] These computer program instructions may also be stored in a computer-readable memory capable of directing a computer or other programmable data processing device to operate in a specific manner, so that the instructions stored in the computer-readable memory produce an article of manufacture comprising an instruction device, which implements the process Figure 1 A process or multiple processes and / or boxes Figure 1 A function specified in one or more boxes.

[0193] These computer program instructions can also be loaded onto a computer or other programmable data processing device so that a series of operating steps are executed on the computer or other programmable device to produce a computer-implemented process, thereby providing instructions for implementing the process. Figure 1 A process or multiple processes and / or boxes Figure 1 The steps for the functions specified in one or more boxes.

[0194] Obviously, those skilled in the art can make various changes and modifications to the present application without departing from the spirit and scope of the present application. Thus, if these modifications and variations of the present application fall within the scope of the claims of the present application and their equivalents, the present application is also intended to include these modifications and variations.

Claims

1. An information processing method, characterized in that: The method comprises: Receive target questions input by subjects; Input the target question into a trained classification model to obtain the question intent to which the target question belongs; the classification model is obtained after fine-tuning based on the labeled financial and taxation question and answer data; According to the problem intent, determine the processing flow corresponding to the target problem; wherein, if the problem intent is the first type of intent, the processing flow is to perform intermediate processing on the target problem and then input the intermediate processing result into the finance and taxation model; if the problem intent is the second type of intent, the processing flow is to directly input the target problem into the finance and taxation model; The output result of the financial and taxation model is used as the answer to the target question.

2. The method according to claim 1, characterized in that The first type of intent includes a first sub-intent and a second sub-intent, and determining a processing flow corresponding to the target problem according to the problem intent includes: If the question intent is the first sub-intent, after rewriting the target question and searching the knowledge base, the rewritten question and search results are input into the trained tax model; If the question intention is the second sub-intention, the target question is input into the trained computational agent, and the output result of the computational agent is input into the financial and taxation model.

3. The method according to claim 1, characterized in that After rewriting the target question and searching the knowledge base, the rewritten question and the search results are input into the trained tax model, including: Rewrite the target question, and input the rewritten question into a knowledge base for retrieval to obtain knowledge fragments; The rewritten target question, the knowledge fragment and the preset first task instruction are spliced ​​together and input into the trained finance and taxation model.

4. The method according to claim 2 or 3, characterized in that If the question intent is the first sub-intent, rewriting the target question includes: According to the preset mapping relationship for the time statement, the fuzzy time text in the target question is converted into a specific time text through regular matching; the specific time text is used to limit the relatively accurate time involved in the target question; According to the pre-built unique dictionary, the industry customary text in the target question is converted into industry standard text.

5. The method according to claim 4, characterized in that If the target question is in a multi-round dialogue, the step of rewriting the target question further includes: If the number of rounds of the multi-round dialogue does not exceed a preset value, rewriting the target question according to dialogue information of a preceding round in the multi-round dialogue; If the number of rounds of the multiple conversations exceeds the preset value, a preceding conversation of the specified round is selected, and based on the relevance of other conversations in the multiple conversations except the target question and the preceding conversation to the target question, a preset number of other conversations are selected, and the target question is rewritten based on the other conversations and the preceding conversation.

6. The method according to claim 2, characterized in that The step of inputting the target problem into a trained computing agent and inputting the output result of the computing agent into the financial and taxation model comprises: Inputting the target problem into the computational agent; Based on the computing agent, decompose the target problem into complex tasks to obtain at least one sub-problem and a decomposition process; Based on the computing agent, determine a processing solution corresponding to each of the sub-problems, and execute the corresponding processing solution on each of the sub-problems to obtain a corresponding processing result; the sub-problems correspond to the processing solutions one by one; Input at least one of the processing solutions, at least one of the processing results, the disassembly process and the second task instruction into the financial and tax model.

7. The method according to claim 1, characterized in that The classification model is a general text classification model; the financial and taxation question and answer data is mixed data of various question intentions.

8. An information processing device, characterized in that: include: A receiving unit, used for receiving a target question input by a subject; An acquisition unit, used for inputting the target question into a trained classification model to acquire the question intent to which the target question belongs; The classification model is obtained after fine-tuning based on the labeled financial and taxation question and answer data; A determination unit is used to determine a processing flow corresponding to the target problem according to the problem intention; wherein, if the problem intention is the first type of intention, the processing flow is to perform intermediate processing on the target problem and then input the intermediate processing result into the finance and taxation model; if the problem intention is the second type of intention, the processing flow is to directly input the target problem into the finance and taxation model; An output unit is used to use the output result of the financial and tax model as the answer to the target question.

9. An electronic device, characterized in that: It comprises a processor and a memory, wherein the memory stores a computer program, and when the computer program is executed by the processor, the processor executes the steps of any one of the methods of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: It includes a computer program. When the computer program is run on an electronic device, the computer program is used to enable the electronic device to execute the steps of any method described in claims 1 to 7.

11. A computer program product, characterized in that The method comprises a computer program stored in a computer-readable storage medium; when a processor of an electronic device reads the computer program from the computer-readable storage medium, the processor executes the computer program, so that the electronic device executes the steps of any one of the methods described in claims 1 to 7.

Images