Talk skill generation method and device, equipment, storage medium and computer program product

Abstract

The invention provides a verbal skill generation method and device, equipment, a storage medium and a computer program product, and the method comprises the steps: carrying out the intention recognition of a to-be-replied question of a user, and obtaining the intention information of the user; the intention recognition comprises the steps of recognizing a first-level intention category of the to-be-replied question, and recognizing a second-level intention category under the first-level intention category; performing parameter extraction on the to-be-replied question to obtain question parameter information; generating reply verbal skills corresponding to the to-be-replied questions based on the user intention information and the question parameter information by using a target algorithm; wherein the target algorithm is an algorithm corresponding to the user intention information and the problem parameter information. According to the scheme, accurate recognition of the user intention can be supported, the target algorithm is called according to the user intention to generate the reply verbal skill, excessive dependence on text vectors can be avoided, the accuracy of generating the reply verbal skill is improved, and the problem that in the prior art, a verbal skill generation scheme is low in reply accuracy is well solved.

Description

Technical Field

[0001] This application relates to the field of script generation technology, and in particular to a script generation method, apparatus, device, storage medium, and computer program product. Background Technology

[0002] Existing automated reply assistants (such as investment and financial management assistants) are mainly based on traditional summary question and answer methods such as FAQs (Frequently Asked Questions) to generate reply scripts for user questions (such as investment and financial management scripts). However, this method cannot achieve accurate replies and relies too much on text vectors, which can easily lead to irrelevant answers.

[0003] As shown above, existing script generation solutions suffer from problems such as low response accuracy. Summary of the Invention

[0004] The purpose of this application is to provide a method, apparatus, device, storage medium, and computer program product for generating dialogue scripts, so as to solve the problem of low response accuracy in existing dialogue script generation solutions.

[0005] To address the aforementioned technical problems, embodiments of this application provide a script generation method, including:

[0006] The intent of a user's pending question is identified to obtain user intent information; the intent identification includes: identifying a first-level intent category of the pending question, and identifying a second-level intent category under the first-level intent category;

[0007] The parameters of the question to be answered are extracted to obtain the question parameter information;

[0008] Using a target algorithm, a response script corresponding to the question to be answered is generated based on the user intent information and question parameter information; wherein, the target algorithm is an algorithm corresponding to the user intent information and question parameter information.

[0009] Optional, also includes:

[0010] Obtain the user's problem information;

[0011] If the question information belongs to a preset script, obtain the preset response corresponding to the question information and use it as the response script corresponding to the question information;

[0012] If the question information does not belong to the preset script, the question information will be treated as the question to be answered.

[0013] Optional, also includes:

[0014] Configure auxiliary parameters for intent recognition; these auxiliary parameters are used to assist in intent recognition.

[0015] Obtain the parameter information of the auxiliary parameter items in the question to be answered;

[0016] The process of identifying the user's unanswered question to obtain user intent information includes:

[0017] Based on the parameter information of the acquired auxiliary parameter items, intent recognition is performed on the user's question to be answered, and user intent information is obtained.

[0018] Optional, also includes:

[0019] Set parameter extraction mapping relationships; the parameter extraction mapping relationships include: character extraction mapping relationships under the first extraction type;

[0020] The step of extracting parameters from the question to be answered to obtain question parameter information includes:

[0021] Parameters are extracted from the question to be answered to obtain initial parameter information under the first extraction type;

[0022] Based on the character extraction mapping relationship, the initial parameter information is mapped to information in the form of target characters, which is then used as the extracted information;

[0023] Based on the extracted information, the question parameter information corresponding to the question to be answered is obtained.

[0024] Optional, also includes:

[0025] The problem parameter information is subjected to a first data processing to obtain the processed problem parameter information; wherein, the first data processing includes at least one of the following: numerical conversion, numerical scaling, numerical verification, numerical mapping, and numerical filtering;

[0026] The step of using the target algorithm to generate the response script corresponding to the question to be answered, based on the user intent information and question parameter information, includes:

[0027] Using the target algorithm, a response script is generated based on the user intent information and the processed question parameter information to address the question to be answered.

[0028] Optionally, the step of performing intent recognition on the user's pending question to obtain user intent information includes:

[0029] The intent of the user's pending questions is identified to obtain preliminary intent information;

[0030] The preliminary intent information is corrected using intent rules to obtain user intent information.

[0031] Optional, also includes:

[0032] The response script is segmented to obtain at least one script fragment;

[0033] Obtain source document fragments that satisfy the first similarity condition with each of the aforementioned speech fragments to obtain a fragment set;

[0034] Obtain the second similarity between each of the stated speech fragments and each source document fragment in the fragment set;

[0035] Based on the second similarity, the target source document fragment corresponding to each of the above-mentioned speech fragments is determined;

[0036] Establish the correspondence between each of the aforementioned script segments and the corresponding target source document segments, and then annotate them.

[0037] Optionally, the at least one speech fragment includes a first speech fragment, and the set of fragments includes a first source document fragment;

[0038] The step of obtaining the second similarity between each of the stated speech segments and each source document segment in the segment set includes:

[0039] Based on the length of the sliding window, the first speech segment is sliced ​​to obtain at least one speech slice.

[0040] Based on the sliding window length, the first source document segment is sliced ​​to obtain at least one document slice;

[0041] Obtain the third similarity between each of the stated speech slices and each of the stated document slices;

[0042] Based on the third similarity, a second similarity is obtained between the first speech fragment and the first source document fragment.

[0043] Optional, also includes:

[0044] Obtain the similarity weight coefficient between each of the speech slices and each of the document slices;

[0045] The step of obtaining the third similarity between each of the speech segments and each of the document segments includes:

[0046] Obtain the initial similarity between each of the stated speech slices and each of the stated document slices;

[0047] Based on the initial similarity and the corresponding similarity weight coefficient, the third similarity between each of the utterance slices and each of the document slices is obtained.

[0048] This application also provides a script generation device, including:

[0049] The first identification module is used to identify the user's intent for the question to be answered, and to obtain the user's intent information; the intent identification includes: identifying the first-level intent category of the question to be answered, and identifying the second-level intent category under the first-level intent category;

[0050] The first extraction module is used to extract parameters from the question to be answered, and obtain question parameter information;

[0051] The first generation module is used to generate a response script corresponding to the question to be answered based on the user intent information and question parameter information using a target algorithm; wherein, the target algorithm is an algorithm corresponding to the user intent information and question parameter information.

[0052] Optional, also includes:

[0053] The first acquisition module is used to acquire the user's problem information;

[0054] The second acquisition module is used to acquire a preset response corresponding to the question information when the question information belongs to a preset script, and use it as the response script corresponding to the question information;

[0055] The first processing module is used to treat the question information as the question to be answered when the question information does not belong to the preset script.

[0056] Optional, also includes:

[0057] The first setting module is used to set auxiliary parameters for intent recognition; the auxiliary parameters are used to assist in intent recognition.

[0058] The third acquisition module is used to acquire parameter information of the auxiliary parameter items in the question to be answered;

[0059] The process of identifying the user's unanswered question to obtain user intent information includes:

[0060] Based on the parameter information of the acquired auxiliary parameter items, intent recognition is performed on the user's question to be answered, and user intent information is obtained.

[0061] Optional, also includes:

[0062] The second setting module is used to set the parameter extraction mapping relationship; the parameter extraction mapping relationship includes: the character extraction mapping relationship under the first extraction type;

[0063] The step of extracting parameters from the question to be answered to obtain question parameter information includes:

[0064] Parameters are extracted from the question to be answered to obtain initial parameter information under the first extraction type;

[0065] Based on the character extraction mapping relationship, the initial parameter information is mapped to information in the form of target characters, which is then used as the extracted information;

[0066] Based on the extracted information, the question parameter information corresponding to the question to be answered is obtained.

[0067] Optional, also includes:

[0068] The second processing module is used to perform first data processing on the problem parameter information to obtain processed problem parameter information; wherein, the first data processing includes at least one of: numerical conversion, numerical scaling, numerical verification, numerical mapping, and numerical filtering;

[0069] The step of using the target algorithm to generate the response script corresponding to the question to be answered, based on the user intent information and question parameter information, includes:

[0070] Using the target algorithm, a response script is generated based on the user intent information and the processed question parameter information to address the question to be answered.

[0071] Optionally, the step of performing intent recognition on the user's pending question to obtain user intent information includes:

[0072] The intent of the user's pending questions is identified to obtain preliminary intent information;

[0073] The preliminary intent information is corrected using intent rules to obtain user intent information.

[0074] Optional, also includes:

[0075] The first segmentation module is used to segment the text of the reply script to obtain at least one script fragment.

[0076] The fourth acquisition module is used to acquire source document fragments that meet the first similarity condition with each of the aforementioned speech fragments, thereby obtaining a fragment set;

[0077] The fifth acquisition module is used to acquire the second similarity between each of the aforementioned speech fragments and each source document fragment in the fragment set;

[0078] The first determining module is used to determine the target source document fragment corresponding to each of the aforementioned speech fragments based on the second similarity.

[0079] The third processing module is used to establish the correspondence between each of the aforementioned speech fragments and the corresponding target source document fragments, and to annotate them.

[0080] Optionally, the at least one speech fragment includes a first speech fragment, and the set of fragments includes a first source document fragment;

[0081] The step of obtaining the second similarity between each of the stated speech segments and each source document segment in the segment set includes:

[0082] Based on the length of the sliding window, the first speech segment is sliced ​​to obtain at least one speech slice.

[0083] Based on the sliding window length, the first source document segment is sliced ​​to obtain at least one document slice;

[0084] Obtain the third similarity between each of the stated speech slices and each of the stated document slices;

[0085] Based on the third similarity, a second similarity is obtained between the first speech fragment and the first source document fragment.

[0086] Optional, also includes:

[0087] The sixth acquisition module is used to acquire the similarity weight coefficient between each of the speech slices and each of the document slices;

[0088] The step of obtaining the third similarity between each of the speech segments and each of the document segments includes:

[0089] Obtain the initial similarity between each of the stated speech slices and each of the stated document slices;

[0090] Based on the initial similarity and the corresponding similarity weight coefficient, the third similarity between each of the utterance slices and each of the document slices is obtained.

[0091] This application also provides a script generation device, including a memory, a processor, and a program stored in the memory and executable on the processor; when the processor executes the program, it implements the above-described script generation method.

[0092] This application also provides a readable storage medium storing a program that, when executed by a processor, implements the steps in the above-described speech generation method.

[0093] This application also provides a computer program product, including computer instructions, which, when executed by a processor, implement the steps of the above-described speech generation method.

[0094] The beneficial effects of the above technical solution in this application are as follows:

[0095] In the above scheme, the script generation method obtains user intent information by performing intent recognition on the user's question to be answered; the intent recognition includes: identifying the first-level intent category of the question to be answered, and identifying the second-level intent category under the first-level intent category; extracting parameters from the question to be answered to obtain question parameter information; and using a target algorithm, generating a response script corresponding to the question to be answered based on the user intent information and question parameter information; wherein, the target algorithm is an algorithm corresponding to the user intent information and question parameter information; it can support accurate identification of user intent and call the target algorithm to generate response script accordingly, which can avoid excessive reliance on text vectors, improve the accuracy of generated response scripts, and effectively solve the problem of low response accuracy in existing script generation schemes. Attached Figure Description

[0096] Figure 1 This is a schematic diagram of the speech generation method according to an embodiment of this application;

[0097] Figure 2 This is a schematic diagram illustrating the specific implementation process of the script generation method in an embodiment of this application;

[0098] Figure 3 This is a schematic diagram of the intent recognition framework in an embodiment of this application;

[0099] Figure 4 This is a schematic diagram of the rules for an embodiment of this application. Figure 1 ;

[0100] Figure 5 This is a schematic diagram of the rules for an embodiment of this application. Figure 2 ;

[0101] Figure 6 This is a schematic diagram of the rules for an embodiment of this application. Figure 3 ;

[0102] Figure 7 This is a schematic diagram of the rules for an embodiment of this application. Figure 4 ;

[0103] Figure 8 This is a schematic diagram of the rules for an embodiment of this application. Figure 5 ;

[0104] Figure 9 This is a schematic diagram of the speech generation device according to an embodiment of this application. Detailed Implementation

[0105] To make the technical problems, technical solutions and advantages of this application clearer, a detailed description will be provided below in conjunction with the accompanying drawings and specific embodiments.

[0106] This application addresses the problem of low response accuracy in existing script generation solutions by providing a script generation method, such as... Figure 1 As shown, it includes:

[0107] Step 11: Perform intent recognition on the user's question to be answered to obtain user intent information; the intent recognition includes: identifying the first-level intent category of the question to be answered, and identifying the second-level intent category (of which the question to be answered is) the first-level intent category;

[0108] Step 11 may include identifying the broad category of intent for the question to be answered (first-level intent category), and then identifying the sub-category of intent under the identified broad category (second-level intent category under the first-level intent category), but is not limited to this. The broad category and sub-category are relative terms; the sub-category is a further refinement of the broad category. For example, if the broad category is "asset allocation," the sub-categories might be "pension asset allocation," "public fund asset allocation," and "fund allocation," etc., under "asset allocation."

[0109] Step 12: Extract parameters from the question to be answered to obtain question parameter information;

[0110] Step 12 may include: extracting parameters and processing the extracted parameters to obtain problem parameter information, but is not limited to this.

[0111] Step 13: Using the target algorithm, generate the response script corresponding to the question to be answered based on the user intent information and question parameter information; wherein, the target algorithm is the algorithm corresponding to the user intent information and question parameter information.

[0112] Step 13 can be understood as executing function_call to generate scripts, but it is not limited to this.

[0113] The script generation method provided in this application embodiment obtains user intent information by performing intent recognition on the user's question to be answered; the intent recognition includes: identifying a first-level intent category of the question to be answered, and identifying a second-level intent category under the first-level intent category; extracting parameters from the question to be answered to obtain question parameter information; and using a target algorithm, generating a response script corresponding to the question to be answered based on the user intent information and the question parameter information; wherein, the target algorithm is an algorithm corresponding to the user intent information and the question parameter information; it can support accurate identification of user intent and call the target algorithm to generate response script accordingly, which can avoid excessive reliance on text vectors, improve the accuracy of generated response scripts, and effectively solve the problem of low response accuracy in existing script generation schemes.

[0114] Furthermore, the script generation method further includes: obtaining the user's question information; if the question information belongs to a preset script, obtaining a preset reply corresponding to the question information as the reply script corresponding to the question information; if the question information does not belong to the preset script, treating the question information as the question to be replied to. This allows for providing accurate fixed replies directly for questions with fixed scripts, avoiding the need to execute the same reply script for all questions. Figure 1 The proposed solutions aim to save resources. The preset dialogue may include: predefined fixed dialogue, such as querying certain data or performing certain operations (e.g., market analysis, etc.); and / or, preset responses may include: predefined fixed responses, such as directly calling corresponding fixed data as a response to the preset dialogue (e.g., directly calling market analysis data, market analysis data, or fixed dialogue to provide a response), but are not limited to these.

[0115] In this embodiment of the application, the script generation method further includes: setting auxiliary parameter items for intent recognition; the auxiliary parameter items are used to assist in intent recognition; obtaining parameter information of the auxiliary parameter items in the question to be answered; wherein, the step of performing intent recognition on the user's question to be answered to obtain user intent information includes: performing intent recognition on the user's question to be answered based on the obtained parameter information of the auxiliary parameter items to obtain user intent information. This can improve the accuracy of intent recognition. Auxiliary parameter items (e.g., parameters for elderly care scenarios: retirement, advanced age, mobility impairment, etc.) can be preset based on experience information, but are not limited to this.

[0116] Furthermore, the script generation method further includes: setting a parameter extraction mapping relationship; the parameter extraction mapping relationship includes: a character extraction mapping relationship under a first extraction type; wherein, extracting parameters from the question to be answered to obtain question parameter information includes: extracting parameters from the question to be answered to obtain initial parameter information under the first extraction type; mapping the initial parameter information into target character form information based on the character extraction mapping relationship, as extracted information; and obtaining the question parameter information corresponding to the question to be answered based on the extracted information. This can support adding restrictive conditions for parameter extraction to achieve accurate parameter extraction. The initial parameter information may include parameter information in its original form from the question to be answered (e.g., low risk), and / or, characters may include numbers, letters, etc., and / or, regarding the character extraction mapping relationship, for example, risk level: {enum:[R1: low risk, R2: medium risk, R3: high risk], but is not limited to this.

[0117] In this embodiment, the script generation method further includes: performing a first data processing on the question parameter information to obtain processed question parameter information; wherein, the first data processing includes at least one of the following: numerical conversion (e.g., converting "zero" to "0"), numerical scaling (e.g., scaling by a ratio of 0.1 to 10), numerical verification (e.g., verifying whether it belongs to the numerical type - floating-point), numerical mapping (e.g., mapping a low level to R1), and numerical filtering (e.g., filtering certain special values, such as None values); wherein, generating the response script corresponding to the question to be answered based on the user intent information and the question parameter information using the target algorithm includes: generating the response script corresponding to the question to be answered based on the user intent information and the processed question parameter information using the target algorithm. This is intended to improve the success rate of the target algorithm (function call). This solution may also include similar processing to the first data processing described above on the user intent information, but is not limited thereto.

[0118] The step of identifying the user's intent regarding the pending question to obtain user intent information includes: identifying the user's intent regarding the pending question to obtain preliminary intent information; and applying intent rules to correct the preliminary intent information to obtain the final user intent information. This can improve the accuracy of intent identification.

[0119] Furthermore, the script generation method further includes: performing text segmentation on the response script to obtain at least one script fragment; acquiring source document fragments that satisfy a first similarity condition with each of the script fragments (e.g., acquiring source document fragments with a first similarity greater than a preset threshold with the script fragments), to obtain a fragment set; acquiring a second similarity between each of the script fragments and each source document fragment in the fragment set; determining the target source document fragment corresponding to each of the script fragments based on the second similarity; establishing a correspondence between each of the script fragments and the corresponding target source document fragments, and labeling them. This can support the implementation of a precise research report viewpoint tracing scheme, that is, based on a given answer, it can automatically locate and trace the source of research report viewpoints and automatically implement serial number labeling, solving the problem of insufficient support or supporting theoretical basis for research report viewpoints. The source document can include, but is not limited to, documents retrieved for the script fragments.

[0120] Wherein, the at least one speech segment includes a first speech segment, and the segment set includes a first source document segment; wherein, obtaining the second similarity between each of the speech segments and each source document segment in the segment set includes: slicing the first speech segment based on the sliding window length to obtain at least one speech segment; slicing the first source document segment based on the sliding window length to obtain at least one document segment; obtaining a third similarity between each of the speech segments and each of the document segments; and obtaining the second similarity between the first speech segment and the first source document segment based on the third similarity. This can specifically achieve obtaining the second similarity. Wherein, obtaining the third similarity between each of the speech segments and each of the document segments may include: obtaining the ratio of the number of intersection characters to the number of union characters of the speech segment and the document segment to obtain the third similarity; it can be implemented as a Jaccard (similarity) score, but is not limited to this. Wherein, the sliding window length can be set to be less than or equal to the length of a document segment, but is not limited to this.

[0121] Furthermore, the script generation method further includes: obtaining similarity weight coefficients between each script segment and each document segment; wherein, obtaining a third similarity between each script segment and each document segment includes: obtaining an initial similarity between each script segment and each document segment; and obtaining a third similarity between each script segment and each document segment based on the initial similarity and the corresponding similarity weight coefficients. This allows for the specific implementation of obtaining the third similarity.

[0122] The following is an example of the script generation method provided in the embodiments of this application, with the financial management scenario as an example.

[0123] To address the aforementioned technical problems, this application provides a script generation method, specifically a large-scale financial assistant technical solution, which supports:

[0124] 1) Precise Q&A Assistant: Precise Q&A is achieved through layered intent recognition, parameter extraction, and script generation;

[0125] 2) Accurate source tracing of research reports: The source tracing algorithm provided in the embodiments of this application enables accurate source tracing of fragments.

[0126] This solution can also be understood as involving two main parts: precise question and answer for assistants and precise source tracing for research reports. Precise question and answer for assistants mainly involves intent recognition, parameter extraction, function call, and script generation. Precise source tracing for research reports mainly involves achieving precise source tracing of fragments through the source tracing algorithm provided in the embodiments of this application.

[0127] Specifically, such as Figure 2 As shown, the solutions provided in this application embodiment may include the following:

[0128] 1. Script Templates: For user questions (which can correspond to the user's question information above), determine whether to provide a fixed response. This allows for precise responses by directly calling upon data from market analysis, market analysis, or fixed scripts for specific scenarios (such as market analysis, market research, and script polishing).

[0129] Specifically, it can be determined whether the user's question is a guiding script (i.e., whether it is a fixed script or a preset script); if yes (i.e., N), a fixed reply (i.e., a preset reply) is provided; if no (i.e., N), the process proceeds to the next part for a precise reply; this corresponds to obtaining the user's question information as described above; if the question information belongs to a preset script, the preset reply corresponding to the question information is obtained and used as the reply script corresponding to the question information; if the question information does not belong to the preset script, the question information is used as the question to be replied to.

[0130] II. Precise Response, including: intent recognition, parameter extraction, function call, and dialogue generation. Intent recognition and parameter extraction can be implemented based on LLM (Large Language Model). The intent mapping in the diagram may include determining the type of function corresponding to the recognized intent, and the parameter mapping may include parameter information processing (such as numerical conversion).

[0131] 1. Intent Recognition: The intent recognition in this embodiment differs significantly from traditional intent recognition. This embodiment employs hierarchical intent recognition and parameter assistance. Traditional intent recognition only segments the text content or relies solely on a large model for intent recognition, which is difficult to distinguish when there are many types of intents or when intents are similar. Specific examples are as follows:

[0132] 1) Traditional intent recognition: Identifying "What assets are better to allocate after retirement?" as retirement asset allocation;

[0133] 2) Layered intent recognition and parameter assistance:

[0134] a) The intention to "What assets are better to allocate after retirement?" was categorized into two levels: [Asset allocation, non-asset allocation] -> [Retirement asset allocation, mutual fund asset allocation, fund allocation]; specifically:

[0135] The first layer of intent is identified and determined based on the overall model to be: asset allocation;

[0136] The second layer of intent was identified based on the larger model as: retirement asset allocation.

[0137] The hierarchical intent recognition here can correspond to the intent recognition of the user's question to be answered, as described above, to obtain user intent information; the intent recognition includes: recognizing the first-level intent category of the question to be answered, and recognizing the second-level intent category under the first-level intent category.

[0138] b) Parameter Assistance: Auxiliary parameters such as retirement, advanced age, and mobility impairment can be prepared in advance (corresponding to the auxiliary parameters for intent recognition mentioned above; these auxiliary parameters are used to assist in intent recognition). During the first and second layers of intent recognition, these parameters can serve as auxiliary parameters for intent recognition. If the user's question contains relevant descriptions of the auxiliary parameters (corresponding to obtaining the parameter information of the auxiliary parameters in the question to be answered), it helps the model to perform accurate intent recognition (corresponding to the aforementioned intent recognition of the user's question to be answered, obtaining user intent information, including: based on the obtained parameter information of the auxiliary parameters, performing intent recognition of the user's question to be answered, obtaining user intent information). Parameter assistance can be implemented through manual annotation and / or predefinition. For example, in asset allocation for retirement assets, terms like "retirement" can be written as auxiliary parameters in the prompt words. Other cases are similar, but not limited to this.

[0139] This solution utilizes intent recognition to address the difficulty of distinguishing between pension funds and mutual funds in asset allocation, a problem inherent in traditional intent recognition methods. For example, the difference between mutual funds and pension funds is very small, primarily depending on age restrictions and retirement intentions. This highly ambiguous intent requires a layered intent recognition approach, as described in this solution, to achieve better results. Specifically, for example… Figure 3 The example shown illustrates how intent recognition can be achieved based on prompt words. These prompt words can be based on the tool name (tool_name) and format (format) to implement related functions, such as... Figure 3 The prompt word structure shown:

[0140] 1) role:

[0141] Your task is: You are a senior financial advisor. Please determine whether the user has the following intentions and call the appropriate tools.

[0142] 2) tool_name:

[0143] {

[0144] tool_name: 'name': 'tool_name', 'description': 'tool desc'

[0145] 'parameters(scope)':{'properties(property)':{'every properties(each property)':{'description:'type(type)':'string(string)'}}

[0146] 'required':['required field'].'type':'object'}

[0147] }

[0148] 3) Format:

[0149] Action:

[0150]

[0151] 2. Parameter Extraction: This solution employs a type and numerical-assisted extraction process. Compared to traditional parameter extraction methods, which are prone to errors such as incorrect extraction format and type, this solution achieves accurate parameter extraction by adding restrictive conditions (such as parameter extraction mapping relationships). It can be used to extract parameters for the aforementioned questions to be answered, obtaining the question parameter information. A specific example is provided below:

[0152] 1) Regarding traditional extraction: For (I want assets with lower risk, higher target volatility, and an annualized return of 2%, the parameters [low risk, high target volatility, and 2% annualized return] can be extracted;

[0153] 2) The precise parameter extraction in this solution, for assets with lower risk, higher target volatility, and an annualized return of 2%, can be performed using the following process:

[0154] The parameters are defined as follows: Risk Level: {enum:[R1: Low Risk, R2: Medium Risk, R3: High Risk], Data Type: string, Data Description: Numerical range R1~R3, other ambiguous ranges need to be mapped, lower, higher, and general are all mapped to fixed ranges R1, R3, and R2}, Volatility: {Number Range:[0-100], Data Type: int, Data Description: 0-100 integer, high, medium, and low are equally divided: [0-33, 34-67, 68-100]}, Annualized Return: {Number Range:[0-100], Data Type: int, Data Description: 0-100 integer, high, medium, and low are equally divided: [0-33, 34-67, 68-100]}}; Based on this, the extracted parameters are [R1, 68 (example), 20], which corresponds to the parameter information in the above question. The above process corresponds to the parameter extraction mapping relationship set above; the parameter extraction mapping relationship includes: character extraction mapping relationship under the first extraction type (e.g., lower, higher, and generally using fixed range mapping to R1, R3, and R2); wherein, the step of extracting parameters from the question to be answered to obtain question parameter information includes: extracting parameters from the question to be answered to obtain initial parameter information under the first extraction type (e.g., low risk); based on the character extraction mapping relationship, mapping the initial parameter information to information in the form of target characters (e.g., R1) as extraction information; and based on the extraction information, obtaining the question parameter information corresponding to the question to be answered.

[0155] 3. function_call (corresponding to) Figure 2(Interface calls in the code): function_call can include binding functions (i.e., determining the target algorithm) based on accurate intent recognition and parameter extraction. Compared with traditional methods, this solution involves not only accuracy but also a large amount of numerical processing. For example, the traditional approach is to directly call large models and document slices or search engines, and then use the large model's summarization capabilities to generate the script. However, the accuracy of function_call calls in this application is far more difficult than the traditional approach because this part calls an interface. The fault tolerance of interfaces is much lower than that of document retrieval and search engines. For example, the interface name, parameter name / value, parameter type, number of parameters, and parameter range must all meet the requirements; otherwise, the interface call will fail directly. To address this, this solution can perform numerical processing and validation on all parameters, specifically, at least one of the following:

[0156] 1) Numerical conversion: (zero, one, two, three, four, five, six, seven, eight, nine, ten) is converted to (0, 1, 2, 3, 4, 5, 6, 7, 8, 9, ten);

[0157] 2) Numerical scaling: (original value, scale = [1, 100]) is converted to (original value × scale (ratio)). For certain ratio parameters (such as percentage, proportion, rate, etc.), the following settings can be made: if the value is between 1 and 10, then scale = 10; if the value is between 0 and 1, then scale = 100.

[0158] 3) Numerical validation: (original numerical value, type = [float(floating point), int, string]);

[0159] a) If the original numerical type is one of the above types, a forced conversion can be performed; for example, a low-risk conversion to R1;

[0160] b) If the original numerical type does not belong to the above types, the reason can be recorded for later verification; and a message indicating that the script generation failed should be displayed, and a log should be generated for later improvement.

[0161] 4) Numerical Mapping: Similar to 3), if the numerical type is string, these parameters are mostly mapping parameters and can be mapped according to rules. For example, risk level [high, medium, low] -> [R1, R2, R3], bonus product [yes, no] -> [0, 1], etc. This part can perform a second verification on the results of the previous parameter extraction to ensure that the parameter type and numerical range are not incorrect; it can be understood as performing mapping again to avoid the previous mapping still having problems.

[0162] 5) Numerical filtering: This function can filter input parameters. For example, after the previous series of processing, the parameters are correct in type and format, but some may still be incorrect (such as other issues that do not meet the above requirements). These parts can be assigned the None value. In the numerical filtering part, the None value can be filtered out, and the remaining parameters are all valid.

[0163] The above intent recognition, parameter extraction, and function call can be processed using, for example, the following format:

[0164] 1) data:

[0165] intent (intent): str (characters),

[0166] args(parameters): list(list);

[0167] 2) Interface:

[0168] llm (Large Language Model) _api (Application Programming Interface).

[0169] In addition, the above function_call can include Figure 2 The solution further includes interface mapping and interface calling, and can further include script generation (which can correspond to the above-mentioned use of the target algorithm to generate the response script corresponding to the question to be answered based on the user intent information and question parameter information; wherein the target algorithm is the algorithm corresponding to the user intent information and question parameter information), and then responding; wherein, interface mapping includes determining a specific function based on intent mapping and parameter mapping; interface calling includes calling the determined function, and thus scripts can be generated based on intent mapping and parameter mapping, but is not limited thereto.

[0170] The above-mentioned numerical processing and verification can be correlated with the first data processing performed on the question parameter information to obtain the processed question parameter information; wherein, the first data processing includes at least one of: numerical conversion, numerical scaling, numerical verification, numerical mapping, and numerical filtering; wherein, the step of generating the reply script corresponding to the question to be replied to based on the user intent information and the question parameter information using the target algorithm includes: generating the reply script corresponding to the question to be replied to based on the user intent information and the processed question parameter information using the target algorithm.

[0171] 4. Intent rewriting can be performed before function_call: Intent rewriting is a supplement to accurate intent recognition. Although layered intent recognition has been performed beforehand, it is still difficult to avoid a small number of intent recognition omissions or errors. In order to solve this problem, intent rewriting can be performed in this solution, for example, based on the following 5 rules to determine the relevant intent.

[0172] 1) Rule 1: Able to match Figure 4 The problem with the pattern shown is related to financial political intentions;

[0173] 2) Rule Two: Able to match Figure 5 The issue presented in the pattern pertains to the intent of an industry sector.

[0174] 3) Rule 3: Able to match Figure 6 The problem with the pattern shown is related to the intention of a retirement asset portfolio;

[0175] 4) Rule 4: Able to match Figure 7 The problem with the pattern shown pertains to portfolio intent;

[0176] 5) Rule Five: Able to match Figure 8 The issue shown in the pattern pertains to fund screening intentions.

[0177] in, Figures 4 to 8 The content within the boxes in the rules shown is written using regular expressions, but other methods can also be used, which are not limited here.

[0178] The above-mentioned intent rewriting can correspond to the intent recognition of the user's pending question to obtain user intent information, including: performing intent recognition on the user's pending question to obtain preliminary intent information (such as the information obtained by precise intent recognition as described above); and using intent rules (such as rule one as described above) to correct the preliminary intent information (which may also involve supplementary intent rewriting operations, etc., which are not limited here) to obtain user intent information.

[0179] III. Accurate Research Report Source Tracing;

[0180] This scheme can further include precise research report tracing to better achieve accurate responses. This part can provide the theoretical support for the research report's viewpoints in Part Two above. The research report tracing scheme in this application differs from traditional tracing methods; it does not rely on vector similarity and text block comparison, but instead uses a sliding window and Jaccard (similarity) scoring. The following steps (abc) are sequential: sentence segmentation, document recall, and precise matching based on the recall results.

[0181] 1. Sentence Segmentation: Before performing precise source tracing of dialogue, sentences can be segmented first, followed by document retrieval; examples of segmentation are as follows:

[0182] 1) Assume text = (Recently, investment and wealth management topics have become trending topics. Investment in large-scale modeling is extremely popular, a consensus within the investment community.)

[0183] 2) Sentence segmentation is mainly based on sentence-ending punctuation, including [.。 ? ? ! ! ], etc. This part can be defined according to the actual situation. After the above segmentation, the sentence text will be divided into two parts: [(Recently, the topic of investment and financial management has become a hot topic.), (Investment in the field of large models is very popular and has become a consensus in the investment field.)].

[0184] The above sentence segmentation corresponds to the text segmentation of the aforementioned reply script, resulting in at least one script fragment (such as the above: "Recently, the topic of investment and financial management has become a trending topic.").

[0185] 2. Document Retrieval: The document retrieval part can adopt the current method, recalling document fragments based on sentence similarity, and then proceeding to precise source tracing; regarding document retrieval, let's assume the recalled document fragments are as follows:

[0186] 1) `cite_list` = [(Large models have become investment hotspots; models A and B have attracted a large amount of investment, and the future value brought by this new technology track is immeasurable.), (Is there a consensus in the investment field? In the context of a weak economy, many investors do not know what to invest in.)]; This corresponds to the source document fragments that meet the first similarity condition with each of the above-mentioned text fragments (such as the above (Large models have become investment hotspots; models A and B have attracted a large amount of investment, and the future value brought by this new technology track is immeasurable.)), resulting in a fragment set (such as the above `cite_list`).

[0187] 3. Precise Source Tracing: Precise source tracing can be implemented using the algorithm provided in the embodiments of this application. In addition to source tracing, this part can also perform sequence number annotation, realizing the mapping between sequence numbers and document fragments. Specifically, this part may include the following:

[0188] 1) Traverse text and cite_list, and use the Jaccard sliding window to compare each passage in text with each passage in cite_list for similarity in turn. For each passage, select the document fragment in cite_list with the highest similarity (corresponding to the second similarity mentioned above) (corresponding to the target traceability document fragment mentioned above), which can be recorded as <cite_id (identifier), cite_content (content), score (score)> to represent the serial number, traceability content, and similarity score respectively. This part of the content can correspond to obtaining the second similarity between each of the above-mentioned speech segments and each traceability document segment in the segment set; based on the second similarity, determine the target traceability document segment corresponding to each of the above-mentioned speech segments.

[0189] Among them, the algorithm of the sliding window involved above can be understood as including: implementing the comparison of each passage through the sliding window algorithm to find the document fragment with the highest similarity; specific examples are as follows. Suppose the similarity between fragment a (original speech) and fragment b is obtained, and the sliding window algorithm can include the following operations:

[0190] a) If fragment a (original speech) or fragment b (traceability text) is empty, return an exception;

[0191] b) If the length of fragment a or fragment b is 0, return an exception;

[0192] c) Perform a sliding window on fragment b. If the window length w is greater than or equal to the length of fragment b, then set the window length w = b;

[0193] d) For fragment a and fragment b, set the step size to w, split fragment a, and the splitting range is 0 - w, and split fragment b, and the splitting range is 0 - w; that is, slice fragment a and fragment b according to the step size w to obtain the slices of fragment a and fragment b;

[0194] e) Calculate the similarity between slice a_cut1 of segment a and slice b_cut1 of segment b. Specifically, this can be achieved by calculating the ratio of the number of characters in the intersection of a_cut1 and b_cut1 to the number of characters in their union (corresponding to the third similarity mentioned above), but is not limited to this method. In this scheme, the obtained slice similarity can be weighted and scored. For example, for a_cut1 and b_cut1, 0-2 characters are first scored (i.e., the initial similarity is obtained), and then 0-3 characters are scored. The scores are then placed into a results list (let's say [0.8, 1]). Finally, a weighted score is applied (let's say the weights are [0.6, 0.4]). The score then becomes [0.8 × 0.6, 0.4 × 1] = [0.48, 0.4]. The overall score is the ratio of the sum of scores to the length of the list (i.e., the ratio of the sum to the number of results in the results list, which corresponds to the third similarity mentioned above), which is (0.48 + 0.4) / 2 = 0.44. Therefore, the final similarity between slice a_cut1 and b_cut1 is 0.44. This part corresponds to the above-mentioned acquisition of the similarity weight coefficient between each of the utterance segments and each of the document segments (this coefficient can be preset, but is not limited to this); wherein, the acquisition of the third similarity between each of the utterance segments and each of the document segments includes: acquiring the initial similarity between each of the utterance segments and each of the document segments; based on the initial similarity and the corresponding similarity weight coefficient (such as the weight [0.6, 0.4] above), obtaining the third similarity between each of the utterance segments and each of the document segments (such as 0.44 above).

[0195] f) Return the combined score of segments a_cut1 and b_cut1, and calculate other scores in sequence (for example, it may include calculating the combined score of each segment other than b_cut1 of segment b) to obtain the final score (i.e. similarity) of segment a and segment b; it may be to sum the combined scores of each segment to obtain the similarity between segment a and segment b, but it is not limited to this.

[0196] The above sliding window algorithm can correspond to the at least one speech segment mentioned above, including a first speech segment, and the segment set including a first source document segment; wherein, obtaining the second similarity between each speech segment and each source document segment in the segment set includes: slicing the first speech segment based on the sliding window length to obtain at least one speech segment; slicing the first source document segment based on the sliding window length to obtain at least one document segment; obtaining a third similarity between each speech segment and each document segment; and obtaining the second similarity between the first speech segment and the first source document segment based on the third similarity.

[0197] 2) Sequence Numbering: After completing the sliding window tracing process described above, sequence numbering can be applied to segment a. Based on the one-to-one correspondence between the dialogue and the tracing content already achieved during the tracing process (e.g., segment a corresponds to segment b), the correspondence between segment a and segment b can be directly established, for example, labeled as (a [1] [1]b), that is, the source text of fragment a is b, and the label number is 1. The following can be deduced in the same way, but not limited to this. This part can correspond to the above-mentioned establishment of the correspondence between each of the above-mentioned speech fragments and the corresponding target source document fragments, and be labeled.

[0198] Therefore, the solution provided in this application embodiment can support the realization of accurate question and answer by assistants and the tracing of research reports.

[0199] This application also provides a script generation device, such as... Figure 9 As shown, it includes:

[0200] The first identification module 91 is used to identify the user's question to be answered and obtain the user's intent information; the intent identification includes: identifying the first-level intent category of the question to be answered, and identifying the second-level intent category under the first-level intent category;

[0201] The first extraction module 92 is used to extract parameters from the question to be answered, and obtain question parameter information;

[0202] The first generation module 93 is used to generate a response script corresponding to the question to be answered based on the user intent information and question parameter information using a target algorithm; wherein, the target algorithm is an algorithm corresponding to the user intent information and question parameter information.

[0203] The script generation device provided in this application embodiment obtains user intent information by performing intent recognition on the user's question to be answered; the intent recognition includes: identifying a first-level intent category of the question to be answered, and identifying a second-level intent category under the first-level intent category; extracting parameters from the question to be answered to obtain question parameter information; and using a target algorithm, generating a response script corresponding to the question to be answered based on the user intent information and the question parameter information; wherein, the target algorithm is an algorithm corresponding to the user intent information and the question parameter information; it can support accurate recognition of user intent and call the target algorithm to generate response script accordingly, which can avoid excessive reliance on text vectors, improve the accuracy of generated response scripts, and effectively solve the problem of low response accuracy in existing script generation schemes.

[0204] Furthermore, the script generation device further includes: a first acquisition module for acquiring the user's question information; a second acquisition module for acquiring a preset reply corresponding to the question information as the reply script corresponding to the question information when the question information belongs to a preset script; and a first processing module for treating the question information as the question to be replied to when the question information does not belong to the preset script.

[0205] In this embodiment of the application, the script generation device further includes: a first setting module, used to set auxiliary parameter items for intent recognition; the auxiliary parameter items are used to assist in intent recognition; a third acquisition module, used to acquire parameter information of the auxiliary parameter items in the question to be answered; wherein, the step of performing intent recognition on the user's question to be answered to obtain user intent information includes: performing intent recognition on the user's question to be answered based on the acquired parameter information of the auxiliary parameter items to obtain user intent information.

[0206] Furthermore, the script generation device further includes: a second setting module, used to set a parameter extraction mapping relationship; the parameter extraction mapping relationship includes: a character extraction mapping relationship under a first extraction type; wherein, the step of extracting parameters from the question to be answered to obtain question parameter information includes: extracting parameters from the question to be answered to obtain initial parameter information under the first extraction type; based on the character extraction mapping relationship, mapping the initial parameter information into information in the form of target characters as extraction information; and based on the extraction information, obtaining the question parameter information corresponding to the question to be answered.

[0207] In this embodiment of the application, the script generation device further includes: a second processing module, used to perform first data processing on the question parameter information to obtain processed question parameter information; wherein, the first data processing includes at least one of: numerical conversion, numerical scaling, numerical verification, numerical mapping, and numerical filtering; wherein, the step of generating the reply script corresponding to the question to be replied to based on the user intent information and the question parameter information using the target algorithm includes: generating the reply script corresponding to the question to be replied to based on the user intent information and the processed question parameter information using the target algorithm.

[0208] The step of identifying the user's unanswered question to obtain user intent information includes: identifying the user's unanswered question to obtain preliminary intent information; and using intent rules to correct the preliminary intent information to obtain user intent information.

[0209] Furthermore, the script generation device further includes: a first segmentation module, used to segment the text of the response script to obtain at least one script fragment; a fourth acquisition module, used to acquire source document fragments that satisfy a first similarity condition with each of the script fragments, to obtain a fragment set; a fifth acquisition module, used to acquire a second similarity between each of the script fragments and each source document fragment in the fragment set; a first determination module, used to determine the target source document fragment corresponding to each of the script fragments based on the second similarity; and a third processing module, used to establish a correspondence between each of the script fragments and the corresponding target source document fragments, and to annotate them.

[0210] Wherein, the at least one speech segment includes a first speech segment, and the segment set includes a first source document segment; wherein, obtaining the second similarity between each of the speech segments and each source document segment in the segment set includes: slicing the first speech segment based on the sliding window length to obtain at least one speech segment; slicing the first source document segment based on the sliding window length to obtain at least one document segment; obtaining a third similarity between each of the speech segments and each of the document segments; and obtaining the second similarity between the first speech segment and the first source document segment based on the third similarity.

[0211] Furthermore, the script generation device further includes: a sixth acquisition module, used to acquire similarity weight coefficients between each script slice and each document slice; wherein, acquiring the third similarity between each script slice and each document slice includes: acquiring the initial similarity between each script slice and each document slice; and obtaining the third similarity between each script slice and each document slice based on the initial similarity and the corresponding similarity weight coefficients.

[0212] The implementation embodiments of the above-mentioned script generation method are all applicable to the embodiments of the script generation device, and can achieve the same technical effect.

[0213] This application also provides a script generation device, including a memory, a processor, and a program stored in the memory and executable on the processor; when the processor executes the program, it implements the above-described script generation method.

[0214] The implementation embodiments of the above-mentioned script generation method are all applicable to the embodiments of the script generation device and can achieve the same technical effect.

[0215] This application also provides a readable storage medium storing a program that, when executed by a processor, implements the steps in the above-described speech generation method.

[0216] The implementation embodiments of the above-described script generation method are all applicable to the embodiments of the readable storage medium and can achieve the same technical effect.

[0217] This application also provides a computer program product, including computer instructions. When the computer instructions are executed by a processor, they implement the various processes of the above-described speech generation method and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0218] It should be noted that many of the functional components described in this specification are referred to as modules in order to more specifically emphasize the independence of their implementation.

[0219] In this embodiment, the module can be implemented in software so that it can be executed by various types of processors. For example, an identified executable code module may include one or more physical or logical blocks of computer instructions, which may be constructed as objects, procedures, or functions. Nevertheless, the executable code of the identified module does not need to be physically located together, but may include different instructions stored in different bits, which, when logically combined, constitute the module and achieve the module's intended purpose.

[0220] In practice, an executable code module can be a single instruction or many instructions, and can even be distributed across multiple different code segments, different programs, and across multiple memory devices. Similarly, operational data can be identified within the module and can be implemented in any suitable form and organized within any suitable type of data structure. This operational data can be collected as a single dataset or distributed across different locations (including different storage devices), and can exist, at least in part, solely as electronic signals within the system or network.

[0221] When a module can be implemented using software, considering the current level of hardware technology, modules that can be implemented in software can be implemented using hardware circuits by those skilled in the art to achieve the corresponding functions, without considering cost. These hardware circuits include conventional very-large-scale integrated circuits (VLSI) or gate arrays, as well as existing semiconductors such as logic chips and transistors, or other discrete components. Modules can also be implemented using programmable hardware devices, such as field-programmable gate arrays, programmable array logic, and programmable logic devices.

[0222] The above describes the preferred embodiments of this application. It should be noted that those skilled in the art can make several improvements and modifications without departing from the principles described in this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for generating dialogue scripts, characterized in that, include: Perform intent recognition on the user's pending questions to obtain user intent information; The intent recognition includes: identifying a first-level intent category of the question to be answered, and identifying a second-level intent category under the first-level intent category; The parameters of the question to be answered are extracted to obtain the question parameter information; Using a target algorithm, a response script corresponding to the question to be answered is generated based on the user intent information and question parameter information; wherein, the target algorithm is an algorithm corresponding to the user intent information and question parameter information.

2. The script generation method according to claim 1, characterized in that, Also includes: Configure auxiliary parameters for intent recognition; The auxiliary parameter item is used to assist in intent recognition; Obtain the parameter information of the auxiliary parameter items in the question to be answered; The process of identifying the user's unanswered question to obtain user intent information includes: Based on the parameter information of the acquired auxiliary parameter items, intent recognition is performed on the user's question to be answered, and user intent information is obtained.

3. The script generation method according to claim 1, characterized in that, Also includes: Set parameters to extract mapping relationships; The parameter extraction mapping relationship includes: the character extraction mapping relationship under the first extraction type; The step of extracting parameters from the question to be answered to obtain question parameter information includes: Parameters are extracted from the question to be answered to obtain initial parameter information under the first extraction type; Based on the character extraction mapping relationship, the initial parameter information is mapped to information in the form of target characters, which is then used as the extracted information; Based on the extracted information, the question parameter information corresponding to the question to be answered is obtained.

4. The script generation method according to claim 1, characterized in that, The process of identifying the user's pending question to obtain user intent information includes: The intent of the user's pending questions is identified to obtain preliminary intent information; The preliminary intent information is corrected using intent rules to obtain user intent information.

5. The script generation method according to claim 1, characterized in that, Also includes: The response script is segmented to obtain at least one script fragment; Obtain source document fragments that satisfy the first similarity condition with each of the aforementioned speech fragments to obtain a fragment set; Obtain the second similarity between each of the stated speech fragments and each source document fragment in the fragment set; Based on the second similarity, the target source document fragment corresponding to each of the above-mentioned speech fragments is determined; Establish the correspondence between each of the aforementioned script segments and the corresponding target source document segments, and then annotate them.

6. The script generation method according to claim 5, characterized in that, The at least one speech segment includes a first speech segment, and the set of segments includes a first source document segment; The step of obtaining the second similarity between each of the stated speech segments and each source document segment in the segment set includes: Based on the length of the sliding window, the first speech segment is sliced ​​to obtain at least one speech slice. Based on the sliding window length, the first source document segment is sliced ​​to obtain at least one document slice; Obtain the third similarity between each of the stated speech slices and each of the stated document slices; Based on the third similarity, a second similarity is obtained between the first speech fragment and the first source document fragment.

7. A script generation device, characterized in that, include: The first identification module is used to identify the user's intent for the question to be answered, and to obtain the user's intent information. The intent recognition includes: identifying a first-level intent category of the question to be answered, and identifying a second-level intent category under the first-level intent category; The first extraction module is used to extract parameters from the question to be answered, and obtain question parameter information; The first generation module is used to generate a response script corresponding to the question to be answered based on the user intent information and question parameter information using a target algorithm; wherein, the target algorithm is an algorithm corresponding to the user intent information and question parameter information.

8. A script generation device, comprising a memory, a processor, and a program stored in the memory and executable on the processor; characterized in that, When the processor executes the program, it implements the script generation method as described in any one of claims 1 to 6.

9. A readable storage medium having a program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the script generation method as described in any one of claims 1 to 6.

10. A computer program product, characterized in that, It includes computer instructions, which, when executed by a processor, implement the steps of the script generation method as described in any one of claims 1 to 6.

Images