CN117493530A

CN117493530A - Resource demand analysis method, device, electronic equipment and storage medium

Info

Publication number: CN117493530A
Application number: CN202311824930.4A
Authority: CN
Inventors: 刘陆阳; 林群阳; 张闯; 王敏
Original assignee: Suzhou Metabrain Intelligent Technology Co Ltd
Current assignee: Suzhou Metabrain Intelligent Technology Co Ltd
Priority date: 2023-12-27
Filing date: 2023-12-27
Publication date: 2024-02-02
Anticipated expiration: 2043-12-27
Also published as: CN117493530B

Abstract

The embodiment of the invention provides a resource demand analysis method, a device, electronic equipment and a storage medium, and relates to the technical field of computers. Based on the dialogue management model and the history dialogue data, determining a dialogue state corresponding to the current dialogue; under the condition that the dialogue state characterizes the end of the dialogue, determining initial structural information based on a target model in the demand analysis model and historical session data; verifying the initial structural information based on a verification module in the demand analysis model, and determining user demand information conforming to a standard output format; the user demand information is used for representing demand information of the user on the target resource. The whole resource demand analysis process realizes full automation, saves operation cost to a certain extent and improves analysis efficiency of resource demand analysis for users.

Description

Resource demand analysis method, device, electronic equipment and storage medium

Technical Field

The invention belongs to the technical field of computers, and particularly relates to a resource demand analysis method, a device, electronic equipment and a storage medium.

Background

With the development progress of society, high-efficiency computing has become a key element for supporting the development of an intelligent society. At present, the application of cloud computing and distributed computing mainly concentrates computing power in computing clusters of all computing power providers, and when a user purchases computing power, the user needs to communicate with a plurality of computing power providers and consult with the computing power providers, so that the use cost of the user is increased to a certain extent.

In the related art, the user needs are often determined by form filling in a power transaction platform or based on the inquiry of customer service personnel to the user, so that the operation cost is high and the efficiency is low.

Disclosure of Invention

In order to overcome the problems in the related art, the invention provides a resource demand analysis method, a device, an electronic device and a storage medium.

In a first aspect, the present invention provides a resource demand analysis method applied to a server running a resource processing platform, where the method includes:

obtaining, by a processor in the server, reply data output by a target dialog model based on user question data input by a user, and determining historical session data based on the user question data and the reply data;

Based on the dialogue management model and the history dialogue data, determining a dialogue state corresponding to the current dialogue;

determining initial structural information based on a target model in a demand analysis model and the historical session data under the condition that the dialogue state characterizes the end of the dialogue;

verifying the initial structured information based on a verification module in the demand analysis model, determining user demand information conforming to a standard output format, and storing the user demand information into a memory in the server; the user demand information is used for representing demand information of the user on target resources.

Optionally, the determining, based on the session management model and the historical session data, a session state corresponding to the current session includes:

inputting the historical session data into the dialog management model; the dialogue management model comprises a target language model and a target classification model;

determining a feature vector corresponding to the historical session data based on the target language model and the historical session data;

determining a classification result based on the feature vector and the target classification model;

and determining, by the processor, a dialog state corresponding to the current dialog based on the classification result.

Optionally, the determining, by the processor, a session state corresponding to the current session based on the classification result includes:

the processor determines the dialogue state as the dialogue end under the condition that the classification result is larger than a preset threshold value;

and the processor determines the dialogue state as being in the dialogue under the condition that the classification result is smaller than or equal to the preset threshold value.

Optionally, the method further comprises:

and repeating the operation of acquiring, by the processor in the server, response data output by the target dialog model based on the user question data input by the user and determining historical session data based on the user question data and the response data, in the case where the dialog state is in a dialog.

Optionally, the determining initial structured information based on the target model in the demand analysis model and the historical session data includes:

defining a standard output format corresponding to the demand analysis model based on a preset configuration file;

inputting the historical session data into the demand analysis model;

invoking the target model based on a monitoring module in the demand analysis model, and acquiring initial structural information output by the target model;

The verifying module in the demand analysis model is used for verifying the initial structured information to determine user demand information conforming to a standard output format, and the verifying module comprises:

correcting the structured information which does not accord with the standard output format in the initial structured information based on the verification module until the initial structured information accords with the standard output format;

and acquiring user demand information output by the demand analysis model based on the structured information conforming to the standard output format and the corrected structured information.

Optionally, the defining, based on a preset configuration file, a standard output format corresponding to the requirement analysis model includes:

transmitting the preset configuration file into the monitoring module, and packaging the target model and the preset configuration file;

initializing the monitoring module.

Optionally, the correcting, based on the verification module, the structured information that does not conform to the standard output format in the initial structured information includes:

based on the verification module, verifying whether each piece of structured information in the initial structured information accords with the standard output format;

Generating an error correction instruction in the presence of structured information that does not conform to the standard output format;

inputting the error correction instruction and initial structural information with error labels into the monitoring module;

the error correction instruction is used for indicating the target model to acquire the corrected structured information corresponding to the structured information which does not accord with the standard output format.

Optionally, the obtaining the user requirement information output by the requirement analysis model based on the structured information conforming to the standard output format and the corrected structured information includes:

and under the condition that a preset stop condition is met, acquiring user demand information output by the demand analysis model based on the structured information conforming to the standard output format and the corrected structured information.

Optionally, the preset stopping condition includes that the initial structured information is not in accordance with the standard output format after being corrected, and the initial structured information is in accordance with the standard output format.

Optionally, the obtaining the user requirement information output by the requirement analysis model based on the structural information conforming to the standard output format and the modified structural information includes:

And matching and filling the information templates in the preset configuration file based on the structured information conforming to the standard output format and the corrected structured information to obtain the user demand information.

Optionally, the obtaining, by the processor in the server, reply data output by the target dialogue model based on user question data input by a user, and determining historical session data based on the user question data and the reply data includes:

acquiring user problem data input by a user, and inputting the user problem data into the target dialogue model;

obtaining reply data corresponding to the user problem data output by the target dialogue model;

and correspondingly storing the user question data and the reply data into the memory, and determining the user question data and the reply data corresponding to one dialogue as the historical session data.

Optionally, the target dialogue model is obtained based on a training function provided by the processor, and the target dialogue model is obtained through training in the following way:

acquiring a first sample data set; the first sample data set comprises a plurality of groups of question-answer sentence pairs corresponding to the target field, and the question-answer sentence pairs comprise sample question data;

Taking the sample question data and the target prompt word as the input of a dialogue model to be trained, and obtaining predicted reply data output by the dialogue model to be trained;

and carrying out parameter adjustment on the dialogue model to be trained based on the predicted reply data, and determining the target dialogue model.

Optionally, the question-answer sentence pair further includes sample answer data; the step of carrying out parameter adjustment on the dialogue model to be trained based on the predicted reply data and determining the target dialogue model comprises the following steps:

based on the predicted reply data and the sample reply data, carrying out parameter adjustment on the dialogue model to be trained; the sample reply data is used for representing real reply content corresponding to the sample question data;

and determining the dialogue model to be trained as the target dialogue model under the condition that the first stopping condition is reached.

Optionally, the performing parameter adjustment on the dialogue model to be trained based on the predicted reply data and the sample reply data includes:

determining a probability loss value based on the predicted reply data and the sample reply data;

and carrying out parameter adjustment on the dialogue model to be trained based on the probability loss value.

Optionally, the probability loss value is derived based on a first formula; the first formula includes:

wherein SFT_Loss is used to represent the probability Loss value; d is used to represent the first sample dataset;for representing the size of said first sample dataset,/or->A set of question-answer sentence pairs for representing said first sample data set->For representing said sample question data, +.>For representing said sample reply data, +.>For representing +.>Fields of individual positions>For representing the sequence of the first i fields in the sample reply data.

Optionally, the dialog management model is derived based on training functions provided by the processor, and the dialog management model is trained by:

acquiring a second sample dataset; the second sample data set comprises a plurality of groups of dialogue histories marked with role information, and the dialogue histories comprise a plurality of interactive operations;

aiming at any dialogue history record in the multiple groups of dialogue history records, acquiring a target feature vector based on a target language model;

taking the target feature vector as input of a classification model to be trained, and acquiring a prediction state result output by the classification model to be trained;

Based on the prediction state result and the state label corresponding to the dialogue history record, carrying out parameter adjustment on the classification model to be trained;

and when the second stopping condition is met, determining the classification model to be trained as a target classification model, and determining the target language model and the target classification model as the dialogue management model.

Optionally, the obtaining, for any one of the multiple sets of conversation histories, a target feature vector based on a target language model includes:

splicing the role information, the text content corresponding to each role information and the state label corresponding to the dialogue history record aiming at any dialogue history record in the multiple groups of dialogue history records to obtain a spliced text;

and inputting the spliced text into a target language model, and acquiring a target feature vector output by the target language model.

In a second aspect, the present invention provides a resource demand analysis apparatus for use with a server running a resource processing platform, the apparatus comprising:

a first acquisition module for acquiring, by a processor in the server, reply data output by a target dialogue model based on user question data input by a user, and determining historical session data based on the user question data and the reply data;

The first determining module is used for determining a conversation state corresponding to the current conversation based on the conversation management model and the historical conversation data;

the second determining module is used for determining initial structural information based on a target model in the demand analysis model and the historical session data under the condition that the dialogue state represents the end of the dialogue;

the third determining module is used for verifying the initial structural information based on the verification module in the demand analysis model, determining user demand information conforming to a standard output format and storing the user demand information into a memory in the server; the user demand information is used for representing demand information of the user on target resources.

In a third aspect, the present invention provides an electronic device comprising: a processor, a memory and a computer program stored on the memory and executable on the processor, characterized in that the processor implements the resource demand analysis method of any of the above first aspects when executing the program.

In a fourth aspect, the present invention provides a readable storage medium, which when executed by a processor of an electronic device, enables the electronic device to perform the steps of the resource demand analysis method as in any of the embodiments of the first aspect described above.

In the embodiment of the invention, the corresponding reply content can be automatically determined according to the user questions by acquiring the reply data output by the target dialogue model based on the user question data input by the user, so that customer service personnel do not need to reply the user question data manually, the automation in the question reply process is improved, and the operation cost is saved to a certain extent. Further, based on the dialogue management model and the historical session data, determining a dialogue state corresponding to the current dialogue, and under the condition that the dialogue state characterizes the dialogue to be ended, further determining user demand information based on the demand analysis model and the historical session data, further analyzing the historical session data through a target model in the demand analysis model under the condition that the dialogue is ended, and checking based on a checking module in the demand analysis model, so that the demand information of a user meeting a standard output format on target resources can be obtained, and the actual resource demand of the user can be obtained. In this way, through the intelligent judgment of the dialogue state of the dialogue management model, the user demand information is automatically determined based on the demand analysis model, and compared with the manual resource demand analysis based on customer service personnel, the embodiment of the invention automatically performs the resource demand analysis through different models, improves the specialty of the resource demand analysis, simultaneously, realizes full automation of the whole resource demand analysis process, saves the operation cost to a certain extent and improves the analysis efficiency of the resource demand analysis to the user.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions of the prior art, the following description will briefly explain the drawings used in the embodiments or the description of the prior art, and it is obvious that the drawings in the following description are some embodiments of the present invention, and other drawings can be obtained according to these drawings without inventive effort for a person skilled in the art.

FIG. 1 is a flow chart of steps of a method for analyzing resource requirements according to an embodiment of the present invention;

FIG. 2 is a flowchart illustrating steps for performing demand analysis based on a demand analysis model according to an embodiment of the present invention;

FIG. 3 is a schematic diagram of a training process of a target dialogue model according to an embodiment of the present invention;

FIG. 4 is a schematic diagram of a training process of a dialog management model according to an embodiment of the present invention;

FIG. 5 is a flowchart illustrating steps of a method for analyzing resource requirements according to an embodiment of the present invention;

FIG. 6 is a block diagram of a resource demand analysis device according to an embodiment of the present invention;

fig. 7 is a block diagram of an electronic device according to an embodiment of the present invention.

Detailed Description

The following description of the embodiments of the present invention will be made clearly and fully with reference to the accompanying drawings, in which it is evident that the embodiments described are some, but not all embodiments of the invention. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

Fig. 1 is a flow chart of steps of a method for analyzing resource requirements, which is provided in an embodiment of the present invention, and is applied to a server running a resource processing platform.

In the embodiment of the invention, the resource processing platform can be a network service platform related to resource transaction, resource allocation and resource retrieval, and can provide functions of automatic dialogue, demand analysis, resource retrieval, resource matching, resource scheduling and the like related to the resource. The resource processing platform may be deployed with a target dialog model, a dialog management model, and a demand analysis model. The resource processing platform runs in a server, and training operations of the model and reasoning operations based on the model (resource demand analysis operations) can be performed by the server, which can be used to operate based on embodiments of the methods described below.

For example, the target session model may be trained by using a method of supervised fine tuning such that the target session model grasps knowledge related to resource traffic. Then judging the dialogue state based on the dialogue content through the trained dialogue management model, and sending an instruction to enable the target dialogue model to continuously acquire the requirements of the user when the dialogue management model judges that the dialogue is not finished, so as to seek the history dialogue data with richer content, deeper level and richer detail through the dialogue with the user; if the dialogue is completed, the historical session data is input into a demand analysis model to obtain the structural representation of the demand information of the user on the target resource, namely the demand information of the user, and the demand information of the user can be input into a resource retrieval module of a resource processing platform as query information to match the required resource for the user.

As shown in fig. 1, the method may include:

step 101, obtaining, by a processor in the server, reply data output by a target dialogue model based on user question data input by a user, and determining historical session data based on the user question data and the reply data.

Alternatively, step 101 may comprise the steps of:

step 1011, obtaining user problem data input by a user, and inputting the user problem data into the target dialogue model.

Step 1012, obtaining reply data corresponding to the user question data output by the target dialogue model.

Step 1013, store the user question data and the reply data to the memory correspondingly, and determine the user question data and the reply data corresponding to a dialogue as the historical session data.

In the embodiment of the invention, the processor in the server acquires the reply data output by the target dialogue model based on the user question data input by the user, specifically, the target dialogue model can be a model for acquiring corresponding reply content based on the question data input by the model, and the target dialogue model can be constructed by mainly relying on the existing trained base model, such as ChatGLM2, llama2-Chinese-chat and the like. The target dialogue model can be used for different scenes in different fields, and correspondingly, the corresponding target dialogue model is required to be obtained based on training data in different scenes in different fields according to different scenes in different fields. By way of example, the target dialog model may be applied in the field of computational force analysis, for example: the target dialogue model may output corresponding reply data according to question data input by the user. The user question data may include a user's computational power requirements, and the response data may provide the user with a corresponding computational power resource analysis based on the user's computational power requirements, and may include, in particular, computational power solutions that satisfy the computational power requirements involved in the user question data. The target dialog model can also be used in other resource analysis fields, such as business resources, human resources, etc.

And taking the user question data and the basic prompt words input by the user as the input of the target dialogue model, and acquiring the reply data output by the target dialogue model. The user problem data and the basic prompt word can be spliced and then input into the target dialogue model, the basic prompt word can be used as a system prompt word for starting dialogue, and the task for indicating the target dialogue model is to dialogue with the user and solve the problem in the user problem data, and simultaneously respond to the requirement proposed by the user. For example: the basic prompt word sample may be: "you are a question and answer robot that can analyze and solve the problem, require careful solution of the problem posed by the user, and cannot contain illegal, offensive, and offensive replies. ". And taking the user question data and the corresponding reply data input by the user as a group of interaction records, wherein the interaction records from the beginning of a dialogue until the end of the dialogue can be taken as the history session data corresponding to the dialogue. The historical session data may include multiple sets of interaction records. It will be appreciated that a session may refer to an intermediate process starting with a user beginning to input user question data to the target session model, until the target session model receives an end instruction sent by the session management model for the present session to characterize the end of the session.

Illustratively, the user issue data may be: "what resource allocation is at least needed to run the BERT-Large model? "corresponding, the reply data corresponding to the user question data may be: "compute instance of GPU requiring more than 4 cores of processor resources, 128GB of system memory, more than 30GB of hard disk space, and at least 40GB of video memory. "

And under the condition that the answer data corresponding to the user question data is obtained, storing the user question data and the corresponding answer data corresponding to each round of interaction in a memory of a server, and determining all the user question data and the answer data corresponding to one dialogue in the memory as historical session data.

Step 102, determining a session state corresponding to the current session based on the session management model and the historical session data.

In the embodiment of the invention, the dialogue state corresponding to the current dialogue is determined based on the dialogue management model and the historical dialogue data. The session management model is used for judging and managing session states corresponding to the session, and specifically, the session management model can determine the session state corresponding to the current session based on historical session data corresponding to the current session. The session state may include session and session end, among other things. The dialogue state can be used for determining whether the information represented by the historical dialogue data accords with the dialogue information requirement for resource requirement analysis, the dialogue information requirement can comprise whether the content of the historical dialogue data is enough, the dialogue management model is used for analyzing the intention, the semanteme and the action of two parties in a plurality of groups of interaction records in the historical dialogue data, and judging the dialogue state and the strategy of the next action to be executed.

Step 103, determining initial structural information based on a target model in a demand analysis model and the historical session data under the condition that the dialogue state represents the end of the dialogue.

In the embodiment of the present invention, the demand analysis model may include a target model, which may be a GPT3.5 model of OpenAI, or may be another pre-training language model, which is not limited in this embodiment of the present invention. In the case where the dialog state characterizes the end of the dialog, the initial structured information output by the target model may be determined based on the target model in the demand analysis model and the historical session data. The initial structural information is the demand information of the user on the target resource, which is obtained by primarily analyzing the historical session data by the target model.

Step 104, verifying the initial structured information based on a verification module in the demand analysis model, determining user demand information conforming to a standard output format, and storing the user demand information into a memory in the server; the user demand information is used for representing demand information of the user on target resources.

In the embodiment of the invention, based on the verification module in the demand analysis model, the format verification can be performed on the structured information in the initial structured information, and the user demand information meeting the standard output format is determined. The verification module may verify the initial structured information based on a preset verification rule, where the preset verification rule may include whether structured information in the initial structured information conforms to a standard output format. The standard output format may include a basic file format of a model output class, and constraint formats of different types such as constraint conditions of parameters such as field types and the like for contents output by the model, and the standard output format may be a JSON format. For example, after the verification module verifies the initial structured information, the verification module may retain the structured information that conforms to the standard output format, and perform format correction on the structured information that does not conform to the standard output format, so as to determine the corresponding structured information that conforms to the standard output format, and determine the user requirement information based on the structured information that conforms to the standard output format and the corrected structured information that does not conform to the standard output format. Under the condition that the dialogue state characterizes the end of the dialogue, the characterization is based on the target dialogue model, the detailed and sufficient reply data is obtained according to the user problem data input by the user, and at the moment, the demand analysis model can analyze based on the sufficient demand information content contained in the historical dialogue data, so that the demand analysis model can be utilized to analyze the resource demand of the historical dialogue data and acquire the user demand information. The user demand information may be information populated based on the target template in order to enhance visualization of the information and conciseness of the information. After the user demand information is obtained, corresponding target resources can be allocated to the user based on the user demand information.

For example, in the case where the target resource is an computing power resource, the computing power resource required by the user may be determined based on the user demand information, and the required computing power resource may be allocated to the user to satisfy the computing power demand of the user.

In one possible implementation manner, the resource processing platform may be an account transaction platform, and the corresponding target dialogue model, dialogue management model and demand analysis model may be deployed on the account transaction platform, where the account transaction platform is used to open and connect each cloud computing service provider through a dedicated network infrastructure, the user puts forward the demand to the transaction platform, and the account transaction platform automatically performs price inquiry and matching with each computing service provider based on the demand of the user, and then provides an account solution to the user according to the user demand and preference. The core functions of the computing power trading platform mainly comprise computing power network interconnection, trade matching, resource allocation, identity authentication, safe operation and maintenance, audit and the like. The power transaction platform is connected with an accessed computing service provider through a unified service access interface. The trade matching mainly means that after the power trading platform acquires the power resource requirements required by the user, the platform extracts the core requirements of the user and forwards the core requirements to the computing service providers of each access platform, the computing service providers provide corresponding power resource allocation schemes and quotations to the platform according to the matching condition of the self resources and the user requirements, the platform synthesizes the schemes returned by the computing service providers and the matching degree of the quotations and the user requirements to perform priority ranking, and finally the priority ranking is presented to the user to select the proper computing service provider so as to meet the power requirements of the user. In the embodiment of the invention, after the user demand information is acquired based on the demand analysis model, the user demand information can be forwarded to the computing service provider based on the computing power transaction platform, the computing power resource allocation scheme and the quotation fed back by the computing service provider based on the user demand information are acquired, and the computing power resource allocation scheme and the quotation are fed back to the user for selection by the user.

In summary, in the embodiment of the invention, by acquiring the reply data output by the target dialogue model based on the user question data input by the user, the corresponding reply content can be automatically determined according to the user question, so that the customer service personnel does not need to reply the user question data manually, the automation in the question reply process is improved, and the operation cost is saved to a certain extent. Further, based on the dialogue management model and the historical session data, determining a dialogue state corresponding to the current dialogue, and under the condition that the dialogue state characterizes the dialogue to be ended, further determining user demand information based on the demand analysis model and the historical session data, further analyzing the historical session data through a target model in the demand analysis model under the condition that the dialogue is ended, and checking based on a checking module in the demand analysis model, so that the demand information of a user meeting a standard output format on target resources can be obtained, and the actual resource demand of the user can be obtained. In this way, through the intelligent judgment of the dialogue state of the dialogue management model, the user demand information is automatically determined based on the demand analysis model, and compared with the manual resource demand analysis based on customer service personnel, the embodiment of the invention automatically performs the resource demand analysis through different models, improves the specialty of the resource demand analysis, simultaneously, realizes full automation of the whole resource demand analysis process, saves the operation cost to a certain extent and improves the analysis efficiency of the resource demand analysis to the user.

Alternatively, step 102 may comprise the steps of:

step 201, inputting the history session data into the dialogue management model; the dialogue management model comprises a target language model and a target classification model.

In the embodiment of the invention, the historical session data is input into a dialogue management model, wherein the historical session data can comprise text interaction records marked with role information, such as: roles may include users and systems. The character information is used to characterize whether the text interaction record corresponding to the character information is input by a user or output by a model. The dialogue management model may include a target language model and a target classification model, where the target language model may be a pre-trained language model, and the language model may be a base model such as ilama 2-Chinese, chatGLM2, which is not limited in this embodiment of the present invention. The target language model is used for acquiring corresponding representation, namely the feature vector, based on the historical session data. The object classification model is used for classifying dialog states based on feature vectors. The target classification model may output dialog states, and it may be appreciated that the target language model may be connected to the target classification model, i.e., the output of the target language model may be used as input content for the target classification model. For example, historical session data may be entered into a target language model in a dialog management model.

Step 202, determining a feature vector corresponding to the historical session data based on the target language model and the historical session data.

In the embodiment of the invention, the feature vector corresponding to the historical session data can be determined based on the target language model and the historical session data. Specifically, the target language model may acquire feature vectors corresponding to the historical session data based on the historical session data. The character information and the corresponding text are spliced according to the time sequence of the dialogue based on character information marked in the historical session data and text interaction records corresponding to the character information, data to be processed are obtained, the data to be processed and task prompt words are spliced and input into a target language model, and the target language model is utilized to gradually decode in an autoregressive mode to generate a representation of the corresponding position of the original sequence of the data to be processed, so that the corresponding feature vector is obtained.

And 203, determining a classification result based on the feature vector and the target classification model.

In the embodiment of the invention, the feature vector is used as the input of the target classification model, wherein the target classification model can be a multi-layer perceptron network, and the target classification model can also comprise an activation function, such as a Sigmoid activation function. And classifying the target classification model based on the feature vector to obtain a classification result. The classification result may be a probability value, and the classification result may be used to characterize the dialog state as a probability in the dialog. Illustratively, the classification result may be obtained as follows:

Wherein, prob is used for representing the classification result, sigmoid is used for representing the activation function in the target classification model, MLP is used for representing the multi-layer perceptron network in the target classification model, and STAT is used for representing the feature vector.

For example, the feature vector may be input into a multi-layer perceptron network with Sigmoid activation to perform state classification, so as to obtain a classification result Sigmoid classification result prob. The classification result prob may be a scalar for characterizing the dialog state as a probability in the dialog, and the dialog state is a probability of ending in the dialog is "1.0-prob", accordingly.

Step 204, based on the classification result, determining, by the processor, a session state corresponding to the current session.

In the embodiment of the invention, the processor can determine the dialogue state corresponding to the current dialogue based on the size of the classification result. Because the classification result is a probability value, the dialog state corresponding to the current dialog can be determined according to the size of the probability value.

Optionally, step 204 may include the steps of:

step 2041, the processor determines the session state as the session end if the classification result is greater than a preset threshold.

In the embodiment of the invention, under the condition that the classification result is larger than the preset threshold value, the dialogue state is determined as the end of the dialogue. The preset threshold may be set according to the requirement, and the preset threshold may be a value greater than or equal to 0.5 and less than 1, for example, the preset threshold may be 0.5. And under the condition that the classification result is larger than a preset threshold value, the probability of representing the dialogue state as the dialogue end is larger than the probability of representing the dialogue state as the dialogue, and the dialogue state can be determined as the dialogue end.

Step 2042, the processor determines the session state as being in a session when the classification result is less than or equal to the preset threshold.

In the embodiment of the invention, the dialogue state is determined as being in the dialogue under the condition that the classification result is smaller than or equal to the preset threshold value. And under the condition that the classification result is larger than the preset threshold value, the probability of representing the dialogue state as the dialogue is smaller than the probability of representing the dialogue state as the dialogue is ended, and the dialogue state can be determined as the dialogue.

In a possible embodiment, the dialogue state corresponding to the result with the highest probability value in the classification result prob and 1-prob may be directly determined as the dialogue state, that is, the output result of the dialogue management model. Exemplary, may specifically be:. Wherein out is used to represent the output result of the dialog management model and prob is used to characterize the classification result.

In the embodiment of the invention, the dialogue state corresponding to the current dialogue can be judged based on the historical dialogue data through the dialogue management model, so that the target dialogue model can judge whether the dialogue needs to be continued or not, and the information acquisition efficiency is improved.

Optionally, the embodiment of the invention can comprise the following steps:

step 301, in the case that the dialogue state is in a dialogue, repeating the operation of obtaining the reply data output by the target dialogue model based on the user question data input by the user, and determining the history dialogue data based on the user question data and the reply data.

In the embodiment of the invention, under the condition that the dialogue state is in the dialogue, the information obtained by characterizing the current dialogue is insufficient, and the reply data corresponding to the question data needs to be obtained continuously based on the target dialogue model so as to determine the user resource requirement. If the session state is in a session, the above step 101 is repeatedly executed to further obtain the reply data based on the target session model, that is, further analyze the resource requirement of the user, to obtain the corresponding resource requirement information.

For example, in the case where the session state is session end, an end instruction may be sent by the session management model to the target session model to end the operation of acquiring the new round of reply data based on the target session model; in the case that the session state is in a session, a continuation instruction may be sent by the session model to the target session model to instruct the target session model to further acquire the new round of reply data, or the session management model may not perform any action, and directly perform the operation of acquiring the new round of reply data when the target session model does not receive the instruction sent by the session management model within a preset time.

In the embodiment of the invention, whether the current dialogue needs to be continued or not can be judged according to the actual dialogue state of the current dialogue, and different operations are executed according to different conditions, so that the automation of the whole dialogue flow is realized while the network resources are saved.

Alternatively, step 103 may comprise the steps of:

step 401, defining a standard output format corresponding to the demand analysis model based on a preset configuration file.

In the embodiment of the invention, the standard output format corresponding to the demand analysis model is defined based on the preset configuration file. The demand analysis model is used for generating a structured representation corresponding to the user demand information based on the historical session data. The demand analysis model may include a target model, and in order to ensure that the output result of the target model is controllable, the target model may be combined with guard rail (guard rail) technology in order to ensure that the target model responds with accurate and proper information. To apply guard rail (guard rail) technology, 3 important components can be determined for customization in a demand analysis model according to the output form (structured representation) of the required user demand information, including: a preset configuration file (Rail), a monitoring module (Guard) and a verification module (valve). It is understood that the demand analysis model may include a target model, a preset configuration file (Rail), a monitoring module (Guard), and a verification module (verifier).

The basic format of the user requirement information output by the target model and the requirement analysis model is defined in the preset configuration file, and specific content of the basic format can be set according to requirements, for example, the JSON format can be used as a standard output format for structured output of the user requirement information. The monitoring module is used for butting and packaging the target model and the preset configuration file so as to ensure that the structural information output by the target model and the user demand information output by the demand analysis model meet the basic format defined in the preset configuration file. And the monitoring module is also used for transmitting the API related authentication authorization information to the target model so as to call the target model based on the target framework. The verification module is used for carrying out format verification and validity verification on the output result of the target model.

And defining a standard output format corresponding to the demand analysis model based on a basic format defined in a preset configuration file, wherein the standard output format is used for representing an output format of an output result of the demand analysis model. The method includes the steps that a preset configuration file which is defined in advance can be used as parameters through a construction function defined in a guardrail library to be transmitted into a monitoring module, and the monitoring module is used for butting and packaging the target model and the preset configuration file, so that the output results of the target model and the demand analysis model are in a standard output format.

Step 402, inputting the historical session data into the demand analysis model.

Step 403, calling the target model based on a monitoring module in the demand analysis model, and acquiring initial structural information output by the target model.

In the embodiment of the invention, historical session data, prompt words and a demand analysis instruction are used as the input of a demand analysis model, and a target model is called based on a monitoring module in the demand analysis model to acquire initial structural information output by the target model. The prompt word may be used to instruct the target model to generate a structured code according to submitted historical session data, and select attributes and values related to computing power configuration from the attributes and values. The demand analysis instructions are used for instructing the target model to conduct demand analysis based on historical session data. The location of the occurrence of the nature and magnitude of the initial structured information may be fixed. And calling the target model through the monitoring module, and transmitting the historical session data into the target model through the monitoring module to acquire initial structural information corresponding to the historical session data output by the target model.

Accordingly, step 104 may include the steps of:

and step 404, correcting the structured information which does not accord with the standard output format in the initial structured information until the initial structured information accords with the standard output format.

In the embodiment of the invention, the initial structured information can be checked based on the check module in the demand analysis model, and under the condition that the initial structured information contains the structured information which does not accord with the standard output format, the structured information which does not accord with the standard output format is required to be corrected until the initial structured information completely accords with the standard output format.

And 405, acquiring user demand information output by the demand analysis model based on the structural information conforming to the standard output format and the corrected structural information.

In the embodiment of the invention, in the process of checking by the checking module, the standard structured information can be judged to comprise structured information conforming to the standard output format and structured information not conforming to the standard output format. And determining corresponding user demand information by using a demand analysis model based on the standard structured information conforming to the standard output format, namely the structured information conforming to the standard output format and the corrected structured information conforming to the standard output format. Wherein, the output format of the user demand information accords with the standard output format.

In the embodiment of the invention, through each module in the demand analysis model and the target model, the demand analysis model can output the user demand information conforming to the standard output format, so that the user demand information can be convenient for a computing service provider to perform operations such as resource matching and the like in a subsequent flow.

Optionally, step 401 may include the steps of:

step 4011, transmitting the preset configuration file into the monitoring module, and packaging the target model and the preset configuration file.

Step 4012, initializing the monitoring module.

In the embodiment of the invention, the preset configuration file is transmitted into the monitoring module, the target model and the preset configuration file are in butt joint and packaged, and the monitoring module is initialized after the packaging is completed. After the monitoring module calls the target model, the output result of the target model conforms to a standard output format.

In the embodiment of the invention, the output format of the output result of the target model can conform to the standard output format under the condition that the target model is called by the monitoring module through packaging the target model and the preset configuration file.

Optionally, step 404 may include the steps of:

step 501, based on the verification module, it is verified whether each structured information in the initial structured information conforms to the standard output format.

In the embodiment of the invention, under the condition that the initial structured information output by the target model is obtained, the initial structured information is input into a verification module in the demand analysis model, and whether each structured information in the initial structured information accords with a standard output format is verified based on the verification module. Wherein, the initial structured information may include a plurality of structured information. Illustratively, in a computing power resource demand analysis scenario, the structured information may include: "CPU:36"," GPU:4".

In one possible implementation, a predetermined JSON format checker in the guardrail library may be selected to check the initial structural information.

Step 502, generating an error correction instruction in the case that the structured information which does not conform to the standard output format exists.

Step 503, inputting the error correction instruction and the initial structural information with error labels into the monitoring module;

In the embodiment of the invention, under the condition that the initial structured information does not accord with the structured information of the standard output format, the structured information which characterizes the existence of the structured information which does not accord with the text format requirement and/or the constraint condition in the initial structured information is required to be corrected, and the verification module can mark the structured information which does not accord with the standard output format in the initial structured information and correspondingly generate the error correction instruction. The error correction instructions are used for instructing the target model to correct the structured information which does not conform to the standard output format. After the error correction instruction is generated, the error correction instruction and the initial structured information with the error label can be input into the monitoring module again, and the target model is called again under the condition that the error correction instruction and the initial structured information with the error label are received by the monitoring module, and the target model determines the corrected structured information corresponding to the structured information which does not accord with the standard output format again based on the error correction instruction. And correcting the structural information to be the corrected structural information which does not accord with the standard output format. For example, the monitoring module may recall the target model again, input the error correction instruction and the initial structural information with the error label into the target model, and obtain the target structural information output again by the target model. The target structural information is corrected initial structural information. And (3) outputting the structured information corresponding to the error marked position in the target structured information again by the target model, namely, correcting the structured information corresponding to the structured information which does not accord with the standard output format.

It will be appreciated that in the case where the target structural information is obtained again, the operations of steps 501 to 503 are repeatedly performed until the target structural information output by the target model completely conforms to the standard output format. For example, fig. 2 shows a flow chart of steps for performing demand analysis based on a demand analysis model, and as shown in fig. 2, historical session data, a prompt word and a demand analysis instruction are used as input of the demand analysis model, a target model is called through a monitoring module (Guard), and the historical session data, the prompt word and the demand analysis instruction are transmitted into the target model through the monitoring module (Guard), so as to obtain initial structural information output by the target model. And checking the initial structured information by using a checking module (Validator), carrying out error marking on the structured information which does not accord with the standard output format in the initial structured information, sending the initial structured information with the error marking and an error correction instruction to a monitoring module (Guard), and calling a target model by the monitoring module again, wherein the target model processes the initial structured information with the error marking based on the error correction instruction so as to achieve a preset stopping condition, namely the structured information which does not accord with the standard output format in the initial structured information accords with the standard output format after being corrected.

In the embodiment of the invention, each piece of structured information of the initial structured information is checked by the checking module to judge whether the initial structured information accords with the standard output format, and the structured information which does not accord with the standard output format is corrected based on the error correction instruction, so that the structured information finally output by the target model can be ensured to accord with the standard output format.

Optionally, step 405 may include the steps of:

and 601, under the condition that a preset stop condition is met, acquiring user demand information output by the demand analysis model based on the structural information conforming to the standard output format and the corrected structural information.

In the embodiment of the invention, under the condition that the preset stop condition is reached, the user demand information output by the demand analysis model is obtained based on the structured information conforming to the standard output format and the corrected structured information not conforming to the standard output format, namely, the corrected structured information. The preset stopping condition may include that the structural information which does not conform to the standard output format in the initial structural information is corrected to conform to the standard output format, and no structural information which does not conform to the standard output format exists in the target structural information which is output by the target model for the last time.

In the embodiment of the invention, the requirement analysis model can acquire the user requirement information based on the structured information (the structured information conforming to the standard output format and the corrected structured information) conforming to the standard output format by setting the preset stop condition.

Alternatively, step 601 may comprise the steps of:

and 701, matching and filling an information template in the preset configuration file based on the structured information conforming to the standard output format and the corrected structured information to obtain the user demand information.

In the embodiment of the invention, the preset configuration file also comprises a predefined information module, and the information template is matched and filled based on the structured information conforming to the standard output format and the corrected structured information to obtain the user demand information. For which the demand information conforms to the standard output format and complies with the content defined in the information template. Illustratively, constraints on field types may be defined in the information template, such as: the fields corresponding to the number of CPUs, the number of memories and the number of hard disks are integers, the GPU calculation force TFLOPS requirement field is a floating point number, the preferred GPU model is a character string and is in the range of [ 'auto', 'legacy', 'main stream', 'flag shift'. Accordingly, the user demand information may include:

{

`CPU`:36;

`RAM`:384,

`RAM_unit`=`GB`;

`SSD`:2,

`SSD_unit`=`TB`;

`GPU`:4,

`GPU_type`=`mainstream`;

}

It can be understood that, field contents matched with the structured information and the corrected structured information conforming to the standard output format in the information template are correspondingly filled and matched, and unmatched field contents can be left as default values. In an exemplary embodiment, in the business requirements of the actual calculation power requirement analysis, as many fields as possible are included in the calculation power business requirement representation, but most users may only pay attention to a plurality of keywords, for example, 20 contents corresponding to the information templates, but a certain user only pays attention to the multi-core performance and the memory quantity of the CPU, and after the information templates are matched and filled based on the structured information and the modified structured information conforming to the standard output format, the fields in the information templates except the CPU quantity and the memory quantity are left as default or automatic.

In the embodiment of the invention, the information template is matched and filled based on the structured information conforming to the standard output format and the corrected structured information, so that the expression form of the user demand information is more standardized and has better visibility.

Optionally, the target session model in the embodiment of the present invention is obtained based on training functions provided by the processor. In particular, the training functions may be provided using a computer processor operatively coupled to a memory and executing appropriate stored computer program instructions. The target dialogue model can be obtained through training the following steps:

Step 801, acquiring a first sample data set; the first sample data set comprises a plurality of groups of question-answer sentence pairs corresponding to the target field, and the question-answer sentence pairs comprise sample question data.

In the embodiment of the invention, the processor in the server can acquire the first sample data set from the preset data source, wherein the first sample data set can be a plurality of groups of question-answer sentence pairs acquired aiming at a vertical field (target field), and the preset data source can be a dialogue resource library under the target field. The question-answer pair may include sample question data and sample answer data, which are similar to the user question data and answer data in step 101, and specifically reference may be made to the description in step 101. The number of question-answer sentence pairs determines the learning ability of the target dialogue model, i.e. the more the number of question-answer sentence pairs is, the stronger the learning ability of the target dialogue model is. The target domain may be determined by a usage scenario of the target dialog model, and by way of example, the target domain may be a computational resource domain, to which embodiments of the present invention are not limited.

Step 802, taking the sample question data and the target prompt word as input of a dialogue model to be trained, and obtaining predicted reply data output by the dialogue model to be trained.

In the embodiment of the invention, the dialogue model to be trained is trained based on the first sample data set. Specifically, the sample question data and the target prompt word can be used as the input of the dialogue model to be trained, and the predicted reply data output by the dialogue model to be trained can be obtained. The dialogue model to be trained can be an LLM model, such as models of ChatGLM2, llama2-Chinese-chat and the like, and the target prompt word can be used as a system prompt word for starting training, and is specifically used for indicating that the task of the dialogue model to be trained is to converse with a user and solve the problem in user problem data, and simultaneously responding to the requirement set by the user. For example: the target prompt word sample may be: "you are a question and answer robot that can analyze and solve the problem, require careful solution of the problem posed by the user, and cannot contain illegal, offensive, and offensive replies. ". And outputting predicted reply data corresponding to the sample question data by the dialogue model to be trained based on the target prompt word and the sample question data.

Step 803, based on the predicted reply data, parameter adjustment is performed on the dialogue model to be trained, and the target dialogue model is determined.

In the embodiment of the invention, the dialogue model to be trained is trained based on the first sample data set, and the trained dialogue model to be trained is determined to be the target dialogue model. Specifically, parameter adjustment can be performed on the dialogue model to be trained based on the predicted reply data and the sample reply data to obtain the target dialogue model. And correspondingly deploying the target dialogue model on the power transaction platform, so that the power transaction platform can execute the resource demand analysis method in the embodiment.

According to the embodiment of the invention, the dialogue model to be trained is trained through the first sample data set, so that the dialogue model to be trained can learn better dialogue capability aiming at the target field, the dialogue capability of the target dialogue model is more excellent, the acquisition cost for acquiring accurate user demands based on the target dialogue model is saved, and the acquisition efficiency of the user demands is improved.

Optionally, the question-answer sentence pair further includes sample answer data; step 803 may include:

step 901, performing parameter adjustment on the dialogue model to be trained based on the predicted reply data and the sample reply data; the sample reply data is used for representing real reply content corresponding to the sample question data.

In the embodiment of the invention, in order to make the similarity of the predicted reply data and the sample reply data as high as possible, the parameters of the dialogue model to be trained can be adjusted based on the predicted reply data and the sample reply data. And continuously adjusting parameters of the dialogue model to be trained so that the similarity of the predicted reply data output by the dialogue model to be trained and the sample reply data corresponding to the sample question data is larger than a first similarity threshold. The value of the first similarity threshold can be set according to the requirement, which is not limited in the embodiment of the invention. The sample answer data is used to characterize the real answer content corresponding to the sample question data, and it is understood that the real answer content may refer to answer content that meets the resource requirement and conforms to the dialog logic, corresponding to the resource requirement involved in the sample question data. For example, random gradient descent (SGD), batch Gradient Descent (BGD), or like optimization algorithms may be used to adjust parameters of the dialog model to be trained.

And step 902, determining the dialogue model to be trained as the target dialogue model under the condition that a first stop condition is reached.

In the embodiment of the present invention, the first stopping condition may include conditions that a loss value of the dialogue model to be trained reaches a first preset threshold, a number of training rounds of the dialogue model to be trained reaches a first preset round number threshold, and the like.

According to the embodiment of the invention, the dialogue model to be trained can learn the universal dialogue capability in the training process by training the dialogue model to be trained, so that the sample problem data can be better replied in the resource demand direction.

Optionally, step 901 may include the steps of:

step 1001, determining a probability loss value based on the predicted reply data and the sample reply data.

In the embodiment of the invention, the probability loss value can be obtained by using a first formula based on the predicted reply data and the sample reply data. The probability loss value is used for representing the gap between predicted reply data predicted by the dialogue model to be trained and real reply content corresponding to the sample question data.

Alternatively, the first formula may include:

wherein SFT_Loss is used to represent the probability Loss value; d is used to represent the first sample dataset;for representing the size of said first sample dataset,/or->A set of question-answer sentence pairs for representing said first sample data set- >For representing said sample question data, +.>For representing said sample reply data, +.>For representing +.>Fields of individual positions>For representing the sequence of the first i fields in the sample reply data.

Step 1002, adjusting parameters of the dialogue model to be trained based on the probability loss value.

In the embodiment of the invention, the model parameters of the dialogue model to be trained are adjusted according to the probability loss value, for example: learning rate, regularization coefficient, etc., to optimize model performance of the dialog model to be trained. In particular, the server may initiate parameter adjustments to the dialog model to be trained based on the probability loss values. The probability Loss value is used for representing an example corresponding to the dialogue model to be trained, the sample problem data is output based on the sample problem data, the probability Loss value (SFT_Loss) can be calculated based on a negative log likelihood function, and parameters of the dialogue model to be trained are adjusted based on the probability Loss value.

For example, as shown in fig. 3, the training process of the target dialogue model may input the target prompt word and the sample question data into the dialogue model to be trained, obtain the predicted reply data output by the dialogue model to be trained, and perform parameter adjustment on the dialogue model to be trained based on the predicted reply data and the probability loss value corresponding to the sample reply data, and re-train the first sample data set until the first stop condition is reached.

In the embodiment of the invention, the difference between the predicted reply data and the sample reply data can be obtained by obtaining the probability loss value, so that the parameters of the dialogue model to be trained can be conveniently adjusted according to the difference. Meanwhile, by carrying out parameter adjustment and retraining on the performance of the dialogue model to be trained, the predicted reply data output by the dialogue model to be trained is more real and reliable, and the dialogue performance of the dialogue model to be trained is further improved.

Optionally, the dialog management model in the embodiment of the invention is derived based on training functions provided by the processor. In particular, the training functions may be provided using a computer processor operatively coupled to a memory and executing appropriate stored computer program instructions. The dialogue management model can be trained by the following steps:

step 1101, obtaining a second sample dataset; the second sample dataset includes multiple sets of conversation histories tagged with role information, the conversation histories containing multiple interactions.

In the embodiment of the invention, the processor in the server can acquire the second sample data set from the appointed data source, the execution data source can be an appointed database or an appointed storage address, the second sample data set comprises a plurality of groups of dialogue history records marked with role information, the dialogue history records can be dialogue records corresponding to history dialogue processes based on different states in the target field, and the dialogue history records can comprise a plurality of dialogue interaction operations, namely, the dialogue history records can comprise dialogue records corresponding to a plurality of rounds of dialogue. The conversation history may be marked with role information, which may be two or more parties participating in the conversation, such as: users and systems.

Step 1102, for any dialogue history record in the multiple groups of dialogue history records, acquiring a target feature vector based on a target language model.

In the embodiment of the invention, the target feature vector is acquired based on the target language model aiming at any dialogue history record in a plurality of groups of dialogue history records. The dialogue management model may include a target language model and a target classification model, where the target language model may be a pre-trained language model, for example, a base model such as ilama 2-Chinese, chatGLM 2. Inputting the dialogue history record into a target language model, and acquiring a corresponding target feature vector based on the target language model. The target feature vector may be used to characterize feature information corresponding to the dialog history.

And step 1103, taking the target feature vector as the input of the classification model to be trained, and obtaining a prediction state result output by the classification model to be trained.

In the embodiment of the invention, the classification model to be trained is trained based on the target feature vector. Specifically, the target feature vector can be used as an input of the classification model to be trained, and a prediction state result output by the classification model to be trained is obtained. The classification model to be trained can be a multi-layer perceptron network.

And 1104, performing parameter adjustment on the classification model to be trained based on the prediction state result and the state label corresponding to the dialogue history record.

In the embodiment of the invention, the classification model to be trained is trained based on the second sample data set, and the trained classification model to be trained is determined to be the target classification model. Specifically, parameter adjustment can be performed on the classification model to be trained based on the prediction state result and the state label corresponding to the dialogue history record, so as to obtain the target classification model. The state label corresponding to the dialogue history record is used for guaranteeing the actual dialogue state corresponding to the dialogue history record. The actual session state may include session and session end, and correspondingly, the state label corresponding to the session history record may also include session and session end.

In order to make the similarity of the predicted state result and the state label corresponding to the dialogue history as high as possible, parameter adjustment can be performed on the classification model to be trained based on the predicted state result and the state label corresponding to the dialogue history. And continuously adjusting parameters of the classification model to be trained, so that the similarity of the predicted state result output by the classification model to be trained and the state label corresponding to the dialogue history record is larger than a second similarity threshold. The value of the second similarity threshold can be set according to the requirement, which is not limited in the embodiment of the invention. For example, optimization algorithms such as random gradient descent (SGD), batch Gradient Descent (BGD), etc. may be used to adjust parameters of the classification model to be trained.

In one possible implementation, a second formula may be used to calculate a first target Loss between the predicted state result and the state label corresponding to the dialog history, which may be a binomial distributed cross entropy Loss (BCE Loss). After computing the binomial distributed cross entropy loss, the model can be fine-tuned over a given dialog history using AdamW algorithm. Specifically, the server may initiate parameter adjustment of the classification model to be trained based on the first target lossThe process of the number adjustment can beWherein Θ is a parameter of the classification model to be trained, label is a state Label corresponding to the dialogue history record, prob is a predicted state result, and x is the dialogue history record corresponding to the predicted state result generated by the model.

Step 1105, determining the to-be-trained classification model as a target classification model and determining the target language model and the target classification model as the dialogue management model when a second stop condition is reached.

In the embodiment of the present invention, the second stopping condition may include conditions that a loss value of the classification model to be trained reaches a second preset threshold, a number of training rounds of the classification model to be trained reaches a second preset round number threshold, and the like. In the case that the second stop condition is reached, the classification model to be trained is determined as a target classification model, and the target language model and the target classification model are determined as a dialogue management model.

According to the embodiment of the invention, the classification model to be trained is trained through the second sample data set, so that the classification model to be trained can learn the general dialogue state classification capability in the training process, and dialogue data can be better classified. The confirmation cost for confirming the dialogue state is saved, and the confirmation efficiency of the dialogue state is improved.

Optionally, step 1102 may include the steps of:

step 1201, for any dialog history record in the multiple groups of dialog history records, splicing the role information, the text content corresponding to each role information and the state label corresponding to the dialog history record to obtain a spliced text.

In the embodiment of the invention, for any dialogue history record in a plurality of groups of dialogue history records, the role information in the dialogue history records, the text content corresponding to each role information and the state label corresponding to the dialogue history record are spliced to obtain a spliced text. Specifically, the role information in the dialogue history record and the text content corresponding to each role information may be spliced according to the dialogue occurrence sequence based on the first splicing rule, and then the state labels are spliced, where the role information, the text content and the state labels may be separated by using a Separator (SEP). Also, to mark the beginning of a dialogue history, a sentence classification flag vector (CLS) may be used to splice at the beginning of the spliced text. Illustratively, assume a dialog history record as: "user: 11111; the system comprises: 22222; the user: 33333; the system comprises: 44444", wherein the user corresponds to the character information with the system, respectively, 11111, 22222, 33333, and 44444 correspond to the text content corresponding to the character information, respectively. The state label corresponding to the dialogue history record is in the dialogue. The character information, the text content corresponding to each character information and the state label corresponding to the dialogue history record can be spliced based on the separator to obtain a spliced text. The spliced text may be: "CLS+user+SEP+11111+SEP+System+SEP+22222+SEP+user +sep+33333+sep+system +sep+44444+sep+in dialog. It can be understood that, in order to simplify the splicing manner, the character information, the text content corresponding to each character information, and the state label corresponding to the dialogue history record may be spliced based on a second splicing rule, where the second splicing rule may be to add a separator between each group of character information and the text content corresponding to the character information, and splice the state label corresponding to the dialogue history record at last. For example: the above spliced text may also be: "CLS+user+11111+SEP+System+22222+SEP+user +33333+sep+system +44444+sep+in dialog). The embodiment of the invention does not limit the splicing mode.

Step 1202, inputting the spliced text into a target language model, and obtaining a target feature vector output by the target language model.

In the embodiment of the invention, the spliced text is input into the target language model, and the target feature vector output by the target language model can be obtained.

For example, fig. 4 shows a schematic diagram of a training process of a dialog management model, and as shown in fig. 4, for any dialog history record in multiple groups of dialog history records, character information, text content corresponding to each character information, and status labels corresponding to the dialog history records are spliced by using separators, so as to obtain a spliced text. Specifically, a separator can be added between each group of character information and text content corresponding to the character information, and a state label corresponding to the last spliced conversation history record is added to obtain a spliced text. And inputting the spliced text into a target language model to obtain a target feature vector output by the target language model. And inputting the target language feature vector into the classification model to be trained, and obtaining a prediction state result output by the classification model to be trained. And calculating a first target loss based on the predicted state result and the state label corresponding to the dialogue history record, and carrying out parameter adjustment on the classification model to be trained based on the first target loss. In the case that the second stop condition is reached, the classification model to be trained is determined as a target classification model, and the target language model and the target classification model are determined as a dialogue management model.

In the embodiment of the invention, the contents of the conversation histories are integrated in the spliced text by splicing the character information, the text contents corresponding to the character information and the state labels corresponding to the conversation histories, so that the target language model can conveniently acquire the characteristic information corresponding to the conversation histories, namely the target characteristic vector, based on the spliced text.

For example, fig. 5 shows a flowchart of specific steps of a resource demand analysis method, as shown in fig. 5, user question data input by a user is obtained, the user question data is input into a target dialogue model, reply data corresponding to the user question data output by the target dialogue model is obtained, and historical session data is determined. Judging the dialogue state based on the historical dialogue data through the dialogue management model, and continuously generating corresponding reply data based on the user question data input by the user through the target dialogue model under the condition that the dialogue state represents the dialogue; in the event that the dialog state characterizes the end of the dialog, user demand information is determined based on the demand analysis model and the historical session record.

Fig. 6 is a schematic structural diagram of a resource demand analysis device according to an embodiment of the present invention, and as shown in fig. 6, the device may specifically include:

A first obtaining module 1301 configured to cause a processor in the server to obtain reply data output by a target dialogue model based on user question data input by a user, and determine historical dialogue data based on the user question data and the reply data;

a first determining module 1302, configured to determine a session state corresponding to the current session based on the session management model and the historical session data;

a second determining module 1303, configured to determine initial structural information based on a target model in a demand analysis model and the historical session data when the session state characterizes a session end;

a third determining module 1304, configured to verify the initial structured information based on a verification module in the requirement analysis model, determine user requirement information that accords with a standard output format, and store the user requirement information to a memory in the server; the user demand information is used for representing demand information of the user on target resources.

The embodiment of the invention provides a resource demand analysis device, which can automatically determine corresponding reply content according to a user question by acquiring reply data output by a target dialogue model based on the user question data input by a user, does not need to manually reply the user question data by customer service personnel, improves the automation in the question reply process, and saves the operation cost to a certain extent. Further, based on the dialogue management model and the historical session data, determining a dialogue state corresponding to the current dialogue, and under the condition that the dialogue state characterizes the dialogue to be ended, further determining user demand information based on the demand analysis model and the historical session data, further analyzing the historical session data through a target model in the demand analysis model under the condition that the dialogue is ended, and checking based on a checking module in the demand analysis model, so that the demand information of a user meeting a standard output format on target resources can be obtained, and the actual resource demand of the user can be obtained. In this way, through the intelligent judgment of the dialogue state of the dialogue management model, the user demand information is automatically determined based on the demand analysis model, and compared with the manual resource demand analysis based on customer service personnel, the embodiment of the invention automatically performs the resource demand analysis through different models, improves the specialty of the resource demand analysis, simultaneously, realizes full automation of the whole resource demand analysis process, saves the operation cost to a certain extent and improves the analysis efficiency of the resource demand analysis to the user.

Optionally, the first determining module 1302 includes:

a first input module for inputting the historical session data into the dialog management model; the dialogue management model comprises a target language model and a target classification model;

the first determining submodule is used for determining a feature vector corresponding to the historical session data based on the target language model and the historical session data;

the second determining submodule is used for determining a classification result based on the feature vector and the target classification model;

and the third determining submodule is used for enabling the processor to determine the dialogue state corresponding to the current dialogue based on the classification result.

Optionally, the third determining submodule includes:

a fourth determining submodule, configured to enable the processor to determine the session state as session end if the classification result is greater than a preset threshold;

and a fifth determining submodule, configured to enable the processor to determine the dialogue state as being in a dialogue when the classification result is less than or equal to the preset threshold value.

Optionally, the apparatus may specifically further include:

and the first execution module is used for repeatedly executing the operation of acquiring the reply data output by the target dialogue model based on the user question data input by the user and determining the historical dialogue data based on the user question data and the reply data when the dialogue state is in the dialogue.

Optionally, the second determining module 1303 includes:

the first defining module is used for defining a standard output format corresponding to the demand analysis model based on a preset configuration file;

the second input module is used for inputting the historical session data into the demand analysis model;

and the first calling module is used for calling a target model based on the monitoring module in the demand analysis model and acquiring initial structural information output by the target model.

Accordingly, the third determination module 1304 includes:

the first correction module is used for correcting the structured information which does not accord with the standard output format in the initial structured information until the initial structured information accords with the standard output format;

and the second acquisition module is used for acquiring the user demand information output by the demand analysis model based on the structured information conforming to the standard output format and the corrected structured information.

Optionally, the first defining module includes:

the first packaging module is used for transmitting the preset configuration file into the monitoring module and packaging the target model and the preset configuration file;

and the first initialization module is used for initializing the monitoring module.

Optionally, the first correction module includes:

the first verification module is used for verifying whether each piece of structured information in the initial structured information accords with the standard output format or not based on the verification module in the demand analysis model;

the first generation module is used for generating an error correction instruction under the condition that structured information which does not accord with the standard output format exists;

the third input module is used for inputting the error correction instruction and the initial structural information with error labels into the monitoring module;

Optionally, the second acquisition module includes:

and the third acquisition module is used for acquiring the user demand information output by the demand analysis model based on the structural information conforming to the standard output format and the corrected structural information under the condition that the preset stop condition is reached.

Optionally, the third acquisition module includes:

and the first filling module is used for matching and filling the information templates in the preset configuration file based on the structured information conforming to the standard output format and the corrected structured information to obtain the user demand information.

Optionally, the first obtaining module 1301 includes:

the first acquisition sub-module is used for acquiring user problem data input by a user and inputting the user problem data into the target dialogue model;

the second acquisition sub-module is used for acquiring reply data corresponding to the user problem data output by the target dialogue model;

and the first storage module is used for correspondingly storing the user question data and the reply data into the memory, and determining the user question data and the reply data corresponding to one dialogue as the historical session data.

Optionally, the apparatus may specifically further include:

a fourth acquisition module for acquiring the first sample dataset; the first sample data set comprises a plurality of groups of question-answer sentence pairs corresponding to the target field, and the question-answer sentence pairs comprise sample question data;

the fourth input module is used for taking the sample question data and the target prompt word as the input of the dialogue model to be trained and obtaining the predicted reply data output by the dialogue model to be trained;

and the first adjustment module is used for carrying out parameter adjustment on the dialogue model to be trained based on the prediction reply data and determining the target dialogue model.

Optionally, the question-answer sentence pair further includes sample answer data; the first adjustment module includes:

the first adjustment sub-module is used for carrying out parameter adjustment on the dialogue model to be trained based on the prediction reply data and the sample reply data; the sample reply data is used for representing real reply content corresponding to the sample question data;

and the fourth determining module is used for determining the dialogue model to be trained as the target dialogue model under the condition that the first stopping condition is reached.

Optionally, the first adjustment submodule includes:

a sixth determination sub-module for determining a probability loss value based on the predicted reply data and the sample reply data;

and the second adjustment sub-module is used for carrying out parameter adjustment on the dialogue model to be trained based on the probability loss value.

Optionally, the apparatus may specifically further include:

a fifth acquisition module for acquiring a second sample dataset; the second sample data set comprises a plurality of groups of dialogue histories marked with role information, and the dialogue histories comprise a plurality of interactive operations;

a sixth obtaining module, configured to obtain, for any one of the multiple groups of dialogue histories, a target feature vector based on a target language model;

The seventh acquisition module is used for taking the target feature vector as the input of the classification model to be trained and acquiring a prediction state result output by the classification model to be trained;

the second adjusting module is used for carrying out parameter adjustment on the classification model to be trained based on the prediction state result and the state label corresponding to the dialogue history record;

and a fifth determining module, configured to determine the to-be-trained classification model as a target classification model and determine the target language model and the target classification model as the dialogue management model when the second stopping condition is reached.

Optionally, the sixth acquisition module includes:

the first splicing module is used for splicing the role information, the text content corresponding to each role information and the state label corresponding to the dialogue history record aiming at any dialogue history record in the multiple groups of dialogue history records to obtain a spliced text;

and the third acquisition sub-module is used for inputting the spliced text into a target language model and acquiring a target feature vector output by the target language model.

The embodiment of the invention also provides a server, which comprises: a processor operatively coupled to the memory; a communication interface; the server runs a resource processing platform and is used for:

It will be appreciated that according to various embodiments, the server may also be adapted to operate in accordance with the resource demand analysis method of the previous embodiments.

The present invention also provides an electronic device, see fig. 7, comprising: a processor 1401, a memory 1402 and a computer program 14021 stored on the memory and executable on the processor, which when executed implements the resource demand analysis method of the foregoing embodiments.

The present invention also provides a readable storage medium, which when executed by a processor of an electronic device, enables the electronic device to perform the resource demand analysis method of the foregoing embodiment.

For the device embodiments, since they are substantially similar to the method embodiments, the description is relatively simple, and reference is made to the description of the method embodiments for relevant points.

In some instances, well-known methods, structures and techniques have not been shown in detail in order not to obscure an understanding of this description.

Various component embodiments of the invention may be implemented in hardware, or in software modules running on one or more processors, or in a combination thereof. Those skilled in the art will appreciate that some or all of the functions of some or all of the components in a sorting device according to the present invention may be implemented in practice using a microprocessor or Digital Signal Processor (DSP). The present invention may also be implemented as an apparatus or device program for performing part or all of the methods described herein. Such a program embodying the present invention may be stored on a computer readable medium, or may have the form of one or more signals. Such signals may be downloaded from an internet website, provided on a carrier signal, or provided in any other form.

It should be noted that the above-mentioned embodiments illustrate rather than limit the invention, and that those skilled in the art will be able to design alternative embodiments without departing from the scope of the appended claims. In the claims, any reference signs placed between parentheses shall not be construed as limiting the claim. The word "comprising" does not exclude the presence of elements or steps not listed in a claim. The word "a" or "an" preceding an element does not exclude the presence of a plurality of such elements. The invention may be implemented by means of hardware comprising several distinct elements, and by means of a suitably programmed computer. In the unit claims enumerating several means, several of these means may be embodied by one and the same item of hardware. The use of the words first, second, third, etc. do not denote any order. These words may be interpreted as names.

It should be noted that all actions for obtaining signals, information or data in this application are performed in compliance with the corresponding data protection legislation policy of the country of location and obtaining the authorization granted by the owner of the corresponding device.

The foregoing description of the preferred embodiments of the invention is not intended to be limiting, but rather is intended to cover all modifications, equivalents, and alternatives falling within the spirit and principles of the invention.

The foregoing is merely illustrative of the present invention, and the present invention is not limited thereto, and any person skilled in the art will readily recognize that variations or substitutions are within the scope of the present invention. Therefore, the protection scope of the invention is subject to the protection scope of the claims.

Claims

1. A resource demand analysis method, for use with a server running a resource processing platform, the method comprising:

2. The method of claim 1, wherein determining a session state corresponding to a current session based on the session management model and the historical session data comprises:

3. The method of claim 2, wherein the determining, by the processor, a dialog state corresponding to a current dialog based on the classification result comprises:

4. The method according to claim 1, wherein the method further comprises:

5. The method of claim 1, wherein determining initial structural information based on the historical session data and a target model in the demand analysis model comprises:

inputting the historical session data into the demand analysis model;

6. The method of claim 5, wherein defining the standard output format corresponding to the demand analysis model based on the preset configuration file comprises:

initializing the monitoring module.

7. The method of claim 5, wherein correcting the structured information in the initial structured information that does not conform to the standard output format based on the verification module comprises:

8. The method of claim 7, wherein the obtaining the user demand information output by the demand analysis model based on the structured information conforming to the standard output format and the corrected structured information comprises:

9. The method of claim 8, wherein the preset stop condition includes that the initial structured information is not in accordance with the standard output format after correction.

10. The method of claim 8, wherein the obtaining the user demand information output by the demand analysis model based on the structured information conforming to the standard output format and the revised structured information comprises:

11. The method of claim 1, wherein the obtaining, by a processor in the server, reply data output by a target dialog model based on user-entered user question data, and determining historical session data based on the user question data and the reply data, comprises:

12. The method of claim 1, wherein the target session model is derived based on training functionality provided by the processor, the target session model being trained by:

13. The method of claim 12, wherein the question-answer pair further comprises sample answer data; the step of carrying out parameter adjustment on the dialogue model to be trained based on the predicted reply data and determining the target dialogue model comprises the following steps:

14. The method of claim 13, wherein the parameter adjusting the dialog model to be trained based on the predicted reply data and the sample reply data comprises:

15. The method of claim 14, wherein the probability loss value is derived based on a first formula; the first formula includes:

16. The method of claim 1, wherein the dialog management model is derived based on training functionality provided by the processor, the dialog management model being trained by:

17. The method of claim 16, wherein the obtaining a target feature vector for any of the plurality of sets of dialog histories based on a target language model comprises:

18. A resource demand analysis apparatus for use with a server running a resource processing platform, the apparatus comprising:

19. An electronic device, comprising:

a processor, a memory and a computer program stored on the memory and executable on the processor, the processor implementing the resource demand analysis method of any of claims 1-17 when the program is executed.

20. A readable storage medium, characterized in that instructions in the storage medium, when executed by a processor of an electronic device, enable the electronic device to perform the resource demand analysis method of one or more of claims 1-17.