CN116662533B

CN116662533B - Problem base knowledge point mining method based on AI assistance and teaching service system

Info

Publication number: CN116662533B
Application number: CN202310960956.5A
Authority: CN
Inventors: 黎国权; 朱晖
Original assignee: Guangdong Xinjufeng Technology Co ltd
Current assignee: Guangdong Xinjufeng Technology Co ltd
Priority date: 2023-08-02
Filing date: 2023-08-02
Publication date: 2023-11-03
Anticipated expiration: 2043-08-02
Also published as: CN116662533A

Abstract

The application provides an AI-assisted question library knowledge point mining method and a teaching service system, wherein a first temporary weak knowledge point estimation network generated through competitive training is used for encoding question response data, the question response vector comprises shared vectors corresponding to different training knowledge samples, then a second temporary weak knowledge point estimation network is generated through self-adaptive network training according to the first temporary weak knowledge point estimation network, and a target weak knowledge point estimation network is obtained according to the second temporary weak knowledge point estimation network, so that the target weak knowledge point estimation network can also extract the shared vectors corresponding to different training knowledge samples for the loaded question response data, and the precision of weak knowledge point estimation in different fields can be improved.

Description

Problem base knowledge point mining method based on AI assistance and teaching service system

Technical Field

The embodiment of the application relates to the technical field of artificial intelligence, in particular to an AI-assisted question bank knowledge point mining method and a teaching service system.

Background

Under the condition that the artificial intelligence technology is continuously mature, the personalized self-adaptive learning platform not only can provide personalized resources according to the personalized growth demands of students, but also records, excavates and analyzes the procedural data (such as question answering behavior data and the like) of the students, thereby providing data support for exploring weak knowledge points possibly existing in the students, facilitating the personalized guidance of teachers and schools, improving the learning quality and achieving the aim of reducing burden and enhancing efficiency. For example, in the process of mining weak knowledge points existing in the student question data based on the artificial intelligence technology, network knowledge learning training needs to be performed in advance by combining training samples, then in knowledge learning nodes of the training samples, the situation that part of the monitored information in the field is comprehensive and part of the monitored information in the field is rare may exist, and in this case, how to improve the accuracy of estimating the weak knowledge points in different fields so as to facilitate accurate pushing of the follow-up teaching resource data is a technical problem to be solved currently urgently.

Disclosure of Invention

In order to at least overcome the above-mentioned shortcomings in the prior art, an object of an embodiment of the present application is to provide a method for mining knowledge points of a question bank based on AI assistance and a teaching service system.

According to one aspect of the embodiment of the application, an AI-assisted question bank knowledge point mining method is provided, which comprises the following steps:

acquiring a known training knowledge sample, wherein the known training knowledge sample comprises target field question answering data, and each question answering behavior of the target field question answering data carries associated priori weak knowledge points;

obtaining an unknown training knowledge sample, wherein the unknown training knowledge sample comprises self-adaptive field question answering data, and local question answering behaviors of the self-adaptive field question answering data carry associated priori weak knowledge points;

performing competitive training on a basic weak knowledge point estimation network according to the known training knowledge sample and the unknown training knowledge sample, so that a question answering vector generated by training on the loaded question answering data by the first temporary weak knowledge point estimation network comprises sharing vectors corresponding to different training knowledge samples;

estimating the question answering data in the self-adaptive field according to the first temporary weak knowledge point estimation network, generating estimation data corresponding to each question answering action in the question answering data in the self-adaptive field, and taking the estimation data corresponding to each question answering action as training weak knowledge points corresponding to each question answering action in the question answering data in the self-adaptive field;

Performing competition training on the first temporary weak knowledge point estimation network according to the adaptive field question response data, the priori weak knowledge points corresponding to the local question response behaviors and the training weak knowledge points, and generating a second temporary weak knowledge point estimation network;

and taking the second temporary weak knowledge point estimation network as a basic weak knowledge point estimation network, returning to perform competition training operation on the basic weak knowledge point estimation network according to the known training knowledge sample and the unknown training knowledge sample, and obtaining a target weak knowledge point estimation network according to the second temporary weak knowledge point estimation network in a convergence state when a network convergence condition is reached, wherein the target weak knowledge point estimation network is used for performing weak knowledge point estimation on question answering behaviors in question answering data in various fields so as to push teaching resource data based on weak knowledge point estimation results.

According to an aspect of the embodiment of the present application, there is provided an AI-assisted-based question bank knowledge point mining system, including a processor and a machine-readable storage medium having stored therein machine-executable instructions loaded and executed by the processor to implement the AI-assisted-based question bank knowledge point mining method in any one of the foregoing possible implementations.

According to an aspect of embodiments of the present application, there is provided a computer program product or computer program comprising computer instructions stored in a computer readable storage medium. The computer instructions are read from a computer-readable storage medium by a processor of a computer device, and executed by the processor, cause the computer device to perform the methods provided in the various alternative implementations of the above aspects.

According to the technical scheme provided by the embodiments of the application, competition training is carried out on a basic weak knowledge point estimation network through known training knowledge samples and unknown training knowledge samples, so that a first temporary weak knowledge point estimation network generated by training comprises shared vectors corresponding to different training knowledge samples for question response data coding of a loaded question, then competition training is carried out on the first temporary weak knowledge point estimation network according to the self-adaptive field question response data, prior weak knowledge points corresponding to local question response behaviors and training weak knowledge points, a second temporary weak knowledge point estimation network is generated, the second temporary weak knowledge point estimation network is used as the basic weak knowledge point estimation network to carry out iterative updating, a target weak knowledge point estimation network is generated, and because the first temporary weak knowledge point estimation network generated by the competition training carries out shared vectors corresponding to different knowledge samples for the loaded question response data coding, the second temporary weak knowledge point estimation network is generated by self-adaptive network training according to the first temporary knowledge point estimation network, and the second temporary weak knowledge point estimation network can be used for improving the target weak point estimation network according to the same weak point estimation network, and the target weak point can also be used for improving the shared knowledge data training data.

The method comprises the steps of obtaining target question answering data, loading the target question answering data into a target weak knowledge point estimation network for estimation, wherein the target weak knowledge point estimation network performs competition training on a basic weak knowledge point estimation network according to a known training knowledge sample and an unknown training knowledge sample to generate a first temporary weak knowledge point estimation network, performs weak knowledge point estimation on the unknown training knowledge sample according to the first temporary weak knowledge point estimation network to generate an estimation result, performs competition training on the first temporary weak knowledge point estimation network based on the estimation result and the unknown training knowledge sample to generate a second temporary weak knowledge point estimation network, performs iterative training on the second temporary weak knowledge point estimation network serving as the basic weak knowledge point estimation network, and is obtained according to the second temporary weak knowledge point estimation network in a convergence state; the target weak knowledge point estimation network outputs estimation data corresponding to the target question answering data. The target weak knowledge point estimation network can be used for estimating the target question answering data according to the target weak knowledge point estimation network, and the target weak knowledge point estimation network generated through competition training can be used for extracting shared vectors corresponding to different training knowledge samples from the loaded question answering data, so that even if the target question answering data are the question answering data in different training knowledge samples, the target weak knowledge point estimation network can be used for guaranteeing the accuracy of the estimated data, and the accuracy of estimating the target question answering data is improved.

Drawings

For a clearer description of the technical solutions of the embodiments of the present application, reference will be made to the accompanying drawings, which are needed to be activated in the embodiments, and it should be understood that the following drawings only illustrate some embodiments of the present application, and therefore should not be considered as limiting the scope, and other related drawings can be extracted by those skilled in the art without the inventive effort.

FIG. 1 is a schematic flow chart of a method for mining knowledge points of a question bank based on AI assistance in an embodiment of the application;

fig. 2 is a schematic block diagram of a system for mining knowledge points of a question bank based on AI assistance, which is provided in an embodiment of the present application and is used for implementing the method for mining knowledge points of a question bank based on AI assistance.

Detailed Description

The following description is presented to enable one of ordinary skill in the art to make and use the application and is provided in the context of a particular application and its requirements. It will be apparent to those having ordinary skill in the art that various changes can be made to the disclosed embodiments and that the general principles defined herein may be applied to other embodiments and applications without departing from the principles and scope of the application. Therefore, the present application is not limited to the described embodiments, but is to be accorded the widest scope consistent with the claims.

Fig. 1 is a flowchart of an embodiment of the disclosure of the method for mining knowledge points of a question bank based on AI assistance, and the method for mining knowledge points of a question bank based on AI assistance is described in detail below.

Step S102, a known training knowledge sample is obtained, wherein the known training knowledge sample comprises target field question answering data, and each question answering behavior of the target field question answering data carries associated priori weak knowledge points.

The known training knowledge sample (which may be understood as source domain data) may refer to a question answer data set carrying complete weak knowledge points, that is, each question answer behavior carries corresponding a priori weak knowledge points and a priori weak link positioning data. The topic response data in the target field refers to the topic response data in the known training knowledge sample, namely, each topic response behavior of the topic response data in the target field carries associated prior weak knowledge points. The prior weak knowledge points and prior weak link positioning data which are associated with each question answering behavior refer to training labels which are referred to by known training knowledge samples when network knowledge learning is carried out. The prior weak knowledge points may refer to weak knowledge points existing in the corresponding question answering behaviors, for example, a certain course knowledge chapter, a course knowledge section and the like, and the prior weak link positioning data may refer to specific behavior data existing in the corresponding question answering behaviors, for example, behavior data such as pause, repeated modification and the like existing in the answering process.

Step S104, obtaining an unknown training knowledge sample, wherein the unknown training knowledge sample comprises self-adaptive field question answering data, and the local question answering behavior of the self-adaptive field question answering data carries associated priori weak knowledge points.

The unknown training knowledge sample (which can be understood as target domain data) refers to a question answering data set carrying weak supervision information, that is, only local question answering behaviors carry corresponding priori weak knowledge points and priori weak link positioning data. The self-adaptive field question answering data refers to question answering data in an unknown training knowledge sample, namely local question answering behaviors of the self-adaptive field question answering data carry associated priori weak knowledge points, and the local question answering behaviors can be specific question answering behaviors in the question answering data, for example, question answering behaviors with larger question answering and checking difficulties in the question answering data.

Step S106, performing competition training on the basic weak knowledge point estimation network according to the known training knowledge sample and the unknown training knowledge sample, so that the question answering vector generated by training and used for encoding the loaded question answering data by the first temporary weak knowledge point estimation network comprises shared vectors corresponding to different training knowledge samples.

The competition training refers to performing the competition training in the field of question answering vectors coded by the known training knowledge samples and the unknown training knowledge samples, so that the first temporary weak knowledge point estimation network generated by the training comprises shared vectors corresponding to different training knowledge samples. The basic weak point estimation network refers to the weak point estimation network of the initial network weight. The first temporary weak knowledge point estimation network refers to a weak knowledge point estimation network generated after a round of competition training. The shared vector refers to a vector of invariance of a resolved matching domain when the current first temporary weak knowledge point estimation network extracts the loaded title answer data. Domain invariance refers to the fact that the question response data originate from different knowledge domains, and the question response vector is a common vector.

In some alternative embodiments, the known training knowledge sample and the unknown training knowledge sample may be loaded into the basic weak knowledge point estimation network to perform competition training, that is, the question response data in the loaded known training knowledge sample and the unknown training knowledge sample are respectively encoded, the encoded question response vector is observed, the encoded question response vector is estimated at the same time, the observed data and the estimated data are generated, then the weight information in the basic weak knowledge point estimation network is updated based on the observed data and the estimated data, and when knowledge learning is finished, a first temporary weak knowledge point estimation network is generated, where the first temporary weak knowledge point estimation network encodes the question response vector of the loaded question response data and includes shared vectors corresponding to different training knowledge samples.

The method includes the steps that competition training can be continuously conducted on a basic weak knowledge point estimation network according to known training knowledge samples and unknown training knowledge samples, a target first temporary weak knowledge point estimation network is generated, namely, the question answer data in the loaded known training knowledge samples and the loaded unknown training knowledge samples are respectively encoded, the encoded question answer vector is observed, meanwhile, the encoded question answer vector is estimated, observation data and estimation data are generated, weight information in the basic weak knowledge point estimation network is updated based on the observation data and the estimation data, a basic weak knowledge point estimation network after knowledge learning is generated, the basic weak knowledge point estimation network after knowledge learning is circularly updated as the basic weak knowledge point estimation network, and the target first temporary weak knowledge point estimation network is generated until the addition Loss value corresponding to the observation data and the estimation data is smaller than a set Loss value.

Step S108, estimating the adaptive field question answering data according to the first temporary weak knowledge point estimation network, generating estimation data corresponding to each question answering action in the adaptive field question answering data, and taking the estimation data corresponding to each question answering action as training weak knowledge points corresponding to each question answering action in the adaptive field question answering data.

The training weak knowledge points are weak knowledge points generated by estimating question answering data in the self-adaptive field by using a first temporary weak knowledge point estimation network.

In some alternative embodiments, the adaptive field question answering data in the unknown training knowledge sample may be loaded to the first temporary weak knowledge point estimation network for estimation, so as to generate estimation data corresponding to each question answering action in the adaptive field question answering data, and the estimation data corresponding to each question answering action in the adaptive field question answering data is used as the training weak knowledge point corresponding to each question answering action in the adaptive field question answering data.

And step S110, performing competition training on the first temporary weak knowledge point estimation network according to the prior weak knowledge points and training weak knowledge points corresponding to the self-adaptive field question response data and the local question response behaviors, and generating a second temporary weak knowledge point estimation network.

The second temporary weak knowledge point estimation network is a weak knowledge point estimation network generated by performing competition training on the first temporary weak knowledge point estimation network according to the adaptive field question response data in the unknown training knowledge sample.

In some alternative embodiments, the adaptive field question answering data in the unknown training knowledge sample may be distinguished based on the prior weak knowledge points corresponding to the local question answering behaviors of the adaptive field question answering data and the training weak knowledge points corresponding to the local question answering behaviors in the adaptive field question answering data, so as to generate first class question answering data and second class question answering data which are distinguished and output, and the first temporary weak knowledge point estimation network may be subjected to competitive training based on the first class question answering data and the training weak knowledge points corresponding to the corresponding individual question answering behaviors and the prior weak knowledge points corresponding to the second class question answering data and the corresponding local question answering behaviors, so as to generate the second temporary weak knowledge point estimation network. The method comprises the steps of loading first-class question answering data and second-class question answering data into a first temporary weak knowledge point estimation network for encoding, performing field observation on the encoded question answering vector, analyzing whether the first-class question answering data vector or the second-class question answering data vector to generate observation data, and simultaneously performing weak knowledge point estimation according to the encoded question answering vector to generate weak knowledge point estimation data. And then optimizing weight information in the first temporary weak knowledge point estimation network based on the observation data and the estimation data, and obtaining a second temporary weak knowledge point estimation network after knowledge learning once when the knowledge learning is finished.

For example, the first temporary weak knowledge point estimation network may be circularly competition-trained according to adaptive field topic answering data, a priori weak knowledge point corresponding to local topic answering behavior, and a training weak knowledge point, so as to generate a target second temporary weak knowledge point estimation network. For example: the method comprises the steps of loading adaptive field question response data into a first temporary weak knowledge point estimation network for coding, observing a coded question response vector, generating observation data, estimating the coded question response vector, generating weak knowledge point estimation data, calculating training error information based on the observation data, the estimation data and priori weak knowledge points and training weak knowledge points corresponding to local question response behaviors, backtracking based on the training error information to update weight information in the first temporary weak knowledge point estimation network, generating a first temporary weak knowledge point estimation network after knowledge learning, circularly updating the first temporary weak knowledge point estimation network after knowledge learning as the first temporary weak knowledge point estimation network until the training error information converges, generating a target second temporary weak knowledge point estimation network, and obtaining the target weak knowledge point estimation network according to the target second temporary weak knowledge point estimation network.

And step S112, taking the second temporary weak knowledge point estimation network as a basic weak knowledge point estimation network, returning to perform competition training operation on the basic weak knowledge point estimation network according to the known training knowledge sample and the unknown training knowledge sample, and obtaining a target weak knowledge point estimation network according to the second temporary weak knowledge point estimation network in a convergence state when a network convergence condition is reached, wherein the target weak knowledge point estimation network is used for performing weak knowledge point estimation on question answering behaviors in question answering data in various fields so as to enable teaching resource data to be pushed based on the weak knowledge point estimation result. For example, each target weak knowledge point in the weak knowledge point estimation result may be transmitted to the lecture service system, so that the lecture service system selects lecture resource data corresponding to each target weak knowledge point to push to the corresponding lecture service terminal.

In some alternative embodiments, whether the network is converged may be determined, when the network is converged, the target weak knowledge point estimation network is obtained according to the second temporary weak knowledge point estimation network in the converged state, when the network is not converged, the second temporary weak knowledge point estimation network is used as the basic weak knowledge point estimation network, and the competition training operation is performed on the basic weak knowledge point estimation network according to the known training knowledge sample and the unknown training knowledge sample in a returning manner, so that the network convergence condition is reached. The first temporary weak knowledge point estimation network is adaptively learned to a target domain, a second temporary weak knowledge point estimation network is generated, and then the second temporary weak knowledge point estimation network is adaptively learned to a source domain, so that traversal cycle learning is performed, and when a network convergence condition is reached, the target weak knowledge point estimation network generated after adaptive learning is generated, and further the reliability of the target weak knowledge point estimation network is improved. The target weak knowledge point estimation network carries out weak knowledge point estimation on the question answering behaviors in the question answering data of various fields.

For example, the adaptive field topic response data in the unknown training knowledge sample may not carry any associated a priori weak knowledge points, and a target weak knowledge point estimation network may be generated.

Based on the steps, competition training is carried out on a basic weak knowledge point estimation network through known training knowledge samples and unknown training knowledge samples, so that a first temporary weak knowledge point estimation network generated by training comprises shared vectors corresponding to different training knowledge samples for the question response data coding, then the first temporary weak knowledge point estimation network is subjected to competition training according to the adaptive field question response data, the prior weak knowledge points corresponding to local question response behaviors and training weak knowledge points, a second temporary weak knowledge point estimation network is generated, the second temporary weak knowledge point estimation network is used as the basic weak knowledge point estimation network to carry out iterative updating, a target weak knowledge point estimation network is generated, and the shared vectors corresponding to different training knowledge samples are included in the first temporary weak knowledge point estimation network generated by competition training for the loaded question response data coding, and then the second temporary knowledge point estimation network is generated through the adaptive network according to the first temporary knowledge point estimation network, so that the target knowledge point estimation network can be used for obtaining the target weak point estimation network, and the target knowledge point estimation network can also be used for improving the target weak point estimation network weak point precision, and the target weak point estimation network can be used for improving the target weak point.

Illustratively, step S106, namely performing competitive training on the basic weak point estimation network according to the known training knowledge samples and the unknown training knowledge samples, so that the question response vector generated by training on the loaded question response data by the first temporary weak point estimation network includes the shared vector corresponding to the different training knowledge samples, includes:

step S202, loading the target field question answering data and the self-adaptive field question answering data into a basic weak knowledge point estimation network, respectively extracting a question answering vector by the basic weak knowledge point estimation network according to the target field question answering data and the self-adaptive field question answering data, and carrying out field observation and weak knowledge point estimation according to the question answering vector to generate field observation data and weak knowledge point estimation data.

The field observation data represents whether the question answering data originates from the question answering data in the known training knowledge sample or the question answering data in the unknown training knowledge sample, and comprises observation data corresponding to the target field question answering data and observation data corresponding to the self-adaptive field question answering data. The weak knowledge point estimation data comprises estimation data corresponding to the target field question answering data and estimation data corresponding to the self-adaptive field question answering data.

In some alternative embodiments, the target field question answering data and the adaptive field question answering data may be loaded into the basic weak knowledge point estimation network, and the basic weak knowledge point estimation network extracts the question answering vector according to the target field question answering data and the adaptive field question answering data, generates a vector corresponding to the target field question answering data and a vector corresponding to the adaptive field question answering data, and then performs field observation and weak knowledge point estimation on the vector corresponding to the target field question answering data and the vector corresponding to the adaptive field question answering data, so as to generate estimated data and observed data corresponding to the target field question answering data and obtain estimated data and observed data corresponding to the adaptive field question answering data.

And S204, performing knowledge learning on the basic weak knowledge point estimation network according to the field observation data, the weak knowledge point estimation data, the priori weak knowledge points corresponding to the respective question answering behaviors of the target field question answering data and the priori weak knowledge points corresponding to the local question answering behaviors of the self-adaptive field question answering data, generating a basic weak knowledge point estimation network after knowledge learning, and obtaining a first temporary weak knowledge point estimation network according to the basic weak knowledge point estimation network after knowledge learning.

In some alternative embodiments, the Loss information corresponding to the target domain question answering data may be calculated based on the estimated data corresponding to the target domain question answering data and the a priori weak knowledge points corresponding to the respective question answering behaviors of the target domain question answering data, and the Loss information corresponding to the adaptive domain question answering data may be calculated based on the estimated data corresponding to the adaptive domain question answering data and the a priori weak knowledge points corresponding to the local question answering behaviors of the adaptive domain question answering data. Then, based on the observation data corresponding to the target field question answering data and the target field question answering data, calculating the observation data Loss information corresponding to the target field question answering data for the known training knowledge sample, and meanwhile, based on the observation data corresponding to the self-adaptive field question answering data and the self-adaptive field question answering data for the unknown training knowledge sample, calculating the observation data Loss information corresponding to the self-adaptive field question answering data, and then updating the weight information in the basic weak knowledge point estimation network according to the Loss information corresponding to the target field question answering data, the observation data Loss information and the observation data Loss information corresponding to the self-adaptive field question answering data, and when knowledge learning is finished, generating the basic weak knowledge point estimation network after knowledge learning. The knowledge-learned underlying weak knowledge point estimation network may then be used as a first temporary weak knowledge point estimation network.

Based on the steps, the target field question answering data and the self-adaptive field question answering data are loaded into the basic weak knowledge point estimation network to generate output field observation data and weak knowledge point estimation data, and then the basic weak knowledge point estimation network is updated based on the field observation data and the weak knowledge point estimation data, so that a first temporary weak knowledge point estimation network is generated, and the first temporary weak knowledge point estimation network obtained through competition training can be resolved into shared vectors of different training knowledge samples.

Illustratively, the basic weak knowledge point estimation network includes a basic encoder, a first basic observation unit, and a basic fully connected unit. Step S202 includes:

step S302, inputting the target field question answering data and the self-adaptive field question answering data into a basic encoder for encoding, and generating a basic target field question answering vector and a basic self-adaptive field question answering vector.

The basic encoder refers to an encoder of initial network weight and is used for extracting a vector of the loaded title answer data. The topic answer vector that the encoder in the convergence state can encode includes a common vector of different training knowledge samples. The basic target field question answering vector refers to a vector corresponding to target field question answering data analyzed by the basic encoder. The basic adaptive field question answering vector refers to a vector corresponding to adaptive field question answering data analyzed by the basic encoder.

Illustratively, the base encoder may be made up of a plurality of residual convolutional layers.

And S304, inputting the basic target field question answering vector and the basic self-adaptive field question answering vector into a first basic observation unit for field observation, and generating basic first target field observation data and basic second target field observation data.

The first basic observation unit is an observation unit with initial weight and is used for observing a domain from which an input question answer vector comes, namely a known training knowledge sample or an unknown training knowledge sample. The basic encoder and the first basic observation unit are continuously subjected to competitive training, namely, the learning direction of the basic encoder is that the first basic observation unit cannot judge whether the generated vector is derived from a known training knowledge sample or an unknown training knowledge sample, so that the question response vector which can be encoded by the encoder in a convergence state comprises shared vectors of different training knowledge samples. The basic first target field observation data refers to the result of observing the field question answer vector of the basic target according to the first basic observation unit, and the basic second target field observation data refers to the result of observing the field question answer vector of the basic self-adaption according to the first basic observation unit.

And step S306, inputting the basic target field question answering vector and the basic self-adaptive field question answering vector into a basic full-connection unit to generate basic target field weak knowledge point estimation data and basic self-adaptive field weak knowledge point estimation data.

Wherein the base fully connected unit may be a decoding network.

Step S204 may include:

step S308, calculating the domain training Loss information of the basic target according to the domain weak knowledge point estimation data of the basic target and the prior weak knowledge points corresponding to the response behaviors of the topics, and calculating the basic self-adaptive domain training Loss information according to the domain weak knowledge point estimation data of the basic self-adaptive domain and the prior weak knowledge points corresponding to the response behaviors of the local topics.

The basic target field training Loss information refers to Loss information corresponding to target field question response data obtained when the basic weak knowledge point estimation network performs supervised network learning. The basic self-adaptive field training Loss information refers to Loss information corresponding to self-adaptive field question response data generated when the basic weak knowledge point estimation network performs weak supervision or unsupervised network learning.

And step S310, backtracking propagation is carried out according to the basic target field training Loss information and the basic self-adaptive field training Loss information so as to update the basic encoder and the basic fully-connected unit.

Step S312, calculating target domain observation Loss information according to the known domain corresponding to the basic first target domain observation data and the known training knowledge sample, and calculating adaptive domain observation Loss information according to the adaptive learning domain corresponding to the basic second target domain observation data and the unknown training knowledge sample.

The field question answering data of the known field which is an index is question answering data in a known training knowledge sample, and the self-adaptive learning field is question answering data in an unknown training knowledge sample.

In some alternative embodiments, the target domain observation Loss information may be generated according to a known domain calculation domain Loss value corresponding to the basic first target domain observation data and the known training knowledge sample, and the adaptive domain observation Loss information may be generated according to an adaptive learning domain calculation domain Loss value corresponding to the basic second target domain observation data and the unknown training knowledge sample.

Step S314, updating the basic encoder and the first basic observation unit according to the objective field observation Loss information and the adaptive field observation Loss information, generating a basic weak knowledge point estimation network after knowledge learning, obtaining a first temporary weak knowledge point estimation network according to the basic weak knowledge point estimation network after knowledge learning, and encoding the question response vector of the basic encoder after knowledge learning in the basic weak knowledge point estimation network after knowledge learning for the loaded question response data, wherein the question response vector comprises shared vectors corresponding to different training knowledge samples.

In some alternative embodiments, the objective domain observation Loss information and the adaptive domain observation Loss information may be propagated back according to a gradient descent algorithm to update the base encoder and the first base observation unit, generate a base encoder after knowledge learning, a first base observation unit after knowledge learning, and a base full connection unit after knowledge learning, obtain a base weak point estimation network after knowledge learning according to the base encoder after knowledge learning, the first base observation unit after knowledge learning, and the base full connection unit after knowledge learning, and then use the base weak point estimation network after knowledge learning as a first temporary weak point estimation network. Illustratively, a first temporary weak knowledge point estimation network may be obtained based on the knowledge learned base encoder and the knowledge learned base fully connected unit, and then an initial second temporary weak knowledge point estimation network may be established according to the first temporary weak knowledge point estimation network and the second base observation unit.

Illustratively, step S314 updates the base encoder and the first base observation unit according to the target domain observation Loss information and the adaptive domain observation Loss information, comprising the steps of:

Calculating a target field vector field of target field observation Loss information, calculating an adaptive field vector field of adaptive field observation Loss information, and performing backtracking propagation according to the target field vector field and the adaptive field vector field to update the first basic observation unit; and calculating a first inverse vector field of the target domain vector field, and calculating a second inverse vector field of the adaptive domain vector field, and performing traceback propagation according to the first inverse vector field and the second inverse vector field to update the base encoder.

The target field vector field is obtained by calculating a vector field of target field observation Loss information according to a gradient descent algorithm. The self-adaptive field vector field is obtained by calculating a vector field of the self-adaptive field observation Loss information according to a gradient descent algorithm. The second reverse vector field is generated by inverting the vector field in the target field, and the second reverse vector field is generated by inverting the vector field in the adaptive field.

According to the embodiment, the target field vector field of the target field observation Loss information can be calculated, the adaptive field vector field of the adaptive field observation Loss information is calculated, the first basic observation unit is backtracked and propagated according to the target field vector field and the adaptive field vector field to be propagated, and the first basic observation unit after knowledge learning is generated. Then, a first inverse vector field of the target domain vector field can be calculated, a second inverse vector field of the self-adaptive domain vector field can be calculated, and the base encoder is backtracking propagated according to the first inverse vector field and the second inverse vector field to propagate, so that the base encoder after knowledge learning is generated. Therefore, the basic encoder after knowledge learning can extract the vector with domain invariance when the basic encoder encodes question answer data in different training knowledge samples, and further the knowledge point estimation accuracy is improved.

Illustratively, the base fully-connected unit includes a base positioning subunit and a base output subunit;

step S306, inputting the basic target field question answering vector and the basic self-adaptive field question answering vector into a basic full-connection unit, generating basic target field weak knowledge point estimation data and basic self-adaptive field weak knowledge point estimation data, comprising:

step S402, inputting the field topic answer vector of the basic target into a basic positioning subunit and a basic output subunit respectively, and generating basic first weak link positioning data and basic field weak estimation data.

The basic positioning subunit refers to a positioning subunit of initial network weight, and the positioning subunit is used for positioning weak link data in response data of a subject field topic. The basic output subunit is an output subunit of initial network weight, and the output subunit is used for outputting weak knowledge points in the response data of the subject field topics.

The basic full connection unit comprises a basic positioning subunit and a basic output subunit. And respectively inputting the field topic answer vector of the basic target into a basic positioning subunit and a basic output subunit to generate basic first weak link positioning data and basic field weak estimation data.

And step S404, inputting the basic self-adaptive field question answer vector into a basic positioning subunit and a basic output subunit respectively, and generating basic second weak link positioning data and basic self-adaptive field weak estimation data.

In some alternative embodiments, the basic adaptive field question answer vector may be input into the basic positioning subunit and the basic output subunit respectively at the same time, so as to generate basic second weak link positioning data and basic adaptive field weak estimation data.

Step S306, calculating domain training Loss information of the basic target according to the domain weak knowledge point estimation data of the basic target and the prior weak knowledge points corresponding to the response behaviors of each question, and calculating the basic self-adaptive domain training Loss information according to the domain weak knowledge point estimation data of the basic self-adaptive domain and the prior weak knowledge points corresponding to the response behaviors of the local questions, including:

step S406, calculating a basic first weak link positioning Loss value according to the basic first weak link positioning data and the prior weak link data of each question response behavior of the target field question response data, calculating a basic target field weak estimation Loss value according to the basic target field weak estimation data and the prior weak knowledge points of each question response behavior of the target field question response data, and obtaining basic target field training Loss information according to the basic first weak link positioning Loss value and the basic target field weak estimation Loss value.

The first weak link positioning Loss value of the foundation characterizes the Loss value between the first weak link positioning data of the foundation and the priori weak link data of the corresponding subject field question response data. The basic target field weak estimation Loss value characterizes a Loss value between the basic target field weak estimation data and a priori weak knowledge point of corresponding target field topic response data.

And step S408, calculating a basic second weak link positioning Loss value according to basic second weak link positioning data and priori weak link data of local question answering behaviors of the self-adaptive field question answering data, calculating a basic self-adaptive field weak estimation Loss value according to basic self-adaptive field weak estimation data and priori weak knowledge points of the self-adaptive field question answering data local question answering behaviors, and obtaining basic self-adaptive field training Loss information according to the basic second weak link positioning Loss value and the basic self-adaptive field weak estimation Loss value.

The second weak link positioning Loss value of the foundation characterizes the Loss value between the second weak link positioning data of the foundation and the priori weak link data of the adaptive field question response data, namely the Loss value between the weak link positioning data corresponding to the local question response action and the priori weak link data. The low value of the weak estimation in the basic self-adaptation field characterizes the low value between the weak estimation data in the basic self-adaptation field and the priori weak knowledge points of the question answering data in the self-adaptation field, namely the low value between the weak knowledge point estimation data corresponding to the local question answering action and the priori weak knowledge points.

In some alternative embodiments, basic adaptive domain training Loss information corresponding to adaptive domain question answering data may be calculated, where only local question answering behaviors in the adaptive domain question answering data carry associated a priori weak knowledge points. When the Loss information is calculated, a positioning Loss value is calculated according to weak link positioning data of the question answering behavior of existing prior weak link data and corresponding prior weak link data, an estimated Loss value is calculated according to estimated data of the question answering behavior of existing prior weak knowledge points and corresponding prior weak knowledge points, and then the sum of the positioning Loss value and the estimated Loss value is calculated to generate basic self-adaption field training Loss information.

Illustratively, step S110 performs competitive training on the first temporary weak knowledge point estimation network according to the adaptive field topic answering data, the priori weak knowledge points corresponding to the local topic answering behaviors, and the training weak knowledge points, and generates a second temporary weak knowledge point estimation network, including:

step S502, distinguishing the adaptive field question answering data in the unknown training knowledge sample according to the prior weak knowledge points and the training weak knowledge points corresponding to the local question answering behaviors, and generating positive adaptive question answering data and negative adaptive question answering data.

For example, the adaptive performance (recall rate) corresponding to the adaptive field question response data may be calculated according to the a priori weak knowledge points and the training weak knowledge points corresponding to the local question response behavior, the adaptive field question response data is positive adaptive question response data when the adaptive performance exceeds the set adaptive performance value, and the adaptive field question response data is negative adaptive question response data when the adaptive performance does not exceed the set adaptive performance value.

In some alternative embodiments, the adaptive performance may be calculated according to the a priori weak knowledge points corresponding to the local topic answering behaviors and the training weak knowledge points corresponding to the local topic answering behaviors, that is, the ratio of the number of estimated accuracy in the training weak knowledge points to the sum of the number of estimated accuracy and the number of estimated errors in the training weak knowledge points is calculated.

Step S504, adjusting training weak knowledge points corresponding to the same topic answering behaviors in the training weak knowledge points according to the prior weak knowledge points corresponding to the local topic answering behaviors, generating adjustment weak knowledge points corresponding to the same topic answering behaviors, and obtaining iteration weak knowledge points corresponding to each topic answering behavior in the self-adaptive field topic answering data according to the training weak knowledge points and the adjustment weak knowledge points corresponding to the same topic answering behaviors.

In some alternative embodiments, training weak knowledge points of the same topic response behavior in the adaptive field topic response data may be adjusted according to a priori weak knowledge points of the local topic response behavior in the adaptive field topic response data, so as to generate adjustment weak knowledge points of the local topic response behavior, and then iteration weak knowledge points of each topic response behavior in the adaptive field topic response data are obtained based on the adjustment weak knowledge points of the local topic response behavior and the training weak knowledge points of other non-adjustment topic response behaviors. Namely, a part of weak knowledge points in the iterative weak knowledge points are training weak knowledge points, and a part of weak knowledge points are adjustment weak knowledge points. Thereby enabling the iterative weak knowledge points that can be obtained to be more accurate.

And step S506, performing competition training on the first temporary weak knowledge point estimation network according to the positive adaptive question answering data and the corresponding iterative weak knowledge points and the negative adaptive question answering data and the corresponding priori weak knowledge points of the local question answering behaviors, and generating a second temporary weak knowledge point estimation network.

In some alternative embodiments, the positive adaptive question answering data may be used as the input of the first temporary weak knowledge point estimation network, the iterative weak knowledge point corresponding to the positive adaptive question answering data may be used as the corresponding output weak knowledge point, the negative adaptive question answering data may be used as the input of the first temporary weak knowledge point estimation network, the priori weak knowledge point of the local question answering behavior corresponding to the negative adaptive question answering data may be used as the corresponding output weak knowledge point to perform competition training, and the second temporary weak knowledge point estimation network may be generated.

Therefore, the adaptive field question answering data are distinguished to generate the positive adaptive question answering data and the negative adaptive question answering data, then the first temporary weak knowledge point estimation network is subjected to competition training according to the positive adaptive question answering data and the negative adaptive question answering data to generate the second temporary weak knowledge point estimation network, and therefore the second temporary weak knowledge point estimation network can reduce the distinction between vectors corresponding to the positive adaptive question answering data and vectors corresponding to the negative adaptive question answering data, and the shared vectors between the positive adaptive question answering data and the negative adaptive question answering data can be obtained, and the precision of weak knowledge point estimation can be improved.

Illustratively, performing competitive training on the first temporary weak knowledge point estimation network according to the positive adaptive topic answer data and the corresponding iterative weak knowledge points and the prior weak knowledge points of the negative adaptive topic answer data and the corresponding local topic answer behaviors, and generating a second temporary weak knowledge point estimation network, including:

step S602, positive adaptive question answering data and negative adaptive question answering data are input into a first temporary weak knowledge point estimation network, the first temporary weak knowledge point estimation network encodes the positive adaptive question answering data and the negative adaptive question answering data to generate positive adaptive question answering vectors and negative adaptive question answering vectors, and self-adaptive field observation and self-adaptive weak knowledge point estimation are performed according to the positive adaptive question answering vectors and the negative adaptive question answering data to generate self-adaptive field observation data and self-adaptive weak knowledge point estimation data.

The positive adaptive question answering vector refers to a vector of resolved positive adaptive field question answering data. The negative adaptive topic response vector refers to a vector from which negative adaptive field topic response data is extracted. The adaptive field observation data characterizes whether the question answering data is positive adaptive question answering data or negative adaptive question answering data, including observation data for positive adaptive question answering data and observation data for negative adaptive question answering data. The adaptive weak knowledge point estimation data refers to the estimation data of the object in the adaptive field question answering data, and comprises the estimation data of the negative adaptive question answering data and the estimation data of the positive adaptive question answering data.

In some alternative embodiments, the weight information of the first temporary weak point estimation network may be shared to the second temporary weak point estimation network to generate an initial second temporary weak point estimation network, that is, the first temporary weak point estimation network may be used as an initial second temporary weak point estimation network, then the positive adaptive question answering data and the negative adaptive question answering data are input into the initial second temporary weak point estimation network, the initial second temporary weak point estimation network encodes the positive adaptive question answering data and the negative adaptive question answering data to generate a positive adaptive question answering vector and a negative adaptive question answering vector, and the adaptive field observation and the adaptive weak point estimation are performed according to the positive adaptive question answering vector and the negative adaptive question answering data to generate the adaptive field observation data and the adaptive weak point estimation data.

Step S604, performing knowledge learning on the first temporary weak knowledge point estimation network according to the self-adaptive field observation data, the self-adaptive weak knowledge point estimation data, the iterative weak knowledge points corresponding to the positive adaptive question answering data and the priori weak knowledge points of the local question answering behavior corresponding to the negative adaptive question answering data, generating a first temporary weak knowledge point estimation network after knowledge learning, obtaining a second temporary weak knowledge point estimation network according to the first temporary weak knowledge point estimation network after knowledge learning, wherein the question answering vector of the second temporary weak knowledge point estimation network for encoding the loaded question answering data comprises sharing vectors corresponding to the same type of distinguishing question answering data.

In some alternative embodiments, the network Loss information may be calculated based on the adaptive domain observation data, the adaptive weak knowledge point estimation data, the iterative weak knowledge points corresponding to the positive adaptive question answering data, and the priori weak knowledge points of the local question answering behavior corresponding to the negative adaptive question answering data, and the weight information in the initial second temporary weak knowledge point estimation network may be updated according to the gradient descent algorithm based on the model Loss information, so as to generate the second temporary weak knowledge point estimation network.

Therefore, the initial second temporary weak knowledge point estimation network is subjected to competition training according to the positive adaptive question answering data and the negative adaptive question answering data to generate the second temporary weak knowledge point estimation network, so that the second temporary weak knowledge point estimation network can be used for encoding the loaded question answering data, the question answering vectors of which the types are different and correspond to the question answering data, and weak knowledge point estimation is carried out according to the shared vectors, the distinction between the negative adaptive question answering data and the positive adaptive question answering data is reduced, and the recognition accuracy of the weak knowledge point estimation network is improved.

The first temporary weak point estimation network, illustratively, includes a first encoder, a second base observation unit, and a first fully connected output unit;

step S702, inputting the positive adaptive question answering data and the negative adaptive question answering data into a first temporary weak knowledge point estimation network, the first temporary weak knowledge point estimation network encodes the positive adaptive question answering data and the negative adaptive question answering data to generate a positive adaptive question answering vector and a negative adaptive question answering vector, and performs adaptive field observation and adaptive weak knowledge point estimation according to the positive adaptive question answering vector and the negative adaptive question answering data to generate adaptive field observation data and adaptive weak knowledge point estimation data, including:

Step S702, positive adaptive question answering data and negative adaptive question answering data are input into a first encoder to be encoded, and positive adaptive question answering vectors and negative adaptive question answering vectors are generated.

The first encoder refers to a basic encoder for learning network knowledge in a basic weak knowledge point estimation network. I.e. the weight information in the first encoder is identical to the architecture and weight information of the updated base encoder. The positive adaptive question answering vector refers to a vector corresponding to positive adaptive question answering data analyzed by the first encoder. The negative adaptive question answering vector refers to a vector corresponding to negative adaptive question answering data analyzed by the first encoder.

In some alternative embodiments, the initial second temporary weak point estimation network may be established from the first temporary weak point estimation network, i.e. the first encoder in the first temporary weak point estimation network is the encoder in the initial second temporary weak point estimation network. When the initial second temporary weak knowledge point estimation network is updated, the positive adaptive question response data and the negative adaptive question response data can be input into the first encoder to be encoded, and positive adaptive question response vectors and negative adaptive question response vectors are generated.

Step S704, the positive adaptive question response vector and the negative adaptive question response vector are loaded into the second basic observation unit to conduct field observation, and positive adaptive field observation data and negative adaptive field observation data are generated.

For example, the second basic observation unit refers to an observation unit of initial network weight, that is, an observation unit of initial network weight in the initial second temporary weak knowledge point estimation network, and the architecture of the second basic observation unit may be the same as that of the first basic observation unit, but does not share the updated weight information of the first basic observation unit. The positive adaptation field observation data refers to data obtained by field observation of the positive adaptation question response data according to the second basic observation unit, and the negative adaptation field observation data refers to data obtained by field observation of the negative adaptation question response data according to the second basic observation unit. And then loading the positive adaptive question answering vector and the negative adaptive question answering vector into a second basic observation unit in the initial second temporary weak knowledge point estimation network for field observation, and generating observation data.

Step S706, the positive adaptive question answer vector and the negative adaptive question answer vector are loaded into the first fully-connected output unit to perform weak knowledge point estimation, and positive adaptive weak knowledge point estimation data and negative adaptive weak knowledge point estimation data are generated.

In some alternative embodiments, the first fully-connected output unit may be used as a weak point estimation network in the established initial second temporary weak point estimation network. And then carrying out weak knowledge point estimation on the positive adaptive question answer vector and the negative adaptive question answer vector according to the first full-connection output unit, and generating positive adaptive weak knowledge point estimation data and negative adaptive weak knowledge point estimation data.

Step S604, performing knowledge learning on the first temporary weak knowledge point estimation network according to the self-adaptive field observation data, the self-adaptive weak knowledge point estimation data, the iterative weak knowledge points corresponding to the positive adaptive question answering data, and the priori weak knowledge points corresponding to the local question answering behavior corresponding to the negative adaptive question answering data, to generate a first temporary weak knowledge point estimation network after knowledge learning, and obtaining a second temporary weak knowledge point estimation network according to the first temporary weak knowledge point estimation network after knowledge learning, including:

step S708, positive adaptation training Loss information is calculated according to the iterative weak knowledge points corresponding to the positive adaptation weak knowledge point estimation data and the positive adaptation question response data, and negative adaptation training Loss information is calculated according to the prior weak knowledge points of the local question response behavior corresponding to the negative adaptation weak knowledge point estimation data and the negative adaptation question response data.

The positive adaptation training Loss information characterizes Loss values between iterative weak knowledge points corresponding to the positive adaptation weak knowledge point estimation data and the positive adaptation question answering data. The negative adaptation training Loss information characterizes the estimated data corresponding to the local question answering behavior in the negative adaptation weak knowledge point estimated data and the Loss value before the priori weak knowledge point of the local question answering behavior corresponding to the negative adaptation question answering data.

In some alternative embodiments, loss between iterative weak knowledge points corresponding to the positive adaptive weak knowledge point estimation data and the positive adaptive question answering data can be calculated according to the supervision Loss function, so as to generate positive adaptive training Loss information. Meanwhile, loss between the estimated data of the negative adaptive weak knowledge points and the priori weak knowledge points of the local question answering behaviors corresponding to the negative adaptive question answering data is calculated according to the weak supervision Loss function, and negative adaptive training Loss information is generated, wherein the negative adaptive training Loss information can be generated according to the square sum of Loss values between the estimated data corresponding to the question answering behaviors in the negative adaptive question answering data and the priori weak knowledge points.

Step S800, updating the first encoder and the first full-connection output unit according to the positive adaptation training Loss information and the negative adaptation training Loss information.

In some alternative embodiments, the backward propagation may be performed according to a gradient descent algorithm according to the positive adaptive training Loss information to update the weight information in the first encoder and the first fully-connected output unit, and the backward propagation may be performed according to a gradient descent algorithm according to the negative adaptive training Loss information to update the weight information in the first encoder and the first fully-connected output unit.

Step S802, calculating positive adaptation field observation Loss information according to the positive adaptation field observation data and the positive adaptation field corresponding to the positive adaptation question answering data, and calculating negative adaptation field observation Loss information according to the negative adaptation field observation data and the negative adaptation field corresponding to the negative adaptation question answering data.

The positive adaptation field refers to question response data derived from the positive adaptation question response data in the unknown training knowledge sample. The positive adaptation field observation Loss information characterizes a Loss value between the positive adaptation field observation data and the positive adaptation field. The negative adaptation field refers to the question response data derived from the negative adaptation question response data in the unknown training knowledge sample. The negative adaptation field observation Loss information characterizes Loss values between the negative adaptation field observation data and the negative adaptation field.

In some alternative embodiments, a Loss value between the observation data of each question answering action in the positive adaptive question answering data and the positive adaptive field corresponding to the positive adaptive question answering data may be calculated according to the Loss function and the Loss square value, so as to generate the observation Loss information of the positive adaptive field, and meanwhile, a Loss value between the observation data of the question answering action in the negative adaptive question answering data and the negative adaptive field corresponding to the negative adaptive question answering data may be calculated according to the Loss function and the Loss square value, so as to generate the observation Loss information of the negative adaptive field.

Step S804, updating the first encoder and the second basic observation unit according to the positive adaptation field observation Loss information and the negative adaptation field observation Loss information, generating a first temporary weak knowledge point estimation network after knowledge learning, obtaining a second temporary weak knowledge point estimation network according to the first temporary weak knowledge point estimation network after knowledge learning, and encoding the loaded question answering data by the first encoder after knowledge learning in the first fully-connected output unit after knowledge learning, wherein the question answering vector comprises sharing vectors corresponding to the same type of distinguishing question answering data.

In some alternative embodiments, the weight information in the first encoder and the second basic observation unit may be updated according to the positive adaptation field observation Loss information and the negative adaptation field observation Loss information, the first encoder after knowledge learning and the second basic observation unit after knowledge learning are generated, and the second temporary weak knowledge point estimation network is obtained according to the first encoder after knowledge learning, the second basic observation unit after knowledge learning and the first fully connected output unit after knowledge learning. For example, a second temporary weak knowledge point estimation network may be obtained based on the first encoder after knowledge learning and the first fully connected output unit after knowledge learning, and then the second temporary weak knowledge point estimation network and the first basic observation unit after knowledge learning may be used as the basic weak knowledge point estimation network to continue iterative training.

Illustratively, step S804 updates the first encoder and the second base observation unit according to the positive adaptation field observation Loss information and the negative adaptation field observation Loss information, including:

calculating a positive adaptation field vector field of the positive adaptation field observation Loss information, and calculating a negative adaptation field vector field of the negative adaptation field observation Loss information, and performing backtracking propagation according to the positive adaptation field vector field and the negative adaptation field vector field to update the second basic observation unit; and calculating a positive adaptation reverse vector field of the positive adaptation field vector field, and calculating a negative adaptation reverse vector field of the negative adaptation field vector field, and performing traceback propagation according to the positive adaptation reverse vector field and the negative adaptation reverse vector field to update the first encoder.

In some alternative embodiments, the second basic observation unit may be updated according to the gradient descent algorithm, and then the first encoder is updated after being inverted according to the vector field according to which the second basic observation unit is updated, that is, the competition training is performed, so that the problem response vector of the encoder in the convergence state for encoding the loaded problem response data includes the shared vector of different training knowledge samples and the shared vector of the difference problem response data between the same training knowledge samples, and the precision of weak knowledge point estimation is further improved.

Illustratively, the first fully connected output unit includes a first positioning subunit and a first output subunit;

step S706, loading the positive adaptive topic answer vector and the negative adaptive topic answer vector into the first fully-connected output unit to perform weak knowledge point estimation, generating positive adaptive weak knowledge point estimation data and negative adaptive weak knowledge point estimation data, including:

and step S902, respectively inputting the positive adaptive question answer vector into a first positioning subunit and a first output subunit to generate positive adaptive weak link positioning data and positive adaptive weak estimation data.

And step S904, inputting the negative adaptive question answer vector into the first positioning subunit and the first output subunit respectively, and generating negative adaptive weak link positioning data and negative adaptive weak estimation data.

The first positioning subunit refers to a parameter layer for positioning weak link data in the question response data in the first full-connection output unit, and the first output subunit refers to a parameter layer for outputting weak knowledge points in the question response data in the first full-connection output unit. The positive adaptive weak link positioning data refers to data obtained by positioning weak link data in the response data of the positive adaptive questions, and the positive adaptive weak estimation data refers to data obtained by outputting weak knowledge points in the response data of the positive adaptive questions. The negative adaptation weak link positioning data refers to data obtained by positioning weak link data in the negative adaptation question answering data, and the negative adaptation weak estimation data refers to data obtained by outputting weak knowledge points in the negative adaptation question answering data.

Step S708 may include:

and S906, calculating a positive adaptation weak link positioning Loss value according to the prior weak link data corresponding to the positive adaptation weak link positioning data and the positive adaptation question answering data, calculating a positive adaptation weak estimation Loss value according to the iteration weak knowledge points corresponding to the positive adaptation weak estimation data and the positive adaptation question answering data, and obtaining positive adaptation training Loss information according to the positive adaptation weak link positioning Loss value and the positive adaptation weak estimation Loss value.

In some alternative embodiments, the segmentation Loss value may be calculated according to the supervised Loss function calculation positive-adaptation weak link positioning data and the prior weak link data in the corresponding iterative weak knowledge points, to generate the positive-adaptation weak link positioning Loss value. Meanwhile, a Loss value between the positive adaptive weak estimation data and the priori weak knowledge point is calculated according to the supervised Loss function, the positive adaptive weak estimation Loss value is generated, backtracking can be conducted according to the positive adaptive weak link positioning Loss value to update weight information in the first positioning subunit, and backtracking can be conducted according to the positive adaptive weak estimation Loss value to update weight information in the first output subunit.

Step S908, calculating a weak link locating Loss value of the negative adaptive question answering data according to the weak link locating data of the negative adaptive question and the prior weak link data of the local question answering behavior corresponding to the negative adaptive question answering data, calculating a negative adaptive weak estimation Loss value according to the negative adaptive weak estimation data and the prior weak knowledge point of the local question answering behavior corresponding to the negative adaptive question answering data, and obtaining negative adaptive training Loss information according to the weak link locating Loss value of the negative adaptive question answering data and the negative adaptive weak estimation Loss value.

Illustratively, a specific application embodiment is presented, comprising the steps of:

step S100, obtaining target question response data, loading the target question response data into a target weak knowledge point estimation network for estimation, wherein the target weak knowledge point estimation network performs competition training on a basic weak knowledge point estimation network according to a known training knowledge sample and an unknown training knowledge sample to generate a first temporary weak knowledge point estimation network, performs weak knowledge point estimation on the unknown training knowledge sample according to the first temporary weak knowledge point estimation network to generate an estimation result, performs competition training on the first temporary weak knowledge point estimation network based on the estimation result and the unknown training knowledge sample to generate a second temporary weak knowledge point estimation network, performs iterative training by taking the second temporary weak knowledge point estimation network as the basic weak knowledge point estimation network, and is obtained according to the second temporary weak knowledge point estimation network in a convergence state;

step S200, the target weak knowledge point estimation network outputs estimation data corresponding to the target question answering data.

The target question answering data refers to question answering data which needs to be estimated, and the question answering data can be data in different domains.

In some alternative embodiments, competition training may be performed on the basic weak knowledge point estimation network in advance according to the known training knowledge sample and the unknown training knowledge sample, a first temporary weak knowledge point estimation network is generated, weak knowledge point estimation is performed on the unknown training knowledge sample according to the first temporary weak knowledge point estimation network, an estimation result is generated, competition training is performed on the first temporary weak knowledge point estimation network based on the estimation result and the unknown training knowledge sample, a second temporary weak knowledge point estimation network is generated, the second temporary weak knowledge point estimation network is used as the basic weak knowledge point estimation network to perform iterative training, and a target weak knowledge point estimation network obtained according to the second temporary weak knowledge point estimation network in a convergence state is then deployed according to the target weak knowledge point estimation network. The method comprises the steps that a server obtains target question answering data, the target question answering data are loaded into a target weak knowledge point estimation network to be estimated, the target weak knowledge point estimation network estimates the coded question answering vector through the fact that the coded question answering vector comprises a shared vector, estimated data are generated, and then the estimated data are output.

Fig. 2 illustrates a hardware structural intent of the AI-assisted question bank knowledge point mining system 100 for implementing the AI-assisted question bank knowledge point mining method according to an embodiment of the application, and as shown in fig. 2, the AI-assisted question bank knowledge point mining system 100 may include a processor 110, a machine-readable storage medium 120, a bus 130, and a communication unit 140.

In an alternative embodiment, the AI-based assistance knowledge point mining system 100 can be a single server or a group of servers. The server farm may be centralized or distributed (e.g., the AI-based assisted question bank knowledge point mining system 100 may be a distributed system). In an alternative embodiment, the AI-based auxiliary knowledge point mining system 100 may be local or remote. For example, the AI-assisted-based question bank knowledge point mining system 100 may access information and/or data stored in the machine-readable storage medium 120 via a network. As another example, the AI-based-aided topic library knowledge point mining system 100 can be directly connected to the machine-readable storage medium 120 to access stored information and/or data. In an alternative embodiment, the AI-based-assisted question bank knowledge point mining system 100 may be implemented on a cloud platform. For example only, the cloud platform may include a private cloud, a public cloud, a hybrid cloud, a community cloud, a distributed cloud, an internal cloud, a multi-layer cloud, or the like, or any combination thereof.

The machine-readable storage medium 120 may store data and/or instructions. In an alternative embodiment, the machine-readable storage medium 120 may store data acquired from an external terminal. In an alternative embodiment, the machine-readable storage medium 120 may store data and/or instructions for use by the AI-based assistant database knowledge point mining system 100 to perform or use to accomplish the example methods described herein. In alternative embodiments, machine-readable storage medium 120 may include mass storage, removable storage, volatile read-write memory, read-only memory, and the like, or any combination thereof. Exemplary mass storage devices may include magnetic disks, optical disks, solid state disks, and the like. Exemplary removable memory may include flash drives, floppy disks, optical disks, memory cards, compact disks, tape, and the like.

In a specific implementation, the plurality of processors 110 execute computer executable instructions stored by the machine-readable storage medium 120, so that the processors 110 may execute the method for mining knowledge points of a question bank based on AI assistance according to the above method embodiment, where the processors 110, the machine-readable storage medium 120 and the communication unit 140 are connected through the bus 130, and the processors 110 may be used to control transceiving actions of the communication unit 140.

The specific implementation process of the processor 110 may refer to the above embodiments of the method executed by the system 100 for mining knowledge points of question bank based on AI assistance, and the implementation principle and technical effect are similar, which is not described herein again.

In addition, the embodiment of the application also provides a readable storage medium, wherein computer executable instructions are preset in the readable storage medium, and when a processor executes the computer executable instructions, the method for mining the knowledge points of the question bank based on the AI assistance is realized.

Similarly, it should be noted that in order to simplify the description of the present disclosure and thereby aid in understanding one or more embodiments of the application, various features are sometimes grouped together in a single embodiment, figure, or description thereof. Similarly, it should be noted that in order to simplify the description of the present disclosure and thereby aid in understanding one or more embodiments of the application, various features are sometimes grouped together in a single embodiment, figure, or description thereof.

Claims

1. The method for mining the knowledge points of the question bank based on the AI assistance is characterized by comprising the following steps:

taking the second temporary weak knowledge point estimation network as a basic weak knowledge point estimation network, returning to perform competition training operation on the basic weak knowledge point estimation network according to the known training knowledge sample and the unknown training knowledge sample, and obtaining a target weak knowledge point estimation network according to the second temporary weak knowledge point estimation network in a convergence state when a network convergence condition is reached, wherein the target weak knowledge point estimation network is used for performing weak knowledge point estimation on question answering behaviors in question answering data in various fields so as to push teaching resource data based on weak knowledge point estimation results;

the competition training is performed on the basic weak knowledge point estimation network according to the known training knowledge sample and the unknown training knowledge sample, so that a question response vector generated by training and encoding the loaded question response data by the first temporary weak knowledge point estimation network comprises shared vectors corresponding to different training knowledge samples, and the competition training comprises the following steps:

Loading the target field question answering data and the self-adaptive field question answering data into a basic weak knowledge point estimation network, wherein the basic weak knowledge point estimation network respectively extracts a question answering vector by utilizing the target field question answering data and the self-adaptive field question answering data, and performs field observation and weak knowledge point estimation according to the question answering vector to generate field observation data and weak knowledge point estimation data;

performing knowledge learning on the basic weak knowledge point estimation network according to the field observation data, the weak knowledge point estimation data, the priori weak knowledge points corresponding to the problem response behaviors of the subject field problem response data and the priori weak knowledge points corresponding to the local problem response behaviors of the self-adaptive field problem response data, generating a basic weak knowledge point estimation network after knowledge learning, and obtaining a first temporary weak knowledge point estimation network according to the basic weak knowledge point estimation network after knowledge learning;

the first temporary weak knowledge point estimation network comprises a first encoder, a second basic observation unit and a first full-connection output unit, and performs competition training on the first temporary weak knowledge point estimation network according to the adaptive field question answering data, the priori weak knowledge points corresponding to the local question answering behaviors and the training weak knowledge points, so as to generate a second temporary weak knowledge point estimation network, and the method comprises the following steps:

Distinguishing the adaptive field question answering data in the unknown training knowledge sample according to the prior weak knowledge points and the training weak knowledge points corresponding to the local question answering behaviors, and generating positive adaptive question answering data and negative adaptive question answering data;

adjusting training weak knowledge points corresponding to the same topic answering behaviors in the training weak knowledge points according to the prior weak knowledge points corresponding to the local topic answering behaviors, generating adjustment weak knowledge points corresponding to the same topic answering behaviors, and obtaining iteration weak knowledge points corresponding to each topic answering behavior in the self-adaptive field topic answering data according to the training weak knowledge points and the adjustment weak knowledge points corresponding to the same topic answering behaviors;

loading the positive adaptive question answering data and the negative adaptive question answering data into the first encoder for encoding to generate a positive adaptive question answering vector and a negative adaptive question answering vector; loading the positive adaptive question answering vector and the negative adaptive question answering vector into the second basic observation unit for field observation, and generating positive adaptive field observation data and negative adaptive field observation data;

Loading the positive adaptive question answering vector and the negative adaptive question answering vector into the first fully-connected output unit to perform weak knowledge point estimation, and generating positive adaptive weak knowledge point estimation data and negative adaptive weak knowledge point estimation data;

calculating positive adaptation training Loss information according to the iterative weak knowledge points corresponding to the positive adaptation weak knowledge point estimation data and the positive adaptation question answering data, and calculating negative adaptation training Loss information according to the prior weak knowledge points of the local question answering behaviors corresponding to the negative adaptation weak knowledge point estimation data and the negative adaptation question answering data;

optimizing the first encoder and the first fully-connected output unit according to the positive adaptation training Loss information and the negative adaptation training Loss information;

calculating positive adaptation field observation Loss information according to the positive adaptation field observation data and the positive adaptation field corresponding to the positive adaptation question response data, and calculating negative adaptation field observation Loss information according to the negative adaptation field observation data and the negative adaptation field corresponding to the negative adaptation question response data;

optimizing the first encoder and the second basic observation unit according to the positive adaptation field observation Loss information and the negative adaptation field observation Loss information, generating a first temporary weak knowledge point estimation network after knowledge learning, obtaining a second temporary weak knowledge point estimation network according to the first temporary weak knowledge point estimation network after knowledge learning, and enabling the first encoder after knowledge learning in the first fully-connected output unit after knowledge learning to encode the loaded question response data to obtain a question response vector corresponding to the same type of distinguishing question response data.

2. The AI-assistance-based question bank knowledge point mining method of claim 1, wherein the basic weak knowledge point estimation network comprises a basic encoder, a first basic observation unit and a basic fully-connected unit, wherein the basic fully-connected unit comprises a basic positioning subunit and a basic output subunit;

the method comprises the steps of loading the target field question answering data and the self-adaptive field question answering data into a basic weak knowledge point estimation network, respectively extracting a question answering vector by the basic weak knowledge point estimation network by using the target field question answering data and the self-adaptive field question answering data, carrying out field observation and weak knowledge point estimation according to the question answering vector, and generating field observation data and weak knowledge point estimation data, wherein the method comprises the following steps of:

loading the field question answering data and the self-adaptive field question answering data into the basic encoder for encoding to generate a basic field question answering vector and a basic self-adaptive field question answering vector;

loading the basic target field question answering vector and the basic self-adaptive field question answering vector into the first basic observation unit for field observation, and generating basic first target field observation data and basic second target field observation data;

The field question response vector of the basic target is respectively loaded into the basic positioning subunit and the basic output subunit, and basic first weak link positioning data and basic field weak estimation data are generated;

and loading the basic self-adaptive field question response vector into the basic positioning subunit and the basic output subunit respectively to generate basic second weak link positioning data and basic self-adaptive field weak estimation data.

3. The AI-assisted question bank knowledge point mining method according to claim 2, wherein the learning of knowledge is performed on the basic knowledge point estimation network according to the domain observation data, the weak knowledge point estimation data, the a priori weak knowledge points corresponding to the respective question response behaviors of the subject domain question response data, and the a priori weak knowledge points corresponding to the local question response behaviors of the adaptive domain question response data, to generate a knowledge-learned basic knowledge point estimation network, and the obtaining of the first temporary weak knowledge point estimation network according to the knowledge-learned basic knowledge point estimation network includes:

calculating a basic first weak link positioning Loss value according to the basic first weak link positioning data and the priori weak link data of each question response behavior of the target field question response data, calculating a basic target field weak estimation Loss value according to the basic target field weak estimation data and the priori weak knowledge points of each question response behavior of the target field question response data, and training Loss information according to the basic first weak link positioning Loss value and the basic target field weak estimation Loss value;

Calculating a basic second weak link positioning Loss value according to the basic second weak link positioning data and the prior weak link data of the local question answering behavior of the self-adaptive field question answering data, calculating a basic self-adaptive field weak estimation Loss value according to the basic self-adaptive field weak estimation data and the prior weak knowledge point of the local question answering behavior of the self-adaptive field question answering data, and obtaining basic self-adaptive field training Loss information according to the basic second weak link positioning Loss value and the basic self-adaptive field weak estimation Loss value;

backtracking propagation is carried out according to the basic target field training Loss information and the basic self-adaptive field training Loss information so as to optimize the basic encoder and the basic fully-connected unit;

calculating target domain observation Loss information according to the basic first target domain observation data and the known domain corresponding to the known training knowledge sample, and calculating self-adaptive domain observation Loss information according to the basic second target domain observation data and the self-adaptive learning domain corresponding to the unknown training knowledge sample;

and optimizing the basic encoder and the first basic observation unit by using the target field observation Loss information and the self-adaptive field observation Loss information, generating a basic weak knowledge point estimation network after knowledge learning, and obtaining a first temporary weak knowledge point estimation network according to the basic weak knowledge point estimation network after knowledge learning, wherein a question response vector of the basic encoder after knowledge learning in the basic weak knowledge point estimation network after knowledge learning for coding the answer data of the loaded questions comprises sharing vectors corresponding to different training knowledge samples.

4. The AI-assistance-based question bank knowledge point mining method according to claim 3, wherein the optimizing the base encoder and the first base observation unit using the subject field observation Loss information and the adaptive field observation Loss information includes:

calculating a target field vector field of the target field observation Loss information, calculating an adaptive field vector field of the adaptive field observation Loss information, and performing backtracking propagation according to the target field vector field and the adaptive field vector field to optimize the first basic observation unit;

and calculating a first inverse vector field of the target domain vector field, calculating a second inverse vector field of the adaptive domain vector field, and performing traceback propagation according to the first inverse vector field and the second inverse vector field to optimize the basic encoder.

5. The AI-assistance-based question bank knowledge point mining method according to claim 1, wherein the optimizing the first encoder and the second basic observation unit in accordance with the positive adaptation field observation Loss information and the negative adaptation field observation Loss information includes:

calculating a positive adaptation field vector field of the positive adaptation field observation Loss information, calculating a negative adaptation field vector field of the negative adaptation field observation Loss information, and performing backtracking propagation according to the positive adaptation field vector field and the negative adaptation field vector field to optimize the second basic observation unit;

And calculating a positive adaptation reverse vector field of the positive adaptation field vector field, calculating a negative adaptation reverse vector field of the negative adaptation field vector field, and carrying out backtracking propagation according to the positive adaptation reverse vector field and the negative adaptation reverse vector field so as to optimize the first encoder.

6. The AI-assisted question bank knowledge point mining method of claim 1, wherein the first fully connected output unit comprises a first positioning subunit and a first output subunit; the step of loading the positive adaptive question response vector and the negative adaptive question response vector into the first fully-connected output unit to perform weak knowledge point estimation, and generating the positive adaptive weak knowledge point estimation data and the negative adaptive weak knowledge point estimation data includes:

loading the positive adaptive question answer vector into the first positioning subunit and the first output subunit respectively to generate positive adaptive weak link positioning data and positive adaptive weak estimation data;

and loading the negative adaptive question answer vector into the first positioning subunit and the first output subunit respectively to generate negative adaptive weak link positioning data and negative adaptive weak estimation data.

7. The AI-assisted question bank knowledge point mining method according to claim 6, wherein the computing positive adaptive training Loss information based on the iterative weak knowledge points corresponding to the positive adaptive weak knowledge point estimation data and the positive adaptive question answering data, and computing negative adaptive training Loss information based on the prior weak knowledge points of the local question answering behavior corresponding to the negative adaptive weak knowledge point estimation data and the negative adaptive question answering data, comprises:

calculating a positive adaptation weak link positioning Loss value according to the positive adaptation weak link positioning data and the prior weak link data corresponding to the positive adaptation question answering data, calculating a positive adaptation weak estimation Loss value according to the positive adaptation weak estimation data and the iteration weak knowledge points corresponding to the positive adaptation question answering data, and obtaining the positive adaptation training Loss information according to the positive adaptation weak link positioning Loss value and the positive adaptation weak estimation Loss value;

and calculating a negative adaptive question answering data weak link positioning Loss value according to the negative adaptive weak link positioning data and the prior weak link data of the local question answering behavior corresponding to the negative adaptive question answering data, calculating a negative adaptive weak estimation Loss value according to the negative adaptive weak estimation data and the prior weak knowledge point of the local question answering behavior corresponding to the negative adaptive question answering data, and obtaining the negative adaptive training Loss information according to the negative adaptive question answering data weak link positioning Loss value and the negative adaptive weak estimation Loss value.

8. An AI-assisted-based question bank knowledge point mining system comprising a processor and a machine-readable storage medium having stored therein machine-executable instructions loaded and executed by the processor to implement the AI-assisted-based question bank knowledge point mining method of any of claims 1-7.