CN112069294A

CN112069294A - Mathematical problem processing method, device, equipment and storage medium

Info

Publication number: CN112069294A
Application number: CN202010972954.4A
Authority: CN
Inventors: 邓彬彬; 沙晶; 付瑞吉; 王士进; 魏思
Original assignee: iFlytek Co Ltd
Current assignee: iFlytek Co Ltd
Priority date: 2020-09-16
Filing date: 2020-09-16
Publication date: 2020-12-11
Anticipated expiration: 2040-09-16
Also published as: CN112069294B

Abstract

The application provides a mathematical problem processing method, a device, equipment and a storage medium, wherein the method comprises the following steps: structuring the question stem information of the target mathematic question to obtain a plurality of conditions in specified forms, and forming a condition set by the obtained conditions; determining a structured candidate problem solving step according to the condition set and the rule base, wherein the determined candidate problem solving step forms a candidate problem solving step set, and the structured candidate problem solving step comprises conditions matched with one rule in the condition set and a conclusion corresponding to the conditions; analyzing and predicting each structured problem solving step of the target mathematic problem according to the candidate problem solving step set and the answer of the target mathematic problem; and combining the conditions obtained by structuring the question stem information and the predicted structured problem solving steps into a logic representation of the problem solving process of the target mathematical problem. The method for processing the mathematical problem can obtain the logic representation of the problem solving process of the target mathematical problem, and can obtain better prediction effect by predicting information based on the logic representation.

Description

Mathematical problem processing method, device, equipment and storage medium

Technical Field

The present application relates to the field of natural language processing technologies, and in particular, to a method, an apparatus, a device, and a storage medium for processing a mathematical problem.

Background

In the field of mathematics, there are some mathematical tasks associated with a mathematical problem, such as predicting knowledge points associated with the mathematical problem, predicting the difficulty of the mathematical problem, and the like. At present, the solutions for solving the mathematical tasks related to the mathematical problems are basically based on an information prediction model (neural network model), that is, texts of the mathematical problems are input into the information prediction model established for the mathematical tasks for prediction.

It is easy to find that, when solving the mathematical task, the existing scheme directly uses the text of the mathematical problem as the input of the information prediction model, that is, the information prediction model predicts based on the mathematical problem itself, for example, for the mathematical task of "determining the knowledge points related to the mathematical problem", the information prediction model directly uses the text of the mathematical problem as the basis to predict the knowledge points related to the mathematical problem.

However, when the information prediction model performs prediction based on information included in the mathematical problem itself, the mathematical problem is not actually "understood", and a problem of poor prediction effect may occur when prediction is performed without actually "understanding" the mathematical problem. In order to enable the information prediction model to truly 'understand' the mathematical problem, a problem solving process of the mathematical problem can be used as input, and how to obtain the problem solving process of the mathematical problem is a problem which needs to be solved at present.

Disclosure of Invention

In view of the above, the present application provides a method, an apparatus, a device and a storage medium for processing a mathematical problem, which are used to determine a logical representation of a problem solving process of the mathematical problem, and the technical solution is as follows:

a method of processing mathematical questions comprising:

structuring the question stem information of the target mathematic question to obtain a plurality of conditions in specified forms, and forming a condition set by the obtained conditions;

determining a structured candidate problem solving step according to the condition set and the rule base, wherein the determined candidate problem solving step forms a candidate problem solving step set, the rule base comprises a plurality of rules and conclusions corresponding to the rules respectively, the structured candidate problem solving step comprises a condition matched with one rule in the rule base in the condition set and a conclusion corresponding to the condition, and the conclusion corresponding to the condition is determined according to the conclusion corresponding to the rule matched with the condition;

predicting each structured problem solving step of the target mathematic problem according to the candidate problem solving step set and the answer analysis of the target mathematic problem;

and structuring the question stem information of the target mathematical question to obtain conditions and the predicted solving steps to form a logic representation of the solving process of the target mathematical question.

Optionally, the mathematical problem processing method further includes:

and carrying out information prediction by using the logic representation of the problem solving process of the target mathematical problem and an information prediction model constructed aiming at the specified mathematical task.

Optionally, the step of determining a structured candidate problem solving according to the condition set and the rule base includes:

matching the conditions in the condition set with the rules in the rule base to obtain the rules matched with the conditions in the condition set;

combining the conditions matched with the same rule into a condition subset to obtain a plurality of condition subsets;

and for each condition subset, determining a conclusion corresponding to the condition subset according to the conclusion corresponding to the matched rule, and determining the conditions in the condition subset and the conclusion corresponding to the condition subset as the structured candidate problem solving step.

Optionally, the step of predicting each structured problem solving of the target mathematical problem according to the candidate problem solving step set and the answer analysis of the target mathematical problem includes:

predicting each structured problem solving step of the target mathematic problem by utilizing a pre-established problem solving step prediction model and the candidate problem solving step set and the answer analysis of the target mathematic problem;

and the problem solving step prediction model is obtained by utilizing candidate problem solving steps determined according to the training mathematic problems and the rule base and the answer analyzing training of the training mathematic problems.

Optionally, the step of predicting each structured problem solving of the target mathematical problem by using a pre-established model for problem solving step prediction and the answer analysis of the candidate problem solving step set and the target mathematical problem includes:

and predicting the problem solving steps one by utilizing the prediction model of the problem solving steps, the candidate problem solving step set and the answer analysis of the target mathematical problem, determining whether the problem solving step is the last problem solving step of the target mathematical problem after each problem solving step is predicted, if so, finishing the prediction, if not, determining whether the number of the predicted problem solving steps reaches a preset number threshold, if so, finishing the prediction, and if not, updating the current candidate problem solving step set based on the predicted problem solving step, so that the prediction is carried out based on the updated candidate problem solving step set when the next problem solving step is predicted.

Optionally, the updating the current candidate problem solving step set based on the problem solving step predicted this time includes:

taking the conclusion in the problem solving step predicted this time as a new condition, and adding the new condition into the current condition set;

and updating the current candidate problem solving step set according to the condition set added with the conditions and the rule base.

Optionally, the predicting a problem solving step by using the prediction model of the problem solving step, the current candidate problem solving step set, and the answer analysis of the target mathematical problem, includes:

and determining a target problem solving step serving as a problem solving step predicted at this time from the current candidate problem solving step set by using the problem solving step prediction model, the answer analysis of the target mathematical problem and the previous problem solving step.

Optionally, the step of determining the target problem solution from the current candidate problem solution step set by using the prediction model of the problem solution step, the answer analysis of the target mathematical problem and the previous problem solution step includes:

determining question stem information capable of characterizing the target mathematical question, answer analysis of the target mathematical question and a vector of a previous problem solving step as a first target vector by using the problem solving step prediction model, the question stem characterization vector of the target mathematical question, the characterization vector of the answer analysis of the target mathematical question and the characterization vector of the previous problem solving step;

determining a second target vector by using the problem solving step prediction model, the first target vector and the characterization vector of each analysis step in the answer analysis of the target mathematical problem, wherein the second target vector enhances the answer analysis of the target mathematical problem compared with the first target vector;

and determining a target problem solving step from the current candidate problem solving step set by using the problem solving step prediction model and the second target vector.

Optionally, the determining a second target vector by using the model for predicting the problem solving step, the first target vector, and the characterization vector of each analysis step in the answer analysis of the target mathematical problem includes:

determining weights corresponding to each analysis step in the answer analysis of the target mathematical question by using the problem solving step prediction model, the first target vector and the characterization vector of each step in the answer analysis of the target mathematical question;

weighting and summing the characterization vectors of the analysis steps according to the weights respectively corresponding to the analysis steps by using the problem solving step prediction model to obtain weighted and summed vectors;

and fusing the first target vector and the vector after weighted summation by using the problem solving step prediction model to obtain a second target vector.

Optionally, the step of determining a target problem solution from the current candidate problem solution step set by using the prediction model of the problem solution step and the second target vector includes:

for each candidate problem solving step in the current candidate problem solving step set, determining a feature vector corresponding to the candidate problem solving step by using the prediction model of the problem solving step, the feature vector of the candidate problem solving step and the feature vector of each analysis step in the answer analysis of the target mathematical problem, wherein the feature vector corresponding to the candidate problem solving step is a vector capable of representing the situation of the candidate problem solving step and the situation of correlation between each analysis step in the answer analysis of the target mathematical problem and the candidate problem solving step;

and determining a target problem solving step from the current candidate problem solving step set by utilizing the prediction model of the problem solving step, the feature vector corresponding to each candidate problem solving step in the current candidate problem solving step set and the second target vector.

Optionally, the determining, by using the prediction model of the problem solving step, the feature vector corresponding to each candidate problem solving step in the current candidate problem solving step set, and the second target vector, a target problem solving step from the current candidate problem solving step set includes:

determining weights respectively corresponding to the candidate problem solving steps in the current candidate problem solving step set by utilizing the prediction model of the problem solving steps, the feature vector corresponding to each candidate problem solving step in the current candidate problem solving step set and the second target vector;

and determining a target problem solving step from the current candidate problem solving step set by utilizing the prediction model of the problem solving step and the corresponding weight of each candidate problem solving step in the current candidate problem solving step set.

Optionally, the step of determining the target problem solution from the current candidate problem solution step set by using the prediction model of the problem solution step and the weights corresponding to the candidate problem solution steps in the current candidate problem solution step set includes:

determining a weight screening mode by using the problem solving step prediction model, and screening target weights from weights respectively corresponding to all candidate problem solving steps in the current candidate problem solving step set according to the determined weight screening mode, wherein the weight screening mode comprises random screening and maximum weight screening;

and determining the candidate problem solving step corresponding to the target weight as a target problem solving step.

Optionally, the process of establishing the prediction model of the problem solving step includes:

structuring the question stem information of the training math question to obtain a plurality of conditions in a specified form, and forming a condition set corresponding to the training math question by the obtained conditions;

predicting each structured problem solving step of the training mathematic problems by using a problem solving step prediction model, a condition set corresponding to the training mathematic problems and answer analysis of the training mathematic problems;

and determining the reward corresponding to each problem solving step of the training mathematical problem, and updating the parameters of the problem solving step prediction model according to the reward corresponding to each problem solving step of the training mathematical problem.

Optionally, the determining the reward corresponding to each problem solving step of the training mathematical problem includes:

determining a contribution value of each problem solving step of the training mathematical problem to a prediction result as a first reward according to whether the answer of the training mathematical problem is predicted or not, so as to obtain the first reward corresponding to each problem solving step of the training mathematical problem;

and/or determining the similarity between each solving step of the training mathematical problem and the corresponding analysis step in the answer analysis of the training mathematical problem as a second reward to obtain the second reward corresponding to each solving step of the training mathematical problem;

and/or determining a third reward according to the position information of each problem solving step of the training mathematical problem and the position information of the corresponding analysis step in the answer analysis of the training mathematical problem so as to obtain the third reward corresponding to each problem solving step of the training mathematical problem;

for the solving step of the training mathematic questions, the analyzing step corresponding to the solving step in the answer analysis of the training mathematic questions is the analyzing step most relevant to the solving step.

A mathematical problem processing apparatus comprising: the device comprises a mathematical problem processing module, a candidate problem solving step set determining module, a problem solving step prediction module and a problem solving process determining module;

the mathematical problem processing module is used for structuring the problem stem information of the target mathematical problem to obtain a plurality of conditions in specified forms, and the obtained conditions form a condition set;

the candidate problem solving step set determining module is used for determining a structured candidate problem solving step according to the condition set and the rule base, and the determined candidate problem solving step set is formed by the determined candidate problem solving steps, wherein the rule base comprises a plurality of rules and conclusions corresponding to the rules respectively;

the problem solving step prediction module is used for predicting each structured problem solving step of the target mathematic problem according to the candidate problem solving step set and the answer analysis of the target mathematic problem;

and the problem solving process determining module is used for forming a logic representation of the problem solving process of the target mathematical problem by the conditions obtained by structuring the question stem information of the target mathematical problem and the predicted problem solving steps.

A mathematical problem processing apparatus comprising: a memory and a processor;

the memory is used for storing programs;

the processor is used for executing the program and realizing each step of the mathematical problem processing method.

A readable storage medium, on which a computer program is stored, wherein the computer program, when executed by a processor, implements the steps of the mathematical problem processing method of any one of the above.

According to the scheme, the method for processing the mathematical questions comprises the steps of structuring the question stem information of the target mathematical question to obtain a plurality of conditions in a specified form, forming a condition set by the obtained conditions, determining the structured candidate problem solving steps according to the condition set and a rule base, forming a candidate problem solving step set by the determined candidate problem solving steps, predicting the structured problem solving steps of the target mathematical question according to the candidate problem solving step set and the answer analysis of the target mathematical question, and forming the logic representation of the problem solving process of the target mathematical question by the conditions obtained by structuring the question stem information of the target mathematical question and the predicted problem solving steps. Therefore, the logical representation of the problem solving process of the target mathematical problem can be obtained through the mathematical problem processing method provided by the application, and the logical representation of the problem solving process of the target mathematical problem is a structured text which is easy to understand by the information prediction model (for example, the problem solving steps are all in a form of condition + conclusion), so that the information prediction model can truly understand the target mathematical problem based on the logical representation of the problem solving process of the target mathematical problem, and the information prediction can be more accurately carried out on the target mathematical problem on the basis.

Drawings

In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the embodiments or the prior art will be briefly described below, it is obvious that the drawings in the following description are only embodiments of the present invention, and for those skilled in the art, other drawings can be obtained according to the provided drawings without creative efforts.

FIG. 1 is a schematic flow chart of a mathematical problem processing method according to an embodiment of the present application;

FIG. 2 is a flowchart illustrating a step of determining a structured candidate problem solving according to a condition set and a rule base according to an embodiment of the present application;

FIG. 3 is a flowchart illustrating a step of determining a target problem solution from a current set of candidate problem solution steps using a model for predicting the problem solution step, an answer analysis of the target mathematical problem, and a previous problem solution step according to an embodiment of the present invention;

FIG. 4 is a schematic diagram illustrating a step of determining a target problem solution from a current set of candidate problem solution steps based on a problem solution step prediction model according to an embodiment of the present application;

FIG. 5 is a diagram illustrating a stem characterization vector determination according to an embodiment of the present application;

FIG. 6 is a schematic diagram of a token vector for determining answer resolution provided in an embodiment of the present application;

fig. 7 is a schematic diagram of determining a feature vector of each candidate problem solving step and a feature vector corresponding to each candidate problem solving step according to the embodiment of the present application;

FIG. 8 is a schematic flow chart illustrating a process of creating a prediction model for the problem solving step according to an embodiment of the present application;

FIG. 9 is a schematic structural diagram of a mathematical problem processing apparatus according to an embodiment of the present application;

fig. 10 is a schematic structural diagram of a mathematical problem processing apparatus according to an embodiment of the present application.

Detailed Description

The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments of the present invention, and it is obvious that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present invention.

In order to obtain the problem solving process of the mathematical problem, the inventor carries out research, and the initial thought is as follows:

firstly, a rule system is used for converting a mathematical question described by a natural language into a logic representation, then the logic representation of the mathematical question is input into the logic system, and the logic system automatically deduces the next condition according to the known condition and through a set of priority rules until an answer is solved.

Because the thought adopts a set of manually defined priority algorithm, the thought is controllable and accurate to the seen mathematical questions, but the possibility of misjudgment is greatly increased for the unseen mathematical questions, and the thought is easy to fall into endless loops.

In view of the above problems, the present inventors have continued research and have proposed the following ideas:

and predicting rules used in each step in the answer analysis of the mathematical questions and the deduced conclusion by using a pre-established solution question process prediction model. Wherein, the problem solving process prediction model is a neural network model such as RNN.

During the research process, the inventor finds that the second idea has higher compatibility with new questions than the first idea, but the second idea also has a plurality of defects: firstly, one rule needs to be selected from a complete rule set (generally hundreds of rules), so that the accuracy is reduced; secondly, although the length of the reasoning chain is limited (the length is consistent with the length of the step in answer analysis), an error rule can be predicted, and the result is that the problem solving logic is not smooth; thirdly, in order to obtain a good prediction effect, a large amount of manual labeling is often needed, however, the number of mathematic subjects is millions, and different solutions are provided for each subject, so that the labeling difficulty is very high, and the training samples of the model are usually few due to the large labeling difficulty.

In view of the problems in the above thought, the present inventors have further studied and finally proposed a method for processing a mathematical problem, which can more accurately predict the problem solving process of a target mathematical problem, and the solution of the predicted problem solving process and the solution of answer analysis can be kept as consistent as possible.

The method for processing the mathematical questions can be applied to scenes for information prediction based on the mathematical questions, such as scenes of knowledge point prediction, difficulty prediction and the like, and can be applied to terminals with data processing capacity, such as a PC (personal computer), a notebook computer, a tablet personal computer, a smart phone and the like, and can also be applied to a single server, a plurality of servers or a server cluster. Next, the mathematical problem processing method provided in the present application will be described by the following examples.

First embodiment

Referring to fig. 1, a flow chart of a mathematical topic processing method provided in an embodiment of the present application is shown, where the method may include:

step S101: structuring the question stem information of the target mathematic question to obtain a plurality of conditions in specified forms, and forming a condition set by the obtained conditions.

Specifically, conditions (here, the conditions are described in a mathematical natural language) may be extracted from the target mathematical problem, and then the extracted conditions are converted into a specified form, thereby obtaining conditions in a number of specified forms, which may be regarded as a logical representation of stem information of the target mathematical problem.

It should be noted that the target mathematical topic in the present embodiment is a mathematical natural language, and the purpose of step S101 is to convert the topic stem information described in the mathematical natural language into structured information that is easy for machine understanding.

Illustratively, the stem information of the objective mathematical topic is "known as AB perpendicular to CD, CD perpendicular to DF, AB perpendicular to GH", and the conditions obtained by step S101 are "perpendicular (AB, CD), perpendicular (CD, EF), perpendicular (AB, GH)".

Step S102: and determining a structured candidate problem solving step according to the condition set and the rule base, and forming a candidate problem solving step set by the determined candidate problem solving steps.

The rule base comprises a plurality of rules and conclusions corresponding to the rules respectively. Illustratively, one rule in the rule base is "vertical (arg1, arg2) and vertical (arg2 and arg 3)", the conclusion of the rule is "parallel (arg1 and arg 3)", and the conclusion of the rule indicates "parallel to a straight line vertical to the same straight line". Note that the "specified form" in step S101 may be the same form as the rule in the rule base.

The step of solving the question of the structured candidate comprises a condition matched with one rule in the rule base in the condition set and a conclusion corresponding to the condition, and the conclusion corresponding to the condition is determined according to the conclusion corresponding to the rule matched with the condition.

Illustratively, the conditions in a structured candidate problem solving step are "vertical (AB, CD), vertical (CD, EF)", and the conclusion is "parallel (AB, EF)", the rule is matched with the rule "vertical (arg1, arg2), vertical (arg2, arg 3)" in the rule base, the conclusion in the structured candidate problem solving step is determined according to the conclusion "parallel (arg1, arg 3)" matched with the rule "vertical (arg1, arg2), vertical (arg2, arg 3)", specifically, the conclusion in "parallel (arg1, arg 3)" is replaced by "AB", and "arg 3" is replaced by "EF", so as to obtain the conclusion "parallel (AB, CD)", which corresponds to the condition "vertical (AB, CD), vertical (CD, EF)".

Step S103: and predicting each structured problem solving step of the target mathematic problem according to the candidate problem solving step set and the answer analysis of the target mathematic problem.

The structured solving step of the target mathematical problem also comprises conditions and conclusions corresponding to the conditions.

In one possible implementation, each structured problem solving step of the target mathematical problem can be determined by using a pre-established problem solving step prediction model and the candidate problem solving step set and the answer analysis of the target mathematical problem. The problem solving step prediction model is obtained by utilizing candidate problem solving steps determined according to the training mathematic problems and the rule base and answer analyzing and training of the training mathematic problems.

Step S104: and structuring the question stem information of the target mathematical question to obtain conditions and the predicted problem solving steps to form a logic representation of the problem solving process of the target mathematical question.

In this embodiment, the condition obtained by structuring the question stem information of the target mathematical question and the predicted structured problem solving steps are combined to be used as a logical representation of the problem solving process of the target mathematical question.

Optionally, the mathematical problem processing method provided in this embodiment may further include: and carrying out information prediction by using the logical representation of the problem solving process of the target mathematical problem and an information prediction model constructed aiming at the specified mathematical task. The designated mathematical task may be, but is not limited to, knowledge point prediction, difficulty prediction, and the like, and if the designated mathematical task is knowledge point prediction, the information prediction model is a model for performing knowledge point prediction.

Specifically, considering that the logical representation of the problem solving process of the target mathematical problem includes conditions obtained by structuring the stem information of the target mathematical problem and predicted problem solving steps, the information prediction can be performed by inputting the characterization vector of the conditions obtained by structuring the stem information of the target mathematical problem and the characterization vector of the predicted problem solving steps into the information prediction model constructed for the specified mathematical task.

The method for processing the mathematical questions comprises the steps of structuring the question stem information of the target mathematical question to obtain a plurality of conditions in specified forms, forming a condition set by the obtained conditions, determining a structured candidate problem solving step according to the condition set and a rule base, forming a candidate problem solving step set by the determined candidate problem solving step, predicting each structured problem solving step of the target mathematical question according to the candidate problem solving step set and answer analysis of the target mathematical question, and forming a logic representation of the problem solving process of the target mathematical question by the conditions obtained by structuring the question stem information of the target mathematical question and the predicted problem solving steps. Therefore, the logical representation of the problem solving process of the target mathematical problem can be obtained through the mathematical problem processing method provided by the embodiment of the application, and the logical representation of the problem solving process of the target mathematical problem is a structured text which is easy to understand by the information prediction model (for example, the problem solving steps are all in a form of condition + conclusion), so that the information prediction model can truly understand the target mathematical problem based on the logical representation of the problem solving process of the target mathematical problem, and the information prediction can be more accurately carried out on the basis.

Second embodiment

This embodiment is similar to the "step S102: and determining a structured candidate problem solving step according to the condition set and the rule base, and forming a candidate problem solving step set for introduction by the determined candidate problem solving steps.

Referring to fig. 2, a flow chart illustrating a step of determining a structured candidate problem solving according to a condition set and a rule base is shown, which may include:

step S201: and matching the conditions in the condition set with the rules in the rule base to obtain the rules matched with the conditions in the condition set.

Specifically, for each rule in the rule base, a condition in the condition set may be matched with the rule, and if a condition matched with the rule exists in the condition set, the rule is determined as a target rule, and the condition matched with the rule is determined as a target condition.

It should be noted that the condition matching one rule in the rule base may be a part of the conditions in the condition set, or may be all the conditions in the condition set.

Step S202: and combining the conditions matched with the same rule into a condition subset to obtain a plurality of condition subsets.

Illustratively, the condition set includes a condition a, a condition b and a condition c, wherein the condition a and the condition b match with a rule r1 in the rule base, and the condition b and the condition c match with a rule r2 in the rule base, then the condition a and the condition b matching with the rule r1 constitute a condition subset, and the condition b and the condition c matching with a rule r2 in the rule base constitute a condition subset, so that two condition subsets { a, b } and { b, c } are obtained.

Step S203: and for each condition subset, determining a conclusion corresponding to the condition subset according to the conclusion corresponding to the matched rule, and determining the conditions in the condition subset and the conclusion corresponding to the condition subset as a structured candidate problem solving step.

As with the example above, after the condition subsets { a, b } and { b, c } are obtained, a conclusion corresponding to { a, b } may be determined from a conclusion corresponding to rule r1 matching { a, b } and a conclusion corresponding to { b, c } may be determined from a conclusion corresponding to rule r2 matching { b, c }.

After the conclusion corresponding to { a, b } and the conclusion corresponding to { b, c } are obtained, the conclusion corresponding to the conditions a and b and { a, b } in { a, b } can be determined as a candidate problem solving step, and the conclusion corresponding to the conditions b and c and { b, c } in { b, c } can be determined as a candidate problem solving step.

And forming a candidate problem solving step set by all the obtained candidate problem solving steps.

Third embodiment

The present embodiment introduces the "solution step of predicting each structured target problem by using the pre-established solution step prediction model, the candidate solution step set, and the answer analysis of the target problem" mentioned in the first embodiment.

Specifically, the process of predicting each structured problem solving step of the target mathematical problem by using the pre-established problem solving step prediction model and the candidate problem solving step set and the answer analysis of the target mathematical problem may include: and predicting the problem solving steps one by utilizing a pre-established prediction model of the problem solving steps, a candidate problem solving step set and answer analysis of the target mathematic problem, determining whether the problem solving step is the last problem solving step of the target mathematic problem after each predicted problem solving step, if the problem solving step is the last problem solving step of the target mathematic problem, ending the prediction, otherwise, determining whether the number of the predicted problem solving steps reaches a preset number threshold value, if the number of the predicted problem solving steps reaches the preset number threshold value, ending the prediction, and otherwise, updating the current candidate problem solving step set based on the current predicted problem solving step so as to predict the next problem solving step based on the updated candidate problem solving step set.

It should be noted that, in consideration of the fact that there may be a case where no answer is predicted, in order to avoid entering the dead loop, a number threshold is preset in the present embodiment, and when the number of predicted problem solving steps reaches the preset number threshold, the prediction is ended.

The process of determining whether the problem solving step is the last problem solving step of the target mathematical problem in the process comprises the following steps: and determining whether the conclusion in the problem solving step is the answer of the target mathematical problem according to the problem in the target mathematical problem, if so, determining that the problem solving step is the last problem solving step of the target mathematical problem, otherwise, determining that the problem solving step is not the last problem solving step of the target mathematical problem.

Illustratively, the problem in the objective problem is "prove AB perpendicular to CD", and assuming that the conclusion in a solving step is "perpendicular (AB, CD)", the solving step can be determined as the last solving step of the objective problem.

The process of "updating the current candidate problem solving step set based on the problem solving step predicted this time" in the above process may include: taking the conclusion in the problem solving step predicted this time as a new condition, and adding the new condition into the current condition set; and updating the current candidate problem solving step set according to the condition set and the rule base after the condition is added.

It should be noted that, when a problem solving step is predicted and the problem solving step is not the last problem solving step of the target mathematical problem, the condition set is updated based on the currently predicted problem solving step, and then the candidate problem solving step set is updated based on the updating of the condition set.

Assuming that the predicted first solving step is condition 1+ conclusion 1 and it is not the last solving step of the target mathematic problem, adding conclusion 1 to the initial condition set (the initial condition set is the condition obtained directly from the question stem information of the target mathematic problem), then updating the candidate solving step set based on the condition set after adding conclusion 1, when predicting the second solving step, performing prediction based on the updated candidate solving step set, assuming that the predicted second solving step is condition 2+ conclusion 2 and is not the last solving step of the target mathematic problem, adding conclusion 2 to the current condition set (i.e. the condition set after one updating), then updating the candidate solving step set based on the updated condition set, when predicting the third solving step, performing prediction based on the second updated candidate solving step set, others may be analogized.

Fourth embodiment

In the above embodiment, it is mentioned that "the problem solving steps can be predicted one by using the pre-established problem solving step prediction model, the candidate problem solving step set, and the answer analysis of the target mathematical problem", and since the process of predicting each problem solving step is basically the same, the present embodiment introduces the prediction process by taking one problem solving step as an example.

Using the model for predicting the problem solving step, the current candidate problem solving step set, and the answer analysis of the target mathematical problem, the process of predicting a problem solving step may include: and determining a target problem solving step which is used as the problem solving step predicted at this time from the current candidate problem solving step set by utilizing the prediction model of the problem solving step, the answer analysis of the target mathematical problem and the previous problem solving step.

Specifically, please refer to fig. 3, which shows a flow chart of the step of determining the target problem solution from the current candidate problem solution step set by using the prediction model of the problem solution step, the answer analysis of the target mathematical problem and the previous problem solution step, and the flow chart may include:

step S301: and determining question stem information capable of representing the target mathematical question, answer analysis of the target mathematical question and a vector of the previous problem solving step as a first target vector by using the prediction model of the problem solving step, the question stem characterization vector of the target mathematical question, the characterization vector of the answer analysis of the target mathematical question and the characterization vector of the previous problem solving step.

Specifically, as shown in fig. 4, a stem characterization vector of the target mathematical problem, a characterization vector of the answer analysis of the target mathematical problem, and a characterization vector of the previous solving step may be input into the serialized decoding module of the prediction model of the solving step, so as to obtain stem information capable of characterizing the target mathematical problem, the answer analysis of the target mathematical problem, and the first target vector of the previous solving step. The serialized decoding module of the problem solving step prediction model can be, but is not limited to, a recurrent neural network RNN, a long-term memory network LSTM, and the like.

It should be noted that, if the first problem solving step is predicted, the "feature vector of the previous problem solving step" in step S301 may be a preset vector for characterizing the 0 th problem solving step, and if the predicted first problem solving step is not the first problem solving step, the "feature vector of the previous problem solving step" in step S301 is the feature vector of the last predicted problem solving step.

In addition, in step S301, the "stem characterization vector of the target mathematical problem" is a vector capable of characterizing stem information of the target mathematical problem and is determined based on the problem solving step prediction model and the initial condition set, and the "characterization vector of the answer analysis of the target mathematical problem" is a vector capable of characterizing the answer analysis of the target mathematical problem and is determined based on the problem solving step prediction model and the answer analysis of the target mathematical problem.

Specifically, the process of determining the stem characterization vector of the target mathematical problem based on the problem solving step prediction model and the initial condition set may include:

step a1, determining a characterization vector of each condition in the initial condition set based on the problem solving step prediction model and the initial condition set.

Specifically, for each condition in the initial condition set, as shown in fig. 5, first, a first word segmentation module of the condition input problem solving step prediction model performs word segmentation to obtain each word in the condition, then each word in the condition is input to a first word embedding module of the condition input problem solving step prediction model to obtain word vectors corresponding to each word in the condition, and then, the word vectors corresponding to each word in the condition are input to a condition characterization vector determination module of the condition solving step prediction model to obtain a characterization vector of the condition, so as to obtain a characterization vector of each condition in the initial condition set. The condition characterization vector determination module of the prediction model in the problem solving step can be Tree-LSTM, RNN and the like.

Step a2, determining a question stem characterization vector of the target mathematical question based on the problem solving step prediction model and the characterization vector of each condition in the initial condition set.

Specifically, as shown in fig. 5, the characterization vectors of the conditions in the initial condition set are sequentially input to the stem characterization vector determination module of the problem solving step prediction model, so as to obtain the stem characterization vector of the target mathematical problem.

The process of determining a characterization vector for the answer resolution of the target mathematical question based on the solution step prediction model and the answer resolution of the target mathematical question includes:

and b1, determining a characterization vector of each analysis step in the answer analysis of the target mathematic question based on the prediction model of the question solving step and the answer analysis of the target mathematic question.

Specifically, the implementation process of step b1 may include:

step b 1-1: and step division is carried out on the answer analysis of the target mathematical question based on the question solving step prediction model, and each step in the answer analysis of the target mathematical question is obtained.

Specifically, as shown in fig. 6, firstly, the answer analysis of the target mathematical question is input to the second segmentation module of the solution step prediction model to perform segmentation, so as to obtain each word in the answer analysis of the target mathematical question, then each word in the answer analysis of the target mathematical question is input to the second word embedding module, so as to obtain a word vector corresponding to each word in the answer analysis of the target mathematical question, then the word vector corresponding to each word in the answer analysis of the target mathematical question is input to the segmentation module of the solution step prediction model, so as to determine whether each word in the answer analysis of the target mathematical question is a segmentation point, so as to obtain a word which is a segmentation point in the answer analysis of the target mathematical question, and finally, the answer analysis of the target mathematical question is divided according to the word which is the segmentation point.

Step b 1-2: and determining a characterization vector of each analysis step in the answer analysis of the target mathematical question based on the solution step prediction model and each analysis step in the answer analysis of the target mathematical question.

Specifically, as shown in fig. 6, for each analysis step in the answer analysis of the target mathematical problem, firstly, word vectors corresponding to each word in the analysis step are sequentially input into the global sequence acquisition module of the problem solving step prediction model to obtain context vectors corresponding to each word in the problem solving step, and then, the context vectors corresponding to each word in the problem solving step are sequentially input into the step characterization vector determination module of the problem solving step prediction model to obtain characterization vectors of the analysis step, so as to obtain a characterization vector of each analysis step in the answer analysis of the target mathematical problem.

And b2, determining a characterization vector of the answer analysis of the target mathematic question based on the solution question step prediction model and the characterization vector of each analysis step in the answer analysis of the target mathematic question.

Specifically, as shown in fig. 6, the characterization vectors of each step in the answer analysis of the target mathematical question are sequentially input to the analysis characterization vector determination module of the problem solving step prediction model, so as to obtain the characterization vectors of the answer analysis of the target mathematical question.

Step S302: and determining a second target vector by using the prediction model of the problem solving step, the first target vector and the characterization vector of each analysis step in the answer analysis of the target mathematical problem.

And compared with the first target vector, the second target vector enhances the answer analysis of the target mathematical question. Specifically, the process of determining the second target vector by using the predictive model of the problem solving step, the first target vector and the characterization vector of each analysis step in the answer analysis of the target mathematical problem includes:

step S3021, determining a weight corresponding to each analysis step in the answer analysis of the target mathematical question by using the prediction model of the answer step, the first target vector and the characterization vector of each analysis step in the answer analysis of the target mathematical question.

Specifically, as shown in fig. 4, the first target vector and the characterization vector of each parsing step in the answer parsing of the target mathematical question may be input into the first attention module of the prediction model of the solving step, so as to obtain the weight corresponding to each parsing step in the answer parsing of the target mathematical question.

It should be noted that the token vector of each parsing step in the answer parsing of the target mathematical question is obtained through step b1 described above.

And S3022, weighting and summing the characterization vectors of each analysis step in the answer analysis of the target mathematical question according to the weight corresponding to each analysis step in the answer analysis of the target mathematical question by using the prediction model of the answer step to obtain weighted and summed vectors.

And S3023, fusing the first target vector and the weighted and summed vector by using the prediction model in the problem solving step to obtain a second target vector.

Specifically, the prediction model in the problem solving step can be used for performing linear transformation on the weighted summation backward vector, then the vector after the linear transformation is spliced with the first target vector, and the spliced vector is used as the second target vector.

Step S303: and determining a target problem solving step from the current candidate problem solving step set by using the problem solving step prediction model and the second target vector.

Specifically, the process of determining the target problem solving step from the current candidate problem solving step set by using the problem solving step prediction model and the second target vector may include:

step S3031, determining a feature vector corresponding to each candidate problem solving step in the current candidate problem solving step set by using the problem solving step prediction model, the feature vector of each candidate problem solving step in the current candidate problem solving step set and the feature vector of each analysis step in the answer analysis of the target mathematical problem.

The characterization vector of a candidate problem solving step is a vector capable of characterizing the candidate problem solving step and is determined based on the prediction model of the problem solving step and the candidate problem solving step.

Specifically, as shown in fig. 7, the process of determining the token vector of the candidate problem solving step based on the prediction model of the problem solving step and the candidate problem solving step includes: firstly, a third word segmentation module of a prediction model of the candidate problem solving step is input to carry out word segmentation to obtain each word in the candidate problem solving step, then each word in the candidate problem solving step is input to a third word embedding module of the prediction model of the candidate problem solving step to obtain word vectors corresponding to each word in the candidate problem solving step, and then the word vectors corresponding to each word in the candidate problem solving step are sequentially input to a characterization vector determination module of the candidate problem solving step prediction model to obtain the characterization vectors of the candidate problem solving step.

The feature vector corresponding to a candidate problem solving step is a vector capable of representing the conditions of the candidate problem solving step and the relevant conditions of each analysis step in the answer analysis of the target mathematical problem.

Specifically, the implementation process of step S3031 may include:

for each candidate problem solving step in the current set of candidate problem solving steps, performing:

and c1, determining the weight corresponding to each step in the answer analysis of the target mathematic question by using the prediction model of the question solving step, the characterization vector of the candidate question solving step and the characterization vector of each analysis step in the answer analysis of the target mathematic question.

Specifically, as shown in fig. 7, the characterization vector of the candidate problem solving step and the characterization vector of each step in the answer analysis of the target mathematical problem are input into the second attention module of the prediction model of the problem solving step, so as to obtain the weight corresponding to each analysis step in the answer analysis of the target mathematical problem.

And c2, weighting and summing the characterization vectors of each analysis step in the answer analysis of the target mathematical question by using the prediction model of the question solving step according to the weight determined in the step c1 to obtain weighted and summed vectors.

And c3, fusing the weighted and summed vector obtained in the step c2 with the characterization vector of the candidate problem solving step by using a prediction model of the problem solving step, wherein the fused vector is used as the feature vector corresponding to the candidate problem solving step.

The feature vector corresponding to each candidate problem solving step in the current candidate problem solving step set can be obtained through the steps.

Step S3032, determining a target problem solving step from the current candidate problem solving step set by using the prediction model of the problem solving step, the feature vector corresponding to each candidate problem solving step in the current candidate problem solving step set and the second target vector.

Specifically, the implementation process of step S3032 may include:

and d1, determining the weight corresponding to each candidate problem solving step in the current candidate problem solving step set by using the prediction model of the problem solving step, the feature vector corresponding to each candidate problem solving step in the current candidate problem solving step set and the second target vector.

Specifically, as shown in fig. 4, the feature vector and the second target vector corresponding to each candidate problem solving step in the current candidate problem solving step set may be input into the third attention module of the prediction model of the problem solving step, so as to obtain the weight corresponding to each candidate problem solving step in the current candidate problem solving step set.

And d2, determining a target problem solving step from the current candidate problem solving step set by using the prediction model of the problem solving step and the weight corresponding to each candidate problem solving step in the current candidate problem solving step set.

Specifically, as shown in fig. 4, a weight screening manner may be determined by the expansion module based on the problem solving step prediction model, then target weights may be screened from weights respectively corresponding to each candidate problem solving step set in the current candidate problem solving step set according to the determined weight screening manner, and finally, the candidate problem solving step corresponding to the target weights may be determined as the target problem solving step.

In this embodiment, an exploration hyper-parameter a may be set, which indicates that the probability with a may adopt the maximum weight screening method to screen the target weight, and the probability with 1-a may adopt the random screening method to screen the target weight. Two weight screening methods are provided to increase the exploration of other solution methods.

And if the determined weight screening mode is random screening, randomly screening one weight from the weights respectively corresponding to the candidate problem solving step sets in the current candidate problem solving step set to serve as a target weight, and if the determined weight screening mode is maximum weight screening, screening the maximum weight from the weights respectively corresponding to the candidate problem solving step sets in the current candidate problem solving step set to serve as the target weight.

Fifth embodiment

In the above embodiments, it is mentioned that the structured problem solving steps of the target mathematical problem can be predicted by using the pre-established problem solving step prediction model, and the process of establishing the problem solving step prediction model is described in this embodiment.

Referring to fig. 8, a flow chart of the prediction model for the problem solving step is shown, which may include:

step S801: structuring the question stem information of the training math question to obtain a plurality of conditions in a specified form, and forming a condition set corresponding to the training math question by the obtained conditions.

Step S802: and predicting each structured problem solving step of the training mathematic problem by utilizing the problem solving step prediction model, the condition set corresponding to the training mathematic problem and the answer analysis of the training mathematic problem.

The implementation process of predicting each structured problem solving step of the training mathematical problem by using the prediction model of the problem solving step, the condition set corresponding to the training mathematical problem and the answer analysis of the training mathematical problem is similar to the implementation process of "predicting each structured problem solving step of the target mathematical problem by using the pre-established prediction model of the problem solving step, the candidate problem solving step set and the answer analysis of the target mathematical problem", and predicting each structured problem solving step of the target mathematical problem "in the above embodiments, and specific reference may be made to the relevant parts in the above embodiments, which is not described herein again in this embodiment.

Step S803: and determining the reward corresponding to each problem solving step of the training mathematical problem, and updating the parameters of the prediction model of the problem solving step according to the reward corresponding to each problem solving step of the training mathematical problem.

Specifically, the process of determining the reward corresponding to each problem solving step of the training mathematics problem comprises one or more of the following steps e1 to e3, preferably three steps e1 to e 3:

step e1, based on whether the answer of the training mathematic question is predicted, determining the contribution value of each solving step of the training mathematic question to the prediction result as the first reward, so as to obtain the first reward corresponding to each solving step of the training mathematic question.

Specifically, the contribution value of each problem solving step of the training mathematical problem to the prediction result can be determined based on two preset values and the attenuation rule.

Illustratively, the two preset values are 100 and-100, respectively, and if the answer to the training mathematical problem is predicted, the contribution value of each problem solving step of the training mathematical problem to the prediction result can be determined based on 100 and a preset screening rule, for example, the contribution value of the last problem solving step to the prediction result is 100 (the contribution value of the last problem solving step to the prediction result is the largest), the contribution value of the penultimate problem solving step to the prediction result is 100.9 to 90, the contribution value of the penultimate problem solving step to the prediction result is 100.9 to 0.9 to 81, and so on; if the answer to the training mathematical problem is not predicted, the contribution value of each solving step of the training mathematical problem to the prediction result can be determined based on-100 and a preset screening rule, for example, the contribution value of the last solving step to the prediction result is-100, the contribution value of the penultimate solving step to the prediction result is-100 × 0.9 ═ 90, the contribution value of the penultimate solving step to the prediction result is-100 × 0.9 ═ 81, and so on.

And e2, determining the similarity between each solving step of the training mathematical problem and the corresponding analysis step in the answer analysis of the training mathematical problem as a second reward, so as to obtain the second reward corresponding to each solving step of the training mathematical problem.

It should be noted that the similarity between a question solving step of a training mathematical question and a corresponding analysis step in answer analysis of the training mathematical question can reflect the consistency between the question solving step and the corresponding analysis step in answer analysis, and the better the consistency between a predicted question solving step and the corresponding analysis step in answer analysis is, the better the prediction effect is, for this reason, the similarity between the predicted question solving step and the corresponding analysis step in answer analysis is used as one of the bases for updating the model parameters in this embodiment.

And e3, determining a third reward according to the position information of each solving step of the training mathematical problem and the position information of the corresponding analysis step in the answer analysis of the training mathematical problem so as to obtain the third reward corresponding to each solving step of the training mathematical problem.

It should be noted that the purpose of determining the third reward is to enable the model to ensure consistency of the predicted problem solving steps.

The third reward in this embodiment may include two aspects of the reward:

the reward of the first aspect is a reward that characterizes a difference in position between a predicted problem solving step and its corresponding solving step.

For example, 10 problem solving steps are predicted by using the problem solving step prediction model, the answer analysis includes 8 analysis steps in total, and for the 4 th problem solving step in the 10 predicted problem solving steps, if the analysis step corresponding to the predicted problem solving step in the answer analysis is the 6 th analysis step, the reward representing the position difference may be g |4/10-6/8| ═ 0.35g (g is a super parameter).

The reward of the second aspect is a reward that characterizes relative position.

Assuming that the predicted analysis step corresponding to the ith problem solving step is the jth analysis step, the predicted analysis step corresponding to the (i + 1) th problem solving step is the kth analysis step, and the reward of the second aspect is used for restricting the analysis step corresponding to the ith problem solving step to be before the analysis step corresponding to the (i + 1) th problem solving step, namely the jth analysis step to be before the kth analysis step.

For example, for the predicted 4 th problem solving step, it is assumed that the corresponding analysis step is the 5 th analysis step, if the predicted 3 rd problem solving step corresponds to the 6 th analysis step, the analysis step corresponding to the 3 rd problem solving step is located after the analysis step corresponding to the 4 th problem solving step, and therefore a negative reward should be given, based on which the reward characterizing the relative position is-b (b is a super parameter which is greater than 0), if the predicted 3 rd problem solving step corresponds to the 2 nd analysis step, the analysis step corresponding to the 3 rd problem solving step is located before the analysis step corresponding to the 4 th problem solving step, and therefore, a positive reward should be given, based on which the reward characterizing the relative position is b.

It should be noted that, for a problem solving step of a training mathematical problem, when the problem solving step is a candidate problem solving step, a weight corresponding to each analysis step in the answer analysis of the training mathematical problem is determined through the characterization vector of the candidate problem solving step and the characterization vector of each analysis step in the answer analysis of the training mathematical problem, and the analysis step most relevant to the problem solving step is the analysis step with the largest weight.

Sixth embodiment

The embodiment of the present application further provides a mathematical problem processing apparatus, which is described below, and the mathematical problem processing apparatus described below and the mathematical problem processing method described above can be referred to in correspondence.

Referring to fig. 9, a schematic structural diagram of a mathematical topic processing apparatus provided in an embodiment of the present application is shown, which may include: a mathematical problem processing module 901, a candidate problem solving step set determining module 902, a problem solving step predicting module 903 and a problem solving process determining module 904.

The mathematical problem processing module 901 is configured to structure the problem stem information of the target mathematical problem, obtain a plurality of conditions in a specified form, and form a condition set from the obtained conditions.

And a candidate problem solving step set determining module 902, configured to determine a structured candidate problem solving step according to the condition set and the rule base, and form a candidate problem solving step set by the determined candidate problem solving steps.

The rule base comprises a plurality of rules and conclusions corresponding to the rules respectively, a structured candidate problem solving step comprises a condition matched with one rule in the rule base in the condition set and a conclusion corresponding to the condition, and the conclusion corresponding to the condition is determined according to the conclusion corresponding to the rule matched with the condition.

And the problem solving step prediction module 903 is used for predicting each structured problem solving step of the target mathematic question according to the candidate problem solving step set and the answer analysis of the target mathematic question.

And a problem solving process determining module 904, configured to compose a logical representation of a problem solving process of the target mathematical problem by using conditions obtained by structuring the stem information of the target mathematical problem and the predicted problem solving steps.

Optionally, the mathematical problem processing apparatus provided in this embodiment may further include: and an information prediction module.

And the information prediction module is used for performing information prediction by using the logic representation of the problem solving process of the target mathematical problem and an information prediction model constructed aiming at the specified mathematical task.

Optionally, the candidate problem solving step set determining module 902 is specifically configured to match the conditions in the condition set with the rules in the rule base, so as to obtain rules matched with the conditions in the condition set; combining the conditions matched with the same rule into a condition subset to obtain a plurality of condition subsets; and for each condition subset, determining a conclusion corresponding to the condition subset according to the conclusion corresponding to the matched rule, and determining the conditions in the condition subset and the conclusion corresponding to the condition subset as the structured candidate problem solving step.

Optionally, the problem solving step predicting module 903 is specifically configured to predict each structured problem solving step of the target mathematical problem by using a pre-established problem solving step prediction model and the answer analysis of the candidate problem solving step set and the target mathematical problem.

Optionally, the problem solving step prediction module 903 predicts the model using a pre-established problem solving step prediction model, and the candidate problem solving step set and the answer analysis of the target mathematic problem, when predicting each structured problem solving step of the target mathematic problem, in particular, the method is used for predicting the problem solving steps one by utilizing the problem solving step prediction model, the candidate problem solving step set and the answer analysis of the target mathematic problem, and after each step of solving the problem is predicted, determining whether the step of solving the problem is the last step of solving the problem of the target mathematical problem, if so, ending the prediction, if not, determining whether the number of the predicted problem solving steps reaches a preset number threshold value, if so, ending the prediction, if not, updating the current candidate problem solving step set based on the problem solving step predicted this time, so as to predict the next solving problem step based on the updated candidate solving problem step set.

Optionally, the problem solving step predicting module 903 is specifically configured to add a conclusion in the problem solving step predicted this time to the current condition set as a new condition when updating the current candidate problem solving step set based on the problem solving step predicted this time; and updating the current candidate problem solving step set according to the condition set added with the conditions and the rule base.

Optionally, the problem solving step predicting module 903 is specifically configured to determine, when a problem solving step is predicted by using the problem solving step prediction model, the current candidate problem solving step set, and the answer analysis of the target mathematical problem, a target problem solving step serving as the problem solving step predicted this time from the current candidate problem solving step set by using the problem solving step prediction model, the answer analysis of the target mathematical problem, and the previous problem solving step.

Optionally, the problem solving step predicting module 903 is specifically configured to determine, when determining a target problem solving step from a current candidate problem solving step set by using the problem solving step prediction model, the answer analysis of the target mathematical problem, and a previous problem solving step, the problem solving step prediction model, the stem characterization vector of the target mathematical problem, the characterization vector of the answer analysis of the target mathematical problem, and the characterization vector of the previous problem solving step, stem information capable of characterizing the target mathematical problem, the answer analysis of the target mathematical problem, and the vector of the previous problem solving step as a first target vector; determining a second target vector by using the problem solving step prediction model, the first target vector and the characterization vector of each analysis step in the answer analysis of the target mathematical problem, wherein the second target vector enhances the answer analysis of the target mathematical problem compared with the first target vector; and determining a target problem solving step from the current candidate problem solving step set by using the problem solving step prediction model and the second target vector.

Optionally, the problem solving step prediction module 903 is specifically configured to determine, when determining the second target vector by using the problem solving step prediction model, the first target vector, and the characterization vector of each analysis step in the answer analysis of the target mathematical problem, a weight corresponding to each analysis step in the answer analysis of the target mathematical problem by using the characteristic vector of each step in the problem solving step prediction model, the first target vector, and the answer analysis of the target mathematical problem; weighting and summing the characterization vectors of the analysis steps according to the weights respectively corresponding to the analysis steps by using the problem solving step prediction model to obtain weighted and summed vectors; and fusing the first target vector and the vector after weighted summation by using the problem solving step prediction model to obtain a second target vector.

Optionally, when the problem solving step prediction module 903 determines a target problem solving step from the current candidate problem solving step set by using the problem solving step prediction model and the second target vector, the problem solving step prediction module is specifically configured to determine, for each candidate problem solving step in the current candidate problem solving step set, a feature vector corresponding to the candidate problem solving step by using the problem solving step prediction model, the feature vector of the candidate problem solving step, and the feature vector of each analysis step in the answer analysis of the target mathematical problem, where the feature vector corresponding to the candidate problem solving step is a vector capable of characterizing a situation related to the candidate problem solving step and each analysis step in the answer analysis of the target mathematical problem; and determining a target problem solving step from the current candidate problem solving step set by utilizing the prediction model of the problem solving step, the feature vector corresponding to each candidate problem solving step in the current candidate problem solving step set and the second target vector.

Optionally, when the problem solving step prediction module 903 determines the target problem solving step from the current candidate problem solving step set by using the problem solving step prediction model, the feature vector corresponding to each candidate problem solving step in the current candidate problem solving step set, and the second target vector, the problem solving step prediction module is specifically configured to determine the weight corresponding to each candidate problem solving step in the current candidate problem solving step set by using the problem solving step prediction model, the feature vector corresponding to each candidate problem solving step in the current candidate problem solving step set, and the second target vector; and determining a target problem solving step from the current candidate problem solving step set by utilizing the prediction model of the problem solving step and the corresponding weight of each candidate problem solving step in the current candidate problem solving step set.

Optionally, when determining the target problem solving step from the current candidate problem solving step set by using the prediction model of the problem solving step and the weights corresponding to the candidate problem solving steps in the current candidate problem solving step set by using the prediction module 903 of the problem solving step, the prediction module is specifically configured to determine a weight screening manner by using the prediction model of the problem solving step, and screen out the target weights from the weights corresponding to the candidate problem solving steps in the current candidate problem solving step set according to the determined weight screening manner, where the weight screening manner includes random screening and maximum weight screening; and determining the candidate problem solving step corresponding to the target weight as a target problem solving step.

Optionally, the mathematical problem processing apparatus provided in this embodiment may further include: solving the problem step and predicting the model building module.

The problem solving step prediction model building module is used for structuring the question stem information of the training mathematic question to obtain a plurality of conditions in a specified form, and the obtained conditions form a condition set corresponding to the training mathematic question; predicting each structured problem solving step of the training mathematic problems by using a problem solving step prediction model, a condition set corresponding to the training mathematic problems and answer analysis of the training mathematic problems; and determining the reward corresponding to each problem solving step of the training mathematical problem, and updating the parameters of the problem solving step prediction model according to the reward corresponding to each problem solving step of the training mathematical problem.

Optionally, the problem solving step prediction model building module is specifically configured to determine, based on whether an answer to the training mathematical problem is predicted or not, a contribution value of each problem solving step to the prediction result of the training mathematical problem as a first reward when determining the reward corresponding to each problem solving step of the training mathematical problem, so as to obtain the first reward corresponding to each problem solving step of the training mathematical problem; and/or determining the similarity between each solving step of the training mathematical problem and the corresponding analysis step in the answer analysis of the training mathematical problem as a second reward to obtain the second reward corresponding to each solving step of the training mathematical problem; and/or determining a third reward according to the position information of each problem solving step of the training mathematical problem and the position information of the corresponding analysis step in the answer analysis of the training mathematical problem so as to obtain the third reward corresponding to each problem solving step of the training mathematical problem. For the solving step of the training mathematic questions, the analyzing step corresponding to the solving step in the answer analysis of the training mathematic questions is the analyzing step most relevant to the solving step.

According to the mathematical problem processing device provided by the embodiment of the application, the logical representation of the problem solving process of the target mathematical problem can be obtained, and the logical representation of the problem solving process of the target mathematical problem is a structured text (for example, the problem solving steps are all in a form of condition + conclusion) which is easy to be understood by the information prediction model, so that the information prediction model can really understand the target mathematical problem based on the logical representation of the problem solving process of the target mathematical problem, and the information prediction can be more accurately carried out on the basis.

Seventh embodiment

An embodiment of the present application further provides a mathematical problem processing apparatus, please refer to fig. 10, which shows a schematic structural diagram of the mathematical problem processing apparatus, and the mathematical problem processing apparatus may include: at least one processor 1001, at least one communication interface 1002, at least one memory 1003 and at least one communication bus 1004;

in the embodiment of the present application, the number of the processor 1001, the communication interface 1002, the memory 1003, and the communication bus 1004 is at least one, and the processor 1001, the communication interface 1002, and the memory 1003 complete communication with each other through the communication bus 1004;

the processor 1001 may be a central processing unit CPU, or an application Specific Integrated circuit asic, or one or more Integrated circuits configured to implement embodiments of the present invention, etc.;

the memory 1003 may include a high-speed RAM memory, and may further include a non-volatile memory (non-volatile memory) or the like, such as at least one disk memory;

wherein the memory stores a program and the processor can call the program stored in the memory, the program for:

determining a structured candidate problem solving step according to a condition set and a rule base, wherein the determined candidate problem solving step forms a candidate problem solving step set, the rule base comprises conclusions corresponding to a plurality of rules and the plurality of rules respectively, the structured candidate problem solving step comprises a condition matched with one rule in the rule base in the condition set and a conclusion corresponding to the condition, and the conclusion corresponding to the condition is determined according to the conclusion corresponding to the rule matched with the condition;

predicting each structured problem solving step of the target mathematic problem according to the candidate problem solving step set and the answer analysis of the target mathematic problem; and structuring the question stem information of the target mathematical question to obtain conditions and the predicted solving steps to form a logic representation of the solving process of the target mathematical question.

Alternatively, the detailed function and the extended function of the program may be as described above.

Eighth embodiment

Embodiments of the present application further provide a readable storage medium, where a program suitable for being executed by a processor may be stored, where the program is configured to:

Finally, it should also be noted that, herein, relational terms such as first and second, and the like may be used solely to distinguish one entity or action from another entity or action without necessarily requiring or implying any actual such relationship or order between such entities or actions. Also, the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising an … …" does not exclude the presence of other identical elements in a process, method, article, or apparatus that comprises the element.

The embodiments in the present description are described in a progressive manner, each embodiment focuses on differences from other embodiments, and the same and similar parts among the embodiments are referred to each other.

The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the present invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein may be applied to other embodiments without departing from the spirit or scope of the invention. Thus, the present invention is not intended to be limited to the embodiments shown herein but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A method for processing a mathematical problem, comprising:

2. The method of processing mathematical problems of claim 1, further comprising:

3. The method for processing mathematical questions according to claim 1 or 2, wherein the step of determining the structured candidate solution questions according to the condition set and the rule base comprises:

4. The method for processing the mathematical problems according to claim 1 or 2, wherein the step of predicting each structured solution of the target mathematical problem according to the candidate solution step set and the answer analysis of the target mathematical problem comprises:

5. The method for processing mathematical problems according to claim 4, wherein the step of predicting each structured solution problem of the target mathematical problem by using a pre-established solution problem step prediction model and the solution problem candidate step set and the solution analysis of the target mathematical problem comprises:

6. The method for processing the mathematical problems according to claim 5, wherein the step of updating the current set of candidate problem solving steps based on the currently predicted problem solving step comprises:

7. The method for processing mathematical problems according to claim 5, wherein the step of predicting a problem solving using the model for predicting a problem solving step, the current set of candidate problem solving steps, and the answer analysis of the target mathematical problem comprises:

8. The method of claim 7, wherein the step of determining the target problem solving from the current set of candidate problem solving steps using the prediction model of the problem solving step, the answer analysis of the target mathematical problem, and the previous problem solving step comprises:

9. The method of claim 8, wherein the determining a second target vector using the solution step prediction model, the first target vector, and the characterization vector for each analysis step in the analysis of the answer to the target mathematical question comprises:

10. The method of claim 8, wherein said step of determining a target problem solution from a current set of candidate problem solution steps using said problem solution step prediction model and said second target vector comprises:

11. The method for processing mathematical problems according to claim 10, wherein the step of determining the target problem solution from the current set of candidate problem solution steps using the prediction model of the problem solution step, the feature vector corresponding to each candidate problem solution step in the current set of candidate problem solution steps, and the second target vector comprises:

12. The method of claim 11, wherein the step of determining the target problem solving from the current set of candidate problem solving steps using the model for predicting the problem solving steps and the weights corresponding to the candidate problem solving steps in the current set of candidate problem solving steps comprises:

13. The method of claim 4, wherein the process of creating the predictive model of the problem solving step comprises:

14. The method of processing mathematical problems according to claim 13, wherein the determining of the reward corresponding to each problem solving step of the training mathematical problem comprises:

15. A mathematical problem processing apparatus, comprising: the device comprises a mathematical problem processing module, a candidate problem solving step set determining module, a problem solving step prediction module and a problem solving process determining module;

16. A mathematical problem processing apparatus, comprising: a memory and a processor;

the memory is used for storing programs;

the processor is configured to execute the program to implement the steps of the mathematical problem processing method according to any one of claims 1 to 14.

17. A readable storage medium on which a computer program is stored, the computer program, when executed by a processor, implementing the steps of the mathematical problem processing method according to any one of claims 1 to 14.