Title recommendation method and device, apparatus, and storage medium

Abstract

The embodiment of the application discloses a question recommendation method and device, equipment and a storage medium. At least, the characteristic representation of a question title and the learning condition characteristic representation of a target object in a predetermined historical duration are obtained. The characteristic representation of the question title is obtained by training at least according to the question title and correct answers given by a plurality of objects when answering the question title. The learning condition characteristic representation is obtained by training at least according to a plurality of questions answered by the target object in the predetermined historical duration. The question title is classified according to the characteristic representation of the question title and the learning condition characteristic representation of the target object, and a classification result is obtained. The classification result represents whether the question title is recommended to the target object. The personalized recommendation for the target object is realized, and the accuracy of the question recommendation is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of artificial intelligence, and more particularly, to a question recommendation method and device, equipment and storage medium. BACKGROUND

[0002] The current question recommendation method mainly includes the following two ways: one way is to find the questions related to the title of the user specified goal, for example, obtaining the text features of the title of the goal, finding the questions similar to the text features of the title of the goal for recommendation, or obtaining the knowledge points of the title of the goal, finding the questions with the same or similar knowledge points as the title of the goal for recommendation; another way is to search and click behavior according to the user, the questions frequently viewed by the user are regarded as the questions interested by the user, if two users have the same questions interested, the different questions interested by the two users are recommended to each other.

[0003] The above question recommendation method cannot make personalized question recommendation for users. SUMMARY

[0004] Therefore, the present application provides a question recommendation method, device, equipment and storage medium to make personalized question recommendation for users.

[0005] In order to achieve the above purpose, the present scheme is as follows:

[0006] A question recommendation method, comprising:

[0007] At least obtaining the feature representation of the title of the goal and the learning situation feature representation of the target object in the predetermined historical duration; the feature representation of the title of the goal is trained at least according to the title of the goal and the correct answers given by the answers of the title of the goal by a plurality of objects, and the learning situation feature representation is trained at least according to the plurality of questions answered by the target object in the predetermined historical duration;

[0008] Classifying the title of the goal at least according to the feature representation of the title of the goal and the learning situation feature representation, obtaining the classification result, and the classification result represents whether to recommend the title of the goal to the target object.

[0009] The above method, preferably, the process of training at least according to the title of the goal and the correct answers given by the answers of the title of the goal by a plurality of objects, comprising:

[0010] Randomly covering the target content in the initial training sample to obtain the target training sample; the target content includes the title of the goal and the correct answer given by the answer of the title of the goal by an object;

[0011] inputting the target training sample into the language representation model to obtain a prediction result output by the language representation model, the prediction result representing covered content in the target training sample;

[0012] updating parameters of the language representation model, aiming at the prediction result approaching to covered content in an initial training sample corresponding to the target training sample;

[0013] determining a target feature representation of the target according to feature vectors output by the last N layers of the language representation model at the end of the training.

[0014] The method, preferably, determining the target feature representation according to the features output by the last N layers of the language representation model comprises:

[0015] calculating a mean value of the feature vectors output by the N layers to obtain the target feature representation.

[0016] The method, preferably, the process of obtaining the prediction result by the language representation model comprises:

[0017] obtaining first initial vector representations of each sentence of the target in the target training sample and second initial vector representations of each answer step of the correct answer in the target training sample;

[0018] fusing each second initial vector representation and each first initial vector representation based on an attention mechanism through a first attention module of the language representation model to obtain a first fused feature representation;

[0019] obtaining the prediction result based on at least the first fused feature representation.

[0020] The method, preferably, the initial training sample further comprises first attribute information of an object providing the correct answer; and obtaining the prediction result based on at least the first fused feature representation comprises:

[0021] obtaining a first semantic representation of the first attribute information;

[0022] obtaining the prediction result based on the first fused feature representation and the first semantic representation.

[0023] The method, preferably, obtaining the prediction result based on the first fused feature representation and the first semantic representation comprises:

[0024] performing at least one fusion processing on the first fused feature representation and the first semantic representation to obtain at least one fusion result;

[0025] Based on the at least one fusion result, the prediction result is obtained.

[0026] Preferably, the process of obtaining the learning situation feature representation according to at least the questions answered by the target object within the predetermined historical time length includes:

[0027] The first type of training sample is obtained, and the first type of training sample at least includes the questions answered by the target object within the predetermined historical time length and the answers given by the target object.

[0028] The first type of training sample is input into the scoring model, so that the scoring model obtains third initial vector representations of each sentence of the questions in the first type of training sample and fourth initial vector representations of each answer step of the answers in the first type of training sample; the second attention module is used to fuse the third initial vector representations and the fourth initial vector representations based on an attention mechanism to obtain second fusion feature representations; and a scoring result is obtained based on at least the second fusion feature representations, the scoring result representing the scores of the questions in the first type of training sample; initial parameters of the second attention module are determined based on parameters of the first attention module when the language representation model is trained.

[0029] Parameters of the scoring model are updated with the scoring result tending to be close to actual scores of the questions in the first type of training sample as a target.

[0030] When the training is completed, a learning situation feature representation is determined according to a feature vector output by the last H-layer network of the scoring model.

[0031] Preferably, the first type of training sample further includes second attribute information of the target object; and the scoring result is obtained based on at least the second fusion feature representations, including:

[0032] A second semantic representation of the second attribute information is obtained.

[0033] The scoring result is obtained based on the second fusion feature representations and the second semantic representation.

[0034] Preferably, the method further includes:

[0035] The second type of training sample is obtained, and the second type of training sample at least includes the questions answered by the target object within the predetermined historical time length and the answers given by the target object.

[0036] A random answer step in the answers given by the target object is selected as a target answer step.

[0037] inputting the second type of training sample into the step judgment model, so that the step judgment model obtains a fifth initial vector representation of each sentence of a question in the second type of training sample and a sixth initial vector representation of each answer step of an answer in the second type of training sample; performing attention mechanism-based fusion on each fifth initial vector representation and each sixth initial vector representation through a third attention module to obtain a third fusion feature representation; obtaining a judgment result based at least on the third fusion feature representation, the judgment result representing whether the target answer step is correct; initial parameters of the third attention module are parameters of the first attention module when the language representation model training is completed;

[0038] updating parameters of the step judgment model with the judgment result approaching an actual judgment result of the target answer step as a target;

[0039] initial parameters of the second attention module are parameters of the third attention module after the step judgment model training is completed.

[0040] The above method, preferably, the second type of training sample further includes second attribute information of the target object; the at least based on the third fusion feature representation to obtain the judgment result, comprising:

[0041] obtaining a second semantic representation of the second attribute information;

[0042] obtaining the judgment result based on the third fusion feature representation and the second semantic representation.

[0043] The above method, preferably, further comprising: obtaining a second semantic representation of the second attribute information of the target object;

[0044] The at least according to the feature representation of the target subject and the learning situation feature representation to classify the target subject to obtain a classification result, comprising:

[0045] According to the feature representation of the target subject, the learning situation feature representation and the second semantic representation, the target subject is classified to obtain a classification result.

[0046] The above method, preferably, the target subject is any one of the candidate question set corresponding to the target object; the candidate question set is determined by at least one of the following ways:

[0047] Method one: the question done wrong by the target object is added to the candidate question set as a candidate question;

[0048] The second way is: for each question in which the target object makes a mistake, a first similarity between the question and the first question is calculated according to the feature representation of the question and the feature representation of the first question; if the first similarity meets a condition, the first question is added to the candidate question set as a candidate question;

[0049] The third way is: a second similarity between the target object and a non-target object is calculated according to the feature representation of the target object and the feature representation of the non-target object; if the second similarity meets a condition, a question in a question bank corresponding to the non-target object is added to the candidate question set as a candidate question.

[0050] The above method preferably further comprises:

[0051] obtaining a first score rate of the target object and a second score rate of the non-target object;

[0052] if the second similarity meets a condition, if the second score rate is higher than the first score rate, and a difference between the second score rate and the first score rate meets a condition, a question in a question bank corresponding to the non-target object is added to the candidate question set as a candidate question; otherwise, a question in a question bank corresponding to the non-target object is prohibited from being added to the candidate question set as a candidate question.

[0053] A question recommendation device comprises:

[0054] an obtaining module configured to obtain at least a feature representation of a target question and a learning situation feature representation of a target object within a predetermined historical time length; the feature representation of the target question is trained according to at least the target question and correct answers given by a plurality of objects to the target question, and the learning situation feature representation is trained according to at least a plurality of questions answered by the target object within the predetermined historical time length;

[0055] a classification module configured to classify the target question according to at least the feature representation of the target question and the learning situation feature representation, to obtain a classification result, the classification result representing whether the target question is recommended to the target object.

[0056] A question recommendation device comprises a memory and a processor;

[0057] The memory is configured to store a program;

[0058] The processor is configured to execute the program to implement each step of the question recommendation method according to any one of the above.

[0059] A computer readable storage medium, having stored thereon a computer program, which, when executed by a processor, implements each step of the question recommendation method according to any one of the above.

[0060] From the above technical solution, the question recommendation method, device, equipment and storage medium provided by the embodiment of the application can obtain at least the question characteristic representation of the target question and the learning characteristic representation of the target object in a predetermined historical time length; wherein, the question characteristic representation is trained at least according to the target question and the correct answers given by the answers of the target question by a plurality of objects, and the learning characteristic representation is trained at least according to the plurality of questions answered by the target object in the predetermined historical time length; the target question is classified at least according to the question characteristic representation and the learning characteristic representation of the target object, and a classification result is obtained, which represents whether the target question is recommended to the target object. Based on the application, the target question is classified according to the question characteristic representation and the learning characteristic representation, and a classification result is obtained. When the classification result represents that the target question is recommended to the target object, the target question is recommended to the target object. Moreover, since the classification result is obtained based on the question characteristic representation and the learning characteristic representation of the target object in the predetermined historical time length, the personalized recommendation for the target object is realized, thereby improving the accuracy of the question recommendation. BRIEF DESCRIPTION OF DRAWINGS

[0061] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the drawings needed in the embodiments or the prior art description will be briefly introduced below. Obviously, the drawings in the following description are only embodiments of the application, and those skilled in the art can obtain other drawings according to the provided drawings without creative labor.

[0062] Figure 1 An implementation flowchart of the question recommendation method disclosed by the embodiment of the application;

[0063] Figure 2 An implementation flowchart of training at least according to the target question and the correct answers given by the answers of the target question by a plurality of objects, disclosed by the embodiment of the application;

[0064] Figure 3 An implementation flowchart of obtaining a prediction result by the language representation model disclosed by the embodiment of the application;

[0065] Figure 4 An implementation flowchart of obtaining a prediction result based at least on the first fusion feature representation disclosed by the embodiment of the application;

[0066] Figure 5An implementation flowchart of obtaining a prediction result by a language representation model disclosed by an embodiment of the present application;

[0067] Figure 6 An implementation flowchart of obtaining a learning characteristic representation of a target object by training the target object according to a plurality of questions answered by the target object within a predetermined historical time length, disclosed by an embodiment of the present application;

[0068] Figure 7 An implementation flowchart of obtaining a score result by a scoring model disclosed by an embodiment of the present application;

[0069] Figure 8 An implementation flowchart of obtaining a judgment result by a step judgment model disclosed by an embodiment of the present application;

[0070] Figure 9 An implementation flowchart of obtaining a classification result by a classification model disclosed by an embodiment of the present application;

[0071] Figure 10 A structural schematic diagram of a question recommendation device disclosed by an embodiment of the present application;

[0072] Figure 11 A hardware structural block diagram of a question recommendation device disclosed by an embodiment of the present application. DETAILED DESCRIPTION

[0073] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all other embodiments obtained by a person of ordinary skill in the art without creative labor fall within the scope of protection of the present application.

[0074] The prior art cannot achieve personalized recommendation for a target object (such as a student or other object in need of recommended questions) when recommending questions. In order to achieve personalized recommendation for a target object, the present application is proposed.

[0075] As shown in Figure 1 An implementation flowchart of a question recommendation method provided by an embodiment of the present application can include:

[0076] Step S101: Obtain at least a question characteristic representation and a learning characteristic representation of a target object within a predetermined historical time length.

[0077] The question characteristic representation is trained at least according to a question and correct answers given by a plurality of objects to the question, and the learning characteristic representation is trained at least according to a plurality of questions answered by the target object within a predetermined historical time length.

[0078] The target question can be any question in the question library; the target object is any object to be recommended a question. As an example, the question can be a math question, or a question in other subjects, such as a physics question, a chemistry question, etc.

[0079] Taking the question recommendation for a certain student (denoted as the target student) as an example, a question is usually answered by many students, and when answering the question, all the students can give the correct answer, or only part of the students can give the correct answer. In the present application, when the feature representation of the target question is trained, only the target question and the correct answers given by a plurality of (denoted as M for ease of description) students to the target question are used. Each of the M students gives one correct answer to the target question, and the correct answers given by different students to the target question can be the same (same solving method) or different (different solving method). As an example, the value of M can be 500, or 1000, etc., which is not limited in the present application, and can also be other values, such as 800, 3000, etc. Training the feature representation of the target question based on at least the target question and the correct answers given by a plurality of objects to the target question can make the feature representation of the target question accurately represent the information related to the question, such as the question, the question type, and the solving method, etc.

[0080] Each student will do many questions in a period of time, some of which will be correct and some of which will be wrong. In order to obtain the learning situation of the student in a period of time, in the present application, the learning situation feature representation of the target student in a predetermined historical time period is trained using a plurality of questions (including the questions and the answers given by the target student) answered by the target student in the predetermined historical time period. The plurality of questions answered by the target student in the predetermined historical time period include both correct questions and incorrect questions. Training the learning situation feature representation of the target object based on at least the plurality of questions answered by the target object in the predetermined historical time period can make the learning situation feature representation represent the mastery of the target object to various types of questions in the predetermined historical time period.

[0081] Step S101: classifying the target question based on at least the feature representation of the target question and the learning situation feature representation of the target object, to obtain a classification result, which represents whether to recommend the target question to the target object.

[0082] As an example, the feature representation of the target question and the above learning situation feature representation can be input into a pre-trained classification model to obtain the classification result output by the classification model.

[0083] As an example, the classification model can perform attention mechanism-based fusion on the subject title feature representation and the above-mentioned learning situation feature representation to obtain a fusion feature, input the fusion feature into a feedforward neural network layer of the classification model, and obtain a classification result output by the feedforward neural network.

[0084] Since the subject title feature representation represents information related to the subject title, and the learning situation feature representation of the target object represents the mastery of the target object to various types of problems within a predetermined historical length, whether the subject title is recommended to the target object can be accurately determined according to the subject title feature representation and the above-mentioned learning situation feature representation.

[0085] Based on the problem recommendation method provided in the present application, the subject title is classified according to the subject title feature representation and the above-mentioned learning situation feature representation to obtain a classification result. The subject title can be recommended to the target object when the classification result represents that the subject title is recommended to the target object. Moreover, since the classification result is obtained based on the subject title and the learning situation feature representation of the target object within a predetermined historical length, personalized recommendation for the target object is realized, thereby improving the accuracy of problem recommendation.

[0086] In an optional embodiment, an implementation flowchart of training at least according to the subject title and the correct answer given by the several objects to the subject title is as shown in Figure 2 , which can include:

[0087] Step S201: randomly covering the target content in the initial training sample to obtain a target training sample; the target content includes the subject title and the correct answer given by an object to the subject title.

[0088] For each initial training sample, random covering can be performed at a certain proportion, for example, 20% of the target content in the initial training sample is randomly covered. The 20% of the content can be continuous content or content distributed in different places.

[0089] Step S202: inputting the target training sample into the language representation model to obtain a prediction result output by the language representation model, the prediction result representing the covered content in the target training sample.

[0090] The input of the language representation model is the target training sample, and the output prediction result represents the covered content in the target training sample, that is, the language representation model is used to generate the corresponding initial training sample according to the target training sample.

[0091] Step S203: updating the parameters of the language representation model with the prediction result tending to the covered content in the initial training sample corresponding to the target training sample as a target.

[0092] Step S204: At the end of the training, a feature representation of the target question is determined according to the feature vectors output by the last N layers of the language representation model.

[0093] The language representation model is a multi-layer network. The feature vectors output by the last N layers of the language representation model are extracted and used to determine the feature representation of the target question.

[0094] N can be a positive integer greater than or equal to 1. As an example, N can be equal to 1, or N can be equal to 2, or N can be equal to 3, and so on.

[0095] In the case where N is greater than 1, the mean of the feature vectors output by the N layers can be calculated, and the mean is taken as the feature representation of the target question.

[0096] In an optional embodiment, an implementation flowchart of obtaining a prediction result by the language representation model is shown in Figure 3 As shown in the figure, the implementation flowchart can include the following steps.

[0097] Step S301: Obtain the first initial vector representation of each sentence of the target question in the target training sample, and the second initial vector representation of each answer step of the correct answer in the target training sample.

[0098] For each target training sample, the question in the target training sample can be divided into several sentences according to the comma and the period, and the correct answer in the target training sample can be divided into several answer steps in the form of lines, i.e., each line of content in the correct answer is an answer step.

[0099] In order to avoid modeling difficulties caused by the diversity of data expressions, in the case where there are mathematical expressions in the sentences or answer steps, the mathematical expressions can be converted into prefix expressions (also known as Polish notation).

[0100] For each sentence in the question, obtain the word vectors of each word in the sentence, and concatenate the word vectors of each word in the sentence according to the order of each word in the sentence to obtain the basic semantic representation of the sentence. Similarly, for each answer step in the correct answer, obtain the word vectors of each word in the answer step, and concatenate the word vectors of each word in the answer step according to the order of each word in the answer step to obtain the basic semantic representation of the answer step.

[0101] The basic semantic representations of the sentences in the question can be processed by a bidirectional gated recurrent unit or a bidirectional long short-term memory network to obtain bidirectional semantic features of the sentences in the question. For each sentence in the question, the bidirectional semantic features of the sentence contain not only semantic information of the sentence, but also semantic information of other sentences in the question. Similarly, the basic semantic representations of the answer steps in the correct answer can be processed by a bidirectional gated recurrent unit or a bidirectional long short-term memory network to obtain bidirectional semantic features of the answer steps in the correct answer. For each answer step in the correct answer, the bidirectional semantic features of the answer step contain not only semantic information of the answer step, but also semantic information of other answer steps in the correct answer.

[0102] As an example, for each sentence in the question, the bidirectional semantic features of the sentence can be taken as an initial vector representation of the sentence (denoted as a first initial vector representation of the sentence).

[0103] For each answer step in the correct answer, the bidirectional semantic features of the answer step can be taken as an initial vector representation of the answer step (denoted as a second initial vector representation of the answer step).

[0104] As an example, for each sentence in the question, a position encoding vector of the sentence can also be obtained based on a position of the sentence in the question, and the bidirectional semantic features of the sentence are summed with the position encoding vector of the sentence to obtain the first initial vector representation of the sentence.

[0105] For each answer step in the correct answer, a position encoding vector of the answer step can also be obtained based on a position of the answer step in the correct answer, and the bidirectional semantic features of the answer step are summed with the position encoding vector of the answer step to obtain the second initial vector representation of the answer step.

[0106] Step S302: The first attention module of the language representation model is used to fuse the second initial vector representations and the first initial vector representations based on an attention mechanism to obtain first fused feature representations.

[0107] The present application considers that, for any target training sample, there is a correlation between the sentences in the question, between the answer steps in the correct answer, and between the sentences and the answer steps in the target training sample. Therefore, in the embodiments of the present application, the correlations between the first initial vector representations, between the second initial vector representations, and between the first initial vector representations and the second initial vector representations are calculated, and then the second initial vector representations and the first initial vector representations are fused based on an attention mechanism based on the correlations between the vector representations to obtain the first fused feature representations.

[0108] Optionally, for any one target training sample, for the convenience of description, assume that the s-th vector in the vector set composed of the first initial vector representation of each sentence of the target of the training sample and the second initial vector representation of each answer step of the correct answer in the target training sample is denoted as cs s , cs k is any one vector in the vector set different from the s-th vector, that is, cs k does not include the s-th vector, and as an optional example, cs k may also include the s-th vector. The fusion of each second initial vector representation and each first initial vector representation based on the attention mechanism can be performed in the following manner:

[0109] The correlation between the vector cs s and the vector cs k can be calculated by the following formula:

[0110]

[0111] Where A sk (cs s , cs k ) represents the correlation between the vector cs s and the vector cs k , W2, b represents the trainable parameters in the first attention module, T represents the transpose operation, [cs s , cs k ] represents the concatenation of the vector cs s and the vector cs k .

[0112] Of course, the correlation between two vectors can also be calculated in other ways, which are not limited in the present application.

[0113] Then, according to the correlation between the vectors, the correlation weight between each vector and other vectors is calculated, and then the vectors are weighted and summed according to the correlation weight between the vectors to obtain the fused feature representation (denoted as the first fused feature representation). This can be achieved in the following manner:

[0114]

[0115]

[0116]

[0117] Where v con represents the first fused feature representation.

[0118] Of course, the first fused feature representation can also be calculated by using other fusion manners, which are not limited in the present application.

[0119] Step S303: obtaining a prediction result based on at least the first fused feature representation.

[0120] As an example, the first fused feature representation can be input into a feedforward neural network layer of a language representation model to obtain a prediction result output by the feedforward neural network layer. Based on this, the last N layers of the language representation model can be the last N layers of the feedforward neural network layer.

[0121] Optionally, the prediction result output by the language representation model represents the probability of each possible decoding object (such as a character, an operator, a value of a number, etc.), and the parameters of the language representation model can be updated based on the following loss function:

[0122]

[0123] wherein P c (y g |y1, y2,..., y g-1 , P) represents the probability of correctly predicting y g at the gth coverage position under the given target content, the given objective and the correct answer P (without the covered part). G represents the number of coverage areas.

[0124] Of course, the present application is not limited to the above loss function, and the parameters of the language representation model can also be updated by using other loss functions, which are not limited in the present application.

[0125] In an optional embodiment, the initial training sample can include attribute information (denoted as first attribute information) of the object providing the correct answer in addition to the target content. As an example, the attribute information of the object can at least include the ID information of the object. Taking a student as an example, the ID information can be information for uniquely identifying the student, for example, can be the student's student ID, or can be a unique identification code randomly assigned to the student. As an example, the attribute information of the object can also include other information, such as at least part of the following information: school ID information (for uniquely identifying a school), class ID information (for uniquely identifying a class), gender, age, etc.

[0126] An implementation flowchart of the above step of obtaining a prediction result based on at least the first fused feature representation is shown in Figure 4 , and can include:

[0127] Step S401: obtaining a first semantic representation of the first attribute information, which can be denoted as u s .

[0128] As described above, the first attribute information can include at least one field, each field in the first attribute information can be mapped into a vector of the same dimension, and the vectors of the fields can be spliced to obtain a semantic representation (denoted as a first semantic representation) of the first attribute information.

[0129] Step S402: obtaining a prediction result based on the first fusion feature representation and the first semantic representation.

[0130] Optionally, the first fusion feature representation and the first semantic representation can be fused to obtain a first target fusion feature, and the prediction result can be obtained based on the first target fusion feature.

[0131] As an example, the first fusion feature representation and the first semantic representation can be fused by any one of the following fusion manners:

[0132] Hadamard product, summation operation, difference operation and splicing operation; wherein,

[0133] The Hadamard product refers to multiplying the data at the corresponding positions of the first fusion feature representation and the first semantic representation.

[0134] The summation operation refers to adding the numbers at the corresponding positions of the first fusion feature representation and the first semantic representation.

[0135] The difference operation refers to subtracting the numbers at the corresponding positions of the first fusion feature representation and the first semantic representation.

[0136] The splicing operation refers to splicing the two vectors of the first fusion feature representation and the first semantic representation.

[0137] As an example, at least two fusion processes can be performed on the first fusion feature representation and the first semantic representation to obtain at least two fusion results, and the prediction result can be obtained based on the at least two fusion results.

[0138] Optionally, the at least two fusion processes can be at least two of the four fusion manners (i.e., Hadamard product, summation operation, difference operation and splicing operation).

[0139] Optionally, when the first fusion feature representation and the first semantic representation are fused by any of the above fusion manners, the dimensions of the first fusion feature representation and the first semantic representation need to be consistent. Based on this, the first fusion feature representation can be converted by a first encoding to obtain a conversion result of the first fusion feature representation, and the first semantic representation can be converted by a second encoding to obtain a conversion result of the first semantic representation. The conversion result of the first fusion feature representation and the conversion result of the first semantic representation have the same dimension. Then, at least one fusion processing is performed on the conversion result of the first fusion feature representation and the conversion result of the first semantic representation to obtain at least one fusion result.

[0140] Optionally, the first fusion feature representation can be input into a first multi-layer perception machine to convert the fusion feature representation by a first encoding to obtain a conversion result of the first fusion feature representation. The first semantic representation can be input into a second multi-layer perception machine to convert the first semantic representation by a second encoding to obtain a conversion result of the first semantic representation.

[0141] As an example, as shown in Figure 5 , an implementation flowchart of a language representation model obtaining a prediction result in the case of four fusion processes of the first fusion feature representation and the first semantic representation provided by the embodiments of the present application is provided. Among them, clause represents a sentence. As an example, clause-1 represents the first sentence in the question, and clause-n represents the n-th sentence in the question. MASK represents the coverage area. sep is used to distinguish the question and the answer. step represents the answer step. As an example, step-1 represents the first answer step in the answer, and step-m represents the m-th answer step in the answer. position represents the position, i.e. the position of the sentence or the position of the answer step. That is, in this example, the position information of the sentence is introduced when calculating the first initial vector representation of the sentence, and the position information of the answer step is introduced when calculating the second initial vector representation of the answer step.

[0142] The language representation model trained in the above manner learns how to review the question and answer.

[0143] In an optional embodiment, an implementation flowchart of training the learning feature representation of the target object according to the questions answered by the target object in a predetermined historical time length is as shown in Figure 6 , which can include:

[0144] Step S601: obtaining a first type of training sample, the first type of training sample including at least the question answered by the target object in a predetermined historical time length and the answer given by the target object.

[0145] In the present application, each first type of training sample includes a question answered by the target object and an answer given by the target object to the question, which can be correct or incorrect. The label of the first type of training sample is the actual score of the question. That is, the samples used by the scoring model in the present application include both questions answered correctly by the target object and questions answered incorrectly by the target object.

[0146] Step S602: inputting the first type of training sample into the scoring model, so that the scoring model obtains third initial vector representations of each sentence of the question in the first type of training sample and fourth initial vector representations of each answer step of the answer in the first type of training sample; obtaining second fusion feature representations by performing attention mechanism-based fusion on each third initial vector representation and each fourth initial vector representation through a second attention module; obtaining a scoring result based at least on the second fusion feature representations, the scoring result representing the score of the question in the first type of training sample; and determining initial parameters of the second attention module based on parameters of the first attention module at the end of training of the language representation model.

[0147] The process of obtaining, by the scoring model, the third initial vector representations of each sentence of the question in the first type of training sample can refer to the process of obtaining the first initial vector representations of each sentence of the question in the target training sample, which will not be described in detail here.

[0148] The process of obtaining, by the scoring model, the fourth initial vector representations of each answer step of the answer in the first type of training sample can refer to the process of obtaining the second initial vector representations of each answer step of the correct answer in the target training sample, which will not be described in detail here.

[0149] The process of performing, by the second attention module, attention mechanism-based fusion on each third initial vector representation and each fourth initial vector representation can refer to the process of performing, by the first attention module, attention mechanism-based fusion on each second initial vector representation and each first initial vector representation, which will not be described in detail here.

[0150] The difference between the second attention module and the first attention module is that the initial parameters of the first attention module are randomly set, while the initial parameters of the second attention module are determined based on the parameters of the first attention module at the end of training of the language representation model. As an example, the initial parameters of the second attention module can be the parameters of the first attention module at the end of training of the language representation model.

[0151] Step S603: updating the parameters of the scoring model with the goal of making the scoring result output by the scoring model approach the actual score of the question in the first type of training sample.

[0152] As an example, the parameters of the scoring model can be updated based on the following loss function:

[0153]

[0154] where m represents the number of first-type training samples input to the scoring model, y (i) represents the true score result of the i-th first-type training sample, represents the score result of the i-th first-type training sample predicted by the scoring model.

[0155] Of course, the present application is not limited to the above loss function, and the parameters of the language representation model can also be updated by other loss functions, which are not limited in the present application.

[0156] Step S604: When the training of the scoring model is completed, the feature representation of the learning situation of the target object is determined according to the feature vector output by the last H-layer network of the scoring model.

[0157] The scoring model is a multi-layer network, and the present application extracts the feature vector output by the last H-layer network of the scoring model to determine the feature representation of the learning situation of the target object.

[0158] H can be a positive integer greater than or equal to 1. As an example, H can be equal to 1, or H can be equal to 2, or H can be equal to 3, etc.

[0159] In the case where H is greater than 1, the mean of the feature vector output by the H-layer network can be calculated, and the mean is taken as the feature representation of the learning situation of the target object.

[0160] Optionally, the first-type training samples can also include attribute information (denoted as second attribute information) of the target object. The second attribute information of the target object can at least include ID information of the target object, and in addition, can include at least one of the following: school ID information (used to uniquely identify the school attended by the target object), class ID information (used to uniquely identify the class attended by the target object), gender, age, etc.

[0161] Optionally, one implementation of the scoring model obtaining the score result based on the second fusion feature representation can be as follows:

[0162] Obtaining a second semantic representation of the second attribute information of the target object;

[0163] Obtaining the score result based on the second fusion feature representation and the second semantic representation.

[0164] The second fusion feature representation and the second semantic representation can be fused to obtain a second target fusion feature, and the score result is obtained based on the second target fusion feature.

[0165] As an example, the second fused feature representation and the second semantic representation can be input into a feedforward neural network of the scoring model, the feedforward neural network fuses the second fused feature representation and the second semantic representation to obtain a second target fused feature, and obtains a scoring result based on the second target fused feature. Based on this, the last H-layer network of the scoring model can be the last H-layer network in the feedforward neural network layer of the scoring model.

[0166] As an example, as shown in Figure 7 FIG. 1 is an implementation flowchart of a scoring result obtained by the scoring model provided by the embodiments of the present application. In the figure, clause represents a sentence, as an example, clause-1 represents the first sentence in the question, and clause-n represents the nth sentence in the question; sep is used to distinguish the question and the answer, or to distinguish the second fused feature representation and the second semantic representation; step represents a solution step, as an example, step-1 represents the first solution step in the answer, and step-m represents the mth solution step in the answer; and position represents a position, i.e., the position of a sentence or the position of a solution step. That is, in this example, the position information of a sentence is introduced when calculating the third initial vector representation of the sentence, and the position information of a solution step is introduced when calculating the fourth initial vector representation of the solution step.

[0167] The scoring model obtained by the above training learns the mastering degree of the target object for each type of question.

[0168] In an optional embodiment, the question recommendation method provided by the embodiments of the present application can further include:

[0169] Obtaining a second type of training sample, the second type of training sample including at least a question answered by the target object within a predetermined historical time length and an answer given by the target object.

[0170] In the present application, each second type of training sample includes a question answered by the target object and an answer given by the target object for the question, which can be correct or can be wrong. That is, the samples used by the model in the training step of the present application include both questions answered correctly by the target object and questions answered incorrectly by the target object.

[0171] Randomly selecting a solution step in the answer of the target object as a target solution step.

[0172] The second type of training sample is input into the step judgment model, so that the step judgment model obtains the fifth initial vector representation of each sentence of the question in the second type of training sample and the sixth initial vector representation of each answer step of the answer in the second type of training sample; the third attention module is used to fuse the fifth initial vector representation and the sixth initial vector representation based on an attention mechanism to obtain a third fusion feature representation; and a judgment result is obtained based on at least the third fusion feature representation, and the judgment result represents whether the target answer step is correct. The initial parameters of the third attention module are the parameters of the first attention module when the language representation model training is completed.

[0173] The process in which the step judgment model obtains the fifth initial vector representation of each sentence of the question in the second type of training sample can refer to the process of obtaining the first initial vector representation of each sentence of the question in the target training sample, which will not be described in detail here.

[0174] The process in which the step judgment model obtains the sixth initial vector representation of each answer step of the answer in the second type of training sample can refer to the process of obtaining the second initial vector representation of each answer step of the correct answer in the target training sample, which will not be described in detail here.

[0175] The process in which the third attention module fuses the fifth initial vector representation and the sixth initial vector representation based on the attention mechanism can refer to the process in which the first attention module fuses the second initial vector representation and the first initial vector representation based on the attention mechanism, which will not be described in detail here.

[0176] The difference between the third attention module and the first attention module is that the initial parameters of the first attention module are randomly set, and the initial parameters of the third attention module are the parameters of the first attention module when the language representation model training is completed.

[0177] The parameters of the step judgment model are updated with the target being the actual judgment result of the judgment result tending to the target answer step.

[0178] In the embodiments of the present application, the judgment result has only two categories of correct and incorrect, and the probability that the judgment result is correct and the probability that the judgment result is incorrect can be updated according to the following loss function:

[0179]

[0180] where P c (y g |P) represents the probability that the judgment result of the gth sample is correct under the given question and answer P (not including the target answer step). G is the number of samples input into the step judgment model.

[0181] The initial parameters of the second attention module are the parameters of the third attention module after the step judgment model has been trained.

[0182] Optionally, the second type of training samples may also include the target object's attribute information (denoted as the second attribute information). The second attribute information of the target object may at least include the target object's ID information. In addition, it may include at least one of the following: school ID information (used to uniquely identify the school attended by the target object), class ID information (used to uniquely identify the class attended by the target object), gender, age, etc.

[0183] Optionally, one implementation of the step-based judgment model that obtains the scoring result based at least on the third fusion feature representation can be:

[0184] Obtain the second semantic representation of the second attribute information of the target object;

[0185] The judgment result is obtained based on the third fusion feature representation and the second semantic representation.

[0186] The third fusion feature representation and the second semantic representation can be fused to obtain the third target fusion feature, and the judgment result can be obtained based on the third target fusion feature. The method of fusing the third fusion feature representation and the second semantic representation can refer to the various implementation methods of fusing the first fusion feature representation and the first semantic representation mentioned above, and will not be repeated here.

[0187] As an example, such as Figure 8 The diagram shows a flowchart illustrating an implementation of the step-based judgment model provided in this application to obtain a judgment result. Here, `clause` represents a sentence; for example, `clause-1` represents the first sentence in the question, and `clause-n` represents the nth sentence. `sep` distinguishes between the question and the answer; `step` represents a solution step; for example, `step-1` represents the first solution step in the answer, and `step-m` represents the mth solution step in the answer; `CHOSE` represents the selected solution step; in this example, the kth solution step `step-k` is selected; and `position` represents the location, i.e., the position of the sentence or the position of the solution step. In other words, in this example, the sentence's position information is introduced when calculating the fifth initial vector representation of the sentence, and the solution step's position information is introduced when calculating the sixth initial vector representation of the solution step.

[0188] The step judgment model trained using the above method has learned the types of errors that the target object frequently makes.

[0189] As mentioned above, this application does not require experts to perform additional annotation work on the samples when training the language representation model, scoring model, and step judgment model. Only the information of the questions and answers themselves, or the actual grading results (i.e., the score of the questions and the judgment results of whether each solution step is correct) are needed for modeling. This solves the problem of the difficulty and high cost of obtaining large-scale labeled data, thereby greatly reducing the difficulty of obtaining training samples and making it easier to obtain a large number of training samples. The more sample data collected, the better the effect (e.g., accuracy) and robustness of the constructed system.

[0190] In an optional embodiment, the question recommendation method provided in this application may further include: obtaining a second semantic representation of the second attribute information of the target object. One implementation of classifying the target questions based at least on the feature representation of the target questions and the learning characteristics of the target object to obtain classification results can be:

[0191] Based on the characteristic representation of the target question, the learning characteristics of the target object, and the second semantic representation, the target question is classified to obtain the classification result.

[0192] As an example, the feature representation of the target question, the learning characteristics of the target object, and the second semantic representation can be input into a pre-trained classification model to obtain the classification result output by the classification model.

[0193] As an example, the classification model can use the fourth attention module to fuse the feature representation of the target question, the learning feature representation of the target object, and the second semantic representation based on the attention mechanism (the specific fusion method can be referred to the process of the first attention module fusing the first initial vector representation and the second initial vector representation, which will not be repeated here) to obtain the fused feature. The fused feature is then input into the feedforward neural network layer of the classification model to obtain the classification result output by the feedforward neural network.

[0194] The initial parameters of the fourth attention module can be set randomly.

[0195] As an example, such as Figure 9 The diagram shown illustrates a flowchart of an implementation of the classification model provided in this application to obtain classification results. Wherein, f con Indicate the characteristic representation of the target question; l con Represents the learning characteristics of the target audience; u s The second semantic representation represents the second attribute of the target object; sep is used to distinguish the feature representation of the target question, the learning characteristics of the target object, and the second semantic representation. In the classification results, 0 indicates not recommended and 1 indicates recommended; or, 0 indicates recommended and 1 indicates not recommended.

[0196] As an example, in the embodiments of the present application, the training sample for training the classification model includes the feature representation of the question, the learning characteristic representation of the object (denoted as A), and the semantic representation of the attribute information of the object A. The label of the training sample is that the question is recommended to the object A, or the label is that the question is not recommended to the object A.

[0197] When training the classification model, the training sample is input into the classification model to obtain a classification result output by the classification model.

[0198] The parameters of the classification model are updated with the classification result tending to the label of the training sample as a target.

[0199] In an optional embodiment, the above-mentioned target question can be any one of the candidate question set corresponding to the target object. The candidate question set corresponding to the target object is determined by at least one of the following ways:

[0200] Way one: the question done wrong by the target object is added to the candidate question set corresponding to the target object as a candidate question. That is, the question done wrong by the target object is naturally put into the candidate question set corresponding to the target object.

[0201] Way two: for each question done wrong by the target object, a first similarity between the question and a first question is calculated according to the feature representation of the question and the feature representation of the first question; if the first similarity meets a condition, the first question is added to the candidate question set as a candidate question.

[0202] The first question can be any one of the questions in the database, that is, way two is to find a question similar to the question done wrong by the target object in the database as a candidate question to add to the candidate question set, that is, to find a question similar to the question done wrong by the target object in the database as a candidate question to add to the candidate question set.

[0203] For any two questions, a first question and a second question, it is assumed that the feature representation of the first question is The feature representation of the second question is The similarity between the first question and the second question can be calculated in the following way:

[0204]

[0205] Wherein, The similarity between the first question and the second question is illustrated.

[0206] Of course, the similarity between the first question and the second question can also be calculated in other ways, for example, the Euclidean distance between the first question and the second question is calculated based on the feature representation as the similarity between the first question and the second question.

[0207] Optionally, the first similarity satisfying condition can include that the first similarity is greater than or equal to a threshold, such as the first similarity being greater than or equal to 0.9.

[0208] Method three: according to the learning characteristic representation of the target object and the learning characteristic representation of any one non-target object, a second similarity between the non-target object and the target object is calculated, and if the second similarity satisfies a condition, the questions in the question bank corresponding to the non-target object (i.e. the questions answered by the non-target object) are added to the candidate question set corresponding to the target object as candidate questions.

[0209] For any two objects, a first object and a second object, the learning characteristic representation of the first object is assumed to be the learning characteristic representation of the second object is The similarity between the first object and the second object can be calculated in the following way:

[0210]

[0211] wherein, represents the similarity between the first object and the second object.

[0212] Of course, the similarity between the first object and the second object can also be calculated in other ways, such as calculating the Euclidean distance between the first object and the second object based on the learning characteristic representation as the similarity between the first object and the second object.

[0213] Optionally, the second similarity satisfying condition can include that the second similarity is greater than or equal to a threshold, such as the second similarity being greater than or equal to 0.9.

[0214] In addition, there can be multiple non-target objects similar to the learning characteristic of the target object, and among the multiple non-target objects, a non-target object similar to the learning characteristic of the target object but with a better score than the target object can be screened, and the questions in the question bank corresponding to the screened non-target object are added to the candidate question set as candidate questions.

[0215] Optionally, the score rate of the target object (denoted as a first score rate) and the score rate of each non-target object similar to the learning characteristic of the target object (denoted as a second score rate) can be calculated. For any one object A, the score rate of object A can be calculated in the following way:

[0216]

[0217] wherein, avg_score represents the score rate of object A, score i represents the actual answer score of the i-th question in the n questions answered by object A, total i represents the total score of the i-th question.

[0218] For any non-target object similar to the learning condition characteristics of the target object, if the second score rate of the non-target object is higher than the first score rate, it is determined that the score of the non-target object is better than the score of the target object. Based on this, if the difference between the second score rate and the first score rate meets the condition, the questions in the question bank corresponding to the non-target object are added to the candidate question set corresponding to the target object as candidate questions; otherwise, the questions in the question bank corresponding to the non-target object are prohibited from being added to the candidate question set corresponding to the target object.

[0219] Optionally, the condition that the difference between the second score rate and the first score rate meets can include that the difference between the second score rate and the first score rate is greater than or equal to a target value, which can be a default value or determined by a user (such as the target object).

[0220] In the embodiments of the present application, the difficulty levels of the questions can be defined according to the difference between the score rates of different objects, for example, if the score rate of the first object is within 10% of the score rate of the second object, it is considered that the questions in the question bank of the first object are easy level questions for the second object, if the score rate of the first object exceeds the score rate of the second object by more than 10% but less than 20%, it is considered that the questions in the question bank of the first object are medium level questions for the second object, and if the score rate of the first object exceeds the score rate of the second object by more than 20%, it is considered that the questions in the question bank of the first object are difficult level questions for the second object. Based on this,

[0221] In the case where the target value is determined by the user, the user can select questions of different difficulty levels according to his own needs, and then the corresponding target value can be determined according to the difficulty level selected by the user, and then the candidate questions of the difficulty level selected by the user can be determined according to the target value.

[0222] Based on the third mode, the purpose of people pushing questions can be achieved.

[0223] Further, in order to improve the data processing efficiency, the questions answered by each object can be stored in a structured manner, for example, Table 1 shows an example of structured data provided by the embodiments of the present application.

[0224] Table 1

[0225]

[0226] In the example shown in Table 1, the attribute information of the object is not involved.

[0227] In the case where the attribute information of the object is involved, for example, Table 2 shows another example of structured data provided by the embodiments of the present application.

[0228] Table 2

[0229]

[0230] In summary, the embodiment of the present application proposes an intelligent recommendation scheme based on multi-task common representation of learning situation characteristics and question characteristics of an object, and realizes personalized recommendation for the object.

[0231] Corresponding to the method embodiment, the embodiment of the present application further provides a question recommendation device. A structural schematic diagram of the question recommendation device provided by the embodiment of the present application is shown in Figure 10 and can include:

[0232] The obtaining module 1001 and the classification module 1002; wherein,

[0233] The obtaining module 1001 is configured to obtain at least a question title question characteristic representation and a learning situation characteristic representation of a target object within a predetermined historical time length. The question title question characteristic representation is trained according to at least the question title question and correct answers given by a plurality of objects when answering the question title question. The learning situation characteristic representation is trained according to at least a plurality of questions answered by the target object within the predetermined historical time length.

[0234] The classification module 1002 is configured to classify the question title question according to at least the question title question characteristic representation and the learning situation characteristic representation, to obtain a classification result. The classification result represents whether to recommend the question title question to the target object.

[0235] The question recommendation device provided by the embodiment of the present application classifies the question title question according to the question title question characteristic representation and the learning situation characteristic representation, and obtains a classification result. When the classification result represents that the question title question is recommended to the target object, the question title question is recommended to the target object. Moreover, since the classification result is obtained based on the question title question and the learning situation characteristic representation of the target object within the predetermined historical time length, personalized recommendation for the target object is realized, thereby improving the accuracy of question recommendation.

[0236] In an optional embodiment, the device further includes:

[0237] The first training module is configured to perform a training process according to at least the question title question and correct answers given by a plurality of objects when answering the question title question, and the process includes:

[0238] Randomly covering target content in the initial training sample to obtain a target training sample. The target content includes the question title question and correct answers given by an object when answering the question title question.

[0239] inputting the target training sample into the language representation model to obtain a prediction result output by the language representation model, the prediction result representing covered content in the target training sample;

[0240] updating parameters of the language representation model, aiming at the prediction result approaching to covered content in an initial training sample corresponding to the target training sample;

[0241] determining a target feature representation of the target sample according to feature vectors output by last N layers of the language representation model at the end of the training.

[0242] In an optional embodiment, when the first training module determines the target feature representation of the target sample according to features output by last N layers of the language representation model, the first training module is configured to:

[0243] calculate a mean value of the feature vectors output by the N layers to obtain the target feature representation of the target sample.

[0244] In an optional embodiment, when the first training module obtains the prediction result, the first training module is configured to:

[0245] obtain first initial vector representations of each sentence of the target sample and second initial vector representations of each answer step of the correct answer in the target training sample;

[0246] fuse each second initial vector representation and each first initial vector representation based on an attention mechanism through a first attention module of the language representation model to obtain a first fused feature representation;

[0247] obtain the prediction result based on at least the first fused feature representation.

[0248] In an optional embodiment, the initial training sample further includes first attribute information of an object providing the correct answer; when the first training module obtains the prediction result based on at least the first fused feature representation, the first training module is configured to:

[0249] obtain a first semantic representation of the first attribute information;

[0250] obtain the prediction result based on the first fused feature representation and the first semantic representation.

[0251] In an optional embodiment, when the first training module obtains the prediction result based on the first fused feature representation and the first semantic representation, the first training module is configured to:

[0252] perform at least one fusion processing on the first fused feature representation and the semantic representation of the first attribute information to obtain at least one fusion result.

[0253] obtaining the prediction result based on the at least one fusion result.

[0254] In an optional embodiment, the apparatus further comprises:

[0255] The second training module is configured to train the process of obtaining the learning situation feature representation based on at least a plurality of questions answered by the target object within the predetermined historical time length, and the second training module comprises:

[0256] obtaining a first type of training sample, wherein the first type of training sample at least comprises a question answered by the target object within the predetermined historical time length and an answer given by the target object;

[0257] inputting the first type of training sample into a scoring model, so that the scoring model obtains a third initial vector representation of each sentence of the question in the first type of training sample and a fourth initial vector representation of each answer step of the answer, and obtains a second fusion feature representation by performing attention mechanism-based fusion on each third initial vector representation and each fourth initial vector representation through a second attention module; and obtaining a scoring result based on at least the second fusion feature representation, wherein the scoring result represents a score of the question in the first type of training sample, and an initial parameter of the second attention module is determined based on parameters of the first attention module when the language representation model is trained.

[0258] updating parameters of the scoring model with a target of approaching the actual score of the question in the first type of training sample to the scoring result;

[0259] determining the learning situation feature representation based on a feature vector output by the last H-layer network of the scoring model when the training is completed.

[0260] In an optional embodiment, the first type of training sample further comprises second attribute information of the target object, and when the second training module obtains a scoring result based on at least the second fusion feature representation, the second training module is configured to:

[0261] obtain a second semantic representation of the second attribute information;

[0262] obtain the scoring result based on the second fusion feature representation and the second semantic representation.

[0263] In an optional embodiment, the apparatus further comprises a third training module configured to:

[0264] obtain a second type of training sample, wherein the second type of training sample at least comprises a question answered by the target object within the predetermined historical time length and an answer given by the target object.

[0265] randomly selecting one answer step in the answer given by the target object as a target answer step;

[0266] inputting the second type of training sample into the step judgment model, so that the step judgment model obtains a fifth initial vector representation of each sentence of a question in the second type of training sample and a sixth initial vector representation of each answer step of an answer in the second type of training sample; performing attention mechanism-based fusion on each fifth initial vector representation and each sixth initial vector representation through a third attention module to obtain a third fusion feature representation; obtaining a judgment result based at least on the third fusion feature representation, the judgment result representing whether the target answer step is correct; initial parameters of the third attention module being parameters of the first attention module when the language representation model training is completed;

[0267] updating parameters of the step judgment model with the judgment result tending to an actual judgment result of the target answer step as a target;

[0268] initial parameters of the second attention module being parameters of the third attention module after the step judgment model training is completed.

[0269] In an optional embodiment, the second type of training sample further includes second attribute information of the target object; when the third training module obtains a judgment result based at least on the third fusion feature representation, the third training module is configured to:

[0270] obtain a second semantic representation of the second attribute information;

[0271] obtain the judgment result based on the third fusion feature representation and the second semantic representation.

[0272] In an optional embodiment, the obtaining module 1001 is further configured to obtain a second semantic representation of second attribute information of the target object.

[0273] The classification module 1002 is specifically configured to classify the target subject based on the subject feature representation, the learning situation feature representation, and the second semantic representation to obtain a classification result.

[0274] In an optional embodiment, the target subject is any one of a candidate question set corresponding to the target object; and the apparatus further includes:

[0275] a candidate question set determination module configured to determine a candidate question set through at least one of the following manners:

[0276] The first way is to add the wrong question of the target object as a candidate question into the candidate question set;

[0277] The second way is to calculate a first similarity between each wrong question of the target object and the first question according to the feature representation of the question and the feature representation of the first question, and add the first question as a candidate question into the candidate question set if the first similarity meets a condition;

[0278] The third way is to calculate a second similarity between the target object and a non-target object according to the feature representation of the target object and the feature representation of the non-target object, and add the questions in the question bank corresponding to the non-target object as candidate questions into the candidate question set if the second similarity meets a condition.

[0279] In an optional embodiment, the device further comprises:

[0280] a score rate determination module configured to obtain a first score rate of the target object and a second score rate of the non-target object;

[0281] The candidate question set determination module is configured to: if the second similarity meets a condition, add the questions in the question bank corresponding to the non-target object as candidate questions into the candidate question set if the second score rate is higher than the first score rate and a difference between the second score rate and the first score rate meets a condition; otherwise, prohibit adding the questions in the question bank corresponding to the non-target object as candidate questions into the candidate question set.

[0282] The question recommendation device provided by the embodiments of the present application can be applied to a question recommendation device, such as a PC terminal, a cloud platform, a server, a server cluster, and the like. Optionally, Figure 11 A hardware structure block diagram of the question recommendation device is shown, and the hardware structure of the question recommendation device can include at least one processor 1, at least one communication interface 2, at least one memory 3, and at least one communication bus 4; Figure 7

[0283] In the embodiments of the present application, the number of the processor 1, the communication interface 2, the memory 3, and the communication bus 4 is at least one, and the processor 1, the communication interface 2, and the memory 3 complete communication with each other through the communication bus 4;

[0284] The processor 1 can be a central processing unit CPU, or a specific integrated circuit ASIC (Application Specific Integrated Circuit), or one or more integrated circuits configured to implement the embodiments of the present application, etc. ​

[0285] The memory 3 can comprise a high-speed RAM memory and can also comprise a non-volatile memory, such as at least one disk memory;

[0286] The memory stores a program, and the processor can invoke the program stored in the memory, and the program is used for:

[0287] obtaining at least a feature representation of the target question and a learning situation feature representation of the target object in a predetermined historical time length; the feature representation of the target question is trained at least according to the target question and correct answers given by a plurality of objects in response to the target question, and the learning situation feature representation is trained at least according to a plurality of questions answered by the target object in the predetermined historical time length;

[0288] classifying the target question at least according to the feature representation of the target question and the learning situation feature representation, to obtain a classification result, the classification result representing whether to recommend the target question to the target object.

[0289] Optionally, the refinement function and the extension function of the program can refer to the description above.

[0290] The embodiments of the present application also provide a storage medium, which can store a program suitable for processor execution, and the program is used for:

[0291] obtaining at least a feature representation of the target question and a learning situation feature representation of the target object in a predetermined historical time length; the feature representation of the target question is trained at least according to the target question and correct answers given by a plurality of objects in response to the target question, and the learning situation feature representation is trained at least according to a plurality of questions answered by the target object in the predetermined historical time length;

[0292] classifying the target question at least according to the feature representation of the target question and the learning situation feature representation, to obtain a classification result, the classification result representing whether to recommend the target question to the target object.

[0293] Optionally, the refinement function and the extension function of the program can refer to the description above.

[0294] Those skilled in the art can realize that the units and algorithm steps of the examples described in combination with the embodiments disclosed in the present text can be realized by electronic hardware or a combination of computer software and electronic hardware. Whether the functions are realized in hardware or software mode depends on the specific application and design constraints of the technical solution. The skilled person can use different methods to realize the described functions for each specific application, but such implementation should not be considered beyond the scope of the present application.

[0295] In several embodiments provided in the present application, it should be understood that the disclosed system, device and method can be implemented by other manners. In addition, the coupling or direct coupling or communication connection between the displayed or discussed units can be indirect coupling or communication connection through some interfaces, devices or units, which can be electrical, mechanical or other forms.

[0296] The units described as separate components can or can not be physically separate, and the components displayed as units can or can not be physical units, i.e. they can be located in one place or distributed on multiple network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0297] In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit.

[0298] If the functions are realized in the form of software functional units and sold or used as independent products, they can be stored in a computer readable storage medium. Based on this understanding, the technical solutions of the present application or the parts of the technical solutions that essentially contribute to the prior art can be embodied in the form of software products. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the embodiments of the present application. The aforementioned storage medium includes: a U disk, a mobile hard disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a magnetic disk or an optical disk, and various media that can store program codes.

[0299] Finally, it should be noted that in this document, relationship terms such as first and second are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply that there is any such actual relationship or order between the entities or operations. Moreover, the terms "include", "contain" or any other variants thereof are intended to cover non-exclusive inclusion, so that the processes, methods, articles or devices including a series of elements not only include those elements, but also include other elements not explicitly listed or inherent to such processes, methods, articles or devices. Without more limitations, the element defined by the statement "including a" does not exclude the presence of other identical elements in the process, method, article or device including the element.

[0300] The various embodiments described in this specification are presented by way of example, and each embodiment is not necessarily composed of all features described with respect to other embodiments. Each embodiment describes a distinct aspect of the present application, and each aspect can be used in combination with one or more other aspects.

[0301] The previous description of the disclosed embodiments is provided to enable any person skilled in the art to make or use the application. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the generic principles defined herein can be applied to other embodiments without departing from the spirit or scope of the application. Thus, the present application 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 subject recommendation method characterized by comprising: The method comprises the following steps: obtaining at least a feature representation of a target question and a learning condition feature representation of a target object in a predetermined historical time length; the feature representation of the target question is obtained based on a language representation model according to at least the target question and correct answers given by a plurality of objects to the target question, so that the feature representation of the target question represents information related to the target question and the solution; the learning condition feature representation is obtained based on a scoring model according to at least a plurality of questions answered by the target object in the predetermined historical time length and respective answer steps of answers given by the target object to the questions, so that the learning condition feature representation represents the mastery degree of the target object to the questions in the predetermined historical time length; the scoring model updates parameters with the goal of approaching the actual scores of the questions in the training samples; and the target question is classified according to at least the feature representation of the target question and the learning condition feature representation, to obtain a classification result, which represents whether the target question is recommended to the target object; the method further comprises the following steps: inputting a plurality of questions answered by the target object in the predetermined historical time length and respective answer steps of answers given by the target object to the questions into a step judgment model, to obtain a judgment result, which represents whether the target answer step is correct, the target answer step being a randomly selected answer step among the answers given by the target object to the questions; the step judgment model updates parameters with the goal of approaching the actual judgment result of the target answer step; and the step judgment model learns the types of common errors of the target object; initial parameters of a third attention module of the step judgment model are parameters of a first attention module of the language representation model when the language representation model is trained; and initial parameters of a second attention module of the scoring model are parameters of the third attention module of the step judgment model after the step judgment model is trained.

2. The method of claim 1, wherein, The process of training the language representation model based on at least the target question and correct answers given by a plurality of objects to the target question comprises the following steps: randomly covering target content in initial training samples to obtain target training samples; the target content comprises the target question and correct answers given by an object to the target question; inputting the target training samples into the language representation model to obtain a prediction result output by the language representation model, the prediction result representing the covered content in the target training samples; updating parameters of the language representation model with the goal of approaching the covered content in the initial training samples corresponding to the target training samples; determining the feature representation of the target question according to feature vectors output by the last N layers of the language representation model when the training is completed.

3. The method of claim 2, wherein, The process of determining the feature representation of the target question according to features output by the last N layers of the language representation model comprises the following steps: calculating the mean value of the feature vectors output by the N layers to obtain the feature representation of the target question.

4. The method of claim 2, wherein, The process of obtaining the prediction result by the language representation model comprises: obtaining first initial vector representations of each sentence of a target question in the target training sample and second initial vector representations of each answer step of a correct answer in the target training sample; performing attention mechanism-based fusion on each second initial vector representation and each first initial vector representation by a first attention module of the language representation model to obtain first fused feature representations; obtaining the prediction result based on at least the first fused feature representations.

5. The method of claim 4, wherein, The initial training sample further comprises first attribute information of an object providing the correct answer; and the obtaining of the prediction result based on at least the first fused feature representations comprises: obtaining a first semantic representation of the first attribute information; obtaining the prediction result based on the first fused feature representations and the first semantic representation.

6. The method of claim 5, wherein, The obtaining of the prediction result based on the first fused feature representations and the first semantic representation comprises: performing at least one fusion processing on the first fused feature representations and the semantic representation to obtain at least one fusion result; and obtaining the prediction result based on the at least one fusion result.

7. The method of claim 4, wherein, The process of obtaining the learning situation feature representation based on a scoring model according to at least a plurality of questions answered by the target object within the predetermined historical time length and each answer step of an answer given by the target object to the question comprises: obtaining first training samples, wherein the first training samples at least comprise the questions answered by the target object within the predetermined historical time length and the answer given by the target object; inputting the first training samples into a scoring model, so that the scoring model obtains third initial vector representations of each sentence of the questions in the first training samples and fourth initial vector representations of each answer step of the answers in the first training samples; performing attention mechanism-based fusion on each third initial vector representation and each fourth initial vector representation by a second attention module to obtain second fused feature representations; and obtaining a scoring result based on at least the second fused feature representations, wherein the scoring result represents a score of the questions in the first training samples; and initial parameters of the second attention module are determined based on parameters of the first attention module when the language representation model training is completed. The parameters of the scoring model are updated so that the scoring result approaches an actual score of the questions in the first training samples. When the training is completed, a feature vector output by a last H-layer network of the scoring model is used to determine the learning situation feature representation.

8. The method of claim 7, wherein, The first training samples further comprise second attribute information of the target object; and the obtaining of the scoring result based on at least the second fused feature representations comprises: obtaining a second semantic representation of the second attribute information; obtaining the scoring result based on the second fused feature representations and the second semantic representation.

9. The method of claim 7, wherein, Further comprising: obtaining a second type of training sample, at least including a question answered by the target object within the predetermined historical time length, and an answer given by the target object; randomly selecting one answer step in the answer given by the target object as a target answer step; inputting the second type of training sample into the step judgment model, so that the step judgment model obtains a fifth initial vector representation of each sentence of the question in the second type of training sample, and a sixth initial vector representation of each answer step of the answer in the second type of training sample; obtaining a third fusion feature representation by the third attention module based on attention mechanism for each fifth initial vector representation and each sixth initial vector representation; obtaining a judgment result based on at least the third fusion feature representation, the judgment result representing whether the target answer step is correct; the initial parameters of the third attention module are the parameters of the first attention module when the language representation model training is completed; updating the parameters of the step judgment model with the judgment result tending to the actual judgment result of the target answer step as the target; the initial parameters of the second attention module are the parameters of the third attention module after the step judgment model training is completed.

10. The method of claim 9, wherein, the second type of training sample further includes second attribute information of the target object; the at least based on the third fusion feature representation to obtain the judgment result, comprising: obtaining a second semantic representation of the second attribute information; obtaining the judgment result based on the third fusion feature representation and the second semantic representation.

11. The method according to any one of claims 1 to 10, characterized in that, Further comprising: obtaining a second semantic representation of the second attribute information of the target object; the at least according to the feature representation of the target object and the learning situation characteristic representation classifies the target object, obtains a classification result, comprising: classifying the target object according to the feature representation of the target object, the learning situation characteristic representation and the second semantic representation, and obtaining a classification result.

12. The method according to any one of claims 1 to 10, characterized in that, The target object is any one candidate question in the candidate question set corresponding to the target object; the candidate question set is determined by at least one of the following ways: way one: the question answered incorrectly by the target object is added to the candidate question set as a candidate question; way two: for each of the questions answered incorrectly by the target object, calculate the first similarity between the question and the first question according to the feature representation of the question and the feature representation of the first question; if the first similarity meets the condition, the first question is added to the candidate question set as a candidate question; way three: according to the learning situation characteristic representation of the target object and the learning situation characteristic representation of any non-target object, calculate the second similarity between the target object and the non-target object, if the second similarity meets the condition, the questions in the question bank corresponding to the non-target object are added to the candidate question set as candidate questions.

13. The method of claim 12, wherein, Further comprising: obtaining a first score rate of the target object and a second score rate of the non-target object; If the second similarity meets the condition, if the second score rate is higher than the first score rate, and the difference between the second score rate and the first score rate meets the condition, the question in the question bank corresponding to the non-target object is added to the candidate question set as a candidate question; otherwise, the question in the question bank corresponding to the non-target object is prohibited from being added to the candidate question set as a candidate question.

14. A title recommendation device characterized by comprising: Comprising: An obtaining module configured to obtain at least a feature representation of a target question and a learning condition feature representation of a target object in a predetermined historical duration; The feature representation of the target question is obtained based on a language representation model according to the target question and correct answers given by a plurality of objects to the target question, so that the feature representation of the target question represents information related to solving the target question; the learning condition feature representation is obtained based on a scoring model according to a plurality of questions answered by the target object in the predetermined historical duration and respective answer steps of answers given by the target object to the question, so that the learning condition feature representation represents the mastery degree of the target object to each type of question in the predetermined historical duration; and the scoring model updates parameters with the goal of approaching the actual scores of the questions in the training samples; A classification module configured to classify the target question according to at least the feature representation of the target question and the learning condition feature representation, to obtain a classification result, which represents whether the target question is recommended to the target object; The question recommendation device is further configured to: input a plurality of questions answered by the target object in the predetermined historical duration and respective answer steps of answers given by the target object to the question into a step judgment model, to obtain a judgment result, which represents whether a target answer step is correct, the target answer step being a randomly selected answer step from the answers given by the target object to the question; the step judgment model updates parameters with the goal of approaching the actual judgment result of the target answer step; and the step judgment model learns the types of common errors of the target object. Initial parameters of a third attention module of the step judgment model are parameters of a first attention module when the language representation model is trained; and initial parameters of a second attention module in the scoring model are parameters of the third attention module after the step judgment model is trained.

15. A title recommendation device characterized by comprising: Comprising a memory and a processor; The memory is configured to store a program; The processor is configured to execute the program to implement each step of the question recommendation method according to any one of claims 1-13.

16. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program, when executed by the processor, implements each step of the question recommendation method according to any one of claims 1-13.

Images