A method for identifying coaching intentions and a personalized coaching method based on coaching intentions

Abstract

The present invention proposes a tutoring intention recognition method and a personalized tutoring method based on tutoring intention, including: S1. preprocessing user question data, and constructing a text representation for tutoring intention based on a constructed keyword list; S2. parsing the latent semantic vector of the user question data using a hierarchical neural network structure model; S3. based on the generated latent semantic vector, using a multi-task learning algorithm to perform tutoring intention recognition. The tutoring intention recognition provided by the present invention is based on multi-task learning and a hierarchical network model, and can effectively deal with the problem that the tutoring intention recognition method in the online tutoring scenario ignores the multi-task learning algorithm and does not fully consider the historical question and answer information.

Description

Technical Field

[0001] The present invention relates to the field of intelligent education technology, and in particular to a method for identifying tutoring intentions and a personalized tutoring method. Background Art

[0002] In the field of personalized tutoring, learners will upload their questions to the system, and the system will fully interpret and understand the learners' questions and intelligently push tutoring resources, ultimately achieving the purpose of personalized tutoring for learners. It can be seen that tutoring intention understanding plays an important role in the entire personalized tutoring process, facilitating the improvement of learner portraits, providing learners with personalized learning resources, and reducing learners' cognitive load.

[0003] With the gradual advancement of educational informatization, digital learning resources are becoming more abundant. Learners will receive tutoring resources pushed from smart terminals during the learning process. These resources generally push similar questions based on the similarity between the learner's wrong exercises and the questions in the question bank. This recommendation algorithm does not fully tap the learner's tutoring intention information, which will inevitably lead to the sea of ​​questions and cause more burden on learners' learning. Therefore, how to fully understand the tutoring intention and provide personalized tutoring based on the tutoring intention has become a problem that needs to be solved urgently. Summary of the invention

[0004] In view of the above problems, the present invention proposes a tutoring intention recognition method and system based on multi-task learning and hierarchical network model, which is convenient for fully mining the tutoring intention information of learners.

[0005] According to one aspect of the present invention, a method for identifying tutoring intention is provided, comprising the following steps:

[0006] S1. Preprocess the user question data and construct a text representation for tutoring intention based on the keyword list;

[0007] S2. Use a hierarchical neural network structure model to obtain the latent semantic vector of the user's question data;

[0008] S3. Based on latent semantic vector, multi-task learning algorithm is used to identify tutoring intention;

[0009] Furthermore, the method for constructing the keyword list includes:

[0010] By using word frequency statistics method and combining the experience of expert teachers, we can obtain the subject concept information, relational words and verb information in the question data, and then construct a list of key information of the question data.

[0011] Furthermore, the subject concept information is the core concept in the question data, the relationship words are words that describe the relationship between two entities, and the skill words, ie, verbs, include but are not limited to calculate, draw, describe, and select.

[0012] Furthermore, in step S2, obtaining the latent semantic vector of the user's question data using the hierarchical neural network structure model includes:

[0013] S21. The underlying network is responsible for semantic analysis of user question data;

[0014] S22. The upper network is responsible for combining historical data to generate latent semantic vectors of user question data.

[0015] Furthermore, the underlying network is a BERT model, and the upper network is a double-layer LSTM model.

[0016] Furthermore, in step S3, the tutoring intention includes three dimensions: obstacle type, subject concept and learning ability level.

[0017] Furthermore, the multi-task learning model includes three MLP networks for recognizing tutoring intentions. The training method of the multi-task learning model includes: based on the generated latent semantic vector, simultaneously training three-dimensional data to obtain an MLP network that represents information of different dimensions.

[0018] Furthermore, in step S3, the obstacle type dimension includes: subject concept-oriented learning and question-oriented learning; the subject concept dimension includes subject knowledge involved in the user question data; and the learning ability dimension includes the learning ability level identified from the user question data.

[0019] Furthermore, learning ability is defined as three dimensions: A learning and understanding, B application and practice, and C transfer and innovation, and then the three different levels under each dimension are specifically defined according to different disciplines.

[0020] According to another aspect of the present invention, a personalized tutoring method is provided, which further comprises S4. providing personalized tutoring to the user based on the tutoring intention and the question data, based on the tutoring intention identification.

[0021] The beneficial effects of the present invention are:

[0022] (1) The present invention pre-processes user input questions based on a constructed education keyword list, which can effectively characterize key education information in user question data.

[0023] (2) The present invention proposes a theoretical framework of tutoring intention, which fully describes the user's tutoring intention from the dimensions of obstacle type, subject concept and learning ability, and well describes the user's learning needs.

[0024] (3) Regarding the recognition of tutoring intentions, the present invention proposes to construct a tutoring intention recognition model by combining a hierarchical neural network structure model and a multi-task learning classification model. Combined with user historical data, it can fully capture the relationship between obstacle types, subject concepts and learning abilities, and accurately identify users’ tutoring intentions.

[0025] (4) The present invention proposes a personalized tutoring method based on tutoring intentions. For different tutoring intentions, facing the same or similar questions, it can distinguish the differences in user tutoring intentions in detail, thereby providing users with personalized tutoring and learning plans. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 A schematic diagram of a flow chart of a method for recognizing tutoring intention according to an embodiment of the present invention;

[0027] Figure 2 A schematic diagram of a counseling intention theory framework according to an embodiment of the present invention;

[0028] Figure 3 A schematic diagram of a flow chart of a method for analyzing problem data according to an embodiment of the present invention;

[0029] Figure 4 A flowchart of a method for understanding tutoring intention according to an embodiment of the present invention;

[0030] Figure 5 The figure is a flowchart of a personalized tutoring method based on tutoring intention according to an embodiment of the present invention. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be checked and described in full in combination with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0032] According to one aspect of the present invention, a method for recognizing tutoring intention based on multi-task learning and hierarchical network model is provided. Figure 1 As shown, the following steps are included:

[0033] S1. Preprocess the user question data and construct a text representation for tutoring intention based on the keyword list;

[0034] In step S1, Figure 3 As shown in the figure, preprocessing includes word segmentation and text cleaning, where text cleaning is to remove invalid characters. Invalid characters are characters that are irrelevant to the content of the question and are sorted out after analyzing and counting the question data in the question bank, mainly including "hello, trouble, explain it, help me explain it, send it, this, hope you can help me solve it, how to do the question, how to answer this question", etc.

[0035] Then, based on the constructed keyword list for understanding tutoring intentions, the representation of user questions is simplified. Specifically, based on the word segmentation results of the cleaned user question text, each word in the text is compared to see if it appears in the keyword list. If it does, it is retained, otherwise it is removed.

[0036] An important part involved in this step is the pre-generation of a keyword list. The keyword list includes information such as subject concept information, relational words, and skill words. The present invention adopts a word frequency statistical method, combined with the experience of expert teachers, to obtain knowledge information, relational words, and skill word information in the key information of the question data, and then constructs a key information list of the question data. Specifically, after a large amount of question data is processed by data preprocessing and word segmentation, a word frequency statistical method is used to extract words with higher word frequencies, and combined with the manual analysis of expert teachers, the words with higher word frequencies are divided into subject concept classes, relational word classes, and skill word classes. Among them, the subject concept refers to the core concept in the question, such as the subject concept of the function part of the mathematics subject, including the relationship between a linear function, a quadratic function, an inverse proportional function, and a variable. Relational words refer to words that describe the relationship between two entities, such as words that describe the relationship between two straight lines are parallel and intersecting. The skill words in the question data are verbs, mainly including verification, calculation, drawing, description, and selection, etc. The structure and example information of the key information list are shown in Table 2.

[0037] Table 2 Structure and example of key information list

[0038]

[0039] In one embodiment, the input question data is "Teacher, please help me check. It is known that the equations of the two adjacent symmetry axes of the function are x=-1 and x=2. Then the minimum positive period of f is π and it is a monotonically increasing function on the interval [3,5]. Did I answer this question correctly?". After word segmentation, it becomes "It is known that the equations of the two adjacent symmetry axes of the function and its image are x=-1 and x=2. Then the minimum positive period of f is π and it is a monotonically increasing function on the interval [3,5]. After data preprocessing, it becomes "It is known that the equations of the two adjacent symmetry axes of the function and its image are x=-1 and x=2. Then the minimum positive period of f is π and it is a monotonically increasing function on the interval [3,5]. ". According to the keyword list information, the text representation of the question data is "the minimum positive period of the symmetry axis of the function image is monotonically increasing".

[0040] S2. Use a hierarchical neural network structure model to obtain the latent semantic vector of the user's question data;

[0041] Personalized counseling based on counseling intention is a concept first proposed by the inventor, so there is no work to define and identify counseling intention at present. Although there are similar works based on user questions for automatic answering, these works mainly use the relevant technology of question-answering system, based on the current user questions, using the method of text similarity, find out the same or similar questions as the user questions from the existing database, and return the answers to the users. In this process, the learners are not analyzed and modeled, so when different users ask the same questions, the content returned to the users by the model is exactly the same. Only the current question data information of the questioner is used, and the historical question data information of the questioner is not involved. Therefore, it is difficult to fully mine the counseling intention information of the current question of the user (questioner), and it is difficult to provide personalized answers and counseling. Therefore, in view of this problem, the present invention analyzes the counseling intention of the learner first for the questions raised by the user, and on this basis, further analyzes the content of the question, so as to provide different answers and counseling for different counseling intentions. Therefore, how to identify the counseling intention of the user is an important problem to be solved.

[0042] For the identification of user tutoring intention, it is not enough to analyze only based on the user's current question information, and the consistency and continuity of the user in the learning process must be considered. The present invention combines the user's historical question and answer data and uses a deep learning model to identify the user's tutoring intention. When constructing a tutoring intention recognition model, the first step is to semantically represent the current question data in combination with the user's historical data. This step needs to solve two key problems: one is how to perform a preliminary semantic understanding of each question data, and the other is how to integrate historical data. To address these two problems, the present invention constructs a two-level network neural model to achieve semantic representation of question data.

[0043] like Figure 4 As shown, in the hierarchical neural network, the bottom layer network is responsible for semantically parsing each question currently being asked and generating a semantic vector for the current question. Specifically, the present invention uses an algorithm model in the field of natural language processing to generate a semantic vector for the question being asked at the current moment based on the word segmentation representation of the current question formed by the key information list. Considering the numerous algorithm models currently used to generate semantic vectors, the present invention uses the more popular BERT model to perform semantic vector representation of the question being asked at the current moment.

[0044] The upper network is responsible for combining historical question data, and it is necessary to use a time-series algorithm to further enhance the semantic representation of the questioner's question. Specifically, in the present invention, based on the question semantic vector generated by the underlying network, combined with the historical question data of the questioner, a double-layer LSTM neural network is used to further generate the potential semantic vector of the questioner's current question. In the entire process of generating the potential semantic vector, compared to a single-layer neural network, a double-layer neural network adds the historical question data of the questioner to the question data at the current moment, which can represent more complex semantic information and mine the potential semantic vector in the questioner's current question data.

[0045] Among them, the BERT model is a deep network composed of stacked Transformer units. The Transformer network unit mainly includes an attention layer, a forward propagation layer, and a layer norm layer. After BERT conversion, the word segmentation representation of the question at the current moment formed based on the key information list is further generated to generate a semantic vector of the question at the current moment. Among them, the various parameters here are commonly used in the BERT model and are not described in detail. Using a two-layer LSTM model, on the basis of the semantic vector of the current moment question generated by the BERT model, the historical question and answer data of the questioner are fully utilized to further characterize the semantic vectors at different moments, and finally fully analyze the question information of the questioner at the current moment to generate the potential semantic vector of the question. In one embodiment, the various parameter settings of the LSTM network are shown in Table 3, the LSTM input feature dimension is 300, the number of LSTM hidden units is 128, and the number of LSTM stacked layers is 2 layers, that is, the second LSTM receives the calculation result of the first LSTM, bias indicates whether the hidden layer state has bias, and the default is true; batch_first indicates whether the first dimension of the input and output is batch_size. Since the input data is 300-dimensional data, batch_first is set to True. The Dropout coefficient is set to 0.5. Bidirectional is set to True. The basic idea of ​​LSTM is to apply a forward and backward LSTM network to each training sequence and connect the output layer at the same time. Therefore, this network structure can transmit complete context information to the output layer.

[0046] Table 3 LSTM network parameter settings

[0047]

[0048] S3. Based on latent semantic vector, multi-task learning algorithm is used to identify tutoring intention;

[0049] In this step, based on the semantic vector of the user's current question generated in step S2, a multi-task learning algorithm is used to discriminate different dimensions of tutoring intention. The tutoring intention defined in the present invention includes three different dimensions, namely, the obstacle type dimension, the subject concept dimension, and the learning ability dimension. Figure 2 As shown, the following is a specific explanation.

[0050] Dimension 1: Barrier type dimension

[0051] Generally speaking, the difficulties encountered by learners in the learning process can be summarized into two categories. One is that learners have difficulty understanding the subject concepts themselves; the other is that learners have difficulty answering specific exercises. Therefore, the types of obstacles are divided into two categories: tutoring for subject concepts and tutoring for questions.

[0052] Dimension 2: Disciplinary Concept Dimension

[0053] According to the question data, the subject concept information can be identified, and different questions have different subject concepts. Subject concept information, as another important dimension of tutoring intention, specifically refers to the subject knowledge involved in the learners' difficult questions. Subject concept is the unit of teaching, and subject concept information can effectively help learners understand the unit where there are difficulties. In the present invention, a hierarchical structure method is used to define subject concepts, and subject concepts are divided into first-level subject concepts, second-level subject concepts, and third-level subject concepts. According to the granularity of subject concepts, the granularity of first-level subject concepts is greater than that of second-level subject concepts, and the granularity of second-level subject concepts is greater than that of third-level subject concepts.

[0054] Dimension 3: Learning ability dimension

[0055] According to the question data, the level of learning ability can be identified. Different questions have different levels of learning ability. The level of learning ability represents the level of knowledge that the question requires learners to master. Learning ability is defined as three dimensions: A learning and understanding, B application practice, and C transfer innovation. According to different subjects, each dimension is further defined as three different levels. Taking mathematics as an example, mathematics learning ability is divided into three dimensions, as shown in Table 1. Among them, dimension A is learning and understanding, which specifically includes three levels: A1 recognition and recall, A2 calculation and operation, and A3 explanation and communication; dimension B is application practice, which specifically includes B1 analysis and generalization, B2 reasoning and argumentation, and B3 simple problem solving; dimension C is transfer innovation, which specifically includes C1 comprehensive application, C2 conjecture and discovery, and C3 exploration and modeling.

[0056] Table 1 Learning ability framework for mathematics

[0057]

[0058] The three dimensions are not independent of each other, but have inherent educational correlations. For example, when the obstacle type belongs to subject concept learning, the model can focus on the information in the input data that better reflects the subject concept; at the same time, subject concepts are closely related to learning ability, and different subject concepts have different requirements for learning ability.

[0059] In one embodiment, the original data of the question is "It is known that the slope of a linear function is 2, and the graph passes through the point (1,2), then the function expression is", and after multi-task model recognition, it is known that the obstacle type is "topic tutoring", the subject concept information is "linear function", and the learning ability level is "A2 calculation and operation". If a separate model is constructed for each dimension, that is, separate model recognition, it is known that the obstacle type is "topic tutoring", the subject concept information is "linear function", and the learning ability level is "A1 recognition and recall". The expert annotation results are: the obstacle type is "topic tutoring", the subject concept information is "linear function", and the learning ability level is "A2 calculation and operation". Comparing the recognition results of the two models with the expert annotation results, it can be seen that there is a deviation in the recognition of learning ability in the separate model recognition. Because in the multi-task model, more attention is paid to the close connection between subject concepts and learning ability, the learning of subject concepts is to achieve a certain level of learning ability, and the learning ability is based on specific knowledge, which ultimately represents the combined application of knowledge. Therefore, when the subject concept "linear function" in the question is identified, it is itself a subject concept about calculation expressions, so the "calculation" related questions belong to "A2 calculation and operation"; while in the single model recognition, the subject concept information cannot be obtained, that is, the mutual influence and association between different dimensions cannot be captured, resulting in deviations in learning ability recognition. Similarly, the original data of another question is "Given that line segment AB = 9cm, draw BC on the straight line AB so that BC = 3cm, then the length of line segment AC is cm". After multi-task model recognition, it is known that the obstacle type is "question guidance", the subject concept information is "line segment", and the learning ability level is "B1 analysis and generalization". After single model recognition, it is known that the obstacle type is "question guidance", the subject concept information is "line segment", and the learning ability level is "A2 calculation and operation". The expert annotation results are: the obstacle type is "question guidance", the subject concept information is "line segment", and the learning ability level is "B1 analysis and generalization". Comparing the recognition results of the two models with the expert annotation results, it can be seen that there is a deviation in the learning ability recognition in the single model recognition. In the multi-task model, on the one hand, the close relationship between subject concepts and learning ability is emphasized; on the other hand, the obstacle type will affect the focus of the model, that is, when the obstacle type is question guidance, the model pays more attention to the known conditions, unknown conditions, known quantities and unknown quantities in the question data. Therefore, when it is identified that the obstacle type in the question is "question guidance", the relationship information between known conditions, unknown conditions, known quantities and unknown quantities will be analyzed, and combined with the result of subject concept recognition is "line segment", it is concluded that there is a classified discussion of "line segment" related issues, so the learning ability is identified as "B1 analysis and generalization"; while in the single model recognition, the obstacle type and subject concept information cannot be obtained, that is, the mutual influence and relationship between different dimensions cannot be captured, resulting in deviations in learning ability recognition.

[0060] If a separate model is built for each dimension, the three models will be calculated separately, and the mutual influence and correlation between different dimensions cannot be captured. Therefore, the multi-task learning algorithm used in the present invention determines the three dimensions at the same time, so that the influence and correlation between the three dimensions can be better considered in the determination and calculation process, thereby improving the accuracy of tutoring intention recognition.

[0061] In addition, since deep learning-based models usually use an end-to-end training method, the semantic vector generated in step S2 is affected by the judgment model training of tutoring intention recognition in step S3. If a separate model is used for each dimension, the semantic vector generated in step S2 may be biased towards one dimension, making it difficult to learn a semantic vector with good generalization ability. In contrast, the semantic vector generated by the multi-task learning algorithm has better generalization ability because it supports judgments in three different dimensions at the same time, and can better capture the educational semantic information represented by the user's question data.

[0062] Specifically, in step S3, based on the latent semantic vector generated in S2, combined with the three dimensions of tutoring intention information, namely, obstacle type dimension, subject concept dimension, and learning ability dimension, a multi-task learning model is used to complete multi-dimensional tutoring intention recognition. Figure 4 As shown, the multi-task learning model uses three task learning algorithms, based on the generated latent semantic vector, and uses MLP networks with different parameter settings to classify and discriminate each dimension of tutoring intention. Among them, the number of MLP layers can be selected according to the specific situation and expert experience. In the present invention, a two-layer MLP network is used.

[0063] For the training of the MLP model, the present invention uses pre-labeled data for training, and multiple MLP model training parameter settings are shown in Table 4. The data content is a question data set, and the question data is labeled by manual labeling. In the obstacle type dimension, 1 represents the question obstacle type, and 0 represents the subject concept obstacle type; in the subject concept dimension, 0-39 is used to represent different subject concept information; in the learning ability dimension, 0-8 is used to represent nine different learning ability levels. Then, based on the above dimensional information, the three dimensional data are trained simultaneously, and different MLP models representing different dimensional information are obtained. Finally, the trained MLP model is used to identify the tutoring intention.

[0064] Table 4. Training parameter settings of multiple MLP models

[0065]

[0066]

[0067] The multi-task learning model used here refers to a combination model of multiple classification models. Because it is trained simultaneously, it can handle multiple classification tasks. Therefore, it is only called a "multi-task learning model" in name, but it is actually multiple classification models.

[0068] Compared with independent task learning, the use of multi-task learning algorithm can make the tutoring intention recognition results consistent with the multi-dimensional feature selection results. Next, the effect is verified by comparing the independent model learning process with the multi-task learning process.

[0069] In one embodiment, for the latent semantic vector generated in S2, first, a task-by-task learning algorithm is used to identify the tutoring intention. According to the order of obstacle type task, subject concept learning task and learning ability level learning task, the above three tasks are learned separately, and finally the result of tutoring intention recognition is obtained. Secondly, a multi-task learning algorithm is used at the same time to identify the tutoring intention. That is, the three tasks of obstacle type task, subject concept learning task and learning ability level learning task are learned at the same time, and finally the result of tutoring intention recognition is obtained. As shown in Table 5. It can be seen from the comparison of experimental results that the independent model recognition results have deviations in the recognition of subject concept information and learning ability level information, mainly because the independent model does not fully consider the mutual constraint relationship between subject concepts and learning abilities. On the contrary, the multi-task model judges the three dimensions at the same time, so that the influence and correlation between the three dimensions can be better considered in the process of judgment and calculation, and then the accuracy of tutoring intention recognition can be improved, so that the recognition results of tutoring intention are consistent with the results of expert annotation.

[0070] Table 5 Comparison of tutoring intention recognition effects under independent model and multi-task model

[0071]

[0072]

[0073] The entire process of tutoring intention recognition includes three steps: question data preprocessing, generating latent semantic vectors based on a hierarchical network model, and tutoring intention recognition based on a multi-task learning algorithm.

[0074] In one embodiment, an example of the complete process of tutoring intention recognition based on multi-task learning and hierarchical network model is as follows: When the original question data is "Teacher, please help me look, it is known that the equations of the two adjacent symmetry axes of the function are x=-1 and x=2, then the minimum positive period of f is π and it is a monotonically increasing function on the interval [3,5]. Did I answer this question correctly?". After word segmentation, it becomes "It is known that the equations of the two adjacent symmetry axes of the function and its image are x=-1 and x=2, then the minimum positive period of f is π and it is a monotonically increasing function on the interval [3,5]. After data preprocessing, it becomes "It is known that the equations of the two adjacent symmetry axes of the function image are x=-1 and x=2, then the minimum positive period of f is π and it is a monotonically increasing function on the interval [3,5]. ". According to the keyword list information, the text representation of the question data is "the minimum positive period of the symmetry axis of the function image is monotonically increasing".

[0075] Next, the hierarchical network model is used to characterize the latent semantic vector of the word segmentation data of the question data. The bottom layer uses the BERT model, and the upper layer uses the double-layer LSTM network model. The historical question data of the questioner are added to the double-layer LSTM network model, where the historical question data of the questioner include "If the inverse of a proposition is a true proposition, then the negation of this proposition must be a true proposition. Is my answer correct?", "A car travels along a hillside for 1000m and its vertical height rises by 50m. What is the slope of this hillside?", "Among the following events, the event that the absolute value of the real number is non-negative" "In Rt△ABC, ∠C=90°sinB=5 / 13, then the value of tanA is 12 / 5", "One of the outer angles of a regular polygon is 20°, then the number of sides of this polygon is 18. Is my answer correct?", etc. According to the time series, the historical question data at different times are sequentially input into the double-layer LSTM network model. Through the hierarchical network model, the latent semantic vector of the current question data of the questioner is retrieved.

[0076] Finally, a multi-task learning model is used, combined with the tutoring intention theory framework, to simultaneously characterize and identify tutoring intentions in different dimensions. Three different MLP networks are used to simultaneously generate latent semantic vectors for tutoring intentions in different dimensions. In the obstacle type dimension, 1 represents the problem obstacle type recognition result, and 0 represents the subject concept obstacle type recognition result; in the subject concept dimension, there are 40 subject concept categories for the recognition of subject concept information; in the learning ability dimension, 0-8 represents the recognition results of nine different learning ability levels. Finally, a more comprehensive tutoring intention recognition result that conforms to educational semantic information is obtained.

[0077] According to another aspect of the present invention, a personalized tutoring method based on tutoring intention recognition is provided, which includes providing personalized tutoring to a user according to the tutoring intention of the user's question data.

[0078] Specifically, if Figure 5 As shown, it may include but is not limited to: matching subject concept information with micro-course resources, matching subject concept information with similar question-and-answer resources, and associating subject concept information with excellent teachers, so as to provide learners with navigation-based step-by-step subject concept guidance at the subject concept guidance level, such as 1) Recognizing subject concepts: explaining the basic information of subject concepts to learners and helping them to recognize subject concepts; 2) Explaining subject concepts: recommending micro-course resources that explain subject concepts to learners. The system explains subject concepts by recommending micro-course resources, thereby further helping learners to fully learn and master subject concepts; 3) Expanding subject concepts: recommending similar question-and-answer resources and excellent teachers to learners. It can also understand information based on the learners' intentions and provide learners with navigation-based step-by-step question guidance, such as 1) Question review: explain the basic information of the question to the learners, including known conditions, known quantities and unknown quantities, so as to help them understand the meaning of the question; 2) Problem solving: recommend solutions to similar questions to learners. The system provides solutions to similar questions and detailed analysis of the questions, allowing learners to understand the solution process of similar questions and further help learners complete the answers to the questions they have questions about; 3) Expansion: recommend high-quality interactive resources and excellent teachers to learners. The system provides learning resources corresponding to the question information, further helping learners plan the next learning path and providing richer learning resources for learners with spare time.

[0079] The technical contents not elaborated in detail in the present invention belong to the common knowledge of those skilled in the art.

[0080] Although the above describes the illustrative specific embodiments of the present invention to facilitate the understanding of the present invention by those skilled in the art, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations using the concept of the present invention are protected.

Claims

1. A method for identifying tutoring intention, characterized in that: The following steps are involved: S1. Preprocess the user question data and construct a text representation for tutoring intention based on the keyword list; S2. Use the hierarchical neural network structure model to obtain the latent semantic vector of the user's question data; S3. Based on latent semantic vector, use multi-task learning model to identify tutoring intention; In step S2, obtaining the latent semantic vector of the user's question data using the hierarchical neural network structure model includes: S21. The underlying network is responsible for semantic analysis of user question data; S22. The upper network is responsible for combining historical data to generate latent semantic vectors of user question data; In step S3, the tutoring intention includes three dimensions: obstacle type, subject concept and learning ability level; The multi-task learning model includes three MLP networks for recognizing tutoring intentions. The training method of the multi-task learning model includes: based on the generated latent semantic vector, simultaneously training three-dimensional data to obtain an MLP network that represents information of different dimensions.

2. The method for recognizing tutoring intention according to claim 1, characterized in that: The method for constructing the keyword list includes: By using word frequency statistics method and combining the experience of expert teachers, we can obtain the subject concept information, relationship word and skill word information in the question data, and then construct a key information list of the question data.

3. The method for recognizing tutoring intention according to claim 2, characterized in that: The subject concept information is the core concepts in the question data, the relationship words are words that describe the relationship between two entities, and the skill words are verbs, including but not limited to calculate, draw, describe, and select.

4. The method according to claim 1, characterized in that: The bottom layer network is a BERT model, and the top layer network is a double-layer LSTM model.

5. The method according to claim 1, characterized in that: In step S3, the obstacle type dimension includes: subject concept-oriented tutoring and question-oriented tutoring; the subject concept dimension includes subject knowledge involved in the user question data; and the learning ability dimension includes the learning ability level identified from the user question data.

6. The method according to claim 5, characterized in that Learning ability is defined as: A learning and understanding, B application and practice, C transfer and innovation, and the levels are then defined specifically according to different disciplines.

7. A personalized tutoring method based on any one of the methods described in claims 1-6, characterized in that: Also includes: S4. Provide personalized coaching to users based on coaching intentions and question data.

Images