A fine-grained knowledge tracking method for learning trajectories

Through a fine-grained knowledge tracking method for learning trajectory, using multi-head self-attention network and time series modeling, combined with multi-task prediction, the problem that the existing technology cannot capture the fine-grained changes in the learner's knowledge state and the complex relationship between the test questions is achieved, Fine-grained tracking and accurate prediction of learners' knowledge status and abilities.

CN114861914BActive Publication Date: 2024-11-08HUAZHONG NORMAL UNIV
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202210326235.4
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-03-29
Publication Date
2024-11-08
Estimated Expiration
2042-03-29

AI Technical Summary

Technical Problem

The existing knowledge tracking model cannot effectively capture the fine-grained changes in the learner's knowledge state, and cannot effectively handle the complex relationships involving multiple knowledge points in a test question. The existing model only focuses on coarse-grained reaction results when predicting, and ignores Differences between reaction time and test difficulty.

Method used

Using a fine-grained knowledge tracking method for learning trajectory, a general knowledge tracking framework is built through multi-head self-attention network and time series modeling, combined with multi-task prediction, to obtain learners' history and current potential knowledge status, and to pay attention through time The force mechanism simulates cognitive laws and predicts learners' reaction speed and results on the test questions.

Benefits of technology

It realizes fine-grained tracking of learners' knowledge status and ability levels, improves prediction accuracy and universality and expansion of models, and can more accurately simulate the complex changes in knowledge status and the relationship between test questions during the learning process.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN114861914B_ABST
    Figure CN114861914B_ABST
Patent Text Reader

Abstract

The present invention relates to the field of educational big data mining and knowledge tracking, and provides a fine-grained knowledge tracking method for learning trajectories, including: (1) constructing a fine-grained knowledge tracking framework for learning trajectories; (2) obtaining the learner's historical potential knowledge state through historical interactions in the learning trajectory; (3) obtaining the learner's current potential ability representation through current interactions in the learning trajectory; (4) fine-grained prediction of the current reaction state. The present invention uses technical methods such as multi-head self-attention networks, multi-task prediction, and time series modeling to model the learner's learning trajectory, track the learner's knowledge state, and predict the learner's future performance, so as to assist in the implementation of precision teaching and personalized learning.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the fields of educational big data mining and knowledge tracking, and in particular to a fine-grained knowledge tracking method oriented to learning trajectories. Background Art

[0002] The development of artificial intelligence technology is conducive to the development of personalized and precise teaching models. Knowledge tracking can analyze the changes in learners' knowledge status during the learning process by modeling the trajectory and results of learners' test answers, and predict learners' performance in answering questions in the future.

[0003] The current deep knowledge tracking models can be roughly divided into three categories according to the models used: knowledge tracking based on long short-term memory networks, knowledge tracking based on attention networks, and knowledge tracking based on custom neural networks.

[0004] The most advanced knowledge tracking method based on long short-term memory network is learning process consistent knowledge tracking (LPKT). LPKT decomposes the learner's learning process into learning gain and forgetting decay, and uses long short-term memory network to simulate the change of knowledge state during the learner's learning process. However, LPKT is limited by the linear characteristics of LSMT model and cannot capture long-distance dependencies. The model training speed is slow and inefficient.

[0005] The most advanced knowledge tracking method based on attention network is context-aware knowledge tracking (AKT). AKT obtains the connection between learners' future performance and past responses through monotonic attention network. However, AKT uses the index distance of the answer sequence, which is equally spaced, while in real life, the intervals between learners' answers are often not equal. In addition, AKT does not model the learners' learning trajectory, so it cannot simulate the complex changes in knowledge status during the learning process.

[0006] The most typical knowledge tracking method based on custom neural networks is the dynamic key-value memory network (DKVMN). DKVMN uses an additional memory network to store the knowledge state of the learner during the answering process, which has good explanatory power. However, models such as DKVMN often only model a phenomenon or law in the answering process, and have poor versatility. In addition, they are limited by their custom calculation rules, have high time complexity and space complexity, and poor scalability.

[0007] In addition, in real life, a test question often involves multiple knowledge points, but the existing advanced models in the field of knowledge tracking often ignore or simplify this point, which not only breaks the connection between test questions, but also simplifies the knowledge status.

[0008] In terms of prediction goals, the current knowledge tracking model aims to predict whether learners can answer questions correctly in the future. However, in the real world, the same reaction results do not mean the same knowledge status and ability level. On the one hand, the difficulty of the test questions and the requirements for the test takers are different; on the other hand, the reaction results of learners answering questions include not only the two states of correct and correct, but also the reaction time, that is, the time spent answering the questions. Existing research has only focused on coarse-grained reaction results, which cannot effectively explore the fine-grained changes in the learners' knowledge status during the answering process.

[0009] Therefore, studying a general, scalable, learning trajectory-oriented, and fine-grained knowledge tracking model that integrates multiple cognitive factors has important research significance and application value for exploring the changes in learners' knowledge status and ability level during the answering process. Summary of the invention

[0010] In response to the shortcomings in the current field of knowledge tracking, the present invention proposes a fine-grained knowledge tracking method for learning trajectories. It comprehensively utilizes technical methods such as multi-head self-attention networks, multi-task prediction, and time series modeling to model the learners' learning trajectories in a more scientific and intuitive way, comprehensively and deeply track the learners' knowledge status and predict the learners' future performance, and assist in the implementation of precise teaching and personalized learning.

[0011] The purpose of the invention is achieved through the following technical measures.

[0012] A fine-grained knowledge tracking method for learning trajectories includes the following steps:

[0013] (1) Construct a fine-grained knowledge tracking framework for learning trajectories; model the learner's learning trajectory and build a general and scalable knowledge tracking framework through encoders, decoders, and predictors;

[0014] (2) Obtain the learner's historical potential knowledge state through historical interactions in the learning trajectory; embed the historical interactions and obtain the potential knowledge state of the corresponding time unit through a one-way encoder;

[0015] (3) Obtain the learner's current potential ability representation through the current interaction in the learning trajectory; use a one-way decoder to obtain the current potential knowledge state through the historical potential knowledge state and the current answer interaction, and simulate cognitive laws through time series characteristics to obtain the potential ability representation;

[0016] (4) Fine-grained prediction of the current reaction state; fine-grained knowledge tracking is achieved through multi-task prediction of multiple reaction features.

[0017] In the above technical solution, step (1) of constructing a fine-grained knowledge tracking framework for learning trajectories includes:

[0018] (1-1) Modeling the learner's answer characteristics in the learning trajectory

[0019] The learner's answering process consists of test questions, learners and time. The learner's answer at time t is abstracted into an interaction triple (E t ,R t ,T t ), where E t is the information of the test questions to be answered, including but not limited to the index information of the test questions, the knowledge points involved in the test questions, etc. t is the response information of the answer, including but not limited to the result of the response, the time of the response, etc. t It is the information of the time when the answer is given. Each time interaction of the learner can be regarded as a time unit, and several time units are connected in series through the time axis to form a time sequence, forming the learner's learning trajectory;

[0020] The answer features of learners in the learning trajectory are composed of two categories, namely time unit features and time series features. Time unit features are the features within each time unit, including the answer question information E t and reaction information R t ,Time series features need to traverse multiple time units to obtain, including the moment information of the learning trajectory and the behavior information generated during the answering process;

[0021] (1-2) Modeling the learning trajectory

[0022] According to step (1-1), the learner's answer characteristics in the learning trajectory are modeled, and the learning trajectory is further modeled, and a knowledge tracking model framework for the learning trajectory is proposed;

[0023] The framework of the knowledge tracking model is decomposed into three steps, namely, first obtaining the potential knowledge state of the time unit in the learning trajectory, then obtaining the current potential ability representation through the current answer questions and the corresponding time series features in the learning trajectory, and finally predicting the learner's reaction to the current test questions; among them, the historical potential knowledge state is obtained by the historical time unit through the intra-unit encoder, and the intra-unit decoder virtually interacts the historical knowledge state with the current test questions to represent the current knowledge state. At the same time, the knowledge state is further converted into potential ability representation, and the reaction predictor is responsible for predicting the relevant reaction state information.

[0024] In the above technical solution, the step (2) of obtaining the learner's historical potential knowledge state through historical interactions in the learning trajectory includes:

[0025] (2-1) Characteristics and Embedded Representation of Historical Time Units

[0026] Assume that the prediction is T t+1 Answer the test questions at any time t+1 The reaction result is T t The question information in the time unit of time includes the question q t , the knowledge point set k involved in the test question t , the reaction information includes the reaction result r t , and the response speed s t , where a test question can contain multiple knowledge points, so the knowledge point set k t For {k t,1 ,k t,2 ,...,k t,m}, reaction speed s t By reaction time rt i The corresponding basic time rt i / bt i The ratio of is mapped to a certain reaction rate category, namely s t =f s (rt t / bt t ), where the basic reaction time bt i is the answer time recommended by experts or the average answer time of all learners, s It is a mapping method. There are many mapping methods here. You can map the ratio into the speed category according to reasonable rules. For example, you can use the ten-equal-division method, that is, divide the reaction speed into ten levels, and the number contained in each level is equal. It is even allowed to directly regard the reaction time as the reaction speed.

[0027] The information of the time unit of historical interaction, that is, {T1,T2,,...,T t The corresponding test questions, knowledge points, reaction results and reaction times are embedded and represented by random embedding and mapped into high-dimensional vectors respectively. If the test questions involve multiple knowledge points, each knowledge point is first mapped into a vector, and then the knowledge point matrix is ​​converted into a knowledge point vector by average pooling. Finally, these four vectors, namely the test question vector, the knowledge point vector, the reaction result vector and the reaction time vector, are added to obtain the vector representation of the corresponding time unit. represents the output of this step, where d k is the dimension of random embedding;

[0028] (2-2) Potential Knowledge State Encoding of Historical Time Units

[0029] The vector representation of the time unit obtained in step (2-1) is encoded through a masked multi-head self-attention network to obtain the potential knowledge state of the corresponding historical time unit;

[0030] The temporal unit encoder consists of three parts: a masked multi-head self-attention network, a residual network, and a feedforward network. The masked multi-head self-attention network uses a mask to mask the context, so that the encoding of each temporal unit can only rely on the temporal unit at the previous moment. The core formula for attention calculation is as follows:

[0031]

[0032]

[0033] Among them, Q, K, and V all come from Input e , is an upper triangular matrix;

[0034] The vector representation of the time unit passes through a masked multi-head self-attention network, and then through a residual network and layer normalization to ensure that the knowledge state will not have a large deviation during the encoding process. Subsequently, the knowledge state vector of each time unit is obtained through a linear transformation of the feedforward network, and finally through a residual network and layer normalization to further improve the encoding ability of the model.

[0035] In the above technical solution, the step (3) of obtaining the learner's current potential ability representation through the current interaction in the learning trajectory includes:

[0036] (3-1) Embedded representation and encoding of current interaction

[0037] By randomly embedding {T1,T2,,...,T t ,T t+1}The corresponding test questions and knowledge points are mapped into high-dimensional vectors, and the vector representation of the corresponding time unit is obtained by summing {T1,T2,,...,T t ,T t+1 The time unit representation of} is passed through a masked multi-head self-attention network, and then through a residual network and layer normalization to encode the vector representation of the current time unit;

[0038] The embedded representation and the encoded vector representation in this step do not contain response information, but only test question information;

[0039] (3-2) Current Potential Knowledge State Decoding Based on Temporal Attention

[0040] Through the correlation between the current time unit and the historical time unit, the knowledge state of the current time unit is obtained by decoding;

[0041] First, the correlation between time units in the time dimension is obtained through time distance. The weight calculation formula is as follows:

[0042]

[0043] Among them, a1, b1, c1 are trainable parameters. Represents T t+1 The time distance correlation between the time unit of the moment and the historical time unit. The larger the weight value, the higher the correlation between the two time units in the time dimension.

[0044] Then, using a masked key-value attention network, T t+1 The vector representation of the moment time unit is regarded as the query, the vector representation of the historical time unit is regarded as the key, and the potential knowledge state of the historical time unit is regarded as the value. The temporal attention network is realized through dot product operation and vector multiplication. The core formula of attention calculation is as follows:

[0045]

[0046]

[0047] in Assign more weight to the historical time units that are more relevant to the current time unit, so as to decode and obtain T t+1 Always in the test questions t+1 Potential knowledge state on

[0048] (3-3) Current potential representation considering temporal characteristics

[0049] Behavioral characteristics elicited by historical answer sequences, i.e., the time interval between the same knowledge l t+1 、Number of knowledge repetitions p t+1 To the potential knowledge state H t+1 Make corrections to obtain potential ability representation;

[0050] The practice proficiency formula is as follows:

[0051]

[0052] Among them, a2, a3, b2, and b3 are all trainable parameters. t+1 ∈(0,1) refers to T t+1 The practice proficiency of the test knowledge points at the time, P t+1 The larger the value is, the higher the learner's proficiency in the knowledge point is. The potential knowledge state is corrected using practice proficiency. The calculation method is as follows:

[0053] C t+1 =H t+1 +P t+1 ×H t+1 ×a4

[0054] Where a4 is a trainable parameter,

[0055] In the above technical solution, the fine-grained prediction of the current reaction state in step (4) includes:

[0056] (4-1) Prediction of reaction rate

[0057] The potential ability vector is passed through the neural network predictor of reaction speed to predict the learner T t+1 Always in the test questions t+1 The reaction speed category on the above, the reaction speed predictor consists of a fully connected layer and an activation function, and its calculation process is as follows:

[0058] s′ t+1 =Softmax(max(0,C t+1 ×W s1 +b s1 )W s2 +b s2 )

[0059] in are all trainable parameters. is the predicted T t+1 The probability matrix of the reaction speed category at the moment, d s is the number of categories of reaction speed;

[0060] (4-2) Prediction of reaction results

[0061] After the potential ability vector is concatenated with the speed vector after the reaction speed predictor, it is predicted by the reaction result predictor. t+1 Always answer the questions correctly t+1 The probability of; the splicing method is as follows:

[0062] R t+1 =concat(s′ t+1 ×W r1 +b r1 ,C t+1 )

[0063] in is a trainable parameter, is the hidden vector used to predict the reaction outcome; the reaction outcome predictor is calculated as follows:

[0064] r′ t+1 =σ(max(0,Rt+1 ×W r2 +b r2 )W r3 +b r3 )

[0065] in are all trainable parameters, r′ t+1 ∈(0,1) is T t+1 The probability of answering the test correctly at a certain time;

[0066] (4-3) Fusion of loss functions

[0067] The loss function of the reaction speed is combined with the loss function of the reaction result by weighted summation. The loss function of the reaction speed is calculated as follows:

[0068]

[0069]

[0070] Where I(s t ) refers to the reaction speed s t Whether it belongs to the corresponding speed category;

[0071] The loss function calculation method of the reaction result is as follows:

[0072]

[0073] The importance of reaction results and reaction speed is adjusted through hyperparameters to balance the impact of multiple tasks on the model. The calculation method of the total loss of fine-grained reaction state prediction is as follows:

[0074] l = kl r +(1-k)*l s

[0075] Among them, k∈[0,1], k is a set constant used to adjust the mixing ratio of the two loss functions.

[0076] In the above technical solution, in step (4), fine-grained knowledge tracking is achieved through multi-task prediction of multiple features, wherein the multi-task prediction of multiple features refers to predicting multiple features of the reaction state, including but not limited to dual-task prediction consisting of reaction speed and reaction result, and can also be combined with other reaction features such as reaction time.

[0077] The fine-grained knowledge tracking method for learning trajectories of the present invention divides the answer features of learners in the learning trajectory into time unit features and time series features according to the time axis; through the time unit encoder and time unit decoder, combined with the learning and forgetting rules in the cognitive process, the learner's potential ability vector is obtained; finally, the fine-grained knowledge tracking is completed through the reaction speed predictor and the reaction result predictor, realizing the knowledge tracking of different learner groups. The present invention can explore the learner's historical answer records, track the changes in knowledge status, predict future learning conditions, and help teachers carry out personalized and precise teaching more efficiently.

[0078] Compared with the prior art, the present invention has the following obvious outstanding substantive features and significant technological advances:

[0079] 1. This invention proposes a modeling method for learning trajectories. The answer features of learners in the learning trajectory are divided into time unit features (within the unit) and time series features (between units) according to their positions on the time axis, and a knowledge tracking model is constructed based on the learning trajectory. Compared with other modeling methods, this method has extremely high scalability and versatility, making up for the shortcomings of the current field in this regard.

[0080] 2. This invention proposes a fine-grained knowledge tracking method for the first time. Compared with other knowledge tracking methods, this method believes that the simple reaction result is not enough to characterize the knowledge state, and introduces the prediction of reaction speed into knowledge tracking for the first time. In addition, this method has high versatility and can effectively improve the performance of most knowledge tracking models.

[0081] 3. The present invention allows a test question to be associated with multiple knowledge points; it simulates the law of learning and forgetting through the temporal attention mechanism, calculates the correlation between time units from the two dimensions of knowledge and time, and allows continuous time distances with unequal spacing; it simulates the impact of historical sequence data of learning trajectories through practice proficiency trends. These are more in line with the real world, conform to the laws of cognitive changes in learning trajectories, and achieve more accurate prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 A diagram of the fine-grained knowledge tracking framework for learning trajectories constructed in the present invention.

[0083] Figure 2 This is an example diagram of the time unit feature and time series feature in the present invention.

[0084] Figure 3 Detailed diagram of the fine-grained knowledge tracking method for learning trajectories in the present invention. DETAILED DESCRIPTION

[0085] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention will be further described in detail below with reference to the accompanying drawings.

[0086] The embodiment of the present invention provides a fine-grained knowledge tracking method for learning trajectory, comprising the following steps:

[0087] (1) Building a fine-grained knowledge tracking framework for learning trajectories

[0088] The purpose of the present invention is to track the changes in potential knowledge status and ability in the learning trajectory through the historical records of learners answering test questions, and to predict the performance of learners in answering test questions in the future.

[0089] The present invention divides the learners' answer characteristics in the learning trajectory into two categories according to the time axis, namely time unit characteristics and time series characteristics, and proposes a universal and extensible model framework. The characteristics and cognitive factors in the learners' answering process can be embedded in the framework in a modular way.

[0090] The process of constructing a fine-grained knowledge tracking framework for learning trajectories includes: modeling the answer characteristics of learners in the learning trajectory and modeling the learning trajectory.

[0091] (1-1) Modeling the learner's answer characteristics in the learning trajectory

[0092] The process of answering by learners mainly consists of test questions, learners and time. Test questions are the object of answering, which is static and does not vary from person to person. Learners are the subject of answering, and the differences and similarities between different learners are often reflected in the reactions to answering test questions. Time is objective and is the information of the moment when learners answer.

[0093] like Figure 1 As shown in the figure, the learner’s answer at time t is abstracted into an interaction triple (E t ,R t ,T t ), where E t is the information of the test questions to be answered, R t is the response information of the answer, T t It is the time information of answering. Each time interaction of the learner can be regarded as a time unit, and several time units form a time sequence through the series connection of the time axis, forming the learner's learning trajectory.

[0094] Therefore, the answer features of learners in the learning trajectory are composed of two categories, namely time unit features and time series features. Time unit features are the features within each time unit, mainly including the answer question information E t and reaction information R tTime series features are meaningful information that can only be obtained by traversing multiple time units, such as the moment information of the learning trajectory and the behavioral information generated during the answering process.

[0095] (1-2) Modeling the learning trajectory

[0096] Based on the modeling of answer features in the previous step, the learning trajectory is further modeled, and a knowledge tracking model framework for learning trajectories that is universal and extensible is proposed.

[0097] like Figure 1 As shown in the figure, the fine-grained knowledge tracking framework for learning trajectories is decomposed into three steps, namely, first obtaining the potential knowledge state of the time unit of historical interaction in the learning trajectory, then obtaining the current potential ability representation through the current answer questions and the corresponding temporal features in the learning trajectory, and finally predicting the learner's reaction to the current question. Among them, the historical potential knowledge state is obtained by the historical time unit through the intra-unit encoder. The intra-unit decoder virtually interacts the historical knowledge state with the current test questions, represents the current knowledge state, and further considers the impact of temporal features on cognition, and further transforms the knowledge state into potential ability representation. The reaction predictor is responsible for predicting the relevant reaction state information.

[0098] This framework has high versatility and extensibility. On the one hand, most of the current knowledge tracking models conform to the structure of this framework, but most of them focus on the design of encoders, and the design of decoders and predictors is relatively one-sided. The present invention greatly makes up for the shortcomings of the current field in this regard. On the other hand, the characteristics of learners' answers in the learning trajectory can almost be decomposed into time units and between time units, and the framework has designed the use positions of these two types of features in the model, which means that in the future, more new features and new cognitive factors can be embedded in the present invention in a modular way, and expanded and iterated on the basis of the present invention.

[0099] (2) Obtain the learner’s historical potential knowledge state through historical interactions in the learning trajectory

[0100] In the time unit of historical interaction, learners answer questions and produce reactions. Therefore, the time unit of history includes question information and reaction information. Assume that the prediction is T t+1 Answer the test questions at any time t+1 The goal of this step is to use the encoder in the unit to t The time unit of each moment is encoded to obtain the learner’s potential knowledge state within the time unit of historical interaction.

[0101] The computational process of obtaining the latent knowledge state of historical interactions includes: the features and embedding representation of historical time units, and the encoding of the latent knowledge state of historical time units.

[0102] (2-1) Characteristics and Embedded Representation of Historical Time Units

[0103] Assume that the prediction is T t+1 Answer the test questions at any time t+1 The reaction result is T t The question information in the time unit of time includes the question q t , the knowledge point set k involved in the test question t , the reaction information includes the reaction result r t , and the response speed s t . One of the test questions can often contain multiple knowledge points, so the knowledge point set k t For {k t,1 ,k t,2 ,...,k t,m Because the reaction time can vary greatly depending on the test questions, we use reaction speed instead of reaction time to eliminate the influence of the test questions. t By reaction time rt i The corresponding basic time rt i / bt i The ratio of is mapped to a certain reaction rate category, namely s t =f s (rt t / bt t ), where the basic reaction time bt i is the answer time recommended by experts or the average answer time of all learners, s It is a mapping method. There are many mapping methods here. You can map the ratio into the speed category according to reasonable rules. For example, you can use the ten-equal-division method, that is, divide the reaction speed into ten levels, and the number contained in each level is equal. It is even allowed to directly regard the reaction time as the reaction speed. Figure 2 A specific example is given, with features within the time unit above the time axis.

[0104] like Figure 3 As shown, the information of the time unit of historical interaction, that is, {T1,T2,,...,T tThe corresponding test questions, knowledge points, reaction results and reaction time are embedded and represented by random embedding and mapped into high-dimensional vectors. In particular, if the test questions involve multiple knowledge points, each knowledge point is first mapped into a vector, and then the knowledge point matrix is ​​converted into a knowledge point vector by average pooling. Finally, these four vectors, namely the test question vector, knowledge point vector, reaction result vector and reaction time vector, are added to obtain the vector representation of the corresponding time unit, and then the vector representation of the corresponding time unit is obtained by using represents the output of this step, where d k is the dimension of random embedding.

[0105] (2-2) Potential Knowledge State Encoding of Historical Time Units

[0106] The vector representation of the time unit obtained in the previous step is encoded through a masked multi-head self-attention network to obtain the potential knowledge state of the corresponding historical time unit.

[0107] The temporal unit encoder consists of three parts: a masked multi-head self-attention network, a residual network, and a feedforward network. The masked multi-head self-attention network uses a mask to mask the context, so that the encoding of each temporal unit can only rely on the temporal unit at the previous moment. The core formula for attention calculation is as follows:

[0108]

[0109]

[0110] Among them, Q, K, and V all come from Input e , is an upper triangular matrix.

[0111] The vector representation of the time unit passes through a masked multi-head self-attention network and then through a residual network and layer normalization. This ensures that the knowledge state will not have large deviations during the encoding process. Subsequently, the knowledge state vector of each time unit is obtained through a linear transformation of the feedforward network. Finally, it passes through a residual network and layer normalization to further improve the encoding ability of the model.

[0112] (3) Obtaining the learner’s current potential ability representation through the current interaction in the learning trajectory

[0113] The previous step obtained the history (i.e. {T1,T2,,...,T t}) The potential knowledge state in the time unit of the interaction. This step decodes the current (i.e., T) by using the historical answering situation and the current interaction in the learning trajectory. t+1 ) Time unit answer questions q t+1The potential knowledge state of is then used to obtain the potential ability representation by considering the temporal characteristics.

[0114] Unlike step (2) which only uses features within the time unit, this step also uses time series features, that is, features that only exist in the sequence. Time series features are divided into two categories. The first category is the time distance feature d t+1 , which is obtained by further processing the timestamp of the time unit; the second type is behavioral features, which are the features of answering behavior obtained by traversing the time unit. Figure 2 As shown in Figure 1, the time series features include time distance features, time intervals between the same knowledge, and the number of knowledge repetitions. The time distance feature refers to the distance between the current time unit and the historical time unit in the time dimension, and the calculation method is shown in the following formula:

[0115] d t+1 =[T t+1 -T1,T t+1 -T2,...,T t+1 -T t ]

[0116] in The same knowledge interval time is the last practice of the current T t+1 The time interval between answering questions and answering questions. The number of knowledge repetitions refers to the number of times the knowledge has been practiced in the past. t+1 The number of knowledge points to answer the test questions at a time. In particular, if a test question involves multiple knowledge points, the same knowledge interval time and knowledge repetition times are averaged. The maximum value of the knowledge interval time is constrained and set to 10 days, in seconds. If the learner practices a certain knowledge point for the first time in the answer sequence, the knowledge interval time is set to the maximum value.

[0117] Figure 2 A calculation example is given. For example, at time T6, two knowledge points, k2 and k3, are involved. Before time T6, k2 has been practiced 2 times, and k3 has been practiced 2 times, so the average number of knowledge repetitions is 2. The last time k2 was practiced was at time T4, and k3 was practiced at time T5, so the knowledge interval is

[0118] The computational process of decoding the latent capability representation of the current interaction includes: embedding representation and encoding of the current interaction, decoding of the current latent knowledge state based on temporal attention, and representation of the current latent capability considering temporal features.

[0119] (3-1) Embedded representation and encoding of current interaction

[0120] T t+1 The time unit feature of the moment only includes the answered questions qt+1 , and the corresponding knowledge point set k t+1 , and does not include reaction information (which needs to be predicted). Similar to the embedding method in step (2-1), {T1,T2,,...,T t ,T t+1} corresponding test questions and knowledge points are mapped into high-dimensional vectors, and the vector representation of the corresponding time unit is obtained by summing. Similar to the method in step (2-2), {T1, T2,, ..., T t ,T t+1 The time unit representation of} passes through a masked multi-head self-attention network, and then through a residual network and layer normalization to encode the vector representation of the current time unit.

[0121] Different from step (2), this step requires T t+1 The time unit of the moment is represented as a vector, and T t+1 The reaction information at the moment is to be predicted, so the embedded representation and the encoded vector representation in this step do not contain reaction information, but only contain test question information.

[0122] (3-2) Current Potential Knowledge State Decoding Based on Temporal Attention

[0123] This step decodes and obtains the knowledge state of the current time unit through the correlation between the current time unit and the historical time unit.

[0124] First, we use the time distance to obtain the correlation between time units in the time dimension. The forgetting curve reveals that human memory will gradually decay over time. Based on this, we assign more weights to time units with closer distances. The weight calculation formula is as follows:

[0125]

[0126] Among them, a1, b1, c1 are trainable parameters. Represents T t+1 The time distance correlation between the time unit of the moment and the historical time unit. The larger the weight value, the higher the correlation between the two time units in the time dimension.

[0127] Then, if Figure 3 As shown, using a masked key-value attention network, T t+1 The vector representation of the moment time unit is regarded as the query, the vector representation of the historical time unit is regarded as the key, and the potential knowledge state of the historical time unit is regarded as the value. The temporal attention network is realized through dot product operation and vector multiplication. The core formula of attention calculation is as follows:

[0128]

[0129]

[0130] in Therefore, more weights can be assigned to the historical time units with higher correlation with the current time unit, so as to decode T t+1 Always in the test questions t+1 Potential knowledge state on

[0131] (3-3) Current potential representation considering temporal characteristics

[0132] The potential knowledge state H obtained in the previous step t+1 Only two factors, knowledge and time distance, are considered. This step uses the behavioral characteristics elicited by the historical answer sequence, i.e., the time interval l between the same knowledge t+1 、Number of knowledge repetitions p t+1 To modify the potential knowledge state and obtain the potential ability representation.

[0133] In the process of answering questions, learners will familiarize themselves with knowledge points through continuous practice. The smaller the interval between the same knowledge points and the more times the knowledge points are repeated, the more likely the learners are to be familiar with the knowledge points, and the better their potential ability in the current test questions may be. The practice proficiency formula based on this design is as follows:

[0134]

[0135] Among them, a2, a3, b2, and b3 are all trainable parameters. t+1 ∈(0,1) refers to T t+1 The practice proficiency of the test knowledge points at the time, P t+1 The larger the value is, the higher the learner's proficiency in the knowledge point is. The potential knowledge state is corrected using practice proficiency. The calculation method is as follows:

[0136] C t+1 =H t+1 +P t+1 ×H t+1 ×a4

[0137] Where a4 is a trainable parameter, This residual connection-like calculation method ensures that the correction bias is reduced.

[0138] (4) Fine-grained prediction of current reaction state

[0139] This step uses latent capability vectors for multi-task prediction and achieves fine-grained knowledge tracking.

[0140] The calculation process includes: prediction of reaction speed, prediction of reaction results, and fusion of loss function.

[0141] (4-1) Prediction of reaction rate

[0142] The potential ability vector is passed through the neural network predictor of reaction speed to predict the learner T t+1 Always in the test questions t+1 The reaction speed predictor consists of a fully connected layer and an activation function, and its calculation process is as follows:

[0143] s′ t+1 =Softmax(max(0,C t+1 ×W s1 +b s1 )W s2 +b s2 )

[0144] in are all trainable parameters. is the predicted T t+1 The probability matrix of the reaction speed category at the moment, d s is the number of categories of reaction speed.

[0145] (4-2) Prediction of reaction results

[0146] This step concatenates the potential ability vector with the speed vector after the reaction speed predictor, and then predicts T through the reaction result predictor. t+1 Always answer the questions correctly t+1 The probability of . The splicing method is as follows:

[0147] R t+1 =concat(s′ t+1 ×W r1 +b r1 ,C t+1 )

[0148] in is a trainable parameter, is the hidden vector used to predict the reaction outcome. The reaction outcome predictor is calculated as follows:

[0149] r′ t+1 =σ(max(0,R t+1 ×W r2 +b r2 )W r3 +b r3 )

[0150] in are all trainable parameters, r′t+1 ∈(0,1) is T t+1 The probability of answering the test correctly at time.

[0151] (4-3) Fusion of loss functions

[0152] The above two steps achieve fine-grained reaction state prediction through multi-task prediction. This step combines the loss function of reaction speed with the loss function of reaction result through weighted summation. The loss function calculation method of reaction speed is as follows:

[0153]

[0154]

[0155] Where I(s t ) refers to the reaction speed s t Whether it belongs to the corresponding speed category.

[0156] The loss function calculation method of the reaction result is as follows:

[0157]

[0158] The importance of reaction results and reaction speed is adjusted through hyperparameters to balance the impact of multiple tasks on the model. The calculation method of the total loss of fine-grained reaction state prediction is as follows:

[0159] l = kl r +(1-k)*l s

[0160] Among them, k is a set constant, k∈[0,1], which is used to adjust the mixing ratio of the two loss functions.

[0161] In the above embodiment, in step (4), fine-grained knowledge tracking is achieved by multi-task prediction of multiple features, wherein the multi-task prediction of multiple features refers to predicting multiple features of the reaction state, including but not limited to dual-task prediction consisting of reaction speed and reaction result, and can also be combined with other reaction features such as reaction time.

[0162] The contents not described in detail in this specification belong to the prior art known to the technical personnel in this field. The fine-grained knowledge tracking task is realized by multi-task prediction of multiple features, which also belongs to the scope of protection of this patent.

[0163] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A fine-grained knowledge tracking method for learning trajectories, characterized by The method comprises the following steps: (1) Construct a fine-grained knowledge tracking framework for learning trajectories; model the learner's learning trajectory and build a general and scalable knowledge tracking framework through encoders, decoders, and predictors; (2) Obtain the learner's historical potential knowledge state through historical interactions in the learning trajectory; embed the historical interactions and obtain the potential knowledge state of the corresponding time unit through a one-way encoder; (3) Obtain the learner's current potential ability representation through the current interaction in the learning trajectory; use a one-way decoder to obtain the current potential knowledge state through the historical potential knowledge state and the current answer interaction, and simulate cognitive laws through time series characteristics to obtain the potential ability representation; (4) Fine-grained prediction of the current reaction state; fine-grained knowledge tracking is achieved through multi-task prediction of multiple reaction features.

2. The fine-grained knowledge tracking method for learning trajectories according to claim 1 is characterized in that The step (1) of constructing a fine-grained knowledge tracking framework for learning trajectories includes: (1-1) Modeling the learner's answer characteristics in the learning trajectory The learner's answering process consists of test questions, learners and time. The learner's answer at time t is abstracted into an interaction triple (E t ,R t ,T t ), where E t is the information of the test questions to be answered, R t is the response information of the answer, T t It is the information of the time when the answer is given. Each time interaction of the learner can be regarded as a time unit, and several time units are connected in series through the time axis to form a time sequence, forming the learner's learning trajectory; The answer features of learners in the learning trajectory are composed of two categories, namely time unit features and time series features; time unit features are the features within each time unit, including the answer question information E t and reaction information R t ,Time series features need to traverse multiple time units to obtain, including the moment information of the learning trajectory and the behavior information generated during the answering process; (1-2) Modeling the learning trajectory Based on the modeling of the answer features in step (1-1), the learning trajectory is further modeled, and a knowledge tracking model framework for the learning trajectory is proposed; The framework of the knowledge tracking model is decomposed into three steps, namely, first obtaining the potential knowledge state of the time unit of historical interaction in the learning trajectory, then obtaining the current potential ability representation through the current answer questions and the corresponding time series features in the learning trajectory, and finally predicting the learner's reaction to the current test questions; among them, the historical potential knowledge state is obtained by the historical time unit through the intra-unit encoder, and the intra-unit decoder virtually interacts the historical knowledge state with the current test questions to represent the current knowledge state. At the same time, the knowledge state is further converted into potential ability representation, and the reaction predictor is responsible for predicting the relevant reaction state information.

3. The fine-grained knowledge tracking method for learning trajectories according to claim 1 is characterized in that The step (2) of obtaining the learner's historical potential knowledge state through historical interactions in the learning trajectory includes: (2-1) Characteristics and Embedded Representation of Historical Time Units Assume that the prediction is T t+1 Answer the test questions at any time t+1 The reaction result is T t The question information in the time unit of time includes the question q t , the knowledge point set k involved in the test question t , the reaction information includes the reaction result r t , and the response speed s t , where a test question can contain multiple knowledge points, so the knowledge point set k t For {k t,1 ,k t,2 ,...,k t,m }, reaction speed s t By reaction time rt i The corresponding basic time rt i / bt i The ratio of is mapped to a certain reaction rate category, namely s t =f s (rt t / bt t ), where the basic reaction time bt i is the answer time recommended by experts or the average answer time of all learners, s It is the mapping method; The information of the time unit of historical interaction, that is, {T1,T2,,...,T t The corresponding test questions, knowledge points, reaction results and reaction times are embedded and represented by random embedding and mapped into high-dimensional vectors respectively. If the test questions involve multiple knowledge points, each knowledge point is first mapped into a vector, and then the knowledge point matrix is ​​converted into a knowledge point vector by average pooling. Finally, these four vectors, namely the test question vector, the knowledge point vector, the reaction result vector and the reaction time vector, are added to obtain the vector representation of the corresponding time unit. represents the output of this step, where d k is the dimension of random embedding; (2-2) Potential Knowledge State Encoding of Historical Time Units The vector representation of the time unit obtained in step (2-1) is encoded through a masked multi-head self-attention network to obtain the potential knowledge state of the corresponding historical time unit; The temporal unit encoder consists of three parts: a masked multi-head self-attention network, a residual network, and a feedforward network. The masked multi-head self-attention network uses a mask to mask the context, so that the encoding of each temporal unit can only rely on the temporal unit at the previous moment. The core formula for attention calculation is as follows: Among them, Q, K, and V all come from Input e , is an upper triangular matrix; The vector representation of the time unit passes through a masked multi-head self-attention network, and then through a residual network and layer normalization to ensure that the knowledge state will not have a large deviation during the encoding process. Subsequently, the knowledge state vector of each time unit is obtained through a linear transformation of the feedforward network, and finally through a residual network and layer normalization to further improve the encoding ability of the model.

4. The fine-grained knowledge tracking method for learning trajectories according to claim 1 is characterized in that The step (3) of obtaining the learner's current potential ability representation through the current interaction in the learning trajectory includes: (3-1) Embedded representation and encoding of current interaction By randomly embedding {T1,T2,,...,T t ,T t+1 }The corresponding test questions and knowledge points are mapped into high-dimensional vectors, and the vector representation of the corresponding time unit is obtained by summing {T1,T2,,...,T t ,T t+1 The time unit representation of} is passed through a masked multi-head self-attention network, and then through a residual network and layer normalization to encode the vector representation of the current time unit; The embedded representation and the encoded vector representation in this step do not contain response information, but only test question information; (3-2) Current Potential Knowledge State Decoding Based on Temporal Attention Through the correlation between the current time unit and the historical time unit, the knowledge state of the current time unit is obtained by decoding; First, the correlation between time units in the time dimension is obtained through time distance. The weight calculation formula is as follows: Among them, a1, b1, c1 are trainable parameters. Represents T t+1 The time distance correlation between the time unit of the moment and the historical time unit. The larger the weight value, the higher the correlation between the two time units in the time dimension. Then, using a masked key-value attention network, T t+1 The vector representation of the moment time unit is regarded as the query, the vector representation of the historical time unit is regarded as the key, and the potential knowledge state of the historical time unit is regarded as the value. The temporal attention network is realized through dot product operation and vector multiplication. The core formula of attention calculation is as follows: in Assign more weight to the historical time units that are more relevant to the current time unit, so as to decode and obtain T t+1 Always in the test questions t+1 Potential knowledge state on (3-3) Current potential representation considering temporal characteristics Behavioral characteristics elicited by historical answer sequences, i.e., the time interval between the same knowledge l t+1 、Number of knowledge repetitions p t+1 To the potential knowledge state H t+1 Make corrections to obtain potential ability representation; The practice proficiency formula is as follows: Among them, a2, a3, b2, and b3 are all trainable parameters. t+1 ∈(0,1) refers to T t+1 The practice proficiency of the test knowledge point at the moment, use the practice proficiency to correct the potential knowledge state, the calculation method is as follows: C t+1 =H t+1 +P t+1 ×H t+1 ×a4 Where a4 is a trainable parameter, 5. The fine-grained knowledge tracking method for learning trajectories according to claim 1 is characterized in that The fine-grained prediction of the current reaction state in step (4) is specifically: (4-1) Prediction of reaction rate The potential ability vector is passed through the neural network predictor of reaction speed to predict the learner T t+1 Always in the test questions t+1 The reaction speed category on the above, the reaction speed predictor consists of a fully connected layer and an activation function, and its calculation process is as follows: s′ t+1 =Softmax(max(0,C t+1 ×W s1 +b s1 )W s2 +b s2 ) in are all trainable parameters. is the predicted T t+1 The probability matrix of the reaction speed category at the moment, d s is the number of categories of reaction speed; (4-2) Prediction of reaction results After the potential ability vector is concatenated with the speed vector after the reaction speed predictor, it is predicted by the reaction result predictor. t+1 Always answer the questions correctly t+1 The probability of; the splicing method is as follows: R t+1 =concat(s′ t+1 ×W r1 +b r1 ,C t+1 ) in is a trainable parameter, is the hidden vector used to predict the reaction outcome; the reaction outcome predictor is calculated as follows: r′ t+1 =σ(max(0,R t+1 ×W r2 +b r2 )W r3 +b r3 ) in are all trainable parameters, r′ t+1 ∈(0,1) is T t+1 The probability of answering the test correctly at a certain time; (4-3) Fusion of loss functions The loss function of the reaction speed is combined with the loss function of the reaction result by weighted summation. The loss function of the reaction speed is calculated as follows: Where I(s t ) refers to the reaction speed s t Whether it belongs to the corresponding speed category; The loss function calculation method of the reaction result is as follows: The importance of reaction results and reaction speed is adjusted through hyperparameters to balance the impact of multiple tasks on the model. The calculation method of the total loss of fine-grained reaction state prediction is as follows: l=kl r +(1-k)*l s Among them, k∈[0,1], k is a set constant used to adjust the mixing ratio of the two loss functions.

Citation Information

Patent Citations

  • Knowledge tracking method based on test question heterogeneous graph representation and learner embedding

    CN113344053A

  • Adaptive learning support device and method based on deep knowledge tracking

    CN113610235A

  • Cognitive description fused attention knowledge tracking method

    CN114021722A