Differential chemical condition modeling method based on long-term and short-term interaction data

Abstract

The invention relates to the technical field of cognitive diagnosis, and discloses a differential chemical condition modeling method based on long-term and short-term interaction data, and the training process of a cognitive model comprises the steps: arranging question samples composed of interaction data and answering results of students on each question into learning records of the students; constructing a recent learning record and a long-term learning record based on the learning record; the recent learning records are sent to a static knowledge tracking module for learning modeling, and recent expressions of the students are obtained; the long-term learning record of the student is sent to a dynamic knowledge tracking module, and the historical cognitive state of the student is captured; obtaining the current cognitive state of the student through a fusion gating network, and predicting the probability that the student correctly answers the question; and taking the cross entropy between the probability and the actual answer result as a loss function for training. According to the method, the cognitive level of the students can be reflected, the future learning achievements of the students can be predicted, and a scientific basis is provided for educational decision making.

Description

Technical Field

[0001] The present invention relates to the technical field of cognitive diagnosis, and in particular to a differential chemical emotion modeling method based on long-term and short-term interaction data. Background Art

[0002] The cognitive model is a mapping of the computer model of the human cognitive process. The technology of establishing cognitive models is often called cognitive modeling. Its purpose is to explore and study the human thinking mechanism, especially the human information processing mechanism, from certain aspects. At the same time, it also provides new architectures and technical methods for designing corresponding artificial intelligence systems.

[0003] For a long time, most cognitive modeling models have mainly used outcome data such as knowledge and learner reactions, with a single data dimension, ignoring the behavioral characteristics of learners in the process of interacting with test questions, such as the order of answering questions, programming data submission data, answering time, multi-round dialogue questions and answers, and learning mouse click behavior, etc., and failed to integrate the existing multi-type multimodal data collected into the cognitive model. Therefore, the speculation of the degree of knowledge mastery is one-sided, and the modeling between students lacks differences; on the other hand, there will be multiple rounds of interaction data in the actual learning process, and the sequence of multiple rounds of interaction data is long. Excessive attention to long-term or short-term interaction data will affect the model prediction performance. Therefore, how to model some students' learning interaction behavior data, while considering the long-term and short-term data of students' learning process, so that students' learning modeling is more differentiated, is still under study. Summary of the invention

[0004] In order to solve the above technical problems, the present invention provides a differential chemical situation modeling method based on long-term and short-term interaction data.

[0005] In order to solve the above technical problems, the present invention adopts the following technical solutions:

[0006] A differentiated sentiment modeling method based on long-term and short-term interaction data is used to model a student's cognitive model. The cognitive model can predict the probability of the student correctly answering a new question based on the student's cognitive state and the interaction data of the new question. The cognitive model includes a static knowledge tracking module and a dynamic knowledge tracking module. The construction and training process of the cognitive model includes the following steps:

[0007] S1, organize the students’ interactive data and answer results on each question into a sample of questions, which is a student’s learning record LP = {(ex1, re1), (ex2, re2), …, (ex T ,re T )};ex T is the interaction data of the Tth question, re T is the answer result of the Tth question, T is the total number of question samples in the learning record LP;

[0008] S2, based on the student's learning record, construct the recent learning record and long-term learning record. If the recent learning record has L question samples, then the recent learning record LP l ={(ex T-L ,re T-L ),…,(ex T ,re T )}, the long-term learning record is the learning record LP;

[0009] S3, send the recent learning records to the static knowledge tracking module based on the language model to learn modeling and obtain the student’s recent performance S l ;

[0010] S4, send the long-term learning records to the dynamic knowledge tracking module based on the language model to capture the student's historical cognitive state S h , and the historical cognitive state S h Save to the historical database;

[0011] S5, through the fusion gating network, S l and S h Adaptively combine to get the student's current cognitive state S c , using the student's current cognitive state S c Interaction data ex with the topic used for training tr To predict the probability that the student answers the question correctly is P;

[0012] S6, using the probability P and the actual answer result re of the student answering the question tr The cross entropy between is used as the loss function to train the cognitive model and realize the modeling of the cognitive model.

[0013] Furthermore, the interactive data of the question includes question content, answering time, number of answering attempts, answering process, and answering tools; the answering results of the question include whether the student answered the question correctly or incorrectly.

[0014] Furthermore, step three specifically includes:

[0015] Record your recent studies l ={(ex T-L ,re T-L ),…,(ex T ,re T )} The whole is taken as a long text and input into the word embedding model to generate words and perform position encoding to obtain the vector x;

[0016] Feed the vector x into a gating network vector G, and the i-th element of the vector x is assigned a weight wi , the weight can affect the processing of each element of vector x by the knowledge tracking expert E. After the vector x passes through the knowledge tracking expert, the student hidden state G(x)E(x) is obtained;

[0017] At the same time, the vector x is also sent to the large language model to obtain the student hidden state LLM(x) of the same dimension;

[0018] Combine the student hidden state 1 and the student hidden state 2 to get the student's recent performance S l :

[0019] S l =G(x)E(x)+LLM(x).

[0020] Furthermore, step four specifically includes:

[0021] All interaction data in the long-term learning record are fed into the large language model in the form of words to form a series of word vectors. After the word vectors pass through the trainable Transformer layer and the feedforward neural network, they are converted into the interaction data representation vector z.

[0022] If there is no old learning situation in the historical learning situation database, the recurrent neural network starts from a randomly initialized initial value h0. If there is old learning situation in the historical learning situation database, the recurrent neural network starts from the old learning situation S′ h Start; input the interaction data representation vector z and the corresponding answer result re corresponding to each interaction data in the long-term learning record into the recurrent neural network, and each step processes a combination of the interaction data representation vector z and the answer result re;

[0023] In the tth step, the input of the recurrent neural network is the interaction data representation vector z t and the corresponding answer results t The combined encoding of the combined input vector for:

[0024]

[0025] in, is the operation of concatenating two vectors;

[0026] The student hidden state h of the previous step t-1 and the combined input vector of the current step Input into the recurrent neural network together, and update the student's hidden state h in the tth step t :

[0027]

[0028]

[0029]

[0030]

[0031] h t =o t ·tan h (c t );

[0032] Among them, i t , f t , o t They represent the input gate, forget gate, and output gate respectively, and c t is a cell memory vector, are the embedding parameters of students;

[0033] After modeling the practice process of each student from the first step to the Tth step, the historical cognitive state S of the current student is obtained. h =h T , the historical cognitive state S h It is stored as new learning situation in the historical learning situation database.

[0034] Furthermore, step five specifically includes:

[0035] Calculate the interaction data ex for the training questions tr Similarity with the interaction data in recent learning records att:

[0036]

[0037] Where σ is the sigmoid activation function, ex′ l Indicates ex l The transpose of

[0038] Through a fusion gate Γ, the recent learning records and long-term learning records are controlled for the student's cognitive state S c Degree of impact:

[0039]

[0040] S c =Γ·S l +(1-Γ)·S h ;

[0041] bΓ represents a matrix and a vector respectively, both of which are trainable parameters of the fusion gate; using the cognitive state S c and interaction data ex tr The inner product of is used to simulate the decision-making process of students when answering questions:

[0042]

[0043] represents the probability of predicting the student's correct answer to the question, Indicates the prediction of the student's correct answer to the question.

[0044] Furthermore, step six specifically includes:

[0045] The loss function loss is:

[0046] loss=-∑(re tr log P+(1-re tr )log(1-P))+λ θ ||θ||2;

[0047] Among them, θ is a trainable parameter; λ is a hyperparameter used to control the regularization strength.

[0048] Compared with the prior art, the beneficial technical effects of the present invention are:

[0049] This invention effectively applies the powerful language understanding ability of the large language model to the modeling of student learning situations; it not only takes into account students' recent academic achievements, but also incorporates their long-term learning history into the analysis, so that it can more comprehensively understand the learning trajectory of each student and highlight the differences between individuals. In this way, it can not only reflect the students' cognitive level, but also help predict their future learning outcomes, providing a scientific basis for educational decision-making. BRIEF DESCRIPTION OF THE DRAWINGS

[0050] Figure 1 Schematic diagram of the model training process of the present invention. DETAILED DESCRIPTION

[0051] A preferred embodiment of the present invention is described in detail below with reference to the accompanying drawings.

[0052] Learning situation is the abbreviation of learning situation. The present invention models the learning situation and obtains a cognitive model that can predict the probability of students correctly answering a new question based on the students' cognitive state and the interaction data of the new question.

[0053] like Figure 1 As shown, the present invention proposes a differentiated learning situation modeling method based on long-term and short-term interaction data, which sends the learning records to the static knowledge tracking module and the dynamic knowledge tracking module respectively, to obtain the students' recent learning situation modeling and long-term learning situation modeling, and the two are integrated to obtain the comprehensive learning situation, thereby realizing the prediction of the answer situation to new questions.

[0054] The differential chemical sentiment modeling method based on long-term and short-term interaction data in the present invention comprises the following steps:

[0055] S1, standardize and organize all the students' interaction data and answer results to obtain the student's learning record LP = {(ex l ,re1),(ex2,re2),…,(ex T ,re T )}.

[0056] Step S1 specifically includes the following steps:

[0057] S11, collect students’ interaction data and answer results on each question.

[0058] The interactive data that can be collected include question content (question_content), answer time (answer_time), number of answer attempts (attempt_count), answer process (answer_process), answer tools (tools_used), etc.; the answer result is a binary value, 1 indicates a correct answer, and 0 indicates an incorrect answer.

[0059] S12, saves students’ interaction data in the form of key-value pairs.

[0060] Students’ interactive data needs to be saved in the form of key-value pairs, where the key represents the interactive data type and the value represents the corresponding value. For example, a student’s answer data for a question can be {″question_content″:″What is the capital of the United Kingdom?”,″answer_time″:″120seconds″,″attempt_count″:″1″,″answer_process″:[″Opened dictionary″,″Searched for′capital′″,″Typed inanswer″],″tools_used″:[″dictionary”]}.

[0061] S13, sorting all answer records into the learning record LP according to the answering order.

[0062] When the number of questions is T, summarize all answer records from 1 to T in chronological order, and the learning record LP = {(ex1, re1), (ex2, re2), ..., (ex T ,re T )}.

[0063] S2, divides students’ learning records into recent learning records and long-term learning records.

[0064] Step S2 specifically includes the following steps.

[0065] S21, differentiate the learning records according to the contextual constraints of the language model;

[0066] If the context length of the language model is limited to k, and the total data size of at most L recent learning records is less than k, then these L learning records are classified as recent learning records, and all answer records, that is, the entire LP, are classified as long-term learning records.

[0067] S22: Sort the recent learning records and long-term learning records in a standardized manner according to the order of answering;

[0068] Recent learning records are LP l ={(ex T-L ,re T-L ),…,(ex T ,re T )};

[0069] Long-term learning record is LP h ={(ex1,re1),(ex2,re2),…,(ex T ,re T )}.

[0070] S3, the recent learning records are sent to the static knowledge tracking module based on the language large model designed by the present invention to learn modeling, and the student’s recent performance S l .

[0071] Step S3 specifically includes the following steps:

[0072] S31, encode the students’ test records into data and position codes.

[0073] First, it is necessary to perform word segmentation on the students' test-taking records to decompose the long text into words or phrases (i.e., tokens), which can be achieved by using a word segmentation tool, such as NLTK; then, the present invention sends these tokens into a word embedding model, which can be a pre-trained Word2Vec model, and each token will be converted into a vector of a fixed length; and an integer is assigned to each element in the sequence to indicate its position in the sequence.

[0074] S32, feeds the encoded data into the knowledge tracking expert model with control gates.

[0075] This part borrows the concepts of gating network and adapter fine-tuning in the hybrid expert model, treating the knowledge tracking model as an expert model, and the gating network controls whether the token passes the knowledge tracking model. This knowledge tracking model can be any model suitable for knowledge tracking, or it can adopt a variety of advanced knowledge tracking algorithms to effectively capture and understand students' behavior patterns and knowledge mastery during the learning process, such as the RNN-based DKT model or the Transformer-based SAKT model; the input of the gating network is the word embedding vector, and the output is a value between 0 and 1, representing whether each token should pass the knowledge tracking model. For example, if the output of the gating network is 1, then the token completely passes the knowledge tracking model; if the output of the gating network is 0, then the token does not pass the knowledge tracking model at all.

[0076] S33, sending the encoded data to the large language model.

[0077] It should be noted that the powerful natural language processing capabilities of the large language model are utilized here, and the large model needs to be configured in a certain way so that its output results are similar to those of the knowledge tracking model, that is, the student's knowledge status.

[0078] S34, combining the results obtained from S32 and S33.

[0079] The two steps both get the student's knowledge status, which needs to be added together to get the final student knowledge status, which is used as the student's recent performance S l ,The advantage here is that the tight coupling of the large language model and ,the knowledge tracking model is achieved.

[0080] S4, the student's complete learning record is sent to the dynamic knowledge tracking module based on the language model designed by the present invention, so as to capture the student's overall historical cognitive state S h , and the historical cognitive state S h Save to the historical database.

[0081] Step S4 specifically includes the following steps:

[0082] S41, use a large language model to encode students' interaction data.

[0083] All user interaction data is fed into the large language model in the form of tokens. The large language model converts these tokens into a series of vectors, each of which corresponds to a token and contains rich semantic and contextual information.

[0084] S42, the encoding result is passed through a Transformer and a feed-forward neural network layer to obtain the interactive data encoding.

[0085] The Transformer layer can capture the intrinsic relationship between students’ problem-solving records. The feedforward neural network layer will further extract and process the output of the Transformer layer to obtain a deeper representation of students’ problem-solving records, and finally obtain a higher-dimensional interaction data representation vector z, which contains all the interaction information of the question.

[0086] S43, all the questions that need to be processed are sent to the recurrent neural network.

[0087] It should be noted that if there is no old learning situation, the recurrent neural network starts from a randomly initialized initial value h0. If there is an old learning situation, the recurrent neural network starts from the old learning situation S′ h Start; in the tth step, the input of the recurrent neural network is the interaction data representation vector z t and the corresponding score re t After one update, we can get the student's hidden state h t .

[0088] S44, save the learning situation to the historical learning situation database.

[0089] In order to facilitate the continuous updating of learning conditions, it is necessary to obtain the historical cognitive state S after the learning condition modeling. h As the learning situation, it is saved in the historical learning situation database, so that the old learning situation can be used as model input when new question data is available.

[0090] S5: S is integrated into the gating network l and S h Adaptively combine to get the student's current cognitive state S c , using the student's current cognitive state S c Interaction data ex with the topic used for training tr , predict the probability that the student answers the question correctly is P.

[0091] Step S5 specifically includes the following steps:

[0092] S51, calculating the similarity between the new interaction data and the old interaction data.

[0093] Calculate the interaction data ex for the training questions t The similarity between r and the interaction data in the recent learning record can be achieved by calculating the dot product between them. The dot product is a common vector operation that can measure the similarity between two vectors:

[0094]

[0095] Where σ is the sigmoid activation function, ex′ lIndicates ex l The transpose of .

[0096] S52, obtains the overall cognitive state of the students through a fusion gate.

[0097] In this embodiment, the student's learning records are divided into two parts: long-term learning records and recent learning records, and two cognitive states S of the student are obtained. l and S h , a fusion gate is needed to combine the two cognitive states. By comparing the similarities between the questions that students have recently answered, the importance of different future questions can be controlled.

[0098] S53, predicting the answer result of the new question.

[0099] In order to make predictions, this embodiment uses the inner product of cognitive state and interaction data to simulate the decision-making process of students when answering questions, and outputs the prediction of the student's answer to the new question, so that the probability of the student answering correctly can be known.

[0100] S6: Use the predicted correct answer probability P and the actual binary answer result re tr The cross entropy between is used as the loss function to train the cognitive model.

[0101] S61: Arrange parameters and set hyperparameters;

[0102] In this step, you need to organize all the parameters required for the model, including the model's weights and biases, etc. At the same time, you also need to set some hyperparameters, such as learning rate, batch size, number of training rounds, etc.

[0103] S62: training model;

[0104] Specifically, this embodiment compares the prediction results of the cognitive model (i.e., the probability that the student will answer the question correctly) with the actual binary answer results and calculates the cross entropy between them. Then, an optimization algorithm (such as gradient descent) can be used to update the parameters of the cognitive model to minimize the loss function. This process is repeated until the performance of the model reaches a satisfactory level or the preset number of training rounds is reached.

[0105] It is obvious to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential features of the present invention. Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting, and the scope of the present invention is defined by the appended claims rather than the above description, and it is intended that all changes falling within the meaning and scope of the equivalent elements of the claims are included in the present invention, and any reference numerals in the claims should not be regarded as limiting the claims involved.

[0106] In addition, it should be understood that although this specification is described in terms of implementation methods, not every implementation method contains only one independent technical solution. This narrative method of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation methods that can be understood by those skilled in the art.

Claims

1. A differential chemical sentiment modeling method based on long-term and short-term interaction data, characterized in that: The method is used to model a cognitive model of a student, wherein the cognitive model can predict the probability of the student correctly answering the new question according to the cognitive state of the student and the interaction data of the new question. The cognitive model includes a static knowledge tracking module and a dynamic knowledge tracking module. The construction and training process of the cognitive model includes the following steps: S1, organize the students’ interactive data and answer results on each question into a sample of questions, which is the students’ learning record LP = {(ex1, re1), (ex2, re2), …, (ex T ,re T )};ex T is the interaction data of the Tth question, re T is the answer result of the Tth question, T is the total number of question samples in the learning record LP; S2, based on the student's learning record, construct the recent learning record and long-term learning record. If the recent learning record has L question samples, then the recent learning record LP l ={(ex T-K ,re T-K ),…,(ex T ,re T )}, the long-term learning record is the learning record LP; S3, send the recent learning records to the static knowledge tracking module based on the language model to learn modeling and obtain the student’s recent performance S l ; S4, send the long-term learning records to the dynamic knowledge tracking module based on the language model to capture the student's historical cognitive state S h , and the historical cognitive state S h Save to the historical database; S5, through the fusion gating network, S l and S h Adaptively combine to get the student's current cognitive state S c , using the student's current cognitive state S c Interaction data ex with the topic used for training tr To predict the probability that the student answers the question correctly is P; S6, using the probability P and the actual answer result re of the student answering the question tr The cross entropy between is used as the loss function to train the cognitive model and realize the modeling of the cognitive model.

2. The differential chemical sentiment modeling method based on long-term and short-term interaction data according to claim 1 is characterized in that: The interactive data of the question includes question content, answering time, number of answering attempts, answering process, and answering tools; the answering results of the question include whether the student answered the question correctly or incorrectly.

3. The differential chemical sentiment modeling method based on long-term and short-term interaction data according to claim 1 is characterized in that: Step three specifically includes: Record your recent studies l ={(ex T-L ,re T-L ),…,(ex T ,re T )} The whole is taken as a long text and input into the word embedding model to generate words and perform position encoding to obtain the vector x; Feed the vector x into a gating network vector G, and the i-th element of the vector x is assigned a weight w i , the weight can affect the processing of each element of vector x by the knowledge tracking expert E. After the vector x passes through the knowledge tracking expert, the student hidden state G(x)E(x) is obtained; At the same time, the vector x is also sent to the large language model to obtain the student hidden state LLM(x) of the same dimension; Combine the student hidden state 1 and the student hidden state 2 to get the student's recent performance S l : S l =G(x)E(x)+LLM(x)。 4. The differential chemical sentiment modeling method based on long-term and short-term interaction data according to claim 1 is characterized in that: Step 4 specifically includes: All interaction data in the long-term learning record are fed into the large language model in the form of words to form a series of word vectors. After the word vectors pass through the trainable Transformer layer and the feedforward neural network, they are converted into the interaction data representation vector z. If there is no old learning situation in the historical learning situation database, the recurrent neural network starts from a randomly initialized initial value h0. If there is an old learning situation in the historical learning situation database, the recurrent neural network starts from the old learning situation S ′ h Start; input the interaction data representation vector z and the corresponding answer result re corresponding to each interaction data in the long-term learning record into the recurrent neural network, and each step processes a combination of the interaction data representation vector z and the answer result re; In the tth step, the input of the recurrent neural network is the interaction data representation vector z t and the corresponding answer results t The combined encoding of the combined input vector for: in, is the operation of concatenating two vectors; The student hidden state h of the previous step t-1 and the combined input vector of the current step Input into the recurrent neural network together, and update the student's hidden state h in the tth step t : h t =o t ·tan h (c t ); Among them, i t ,f t ,o t They represent the input gate, forget gate, and output gate respectively, and c t is a cell memory vector, are the embedding parameters of students; After modeling the practice process of each student from the first step to the Tth step, the historical cognitive state S of the current student is obtained. h =h T , the historical cognitive state S h It is stored as new learning situation in the historical learning situation database.

5. The differential chemical sentiment modeling method based on long-term and short-term interaction data according to claim 1 is characterized in that: Step 5 specifically includes: Calculate the interaction data ex for the training questions tr Similarity with the interaction data in recent learning records att: Where σ is the sigmoid activation function, ex′ l Indicates ex l The transpose of Through a fusion gate Γ, the recent learning records and long-term learning records are controlled for the student's cognitive state S c Degree of impact: S c =Γ·S l +(1-C)·S h ; b Γ Represent a matrix and a vector respectively, both of which are trainable parameters of the fusion gate; using the cognitive state S c and interaction data ex tr The inner product of is used to simulate the decision-making process of students when answering questions: represents the probability of predicting the student's correct answer to the question, Indicates the prediction of the student's correct answer to the question.

6. The differential chemical sentiment modeling method based on long-term and short-term interaction data according to claim 1 is characterized in that: Step 6 specifically includes: The loss function loss is: loss=-∑(re tr log P+(1-re tr )log(1-P))+λ θ ‖θ‖ 2 ; Among them, θ is a trainable parameter; λ is a hyperparameter used to control the regularization strength.

Images