A teaching evaluation sentiment analysis method fusing cognitive transfer

Abstract

The present application relates to a kind of fusion cognitive transfer teaching evaluation sentiment analysis method, belong to the field of educational informationization.The method includes: the data collected is cleaned and preprocessed;Unstructured teaching evaluation text is carried out word embedding using BERT language model, it is input in bidirectional GRU network and learns, the context hidden memory information of output text;Student historical practice record is modeled using neural cognitive diagnosis model, and the personalized cognitive ability vector of student is obtained as prior knowledge representation;The text hidden memory information extracted is fused into cognitive transfer matrix with prior knowledge information, and prior knowledge is transferred into text memory by attention mechanism, and the real emotional characteristics contained in text are enhanced.The present application can accurately classify the sentiment of teaching evaluation text carrying conflictive emotional characteristics, solve the problem that sentiment is misclassified due to the existence of conflictive emotional characteristics in teaching evaluation.

Description

Technical Field

[0001] This invention belongs to the field of educational informatization and relates to a teaching evaluation sentiment analysis method that integrates cognitive transfer. Background Technology

[0002] Text sentiment analysis, a branch of natural language processing, can automatically extract and identify emotions and opinions from text, thereby enabling the analysis, processing, summarization, and reasoning of subjective texts with emotional overtones. In recent years, the introduction of deep learning has empowered text sentiment analysis, making significant progress in solving unstructured and complex semantic data. This technology has been widely used in product reviews, film evaluations, and other fields. However, unlike general applications, the emotions implied in teaching evaluations are often subtle and implicit, increasing the difficulty of emotion recognition. Furthermore, the same teaching evaluation text may contain conflicting emotions, meaning the model may extract inconsistent emotional features from different evaluation entities within the same text. For example, a teaching evaluation might state, "The teacher's lesson content was very good, but the voice was a little soft." This evaluation includes the evaluation entities "lesson content" and "voice." Students might express positive emotions towards the entity "lesson content" but negative emotions towards "voice," yet overall, their assessment of the teaching quality is positive. In other words, there is a conflict in the emotional features of "course content" and "lecture voice" in the evaluation. This conflict can interfere with the model's recognition of the true emotion in the text, leading to misclassification. How to accurately determine the emotional polarity of teaching evaluations carrying conflicting emotional features has become a key problem that urgently needs to be solved in the context of sentiment analysis of teaching evaluations.

[0003] Research in educational psychology and cognitive psychology suggests that students' cognitive outcomes during course learning can potentially influence their emotions and satisfaction with teaching. In other words, students' perception of teaching quality is influenced by their understanding of course knowledge; this phenomenon is termed "cognitive transfer" in this invention. Reference [Huang Changqin, Yu Jianhui, Wang Xizhe. Research on learning recommendation based on emotion analysis in learning cloud space [J]. China Educational Technology, 2018(10):7-14+39.] indicates that students' knowledge level, learning outcomes, and other factors are directly related to emotion generation and are key factors influencing learners' emotions. Reference [Chen WS, Yao A YT. An Empirical Evaluation of Critical Factors Influencing Learner Satisfaction in Blended Learning: A pilot study [J]. Universal Journal of Educational Research, 2016,4(7):1667-1671.] indicates that when assessing students' learning satisfaction, incorporating knowledge-dependent variables such as the learning environment and students' test scores makes the satisfaction assessment results more accurate and meaningful. By capturing students' cognitive ability vectors and integrating them into the sentiment analysis task of teaching evaluation, the authentic emotional characteristics of the text are enhanced, thereby solving the misclassification problem caused by conflicting emotional features in the text. The challenge here is how to obtain students' cognitive ability vectors. Cognitive diagnosis, as a quantitative evaluation method combining psychometrics and pedagogy, mainly models students' historical practice records to discover their potential knowledge status and cognitive level, thus judging their mastery of course knowledge. Therefore, this paper uses cognitive diagnostic technology to extract students' cognitive ability vectors from their historical practice records and integrates them as influencing factors into the sentiment analysis model of teaching evaluation. This allows the model to not only consider the semantic information contained in the text but also be influenced by cognitive factors when extracting emotional features, ultimately improving the accuracy of the model's sentiment classification.

[0004] Research on sentiment analysis in teaching evaluation can be mainly divided into dictionary-based and machine learning-based approaches. Dictionary-based approaches rely on expert knowledge to formulate rules and dictionaries, primarily using NLP techniques to extract sentiment words from teaching evaluations for analysis. Wu Linjing et al. identified sentiment words from MOOC platform course reviews and matched them with a sentiment dictionary, using the difference between positive and negative sentiment words as the sentiment category of the text. Xue Yaofeng et al. performed fuzzy matching between teaching evaluation data and a sentiment lexicon to calculate learners' sentiment polarity. These sentiment classification methods rely on expert-constructed sentiment dictionaries; however, constructing such dictionaries is difficult, and the accuracy of sentiment analysis obtained using dictionary methods is relatively low. Currently, mainstream sentiment analysis is primarily based on machine learning, which embeds unstructured teaching evaluations into structured vectors through language models. These embedded vectors are then input into a machine learning model to capture the sentiment features needed for sentiment classification, ultimately classifying the sentiment of sentences. The literature [Gutiérrez G, Canul-Reich J, Zezzatti AO, et al. Mining: Students comments about teacher performance assessment using machine learning algorithms[J]. International Journal of Combinatorial Optimization Problems and Informatics, 2018, 9(3): 26-40.] explored how to improve the performance of sentiment classification in teaching evaluation using support vector machines and random forest algorithms on teaching evaluation data. The experimental results showed that the study achieved the best performance at the time, but due to the shallow text features extracted by the model, the generalization ability of the model was reduced and the language features learned by the model were not rich enough. The literature [Lin Q, Zhu Y, Zhang S, et al. Lexicalbased automated teaching evaluation via students's short reviews[J]. Computer Applications in Engineering Education, 2019, 27(1): 194-205.] conducted experimental comparisons on student teaching evaluation data using different machine learning models such as Naive Bayes and Logistic Regression, and achieved a more ideal sentiment classification effect. However, this research primarily focuses on sentiment analysis of short evaluation texts, and its ability to recognize sentiment in long texts is limited. Sentiment analysis tasks for teaching evaluations based on traditional machine learning models largely rely on sentiment dictionaries and feature engineering. Manually designing features and building sentiment dictionaries requires significant time investment and prior knowledge.Furthermore, traditional machine learning models have weak modeling capabilities and low accuracy in sentiment classification for complex teaching evaluations. With the development of deep learning technology, the literature [Tseng CW, Chou JJ, Tsai YC. Text mining analysis of teaching evaluation questions for the selection of outstanding teaching faculty members[J].IEEE Access, 2018, 6: 72870-72879.] considers the time-series characteristics of text data and designs an automatic classifier for identifying teaching evaluation sentiment based on an LSTM network, using the classification results as one of the bases for teacher performance evaluation.

[0005] Existing sentiment analysis models for teaching evaluation only consider the emotional features contained in the text. When there are conflicts in the emotional features in the text, simply extracting features from the semantic space of the text to identify the true emotions expressed in the teaching evaluation will reduce the accuracy of the identification and the model will have the problem of misclassification. Summary of the Invention

[0006] In view of this, the purpose of this invention is to provide a teaching evaluation sentiment analysis method that integrates cognitive transfer, thereby making up for the problem of misclassification in sentiment classification models caused by the presence of conflicting emotional characteristics in teaching evaluation, and improving the accuracy of sentiment classification in teaching evaluation.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A teaching evaluation affective analysis method that integrates cognitive transfer, comprising the following steps:

[0009] S1: Data collection stage: Use Python web crawling technology to obtain the required research data, then manually label the data, and clean and preprocess the collected data at the same time.

[0010] S2: Text feature extraction stage: Use the pre-trained language model BERT to embed words in the teaching evaluation, and input the embedding vector into the bidirectional GRU network for training to obtain the hidden memory information contained in the text;

[0011] S3: Prior knowledge extraction stage: Use a neurocognitive diagnostic model to model the student's historical practice records to obtain the student's cognitive ability vector, and use this vector as a representation of prior knowledge.

[0012] S4: Attention Interaction Stage: The attention mechanism is used to interact with the hidden memory information of the text obtained in S2 and the prior knowledge information of students extracted in S3, thereby enhancing the real emotional features in the semantics of the text.

[0013] S5: Sentiment Classification Stage: The sentiment features enhanced by S4 are input into the softmax function for calculation, and the probability distribution of the corresponding sentiment label is output.

[0014] Optionally, in S1, the sentiment analysis task collects teaching evaluation texts as research objects and collects students' historical practice records as prior knowledge to enhance text features.

[0015] Optionally, in S2, a pre-trained language model is used to perform word embedding on the unstructured text, and a structured high-dimensional vector v is used to represent the semantic information of the input text, where v i The embedding representation of each word, i = 1, 2, ..., l, where l represents the length of the evaluation text:

[0016] v = {v1, v2, ..., v} l}

[0017] Then, a bidirectional GRU memory network is used to model the contextual dependencies of the text. The main function of the bidirectional GRU network is to iterate the structured evaluation text vectors output by the text embedding module to form a text memory matrix carrying deep sentiment features.

[0018]

[0019]

[0020]

[0021]

[0022] Where, r t This represents the reset gate, used to control how much of the previous word's hidden state memory should be retained. z t The gate represents the update mechanism, controlling how much information needs to be forgotten from the previous word and how much hidden state information of the current word needs to be added. v t The word embedding vector at position t; σ, tanh represent two types of activation functions.

[0023] Optionally, in step S3, the student's cognitive vector f is extracted by training a neural network. s The value of each element in the vector This represents the students' level of mastery of the course content.

[0024]

[0025] Let the factor matrix of the test items be represented as p. e The difficulty of a knowledge point is represented by k. h The discrimination index of the test items is represented by d. e The corresponding calculation formula is:

[0026] p e =o e ×Q

[0027] f s =σ(o s ×W1)

[0028] k h =σ(o e ×W2)

[0029] d e =σ(o e ×W3)

[0030] in, K represents the total number of knowledge points covered in the test questions; E represents the total number of test questions; n represents the total number of students; o e The one-hot encoding of the test question, o s One-hot encoding representing students, This represents the question-knowledge point association matrix, where each element indicates whether a question contains a specific knowledge point. The representative parameter matrix; the interaction function of the first layer is derived from the multidimensional project response theory model:

[0031]

[0032] The intermediate output result h is iteratively calculated through the fully connected layer:

[0033] h2=σ(W4×h T +b1)

[0034] h3=σ(W5×h2+b2)

[0035] y = σ(W6×h3+b3)

[0036] Where σ represents the activation function, W4, W5, W6 represent parameter matrices, and b4, b5, b6 represent bias; after training of this module, f s The value of f is now updated. s This refers to the required cognitive vector, which represents the prior information about the student's cognition.

[0037] Because the number of knowledge points contained in students' historical practice records is limited, the dimension of the cognitive ability vector extracted from them cannot match the dimension of the text memory matrix, i.e., f s with h t The dimensions are not consistent, so direct computation is not possible; this requires integrating textual hidden memory and cognitive ability vector f. s Before, for f s The cognitive ability matrix h is obtained by performing a nonlinear transformation. f :

[0038]

[0039] in, and b f These represent the parameter matrix and the bias, respectively.

[0040] Optionally, in S4, the text memory matrix h is implemented. t and cognitive ability matrix h f Attention interaction fusion is used to output a feature fusion-level CTM, and the CTM is used to enhance the two features respectively; firstly, h t with h f The input is fed into a neural network for fully connected computation, enabling attentional interaction between teaching evaluation features and cognitive ability features. Then, it is activated by an activation function, thereby constructing a CTM to represent h. t and h f The results of the interaction between them:

[0041] CTM=σ(h t *W CTM *(h f ) T +b fm )

[0042] in, CTM, in conjunction with attention mechanisms, respectively targets h t with h f The attention calculation process for feature enhancement is described as follows:

[0043]

[0044]

[0045] α ij h represents an element at a specific position in the CTM, and is calculated by accumulating the values ​​of each row and each column in the CTM. t and h f Attention score α ij ; It reflects h tThe importance score of the i-th row, It reflects h f The importance score in column j; the attention score assigns different weights to each element in the matrix, making it easier for the model to capture important sentiment features; in h t and h f After assigning attention weights, the enhanced memory representation x is then calculated separately. u and enhanced cognitive representation x v :

[0046]

[0047]

[0048] Optionally, in S5, the two representations enhanced by CTM are used as emotional features, and an emotional classifier is used to classify the reconstructed emotional features; the enhanced memory representation x u and enhanced cognitive representation x v The concatenated data is used as input to the fully connected layer for global sentiment features, and the sentiment category y is calculated using the softmax function. i :

[0049]

[0050]

[0051] Where z represents the number of sentiment tags;

[0052] The training objective of the cognitive extraction module is to minimize the output y. i With real label y i Cross-entropy loss between 'loss1:

[0053]

[0054] Where i represents the index of the sample data; the training objective of the sentiment classification part is also to minimize the cross-entropy loss2 of the classification:

[0055]

[0056] Where λ represents the weight parameters of the L2 regularization penalty term, and θ represents the parameter set; y c Represents the true label, f c (x; θ) represent the labels predicted by the model;

[0057] To prevent overfitting, early stopping and exit rates are set during training; the overall loss function of the model is expressed as follows:

[0058]

[0059] The beneficial effects of this invention are as follows:

[0060] (1) In the teaching evaluation emotion analysis task, in addition to extracting the inherent semantic features in the text, this invention also extracts personalized cognitive ability vectors from students' historical practice records as prior knowledge, and enhances the real emotional features contained in the text through the attention mechanism.

[0061] (2) This invention uses a pre-trained BERT language model for word embedding. Compared with traditional word embedding methods, BERT uses a bidirectional Transformer model for pre-training. The self-attention mechanism in the Transformer structure can better capture language features in teaching evaluation based on embedding.

[0062] (3) This invention uses a bidirectional GRU network instead of a bidirectional LSTM network. Compared with LSTM, GRU has one less gate unit in its internal structure, but can achieve the same functionality as LSTM. In addition, GRU has fewer training parameters and the model fits faster.

[0063] (4) The present invention uses enhanced emotional features for emotional classification tasks, which can effectively solve the problem of misclassification of emotional classification models due to conflicting emotional features in teaching evaluation.

[0064] Other advantages, objectives, and features of the invention will be set forth in part in the description which follows, and in part will be apparent to those skilled in the art from the following examination, or may be learned from practice of the invention. The objectives and other advantages of the invention can be realized and obtained through the following description. Attached Figure Description

[0065] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0066] Figure 1 This is a schematic diagram of the invention;

[0067] Figure 2 This is an attention heatmap of a certain teaching evaluation text in the embodiment. Detailed Implementation

[0068] The following specific examples illustrate the implementation of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic representations of the basic concept of the present invention. Unless otherwise specified, the following embodiments and features can be combined with each other.

[0069] The accompanying drawings are for illustrative purposes only and are schematic diagrams, not actual pictures, and should not be construed as limiting the invention. To better illustrate the embodiments of the invention, some parts in the drawings may be omitted, enlarged, or reduced, and do not represent the actual product dimensions. It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0070] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components. In the description of the present invention, it should be understood that if terms such as "upper," "lower," "left," "right," "front," and "rear" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, they are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, the terms used to describe positional relationships in the drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those skilled in the art, the specific meaning of the above terms can be understood according to the specific circumstances.

[0071] Please see Figure 1 This is a sentiment analysis method for teaching evaluation that integrates cognitive transfer. This method can be applied to scenarios such as product review analysis (understanding user satisfaction with products and developing marketing strategies), film review analysis (understanding audience approval of films and identifying their shortcomings), and educational evaluation (understanding teacher teaching effectiveness and students' learning outcomes and feelings). This method is primarily used in teaching evaluation scenarios to conduct sentiment analysis research on student evaluation data. The technical problem it addresses is the issue of misclassification of sentiments in teaching evaluation due to conflicting emotional characteristics.

[0072] Step 1: Use web scraping technology to obtain the required dataset, manually label the data, and then perform data cleaning and preprocessing. Sentiment analysis tasks not only collect evaluation texts as research objects but also collect students' historical practice records as prior knowledge to enhance text features.

[0073] Step Two: Text Feature Extraction Stage: The BERT language model is used to embed words into the teaching evaluation, and the embedded vectors are input into a bidirectional GRU network for training to obtain the hidden memory information contained in the text. A pre-trained language model is used to embed words into unstructured text, using a structured high-dimensional vector v to represent the semantic information of the input text, where v... i (i = 1, 2, ..., l) represents the embedding representation of each word, and l represents the length of the evaluation text:

[0074] v = {v1, v2, ..., v} l}

[0075] Then, a bidirectional GRU memory network is used to model the contextual dependencies of the text. The main function of the bidirectional GRU network is to iterate the structured evaluation text vectors output by the text embedding module to form a text memory matrix carrying deep sentiment features.

[0076]

[0077]

[0078]

[0079]

[0080] Where, r t This represents the reset gate, used to control how much of the previous word's hidden state memory should be retained. z t The gate represents the update mechanism, controlling how much information needs to be forgotten from the previous word and how much hidden state information of the current word needs to be added. v t The word embedding vector at position t. σ and tanh represent two types of activation functions.

[0081] Step 3: Model students' historical practice records using a neurocognitive diagnostic model to obtain their cognitive ability vectors, which will then be used as representations of prior knowledge. The cognitive vector f of each student will be extracted by training a neural network. s The value of each element in the vector It represents the student's level of mastery of the course knowledge points.

[0082]

[0083] Let the factor matrix of the test items be represented as p. e The difficulty of a knowledge point is represented by k. h The discrimination index of the test items is represented by d. e The corresponding calculation formula is:

[0084] pe =o e ×Q

[0085] f s =σ(o s ×W1)

[0086] k h =σ(o e ×W2)

[0087] d e =σ(o e ×W3)

[0088] in, K represents the total number of knowledge points covered in the test questions, E represents the total number of test questions, and n represents the total number of students. e The one-hot encoding of the test question, o s One-hot encoding representing students, This represents the question-knowledge point association matrix, where each element indicates whether a question contains a specific knowledge point. The representative parameter matrix. The interaction function of the first layer is derived from the multidimensional project response theory model:

[0089]

[0090] The intermediate output result h is iteratively calculated through the fully connected layer:

[0091] h2=σ(W4×h T +b1)

[0092] h3=σ(W5×h2+b2)

[0093] y = σ(W6×h3+b3)

[0094] Where σ represents the activation function, W4, W5, W6 represent the parameter matrices, and b4, b5, b6 represent the bias. After training this module, f s The value of f is now updated. s This refers to the required cognitive vector, which represents the prior information about the student's cognition.

[0095] Because the number of knowledge points contained in students' historical practice records is limited, the dimension of the cognitive ability vector extracted from them cannot match the dimension of the text memory matrix, i.e., f s with h t The dimensions are not consistent, so direct computation is not possible. Therefore, when fusing textual hidden memory and cognitive ability vector f... s Before that, we need to first deal with f. s The cognitive ability matrix h is obtained by performing a nonlinear transformation. f :

[0096]

[0097] in, and b f These represent the parameter matrix and the bias, respectively.

[0098] Step 4: Attention Interaction Stage: Utilizing an attention mechanism, the hidden memory information of the text acquired in Stage S2 is interacted with the prior knowledge information of the students extracted in Step S3 to enhance the authentic emotional features in the text semantics. This achieves the text memory matrix h. t and cognitive ability matrix h f Attention interaction fusion is used to output a feature fusion-level CTM, and the CTM is then used to enhance the two features separately. First, h t with h f The input is fed into a neural network for fully connected computation, enabling attentional interaction between teaching evaluation features and cognitive ability features. Then, it is activated by an activation function, thereby constructing a CTM to represent h. t and h f The results of the interaction between them:

[0099] CTM=σ(h t *W CTM *(h f ) T +b fm )

[0100] in, CTM, in conjunction with attention mechanisms, respectively targets h t with h f The attention calculation process for feature enhancement is described as follows:

[0101]

[0102]

[0103] α ij h represents an element at a specific position in the CTM, and is calculated by accumulating the values ​​of each row and each column in the CTM. t and h f Attention score α ij . It reflects h t The importance score of the i-th row, It reflects h f The importance score in column j. Attention scores can assign different weights to each element in the matrix, making it easier for the model to capture important sentiment features. In the case of h... t and h fAfter assigning attention weights, the enhanced memory representation x is then calculated separately. u and enhanced cognitive representation x v :

[0104]

[0105]

[0106] Step 5: Sentiment Classification Stage: The sentiment features enhanced in step S4 are input into the softmax function for calculation, outputting the probability distribution of the corresponding sentiment labels. The two representations enhanced by CTM are used as sentiment features, and a sentiment classifier is used to classify the reconstructed sentiment features. Enhanced memory representation x u and enhanced cognitive representation x v The concatenated data is used as input to the fully connected layer for global sentiment features, and the sentiment category y is calculated using the softmax function. i :

[0107]

[0108]

[0109] Where z represents the number of sentiment tags.

[0110] like Figure 2 As shown, the method of this invention is applied to a teaching evaluation text. The numerical value in each square represents the model's attention score for words in the text. The higher the score, the higher the semantic attention the model pays to that word. The first row of the heatmap represents the case without cognitive ability. Due to the noise brought to the model by conflicting emotions, the attention scores for "offer" and "many" are high, while the attention scores for "but," "perfect," and "fine" are low. The model's judgment of the text's sentiment category is interfered with, leading to misclassification. The second row of the heatmap represents the case with cognitive ability. After incorporating cognitive ability, the attention score for "many" decreased by 0.03, the attention for "but" increased by 0.04, the attention for "perfect" and "fine" increased by 0.03, and the attention for "study" also decreased. The main reason for this is that cognitive ability enhances the global true sentiment features implied in the text.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A teaching evaluation affective analysis method integrating cognitive transfer, characterized in that: The method includes the following steps: S1: Data collection stage: Use Python web crawling technology to obtain the required research data, then manually label the data, and clean and preprocess the collected data at the same time. S2: Text feature extraction stage: Use the pre-trained language model BERT to embed words in the teaching evaluation, and input the embedding vector into the bidirectional GRU network for training to obtain the hidden memory information contained in the text; S3: Prior knowledge extraction stage: Use a neurocognitive diagnostic model to model the student's historical practice records to obtain the student's cognitive ability vector, and use this vector as a representation of prior knowledge. S4: Attention Interaction Stage: The attention mechanism is used to interact with the hidden memory information of the text obtained in S2 and the prior knowledge information of students extracted in S3, thereby enhancing the real emotional features in the semantics of the text. S5: Sentiment Classification Stage: The sentiment features enhanced by S4 are input into the softmax function for calculation, and the probability distribution of the corresponding sentiment label is output.

2. The teaching evaluation affective analysis method integrating cognitive transfer according to claim 1, characterized in that: In S1, the sentiment analysis task collects teaching evaluation texts as the research object and collects students' historical practice records as prior knowledge to enhance text features.

3. The teaching evaluation emotion analysis method integrating cognitive transfer according to claim 2, characterized in that: In step S2, a pre-trained language model is used to embed words into the unstructured text, and a structured high-dimensional vector v is used to represent the semantic information of the input text, where v i The embedding representation of each word, i = 1, 2, ..., l, where l represents the length of the evaluation text: v={v1,v2,...,v l } Then, a bidirectional GRU memory network is used to model the contextual dependencies of the text. The main function of the bidirectional GRU network is to iterate the structured evaluation text vectors output by the text embedding module to form a text memory matrix carrying deep sentiment features. Where, r t This represents the reset gate, used to control how much of the previous word's hidden state memory should be retained. z t The gate represents the update mechanism, controlling how much information needs to be forgotten from the previous word and how much hidden state information of the current word needs to be added. v t The word embedding vector at position t; σ, tanh represent two types of activation functions.

4. The teaching evaluation emotion analysis method integrating cognitive transfer according to claim 3, characterized in that: In step S3, the student's cognitive vector f is extracted by training a neural network. s The value of each element in the vector This represents the students' level of mastery of the course content. Let the factor matrix of the test items be represented as p. e The difficulty of a knowledge point is represented by k. h The discrimination index of the test items is represented by d. e The corresponding calculation formula is: p e =o e ×Q f s =σ(o s ×W1) k h =σ(o e ×W2) d e =σ(o e ×W3) in, K represents the total number of knowledge points covered in the test questions; E represents the total number of test questions; n represents the total number of students; o e The one-hot encoding of the test question, o s One-hot encoding representing students, This represents the question-knowledge point association matrix, where each element indicates whether a question contains a specific knowledge point. The representative parameter matrix; the interaction function of the first layer is derived from the multidimensional project response theory model: The intermediate output result h is iteratively calculated through the fully connected layer: h3=σ(W5×h2+b2) y = σ(W6×h3+b3) Where σ represents the activation function, W4, W5, W6 represent parameter matrices, and b4, b5, b6 represent bias; after training of this module, f s The value of f is now updated. s This refers to the required cognitive vector, which represents the prior information about the student's cognition. Because the number of knowledge points contained in students' historical practice records is limited, the dimension of the cognitive ability vector extracted from them cannot match the dimension of the text memory matrix, i.e., f s with h t The dimensions are not consistent, so direct computation is not possible; this requires integrating textual hidden memory and cognitive ability vector f. s Before, for f s The cognitive ability matrix h is obtained by performing a nonlinear transformation. f : in, and b f These represent the parameter matrix and the bias, respectively.

5. The teaching evaluation emotion analysis method integrating cognitive transfer according to claim 4, characterized in that: In S4, the text memory matrix h is implemented. t and cognitive ability matrix h f Attention interaction fusion is used to output a feature fusion-level CTM, and the CTM is used to enhance the two features respectively; firstly, h t with h f The input is fed into a neural network for fully connected computation, enabling attentional interaction between teaching evaluation features and cognitive ability features. Then, it is activated by an activation function, thereby constructing a CTM to represent h. t and h f The results of the interaction between them: in, CTM, in conjunction with attention mechanisms, respectively targets h t with h f The attention calculation process for feature enhancement is described as follows: α ij h represents an element at a specific position in the CTM, and is calculated by accumulating the values ​​of each row and each column in the CTM. t and h f Attention score α ij ; It reflects h t The importance score of the i-th row, It reflects h f The importance score in column j; the attention score assigns different weights to each element in the matrix, making it easier for the model to capture important sentiment features; in h t and h f After assigning attention weights, the enhanced memory representation x is then calculated separately. u and enhanced cognitive representation x v :

6. The teaching evaluation emotion analysis method integrating cognitive transfer according to claim 5, characterized in that: In step S5, the two representations enhanced by CTM are used as emotional features, and an emotional classifier is used to classify the reconstructed emotional features; the enhanced memory representation x u and enhanced cognitive representation x v The concatenated data is used as input to the fully connected layer for global sentiment features, and the sentiment category y is calculated using the softmax function. i : Where z represents the number of sentiment tags; The training objective of the cognitive extraction module is to minimize the output y. i With real label y′ i Cross-entropy loss 1: Where i represents the index of the sample data; the training objective of the sentiment classification part is also to minimize the cross-entropy loss2 of the classification: Where λ represents the weight parameters of the L2 regularization penalty term, and θ represents the parameter set; y c Represents the true label, f c (x; θ) represent the labels predicted by the model; To prevent overfitting, early stopping and exit rates are set during training; the overall loss function of the model is expressed as follows:

Images