1. Introduction

With the rapid development of information technology, users express and transmit their emotions through internet platforms. This has generated a lot of valuable comment information on the internet, which expresses people's emotional colors and tendencies in different aspects. Therefore, researching text sentiment analysis has gradually become a new trend. Emotional analysis is an important branch of natural language processing tasks. It mainly refers to the use of data mining algorithms to automatically analyze and determine emotional tendencies in texts with emotional attitudes ^[1,2]. However, in solving aspect level sentiment analysis, domestic and foreign scholars often use attention mechanisms combined with neural networks to construct sentiment analysis models ^[3,4]. These sentiment analysis models overly focus on analysis efficiency, while ignoring the inherent connections between different aspects of the sentence, and global or local attention calculations cannot fully mine key hidden information in the text. Therefore, the study proposes to improve the attention mechanism. It combines global attention and local attention and proposes an sentiment analysis algorithm that integrates multiple attention mechanisms for university course evaluation methods. The combination of sentiment analysis technology and teaching evaluation is conducive to promoting teachers to timely discover students' true feelings and experiences about the curriculum, thereby improving teaching quality. Therefore, on the basis of improving the attention mechanism, the study further integrates the multi-attention mechanism with neural network, and constructs a college student sentiment analysis model that integrates the multi-attention mechanism and neural network. At the same time, the study innovatively utilizes the gating unit to dynamically adjust the weights of the attention mechanism of the model, and introduces the affective resource attention module to enhance the weakened semantic information, aiming to solve the one-sidedness problem of most sentiment analysis methods. The overall structure of the study consists of four parts: The first part summarizes the research achievements and shortcomings of sentiment analysis techniques at home and abroad. In the second part, a student sentiment analysis model integrating multiple attention mechanisms and neural network algorithms was studied and designed. The third part conducted experiments and analysis on the proposed student sentiment analysis model. The fourth part summarizes the experimental results and points out future research directions.

2. Related Works

Emotional analysis technology refers to analyzing or processing text to uncover the semantic information, emotions, or attitudes it contains. The implementation of this process is called sentiment analysis, and domestic and foreign scholars have conducted relevant research on sentiment analysis from different granularity of text processing ^[5,6]. Basiri et al. proposed a bidirectional neural network and convolutional network deep model on the ground of attention mechanism to address the inefficiency of sentiment analysis models in feature extraction. This achieves effective classification and efficient extraction of document information features by the model, and predicts emotional polarity of document semantic information ^[7]. Ray et al. proposed using a seven layer deep convolutional neural network (CNN) to construct an sentiment analysis model on the ground of deep learning algorithms to further study the application value of deep learning algorithms in the field of text mining. This achieved sentiment polarity prediction with an accuracy of 0.87 in aspect level text classification ^[8]. Powerful multimodal sentiment analysis models are typically implemented by deep learning algorithms. Wang et al. proposed an interactive visual analysis system for interpreting and visualizing multimodal models for sentiment analysis. This achieves a multifaceted exploration of model behavior from the perspectives of language, acoustics, and visual modalities ^[9]. Liu et al. addressed the issue of limited recognition of Chinese grammar and semantics in sentiment analysis models and introduced a new personal emotion source to construct an improved expression embedding model on the ground of Bi LSTM. The analysis accuracy of this pair of online Chinese texts has reached 95% ^[10]. Zhang et al. proposed a multi task transformer network on the ground of feature and fine-tuning to achieve simultaneous processing of different emotion analysis tasks. This achieved an F1 score of 77.8% in Twitter's sentiment analysis dataset and improved the importance of overlooked design decisions in searching for contextual features in depth and luminosity spaces ^[11].

With the rapid development of artificial intelligence, neural networks have gradually become an important means to solve various complex problems. As an important module in neural networks, attention mechanism has been widely applied in fields such as natural language processing and computer vision. Lin et al. proposed a spatiotemporal attention long short-term memory model to improve the interpretability of vehicle trajectory prediction models. This reveals the influence of historical trajectories and adjacent vehicles on the target vehicle, and demonstrates the model's capture and interpretation of fine-grained lane changing behavior ^[12]. Zhu et al. proposed a lightweight single image super-resolution grid with expected maximum attention mechanism to apply CNNs based super-resolution methods in low computing power devices. This effectively reduces the number of parameters in quantitative measurement and visual quality ^[13]. Alirezazadeh et al. proposed using convolutional attention modules to improve the classification of CNNs in response to the problem of limited data in plant disease recognition model datasets. This effectively improves the recognition accuracy of pre trained neural networks with limited training data ^[14]. Lee et al. designed an analysis model on the ground of long-term memory and attention mechanisms to analyze the reasons for low production in semiconductor manufacturing processes. This achieves effective prediction of low returns in real production data of semiconductor companies ^[15]. Islam et al. proposed a deep transfer learning model on the ground of convolutional block attention module to improve the accuracy of breast lesion classification. This confirms that the optimization ability of attention mechanisms in specific application tasks can further improve the performance of the model ^[16].

On the ground of the above, domestic and foreign researchers have studied or optimized sentiment analysis models from different perspectives. However, the granularity of text classification is relatively large and the key information hidden in the text is not fully explored. In addition, the application of sentiment analysis models is mostly focused on social or business, and there is relatively little research in the field of education. Therefore, the study proposes to integrate multiple attention mechanisms and neural networks to construct an aspect level student sentiment analysis model for the evaluation of university course teaching methods. And it innovatively utilizes gating units to dynamically adjust the attention mechanism weights of the model, introducing an emotional resource attention module to enhance the weakened semantic information. By studying the application of sentiment analysis technology in the field of education, it is expected to promote the improvement of teaching quality and the efficient transformation of the education system.

3. Construction of A Student Emotion Analysis Model Integrating Attention Mechanism Algorithm and Neural Network Algorithm

It is necessary to identify the emotional characteristics corresponding to each aspect of the sentence in the text. Therefore, this study proposes to further integrate neural networks to construct an emotion analysis model with multiple attention mechanisms on the ground of the construction of a multi attention mechanism emotion analysis model. By dynamically adjusting the weights of global and local attention fusion processes through gate control units, emotional resource attention enhances semantic information weakened by long distances. Then it uses Softmax to output the final sentiment analysis results, thereby achieving the calculation of the intrinsic connections between various aspects of the sentence. This can obtain more accurate emotional analysis results.

3.1 A Sentiment Analysis Model Integrating Multiple Attention Mechanisms

To identify key information hidden in text sentences, this study proposes a combination of global attention mechanism and local attention mechanism. Then it establishes a Multi Attention Fusion Model (MAFM) that integrates multiple attention mechanisms for sentiment analysis. By fully leveraging the advantages of both, the sentiment analysis model can more accurately identify the key information contained in the sentence. In addition, unlike other studies that use cascading methods, the study employs gating mechanisms for the fusion of multiple attention mechanisms. The gating unit can dynamically adjust the weight of global and local attention in the fusion, thereby avoiding excessive reliance on global or local attention and causing significant bias in prediction results. Therefore, the proposed sentiment analysis model that integrates multiple attention mechanisms is shown in Fig. 1.

The MAFM model mainly includes sentence representation in embedding, semantic information mining, attention module, gating unit, and Softmax layer.

Fig. 1. Network architecture of MAFM.

3.1.1 Embedded Aspect Sentence Representation Module

The main function of the embedded aspect sentence representation module is to connect each word in the sentence with a specific aspect. This can make each sentence focus more on a specific aspect each time. The specific definition formula for this module is shown in Eq. (1).

(1)

$ D_{a}=\left\{\left[w_{1};a\right]\left[w_{2};a\right]\ldots \left[w_{n};a\right]\right\} $

In Eq. (1), $D$ represents a sentence. $a$ represents a specific aspect. $w$ represents a word. $n$ represents word order. $\left[w_{1};a\right]$ represents the concatenation of words with specific aspects.

3.1.2 Semantic Information Mining Module

The main function of the semantic information mining module is to feed word vectors into the bidirectional Long Short Term Memory (LSTM) network in the hidden layer to encode contextual information. For LSTM with forward propagation, the formula for defining the hidden state at a certain moment is shown in Eq. (2).

(2)

$ \overset{\rightarrow }{b}_{t}=\overset{\rightarrow }{LSTM}\left(D_{a}\overset{\rightarrow }{b}_{t-1}\right) $

In Eq. (2), $\overset{\rightarrow }{b}_{t-1}$ represents the given hidden state. $t$ represents time. $\overset{\rightarrow }{b}_{t}$ represents the hidden state when the time is $t$. $\overset{\rightarrow }{LSTM}$ represents forward LSTM. The MAFM model concatenates the forward and backward LSTM results, and the concatenated results are processed by the tanh activation function to generate the final hidden state. The specific formula is shown in Eq. (3).

(3)

$ b_{t}=\tanh \left(\left[\overset{\rightarrow }{b}_{t};\overset{\leftarrow }{b}_{t}\right]\right) $

In Eq. (3), $\overset{\leftarrow }{b}_{t}$ represents the output result of the reverse LSTM at time $t$.

3.1.3 Attention Module

The attention module is mainly responsible for learning semantic relationships in specific aspects of sentences using information from global and local views. The expression formula for global attention is shown in Eq. (4).

(4)

$\left\{\begin{array}{l} c_{i}=W_{att2}^{T}\tanh (W_{att1}\left[b_{i};a\right]+d_{att1})\\ \alpha _{i}=\frac{\exp (c_{i})}{\sum _{j=1}^{N}\exp (c_{j})} \end{array}\right.$

In Eq. (4), $c_{i}$ represents the global attention of each word. $b_{i}$ represents a word in LSTM. $\alpha _{i}$ represents attention score. $W_{att2}^{T}$ and $W_{att1}$ represent weight matrices, respectively, mainly obtained by Softmax calculation. $d_{att1}$ represents the offset. According to the attention calculation formula represented by each word, the context representation of the global attention mechanism is shown in Eq. (5).

(5)

$ f_{glo}=\sum _{i=1}^{N}b_{i}a $

In Eq. (5), $f_{glo}$ represents the contextual representation of global attention. Due to the excessive focus of local attention on the subset of words in the context, this study introduces the word distance defined by dependency trees, which is the distance on the ground of grammar, to achieve the target word capture of local attention. The comparison between distance representation on the ground of grammar and distance representation on the ground of position is shown in Fig. 2.

Fig. 2. Comparison of grammar based distance and position based distance.

In Fig. 2, the distance on the ground of syntax is defined as a dependency tree for a given sentence, and each word belongs to a node in the dependency tree. The distance between two connected nodes is defined as 1, so the distance calculation between all remaining words and the current target word can be achieved by traversing the dependency tree. Then it determines local attention words on the ground of the distance on the ground of grammar. The formula for calculating the attention to be allocated to words within the grammatical distance of the target word is shown in Eq. (6).

(6)

$ \left\{\begin{array}{l} g_{i}=W_{att4}^{T}\tanh \left(W_{att3}\left[b_{i};a\right]+d_{att2}\right)\\ \beta _{i}=\frac{\exp \left(g_{i}\right)}{\sum _{j=LS\left(t\right)}\exp \left(g_{j}\right)} \end{array}\right. $

In Eq. (6), $g_{i}$ represents the local attention of each word. $W_{att4}^{T}$ and $W_{att3}$ represent weight matrices, respectively, mainly obtained by Softmax calculation.. $d_{att2}$ represents the offset. $\beta _{i}$ represents attention score. $LS(t)$ represents the words within grammatical distance of the target word. When a target contains multiple words, the words within the synchronization distance of each word in a specific target should be selected. Therefore, the local attention score calculation formula and context representation formula specified for each context are shown in Eq. (7).

(7)

$ \left\{\begin{array}{l} \beta =\left\{\begin{array}{l} 0,i\notin LS\left(t\right)\\ \frac{\exp \left(g_{i}\right)}{\sum _{j=LS\left(t\right)}\exp \left(g_{j}\right)},i\in LS\left(t\right) \end{array}\right.\\ f_{loc}=\sum _{i=1}^{N}b_{i}\beta _{i} \end{array}\right. $

In Eq. (7), $\beta $ represents the specified local attention score. $f_{loc}$ represents the contextual representation of local attention.

3.1.4 Gate Control Unit Module

After obtaining global and local attention context representations, the gate control unit module is responsible for synthesizing the two attention results into specific target information. The way in which the gate control unit integrates global and local attention is shown in Eq. (8).

(8)

$ h=sigmoid\left(W_{gate}\left(f_{glo}+f_{loc}\right)\right) $

In Eq. (8), $h$ represents the gate control unit. $W_{gate}$ represents the parameters, mainly obtained by training. The different dimensions of words reflect the different meanings of words, so the study sets gate units where each dimension controls the various perspectives of the attention vector. The context representation formula obtained from this is shown in Eq. (9).

(9)

$ \left\{\begin{array}{l} f_{final}^{loc}=\sum _{i=1}^{N}b_{i}\odot \lambda _{i}^{loc}\\ f_{final}^{glo}=\sum _{i=1}^{N}b_{i}\odot \lambda _{i}^{glo}\\ f_{final}=f_{final}^{loc}+f_{final}^{glo} \end{array}\right. $

In Eq. (9), $f_{final}^{loc}$ represents the final local attention context representation. $f_{final}^{glo}$ represents the final global attention context representation. $f_{final}$ represents the final contextual representation of multi attention fusion. $\lambda _{i}^{loc}$ represents the total score of local attention in a certain dimension. $\lambda _{i}^{glo}$ represents the total global attention score for a certain dimension. The final representation of a specific target aspect output by the gate control unit module is output by the Softmax classifier, and the specific output formula is shown in Eq. (10).

(10)

$ \hat{y}=soft\max \left(W_{s}f_{final}+d_{s}\right) $

In Eq. (10), $\hat{y}$ represents the polarity prediction result of text sentiment analysis. $W_{s}$ represents the parameters, mainly obtained by training. $d_{s}$ represents the training parameters of the Softmax layer. On the ground of the above, the specific process of the proposed MAFM model is shown in Fig. 3.

Firstly, the MAFM model embeds each word in the text with a specific target aspect, generating sentences with embedded aspects for representation, prompting sentences to focus only on a specific aspect. Secondly, LSTM networks and attention mechanisms are used to calculate global and local attention separately, and the fusion of the two types of attention is carried out through gating units. Finally, it outputs the final emotional polarity of the sentence through the Softmax layer.

Fig. 3. Execution process of MAFM.

3.2 A Sentiment Analysis Model Integrating Multiple Attention Mechanisms and Neural Networks

To fully consider the connections between multiple aspects of a sentence, this study combines neural network algorithms on the basis of MAFM to construct a Multi Aspect Sentiment Attention Fusion Modeling (MASAFM) that integrates multiple attention mechanisms and neural networks. It obtains more accurate sentiment analysis results by calculating the relationships between various aspects of the sentence. MASAFM utilizes gated loop units, MAFM attention calculation method, and LSTM to mine implicit relationships in various aspects of sentences. Then it adds emotional resources to enhance attention and weaken semantic information due to long-distance propagation. The MASAFM model architecture is shown in Fig. 4.

As can be seen from Fig. 4, the MASAFM model mainly consists of the Sentence Representation Module with embedded aspects, the Inter-Aspect Dependency Module, the Sentiment Resource Attention Module and the Softmax Layer. Among them, GUR denotes Gate Recurrent Unit, which is utilized in the study to solve the neural network gradient problem.

Fig. 4. Network architecture of MASAFM.

3.2.1 Embedded Sentence Representation Module

The sentence representation module for embedding aspects in the MASAFM model includes aspect embedding operations and attention fusion operations. The specific network architecture is shown in Fig. 5.

Fig. 5. Embedded sentence representation module network architecture.

The aspect embedding operation is responsible for concatenating each word in a sentence with a specific target aspect, and its definition formula is shown in Eq. (11).

(11)

$ S_{a}=\left\{\left[w_{1};a\right]\left[w_{2};a\right]\ldots \left[w_{n};a\right]\right\} $

In Eq. (11), $S_{a}$ represents the sentence representation of the embedding aspect. The attention fusion operation is responsible for sending the embedded aspect of sentence representation into a bidirectional gated loop unit, to propagate contextual information between sentences. The definition formula of the bidirectional gate control loop unit is shown in Eq. (12).

(12)

$ \left\{\begin{array}{l} k=\sigma \left(x_{t}U^{k}+s_{t-1}W^{k}\right)\\ \lambda =\sigma \left(x_{t}U^{\lambda }+s_{t-1}W^{\lambda }\right)\\ b_{t}=\tanh \left(x_{t}U^{b}+\left(s_{t-1}\ast \lambda \right)W^{b}\right)\\ s_{t}=\left(1-k\right)\ast b_{t}+k\ast s_{t-1} \end{array}\right. $

In Eq. (12), $U$ represents a bidirectional gated loop unit. $s_{t}$ represents the unit state at time $t$. $b_{t}$ represents the hidden output at time $t$. $k$ represents the input of LSTM. $\sigma $ represents the weight parameter. $W$ represents the weight matrix. $\lambda $ represents sum of attention scores. $b$ represents the hidden output. To enhance the connection between specific aspects and corresponding emotional words, a attention layer was added to the MASAFM model. Therefore, the sentence context representation formula for embedding is shown in Eq. (13).

(13)

$ f_{a}=\lambda ^{T}B $

In Eq. (13), $B$ represents the hidden state matrix of the output.

3.2.2 Dependencies Between Aspects Module

The dependency module between aspects is responsible for transmitting the set of sentence context representations embedded in the aspects to the gated loop unit hidden layer. Then it uses LSTM for modeling, converting the representation of the target aspect in the network into internal query states by using fully connected layers. The expression formula for internal query status is shown in Eq. (14).

(14)

$ p=\tanh \left(f_{{a_{t}}}W_{T}+d_{T}\right) $

In Eq. (14), $p$ represents the representation of a specific target. $f_{{a_{t}}}$ represents the representation of the target aspect in the network. $W_{T}$ represents the weight matrix. $d_{T}$ represents the offset.

3.2.3 Emotional Resource Attention Module

The emotional resource attention module increases the emotional resources corresponding to sentences by adding an attention layer in LSTM. Therefore, the input and output expressions of the memory network are shown in Eq. (15).

(15)

$ \left\{\begin{array}{l} k=pP^{T}\\ q=\beta ^{T}P' \end{array}\right. $

In Eq. (15), $P$ represents the information obtained in the gated loop unit. $P'$ represents the implicit information obtained through multiple gated loop units. The final sentiment classification result is output through the Softmax module, as shown in Eq. (16).

(16)

$ \left\{\begin{array}{l} \rho =soft\max \left(\left(p+q\right)W_{s\max }+d_{s\max }\right)\\ \hat{y}=argmax_{i}\left(\rho _{i}\right) \end{array}\right. $

In Eq. (16), $\rho $ represents the new target aspect representation obtained by adding the target aspect representation and LSTM output. On the ground of the above, the specific process of the MASAFM model is shown in Fig. 6.

Fig. 6. Flow chart of MASAFM model.

Firstly, on the ground of the input user comment sample, a gated loop unit and MAFM are used to generate embedded aspect sentence representations, improving the sentence's attention to specific aspects. Secondly, it utilizes LSTM to repeatedly match the representation of the target aspect with other aspects. Before outputting LSTM, it adds corresponding emotional resources to the representation of the target aspect, and then queries emotional resources and sentence hidden information through the gated loop unit. Finally, all semantic information obtained will be processed through the Softmax layer for final emotional polarity output.

4. Verification of A Student Emotion Analysis Model Integrating Attention

To verify the execution of the proposed model, an overview of the sample dataset and parameter settings was first provided, followed by an exploration of the performance comparison between the MAFM model and other models. Finally, the MASAFM was compared with other different models in a comparative experiment, and the comprehensive performance evaluation of the model was conducted.

4.2 Validation of Sentiment Analysis Model Integrating Multiple Attention Mechanisms

To ensure the effectiveness of the student sentiment analysis model proposed in the study, three datasets, Education, Restaurants, and Courses, were used as the data sources for model evaluation. Meanwhile, to ensure the smooth progress of the experiment, this study set the dimension of all word embeddings in the model to 300, with an initial learning rate of 0.001 and a hidden layer size of 300. Firstly, the MAFM model was used to predict emotional polarity for the most typical examples in the Education dataset. The specific probability distribution is shown in Table 2.

On the ground of the specific content of the comment examples, it can be seen that the MAFM model proposed in the study accurately predicts the emotional polarity of two typical examples. Therefore, this study further utilized the Local Attention Emotional Analysis Model (LA-AM), Global Attention Emotional Analysis Model (GA-AM), and MAFM to compare their Area Under Curve (AUC) on two Chinese datasets, Education and Restaurants. The specific comparison results are shown in Fig. 7.

Fig. 7(a) shows that the proposed MAFM model has significantly better AUC values than GA-AM and LA-AM. In the Education dataset, the AUC values of MAFM increased by 1.22% and 2.13 respectively compared to GA-AM and LA-AM. In the comparison of the Restaurants dataset in Fig. 7(b), the AUC value of the MAFM model is still better than the two single attention mechanisms, indicating that the MAFM algorithm is superior. The poor performance of the two single attention mechanism models may be due to the GA-AM model only capturing information on the ground of the distance between words. The LA-AM model only extracts sentence information from a single local perspective. Additionally, The study introduced the CNN sentiment analysis model proposed by Kottursamy et al., the Attention Emotion Enhanced Convolutional Long Short Term Memory (AEC-LSTM) sentiment analysis model proposed by Huang et al., and the Deep Canonical Correlation Analysis (DCCA) and Bimodal Deep Autoencoder (BDAE) proposed by Liu et al. And this study compared their performance ^[17-19]. The comparison results in the Education, Restaurants, and Courses datasets are shown in Fig. 8.

Fig. 8 shows that the MAFM model proposed in the study has certain advantages in performance among the five models. In the Chinese datasets of Education and Courses, the accuracy of MAFM increased by 0.05% -5.23% and 2.55% -5.33%, respectively, compared to other models. This also indicates that the MAFM model performs more stably in Chinese corpora, indicating that the proposed multi attention mechanism fully leverages the advantages of local attention mechanisms. In the Restaurants dataset, the MAFM model has higher accuracy, recall, and F1 value than other models, indicating that the performance of MAFM is more stable. The MAFM model fully demonstrates the advantages of utilizing global information and local information on the ground of syntax, while dynamically adjusting attention weights to improve the evaluation results of the model.

Fig. 7. AUC comparison of three models on two datasets.

Fig. 8. Performance comparison of various models on three datasets.

Table 2. Probability distribution of predicted polarity for typical examples in the Education dataset.

Comment example	Target aspect	Forward direction	Neutrality	Negative direction
Compared to dull content, I really like the teacher's teaching method type	Content	0.81	0.13	0.06
	Teacher	0.99	0.01	0.00
Not understanding the content in the book, the teacher's teaching is relatively poor, unable to gain a deeper understanding and study of this subject	Content	0.00	0.05	0.95
	Teacher	0.00	0.00	1.00

4.3 Validation of Sentiment Analysis Model Integrating Multiple Attention Mechanisms and Neural Networks

To comprehensively evaluate the performance of the MASAFM model constructed by integrating neural networks on the ground of MAFM, the study first compared the performance of MAFM and MASAFM models in three datasets. The specific results are shown in Table 3.

Table 3 shows that MASAFM has improved classification accuracy, precision, and F1 value in the Education dataset compared to MAFM. In the Courses dataset, the classification performance of the MASAFM model is superior to that of MAFM. In both Chinese datasets, the accuracy of the MASAFM model increased by 0.30% and 1.22%, respectively. In the Restaurants dataset, the F1 value of the MASAFM model decreased by 0.27% compared to MAFM. However, the accuracy, precision, and recall of the MASAFM model are still superior to those of the MAFM model. Overall, the performance of the MASAFM model constructed on the basis of the MAFM model has been improved. Especially in the classification processing of Chinese datasets, it fully leverages the advantage of mining the inherent connections between various aspects. Meanwhile, the study further compared the performance of the MASAFM model with other models on the Education Chinese dataset and Restaurants English dataset. The comparison results in the Education dataset are shown in Fig. 9.

Fig. 9 shows that when the dataset is 100%, the accuracy, precision, and recall of the MASAFM model are the highest among all models, with an average increase of 2.45%, 3.01%, and 7.72%, respectively. It can also be seen that the MASAFM model has good generalization performance, while the CNN model has the worst generalization performance among all models, and its accuracy and recall are the lowest among all models. The performance comparison of five models in the English dataset Restaurants is shown in Fig. 10.

Fig. 10 shows that in the Restaurants dataset, the accuracy of MASAFM is lower than CNN, but still higher than other models. The generalization of the MASAFM model is the best among all models, indicating that the performance of MASAFM is superior. Relatively speaking, the classification accuracy, precision, and recall of the MASAFM model have significantly improved compared to AEC-LSTM, DCCA, and BDAE models, indicating that the proposed MASAFM model has better classification performance. In addition, the study further explored the attention weight distribution of CNN and MASAFM in typical examples of the Education dataset, as shown in Fig. 11.

Fig. 11 shows that in terms of "content", the attention weight of the MASAFM model mainly focuses on the word "boring", while the CNN model mainly focuses on the word "like". However, in this example sentence, "boring" corresponds to "content", and "liking" corresponds to "teacher". This indicates that the CNN model has deviated in sentiment category analysis. In Fig. 11(b), regarding the "teacher" aspect, the MASAFM model focuses attention weights on the word "like" and accurately predicts the emotional polarity of the "teacher" aspect in the example sentence. This indicates that the MASAFM model focuses more on the connections between various aspects. This enables more effective differentiation and acquisition of emotional words corresponding to specific aspects of the sentence.

Fig. 9. Performance comparison of models under different proportions in the Education dataset.

Fig. 10. Performance comparison of models under different proportions in the Restaurants dataset.

Fig. 11. The attention weight distribution of CNN and MASAFM models for the same sentence.

Table 3. Performance Comparison of MAFM and MASAFM on Three Datasets.

Index	Education dataset		Courses dataset		Restaurants dataset
	MAFM	MASAFM	MAFM	MASAFM	MAFM	MASAFM
Accuracy	94.63	94.91	81.35	82.34	79.62	79.72
Precision	92.15	92.64	86.26	87.42	76.79	75.77
Recall	93.46	93.01	72.45	73.64	69.03	69.38
F1 score	92.80	92.82	78.75	79.94	72.70	72.43

5. Conclusion

The current evaluation methods for university curriculum teaching are inefficient and require a large amount of classification work, while there is relatively little research on sentiment analysis technology in the field of education. This study proposes the construction of a student sentiment analysis model for teaching evaluation, providing theoretical support for improving teaching quality. On the ground of the research of neural networks on the ground of attention, a sentiment analysis model MAFM is proposed to construct multiple attention mechanisms using global and local attention. At the same time, considering the intrinsic connections between aspects and enhancing the attention of emotional resources on the basis of MAFM, semantic information is propagated over long distances. This constructs a MASAFM model that integrates multiple attention mechanisms and neural networks. The performance verification of the model shows that the proposed MAFM model has superior performance in predicting the emotional polarity of typical example sentences. Compared with a single attention mechanism, the accuracy of MAFM increased by an average of 1.84% in the Chinese dataset and 1.97% in the English dataset. Compared with other models, the F1 value of the MAFM model increased by 1.52% -10.89%, 11.24% -23.80%, and 2.57% -18.58% in the three datasets, respectively. Compared with MAFM, the accuracy of the MASAFM model increased by an average of 1.22%, 0.30%, and 0.13% in the three datasets. The comparison of attention weights in sentences shows that the MASAFM model allocates attention weights more reasonably and accurately predicts the emotional polarity of aspects in sentences. However, the model proposed in the study requires a high degree of relevance to the text. If there is a correlation angle between the text aspects of certain datasets in certain fields, more runtime is required for correlation calculation, which is not conducive to the final classification accuracy. Therefore, the study will further improve the model's dependence on text relevance, achieving performance validation and effective application on multi domain datasets.