https://orcid.org/0009-0000-0915-5787
Figures
Abstract
To help non-native English speakers quickly master English vocabulary, and improve reading, writing, listening and speaking skills, and communication skills, this study designs, constructs, and improves an English vocabulary learning model that integrates Spiking Neural Network (SNN) and Convolutional Long Short-Term Memory (Conv LSTM) algorithms. The fusion of SNN and Conv LSTM algorithm can fully utilize the advantages of SNN in processing temporal information and Conv LSTM in sequence data modeling, and implement a fusion model that performs well in English vocabulary learning. By adding information transfer and interaction modules, the feature learning and the timing information processing are optimized to improve the vocabulary learning ability of the model in different text contents. The training set used in this study is an open data set from the WordNet and Oxford English Corpus data corpora. The model is presented as a computer program and applied to an English learning application program, an online vocabulary learning platform, or a language education software. The experiment will use the open data set to generate a test set with text volume ranging from 100 to 4000. The performance indicators of the proposed fusion model are compared with those of five traditional models and applied to the latest vocabulary exercises. From the perspective of learners, 10 kinds of model accuracy, loss, polysemy processing accuracy, training time, syntactic structure capturing accuracy, vocabulary coverage, F1-score, context understanding accuracy, word sense disambiguation accuracy, and word order relation processing accuracy are considered. The experimental results reveal that the performance of the fusion model is better under different text sizes. In the range of 100–400 text volume, the accuracy is 0.75–0.77, the loss is less than 0.45, the F1-score is greater than 0.75, the training time is within 300s, and the other performance indicators are more than 65%; In the range of 500–1000 text volume, the accuracy is 0.81–0.83, the loss is not more than 0.40, the F1-score is not less than 0.78, the training time is within 400s, and the other performance indicators are above 70%; In the range of 1500–3000 text volume, the accuracy is 0.82–0.84, the loss is less than 0.28, the F1-score is not less than 0.78, the training time is within 600s, and the remaining performance indicators are higher than 70%. The fusion model can adapt to various types of questions in practical application. After the evaluation of professional teachers, the average scores of the choice, filling-in-the-blank, spelling, matching, exercises, and synonyms are 85.72, 89.45, 80.31, 92.15, 87.62, and 78.94, which are much higher than other traditional models. This shows that as text volume increases, the performance of the fusion model is gradually improved, indicating higher accuracy and lower loss. At the same time, in practical application, the fusion model proposed in this study has a good effect on English learning tasks and offers greater benefits for people unfamiliar with English vocabulary structure, grammar, and question types. This study aims to provide efficient and accurate natural language processing tools to help non-native English speakers understand and apply language more easily, and improve English vocabulary learning and comprehension.
Citation: Wang Y (2024) Construction and improvement of English vocabulary learning model integrating spiking neural network and convolutional long short-term memory algorithm. PLoS ONE 19(3): e0299425. https://doi.org/10.1371/journal.pone.0299425
Editor: Mohamed Hammad, Menoufia University, EGYPT
Received: November 8, 2023; Accepted: February 11, 2024; Published: March 22, 2024
Copyright: © 2024 Yunxia Wang. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited.
Data Availability: All relevant data are within the manuscript and its Supporting Information files.
Funding: The author(s) received no specific funding for this work.
Competing interests: The authors have declared that no competing interests exist.
Introduction
As a second language used worldwide, English is of great significance to many non-native English speakers. Mastering English vocabulary is the foundation of language learning, and it is crucial to improve the ability to speak, read, and write. However, the process of learning English vocabulary is often challenging, especially for non-native English speakers [1]. With the rapid development of artificial intelligence (AI) and machine learning, researchers begin to apply these technologies to construct and improve English vocabulary learning models. Spiking Neural Network (SNN) is a new neural network (NN) model, which mimics the pulse transmission mode between neurons in the human brain, and has high biological interpretability and computational efficiency. The Convolutional Long Short-Term Memory (Conv LSTM) algorithm is a deep learning (DL) model that combines Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM), with the ability to process sequence data and image data [2].
This study aims to explore the application of SNN and Conv LSTM algorithms to the construction and improvement of English vocabulary learning models. Firstly, the existing vocabulary learning model is analyzed, and the existing problems and challenges are put forward. Secondly, the principles and characteristics of SNN and Conv LSTM algorithms are studied, and then an English vocabulary learning model integrating SNN and Conv LSTM algorithms is designed and implemented. Finally, the effectiveness of the optimized model in this study is verified by experiments. The results of this study are expected to provide a novel and efficient English vocabulary learning model for non-native English speakers. Meanwhile, this study also helps to promote the application and development of SNN and DL algorithms in the field of education, opening up new possibilities for the combination of language education and AI technology.
Literature review
In traditional research, there have been some studies on constructing and improving English vocabulary learning models. Mandasari and Wahyudin (2021) explored the impact of online contextualization on English vocabulary learning. The results showed that combining vocabulary learning materials and actual context can improve the learner’s vocabulary learning effect. However, this study mainly relied on manual annotation and text processing techniques and lacked the exploration of using the DL algorithm to improve efficiency and accuracy [3]. Zhang et al. (2020) studied the effect of enhancing English vocabulary learning through word association. They found that word association can help learners to remember and use new words better. However, this study had yet to fully utilize the DL algorithm’s potential, and only adopted traditional association technology [4]. Maqsood et al. (2022) comprehensively assessed computer-aided (CA) vocabulary learning. The results denoted that CA learning tools positively improved learners’ vocabulary and vocabulary mastery. However, studies also pointed out some limitations of existing CA learning tools, such as inadequate adaptability to individual differences and insufficient guidance on learning strategies [5]. Chen and Hsu (2020) evaluated the effect of mobile device-assisted vocabulary learning through a meta-analysis of a large number of relevant studies. They found that in the mobile learning environment, learners learned vocabulary using mobile applications and achieved significant learning results. But the study also proposed the challenges of existing mobile-assisted learning methods in adapting to individual needs and creating interactive learning environments [6]. In the study of NNs, Turukame et al. (2022) reviewed the monitoring methods of DL classification techniques. Studies revealed that employing a secure location protocol can improve the recognition performance of NNs in monitoring systems, but the training time was still a problem [7]. Wongchai et al. (2022) improved the monitoring performance of the DL model by adjusting the neural network architecture, changing the position of neurons in the forgetting gate of the model, and enhancing the neural networks’ monitoring performance. However, the problem that the sample size of each input of the model was too small was still not solved [8].
To sum up, traditional research on English vocabulary learning revealed the positive effects of CA, mobile, contextualized, gamified, and multimedia annotation on vocabulary learning. Nevertheless, these approaches faced challenges and limitations, such as inadequate personal adaptability, learner motivation issues, and balancing the relationship between learning and entertainment. In contrast, the research method in this study integrated SNN and Conv LSTM algorithms into the English vocabulary learning model to improve the personalized adaptability, advanced sequence processing ability, and accuracy of the model effect, which provided better learning support and improved the expression ability of non-native English learners.
The application of SNN and Conv LSTM algorithms in English vocabulary learning model
Problems existing in traditional English vocabulary learning model
Traditional English vocabulary learning methods have always been an important way for learners to master and remember English words. These traditional methods are mainly exhibited in Table 1.
The advantage of traditional methods is that they are easy to understand and implement and can be adapted and extended according to the needs of learners. However, they are often not systematic and personalized, and cannot be well adapted to learners’ characteristics and learning goals [9–12]. In addition, due to the large and varied vocabulary, the traditional model is difficult to cover all words, and it is not easy to cope with the diversity of meanings and the complexity of contexts.
Application of SNN and Conv LSTM algorithms in information processing
SNN is a simulation algorithm based on the biological neural network (BNN), which plays an important role in timing information processing and pattern recognition tasks. Unlike the traditional NN model, SNN represents the communication between neurons with discrete events, and transmits and processes information through time interval and frequency coding [13].
SNN has the following advantages in timing information processing. The first is time accuracy. SNN can more accurately capture and process the time correlation in the input data. Since the occurrence time and frequency of pulses play a key role in information encoding, SNN can more accurately perceive and respond to subtle patterns and dynamic changes. The second is energy efficiency. Compared with traditional NN models, SNN consumes less computing resources and energy when processing timing information. This is because the SNN is only activated when a pulse is received and is computed sparsely in time, reducing computation overhead and power consumption. The last is asynchronous communication. Communication between neurons in SNN is carried out asynchronously. That is, neurons autonomously generate pulses based on their internal state and transmit information to other neurons via the pulses. This allows SNN to process parallel and distributed timing information, helping to simulate complex time-dependent tasks in the real world.
Conv LSTM algorithm is a model that combines CNN and LSTM, which is of great significance for sequence data modeling task. Compared with traditional Recurrent Neural Network (RNN) and LSTM models, Conv LSTM takes advantage of convolutional operations’ local connectivity and weight sharing characteristics and can also process and remember long-distance dependencies in sequences [14–17]. In sequence data modeling, its advantages are outlined in Table 2.
Conv LSTM has been extensively used in many fields, such as video analysis, motion recognition, natural language processing (NLP), etc. It can effectively model the spatio-temporal relationship in sequence data and provide more accurate prediction and analysis results [18–21]. Moreover, Conv LSTM is also used for image generation, video prediction, and other tasks, demonstrating its powerful representation and modeling capabilities for sequence data.
Establishment of the English vocabulary learning model based on SNN and Conv LSTM algorithms
SNN is a biologically inspired NN model, which simulates the working mechanism of brain neurons and represents the information transmission between neurons through discrete pulse signals [22]. Neuronal models in SNN are usually based on the Izhikevich model of pulse conduction or the Leaky integrand-fire model. These models can generate a pulsed output based on the input current and the neuron’s membrane potential. SNN uses a time-encoding scheme to represent information [23]. While traditional NNs perform calculations based on the level of neuron activation, SNN encodes information by the arrival time and frequency of the pulses. Higher frequencies indicate stronger activation. When a neuron’s membrane potential exceeds a threshold, it generates a pulse and transmits it to the connected neuron. In this way, the transmission of information between neurons is carried out through discrete pulse signals.
The model structure of Conv LSTM consists of a series of Conv LSTM units. Each Conv LSTM unit includes an input gate, a forget gate, and an output gate, as well as a cell state and a hidden state [24–26]. Different from traditional LSTM, the input, forget, and output gates of Conv LSTM, as well as cell and hidden states, are all two-dimensional matrices [27]. Therefore, this study integrates SNN and Conv LSTM algorithms, and adds information transfer and interaction modules to improve the effect of English vocabulary learning. The new model is displayed in Fig 1.
Using SNN to process temporal information can better capture temporal correlations and long-term dependencies in temporal data than traditional NN models. This makes SNN potentially useful in processing time-related tasks such as speech recognition, time series prediction, and brain signal analysis. The merit of Conv LSTM in processing sequence data is that it can consider both spatial and temporal features, and effectively process image sequence and video data. Conv LSTM has more advantages than traditional LSTM models regarding computational efficiency and modeling capability. This makes Conv LSTM widely used in image timing prediction, video frame interpolation, action recognition, and other tasks.
By integrating SNN and Conv LSTM algorithms, and adding information transfer and interaction modules, the model’s time series modeling ability and feature extraction ability can be further improved. Such a fusion model can exhibit better performance and effect in tasks dealing with timing information, such as text generation, speech recognition, video analysis, etc.
Experimental design
The data sets used here are the public data sets in WordNet and Oxford English Corpus. WordNet is an English dictionary and vocabulary database, which organizes a lot of words and related information about their meanings based on the concept of words. WordNet organizes words into a network of meanings in which words and their various meanings are called "synsets". Each synset represents a word’s meaning; multiple words can be associated with a synset. It divides the relationship between the upper and lower words among different meanings. WordNet data set can be accessed through the official website to obtain data: https://wordnet.princeton.edu/. The Oxford English Corpus is a large and diverse corpus containing massive English text samples from diverse textual sources for the study and analysis of language use in English. It collects samples of texts from different fields and genres, involving books, blogs, newspapers, journals, social media, and more. The samples come from English speakers in different countries and regions, covering a variety of language styles and styles. The Oxford English Corpus data set can be accessed through the official website to obtain data: https://ota.bodleian.ox.ac.uk/repository/xmlui/handle/20.500.12024/2554.
In the experimental environment, the Central Processing Unit (CPU) size is 2.5G, the server’s system version is Windows 7, the hard disk size is 12GB, and the network is Apache-tomecat6. Part of the code used to run the model in the experiment is as follows:
- #Add embedded layer
- Model. add (Embedding (input_dim = vocab_size, output_dim = embedding_dim, input_length = sequence_length))
- #Add one-dimensional convolutional layer
- Model. add (Conv1D (filters = num_filters, kernel_size = filter_size, activation = ’redu ’)
- Model. add (Dropout (rate = 0.2))
The experiment needs to unify the parameters, where the LSTM quantity is 16, the learning rate is 0.001, the regularization parameter is 8, the batch size is 16, and the activation function is Relu.
Performance analysis of English vocabulary learning model using SNN and Conv LSTM algorithms
Performance analysis of English vocabulary learning model after optimization
The optimization model proposed in this study is compared with five traditional models. The comparison models are Logistic Regression (LR), Support Vector Machine (SVM), Decision Tree (DT), Random Forest (RF), and Gradient Boosting (GB). From the learners’ viewpoint, the model’s accuracy, loss, F1-score, training time, polysemy processing accuracy, syntactic structure capturing accuracy, vocabulary coverage, context understanding accuracy, word sense disambiguation accuracy, and word order relation processing accuracy are considered. For the convenience of the experiment, A refers to accuracy; B stands for loss; D means polysemy processing accuracy; E represents syntactic structure capture accuracy; G indicates word coverage; H denotes context understanding accuracy; I displays word sense disambiguation accuracy; J signifies word order relation processing accuracy. When the training set size is 100, 200, and 400, the model comparison is presented in Fig 2.
Comparison results of model performance under the 100–400 training set: ((a), (b), and (c) represent that the size of the training set is 100, 200, and 400).
In Fig 2, the proposed optimization model’s accuracy is 0.75%-0.77 in the 100–400 text volume range, which means the model can accurately predict text classification. In addition, the model’s loss value is less than 0.45, indicating that the model has a high degree of fitting to the training data. The F1-score is more than 0.75, illustrating that the model can balance accuracy and recall rate to achieve a better prediction effect. The comparison results of training time are plotted in Table 3.
Table 3 describes that the training time of the proposed algorithm model is less than 300 seconds, which indicates that the model’s training speed is relatively fast. Besides, the model also achieves good results on other performance indicators, exceeding the 65% level. When the training set size is 500–1000, the performance comparison results of the model are suggested in Fig 3.
Comparison of model performance in the 500–1000 training set: ((a), (b), and (c) indicate that the training set sizes are 500, 800, and 1000).
According to further analysis of Fig 3, it can be observed that in the range of 500–1000 text volume, the accuracy of the proposed model is 0.81–0.83, the loss is no more than 0.40, and the F1-score is no less than 0.78. The model has reached a high level in all indicators showing its strong ability and comprehensive performance. It means that the proposed model fits the training data well, and the comparison results of training time are revealed in Table 4.
Table 4 demonstrates that the proposed model’s training time is within 400s, which indicates that the model has high training efficiency, enabling it to learn and adapt quickly on large-scale data sets, and other performance indicators are over 70%. When the training set size is 1500–3000, the models’ performance comparison results are expressed in Fig 4.
Comparison results of model performance under the 1500–3000 training set (a): The training set size is 1500; (b): The training set size is 2000; (c): The training set size is 3000.
Fig 4 details that the optimized model performs excellently in the text volume range of 1500–3000. The model’s accuracy is stable in the range of 0.82 to 0.84, which means that the model can make classification predictions of text with high accuracy. With such a high degree of accuracy, the model’s predictions will be very reliable. At the same time, the model’s loss value can remain low, no more than 0.28, indicating a high degree of model adaptation to the training data. The model’s excellent fit means it can effectively capture patterns and features in the data. In terms of F1-score, the model’s performance exceeds 0.78, demonstrating that it has a good balance between balanced prediction effects and recall rates. This balancing performance helps the model to efficiently process various categories of text and provide accurate classification results. The comparison results of training time are shown in Table 5.
Table 5 expresses that the training time of the proposed model is also quite considerable, and it only takes 600 seconds to complete the training. This means that the model has a high training efficiency, allowing it to learn and adapt to large text data sets relatively quickly. Ultimately, in addition to accuracy, loss, and F1-score, the model also shows excellent performance on other performance indicators, exceeding the 70% level. These indicators may include accuracy, recall, class-specific accuracy, etc., further confirming the model’s overall performance.
Simulation results of the model
Simulation experiments were conducted to verify the proposed model’s feasibility and effectiveness. LR, SVM, and DT models were selected for comparison. Six types of question types were identified and analyzed after evaluation by professional teachers, including choice, fill-in-the-blanks, spelling, matching, exercises, and synonyms. The experimental results are portrayed in Fig 5.
Fig 5 draws that after evaluation by professional teachers, the mean scores of matching, fill-in-the-blank, exercises, choice, spelling, and synonym questions are 92.15, 89.45, 87.62, 85.72, 80.31, and 78.94, respectively, which are far more than other traditional models. It can explain the feasibility and effectiveness of the proposed model. Additionally, combined with the horizontal comparison data of the model, it can be found that the fusion model’s performance is gradually enhanced with the increase of the text volume, showing lower loss and higher accuracy.
Discussion
According to the experimental results, the performance of the optimized fusion model under different text volumes shows higher superiority. In the range of 100–400 texts, the accuracy of the fusion model is stable between 0.75–0.77, the loss is maintained below 0.45, the F1-score remains above 0.75, the training time is controlled within 300 seconds, and other performance indicators are maintained above 65%. In the range of 500–1000, the fusion model’s accuracy performance is slightly improved, reaching the level of 0.81–0.83, the loss does not exceed 0.40, the F1-score remains more than 0.78, the training time increases to less than 400 seconds, and other performance indicators exceed 70%. When the amount of text is expanded to 1500–3000, the model’s accuracy remains between 0.82–0.84, the loss is significantly reduced to less than 0.28, the F1-score is not less than 0.78, the training time is slightly longer, but it is still controlled within 600 seconds. Other performance indicators are more than 70%. These results manifest that with the rise of text volume, the performance of the fusion model is slowly promoted, reflecting higher accuracy and lower loss. This also shows that this model has strong scalability and generalization ability, and can achieve good performance when dealing with text data of different sizes. The optimized fusion model presents excellent results in English learning tasks in the simulation experiment. According to professional teachers’ evaluation, the fusion model’s average scores in the synonym types, spelling, choice, exercises, filling-in-blank, and matching are 78.94, 80.31, 85.72, 87.62, 89.45, and 92.15. These scores are significantly higher than other traditional models, further verifying the superiority of the fusion model. The study aims to offer accurate and efficient NLP tools to help non-native English speakers understand and apply the language more easily, and improve their English vocabulary learning and comprehension skills. The fusion model’s excellent performance shows its potential and application value, which provides greater help to English learners.
Conclusion
In the past few years, the swift growth of DL and NLP technologies has provided new opportunities for English vocabulary learning. Here, by combining the strengths of SNN and Conv LSTM algorithms in time series information processing, an English vocabulary learning model combining the two algorithms is designed and constructed. The experimental outcomes show that the performance of the fusion model is better under various text sizes. In the text volume range of 100–400, the fusion model reaches a stable accuracy between 0.75 and 0.77, and a loss control below 0.45. In addition, the F1-score achieves above 0.75. The training time does not exceed 300 seconds, while the other performance indicators are more than 65%. As the number of texts increases to 500–1000, the fusion model’s accuracy improves, between about 0.81–0.83. The loss value remains below 0.40, the F1-score stays above 0.78, with training time limited to 400 seconds or less, while other performance indicators exceed 70%. For the range of 1500–3000, the fusion model still maintains a high accuracy of about 0.82–0.84. The loss is reduced to less than 0.28, while the F1-score is no less than 0.78. Although the training time is slightly longer, it is under 600 seconds. Moreover, other performance indicators are also higher than 70%. Additionally, the fusion model can adapt to various types of questions. 85.72, 89.45, 80.31, 92.15, 87.62, and 78.94 correspond to the average scores on question types such as choice, fill-in-the-blank, spelling, matching, exercises, and synonyms, respectively. These scores are significantly higher than other models, illustrating that the fusion model has a good effect in practical application. This study also has many shortcomings. First, the types of questions covered and the scope of data may be limited, failing to cover all possible situations. Second, the experimental results are based on specific data sets and are limited to English learning tasks, so their applicability may be affected by other tasks and fields. Future studies will be further refined and extended to verify the performance and applicability of the fusion model in a broader range of scenarios.
References
- 1. Li Y, Xu H, Bian M, & Xiao J. Attention based CNN-ConvLSTM for pedestrian attribute recognition. Sensors, 2020; 20(3): 811. pmid:32028568
- 2. Agga A, Abbou A, Labbadi M, & El Houm Y. Short-term self consumption PV plant power production forecasts based on hybrid CNN-LSTM, ConvLSTM models. Renewable Energy, 2021; 177(44): 101–112.
- 3. Mandasari B, Wahyudin A Y. Flipped classroom learning model: implementation and its impact on EFL learners’ satisfaction on grammar class. Ethical Lingua: Journal of Language Teaching and Literature, 2021; 8(1): 150–158.
- 4. Zhang J, Lin T J, Liu Y & Nagy W. E. Morphological awareness and reading comprehension: Differential mediation mechanisms in native English speakers, fluent English learners, and limited English learners. Journal of Experimental Child Psychology, 2020; 199(56): 104915. pmid:32653507
- 5. Maqsood S, Shahid A, Afzal M T. Assessing English language sentences readability using machine learning models. PeerJ Computer Science, 2022; 8(1): 818. pmid:35111913
- 6. Chen Y L, Hsu C C. Self-regulated mobile game-based English learning in a virtual reality environment. Computers & Education, 2020; 154(1): 103910.
- 7. Turukame A, Alhebaishi N, Abdulrhman M, Olfat M, Bhardwaj A. Multispectral image analysis for monitoring by IoT based wireless communication using secure locations protocol and classification by deep learning techniques. Optik, 2022; 271(16): 170122.
- 8. Wongchai A, Jengrti D, Priyadarsini I, Deb N, Bhardwaj A. Farm monitoring and disease prediction by classification based on deep learning architectures in sustainable agriculture. Ecological Modelling, 2022; 474(36): 110167.
- 9. Vosta S, Yow K C. A cnn-rnn combined structure for real-world violence detection in surveillance cameras. Applied Sciences, 2022; 12(3): 1021.
- 10. Zheng H, Lin F, Feng X, & Chen Y. A hybrid deep learning model with attention-based conv-LSTM networks for short-term traffic flow prediction. IEEE Transactions on Intelligent Transportation Systems, 2020; 22(11): 6910–6920.
- 11. Agga A, Abbou A, Labbadi M, & Ali I H O. CNN-LSTM: An efficient hybrid deep learning architecture for predicting short-term photovoltaic power production. Electric Power Systems Research, 2022; 208(19): 107908.
- 12. Hou S, Li W, Liu T, Qin R, & Wan Z. D2CL: A dense dilated convolutional LSTM model for sea surface temperature prediction. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2021; 14(9): 12514–12523.
- 13. He K, Ji L, Wu C W D, & Tso K F G. Using SARIMA–CNN–LSTM approach to forecast daily tourism demand. Journal of Hospitality and Tourism Management, 2021; 49(15): 25–33.
- 14. Elaskily M A, Alkinani M H, Sedik A, Sedik A, & Dessouky M M. Deep learning based algorithm (ConvLSTM) for copy move forgery detection. Journal of Intelligent & Fuzzy Systems, 2021, 40(3): 4385–4405.
- 15. Fang W, Chen Y, Xue Q. Survey on research of RNN-based spatio-temporal sequence prediction algorithms. Journal on Big Data, 2021; 3(3): 97.
- 16. Atila O, Şengür A. Attention guided 3D CNN-LSTM model for accurate speech based emotion recognition. Applied Acoustics, 2021; 182(56): 108260.
- 17. Qiao Y, Guo Y, Yu K, & He D. C3D-ConvLSTM based cow behaviour classification using video data for precision livestock farming. Computers and Electronics in Agriculture, 2022; 193(13): 106650.
- 18. Sun H, Chen M, Weng J, Liu Z, & Geng G. Anomaly detection for in-vehicle network using CNN-LSTM with attention mechanism. IEEE Transactions on Vehicular Technology, 2021, 70(10): 10880–10893.
- 19. Singh S P, Sharma M K, Lay-Ekuakille A, Gangwar D, & Gupta S. Deep ConvLSTM with self-attention for human activity decoding using wearable sensors. IEEE Sensors Journal, 2020; 21(6): 8575–8582.
- 20. Wang J. Speech recognition of oral English teaching based on deep belief network. International Journal of Emerging Technologies in Learning (Online), 2020; 15(10): 100.
- 21. Hu X, Liu T, Hao X, & Lin C. Attention-based Conv-LSTM and Bi-LSTM networks for large-scale traffic speed prediction. The Journal of Supercomputing, 2022; 78(10): 12686–12709.
- 22. Chen Y, Wang Y, Dong Z, Han Z, Zhou D,… & Bao Y. 2-D regional short-term wind speed forecast based on CNN-LSTM deep learning model. Energy Conversion and Management, 2021; 244(10): 114451.
- 23. Zhang Z, Luo H, Wang C, Gan C, & Xiang Y. Automatic modulation classification using CNN-LSTM based dual-stream structure. IEEE Transactions on Vehicular Technology, 2020; 69(11): 13521–13531.
- 24. Li T, Hua M, Wu X U. A hybrid CNN-LSTM model for forecasting particulate matter (PM2. 5). Ieee Access, 2020; 8(6): 26933–26940.
- 25. Xu G, Ren T, Chen Y, & Che W. A one-dimensional cnn-lstm model for epileptic seizure recognition using eeg signal analysis. Frontiers in neuroscience, 2020; 14(3): 578126. pmid:33390878
- 26. Lee S W, Kim H Y. Stock market forecasting with super-high dimensional time-series data using ConvLSTM, trend sampling, and specialized data augmentation. expert systems with applications, 2020; 161(32): 113704.
- 27. Hermanto D T, Setyanto A, Luthfi E T. Algoritma LSTM-CNN untuk Binary Klasifikasi dengan Word2vec pada Media Online. Creative Information Technology Journal, 2021; 8(1): 64–77.