Identification of Key Elements and Path Design of Vocational Education Governance Modernization Supported by Deep Learning

Modernization of vocational education governance refers to combining traditional vocational education with modern educational concepts and technical means, and improving the adaptability and quality of vocational education through reform and innovation to meet the needs of modern society for talent training [1-3].

Understanding the goals and characteristics of the modernization of vocational education governance is the basis for identifying its key elements and path design, which mainly involves six aspects of understanding: first, adapting to social needs. The core goal of the modernization of vocational education governance is to cultivate high-quality and high-skilled vocational talents who adapt to the needs of economic and social development [4-6]. By analyzing the social demand for various types of talents, the vocational education curriculum and teaching content adjustment and updating, to achieve the integration of production and education. Second, the introduction of modern education concepts. The modernization of vocational education governance focuses on cultivating students’ innovative thinking and practical ability, and emphasizes students’ initiative and openness [7-8]. Adopting modern educational concepts, students are encouraged to think independently, learn independently and explore and practice. Third, innovative teaching mode. The modernization of vocational education governance requires innovative teaching modes to improve the quality and efficiency of teaching by drawing on and introducing advanced educational concepts and teaching methods [9-11]. Such as the use of multimedia technology, virtual reality technology and other modern teaching methods, the opening of online courses, distance learning and other learning methods to meet the learning needs of students. Fourth, the development of practice base and school-enterprise cooperation. The modernization of vocational education governance requires schools to cooperate closely with practice bases and enterprises, combining school teaching and practical training with actual work scenes, so that students can carry out real practical operations and work experience, and cultivate practical application ability and vocational literacy [12-14]. Fifth, strengthening teacher training and quality improvement. The modernization of vocational education governance requires teachers to have modern education concepts and innovation ability, and to be able to adapt to the development requirements of the modernization of vocational education governance [15-17]. Sixth, combine with internationalization development. The modernization of vocational education governance should be in line with international standards, draw on and absorb international advanced education concepts and practical experience, and cultivate vocational talents that meet the needs of the international market [18-20]. At the same time, it can actively participate in international vocational education exchanges and cooperation to promote the internationalization of vocational education.

In general, the modernization of vocational education governance is a comprehensive reform about education content, teaching method, education environment and teacher quality, aiming to improve the effectiveness and quality of vocational education and cultivate high-quality vocational talents adapted to the needs of the society.

The article firstly describes the relevant definitions and explanations of events and emergencies, then describes in detail the relevant theoretical models and techniques used in the research of this paper, and proposes the research line of identifying the elements of events based on feature fusion. Then, Attention and LSTM, which are the basic technologies for identifying key elements of modernization of vocational education governance, are introduced in detail. On this basis, the article proposes an element recognition model based on improved BiLSTM-CRF, including an improved BERT design method based on Attention mechanism, an improved design (Position) method based on local position information, an improved design (LSTMD) method based on LSTMDecoder, and an improved design based on global attention mechanism (Attention) method. Then the recognition of various basic elements by this paper’s model and other models is analyzed. Subsequently, the article explores the path of modernizing vocational education governance at five levels: concept, system, resources, governance tools, and governance mechanisms. Finally, using the QCA analysis method, the multiple key factors affecting the modernization of vocational education are analyzed.

2

Methodology for identifying key elements for modernizing the governance of vocational education

2.1

Relevant theories and lines of research

2.1.1

Event-related concepts

1)

Events and event elements

An event, as a representation of information, is an event that occurs at a particular time and in a particular context, involves a number of actors, and exhibits a number of action characteristics. An event can usually be represented formally as e, defined as a hexadecimal group: 1 $e = (A, O, T, V, P, L)$

Among them, A, O, T, V, P, and L are all representations of event elements, representing action elements, object elements, time elements, environment elements, assertion elements, and linguistic expression elements, respectively [21]. 2)

Critical incidents

Sudden event can be understood as something that happens suddenly in a broad sense, with two meanings: first, the event occurs very quickly. The second is that the event is handled in a tricky way. Narrowly defined emergencies refer to the sudden occurrence of major or sensitive events, in short, natural and man-made disasters. The term “natural disasters” refers to disasters caused by the natural environment, such as earthquakes, tsunamis, floods, fires, etc., while the term “man-made disasters” refers to human-caused hazards, such as terrorist attacks and social conflicts.

2.1.2

Text representation

1)

One-hot coding

One-hot coding is also known as solo-hot coding, which is based on the vector space model and the algorithm is calculated based on metrics in the vector space. The encoding is done by constructing a word list for the processed dataset, and then each word in the word list is represented by a vector, the dimension of the vector is the size of the word list. Only one position of each word is encoded as 1, and the rest of the positions are all 0, so that each word has and only has its own unique encoding.

For example, there are two words in the corpus, “critical incident extraction” and “event element identification”. The process of using One-hot coding is as follows: (1)

Perform Chinese segmentation on the corpus to obtain the words in the corpus and construct the word list “sudden, event, extraction, element, recognition”.

(2)

Determine the size of the word list to be 5, and represent each word in the order of the word list, with a word vector of $[1, 0, 0, 0, 0]$ for “burst”, $[0, 1, 0, 0, 0]$ for “event”, $[0, 0, 1, 0, 0]$ for “extraction”, $[0, 0, 0, 1, 0]$ for “feature”, and $[0, 0, 0, 0, 1]$ for “recognition”. The vector representation of the text is shown in Table 1.

Table 1.

Text vector representation

Text	Vector representation
Emergency extraction	$[1, 1, 1, 0, 0]$
Event factor recognition	$[1, 0, 0, 1, 1]$

According to the text vector coding table, the matrix representation of the text is obtained as shown in Equation (2). 2 $M a t r i x = [\begin{array}{l} 10000 \\ 01000 \\ 00100 \end{array}]$ 2)

Word2vec

The principle of the CBOW model is to predict the word vectors of the target word given the context words, i.e., to predict the vector representation of W_t given the contexts W_t−2, W_t−1, W_t+1 and W_t+2 of the target word W_t. The model consists of an input layer, a projection layer and an output layer.

The CBOW model trains the word vectors as follows: first, the words in the text are represented by One-hot coding and input as vectors to the input layer of CBOW [22]. Next, it is multiplied with a randomly generated initialization matrix C1 from the projection layer, and the vector obtained from the multiplication is averaged and processed. Then, it is multiplied with a weight matrix C2 to obtain the output vector of the target word. Finally, the probability distribution of the target word is obtained using the Softmax function normalization process, and the vector representation of the target word has the highest probability. By constantly changing the target word, the vector representation of each word in the text can be obtained finally.

Different from the principle of CBOW model, Skip-Gram model is to predict the context vectors of target words given the target words, i.e., to predict W_t−2, W_t−1, W_t+1, and W_t+2 under the condition that target word W_t is known.Skip-Gram model also consists of an input layer, a projection layer, and an output layer. Its training process is also similar to that of CBOW model, the input layer is to encode the target word and send it to the projection layer, after weight matrix operation and Softmax function normalization, the context word vector of the target word is obtained.

3)

BERT pre-training model

BERT pre-training language model has become a popular model in the field of natural language processing. The Bidirectional Transformer is the core structure of BERT, which solves the problem that convolutional neural networks cannot solve serialized text and the problem of memory explosion in recurrent neural networks.

Transformers mainly contain two parts: encoding and decoding. The input of the encoding part includes word vector information and location information, which is input to the decoding part after passing through the multi-head attention mechanism and feed-forward network. The output of the decoding part is obtained by the multi-head attention mechanism, feed-forward network and linear network.The computational efficiency of the Transformer structure is greatly improved compared with the previous models, thanks to the fact that it does not use the classical loop structure, but uses the attention mechanism to compute the implicit representations of the inputs and outputs of the model. The input layer also contains the lexical position information of the text, which can be computed using the sine and cosine functions as shown in Equations (3) and (4). 3 $P_{(p o s, 2 i)} = \sin (p o s / 10000^{2 i / d_{\mod e l}})$ 4 $P_{(p o s, 2 i)} = \cos (p o s / 10000^{2 i / d_{\mod e l}})$

where pos denotes the dimension of the word and d_model denotes the dimension of the model.

2.1.3

Feature fusion based recognition method for event elements

This paper proposes a research on event element recognition method based on feature fusion, which contains two main parts: event element recognition method based on word feature fusion and event element recognition method based on multidimensional feature fusion. In this paper, we use the BERT pre-trained language model, and utilize the Chinese word segmentation tool to slice the words in the corpus to get the word-level semantic information, introduce adversarial training to improve the robustness of the model, and combine the bidirectional recurrent neural unit network and the conditional random field to put forward a feature fusion-based event element recognition method. The research route of the feature-fusion-based Chinese critical incident element recognition method is shown in Fig. 1, which mainly contains five steps, data preprocessing, text vector embedding, adversarial training, semantic feature extraction and event element recognition.

2.2

Element Recognition Related Techniques

2.2.1

Attention mechanism

Attention mechanisms originated from the study of human vision and are now widely used in areas such as machine translation, image generation, and natural language processing. An attention mechanism is proposed when processing information encounters a bottleneck. People will focus on part of the information while intentionally ignoring irrelevant or unimportant information. The attention mechanism is divided into 2 main aspects, one is to decide which part of the input needs to be focused on, and the other is to allocate limited resources to the important part.

2.2.2

Long and short-term memory neural networks

The Long Short-Term Memory (LSTM) network inherits the memory function of a Recurrent Neural Network (RNN), which is able to record current input information and arbitrary length information [23]. LSTM network solves the long term dependency problem existing in RNN by introducing the gating unit and the storage unit and alleviates the problem of the RNN gradient vanishing and the gradient exploding problem.There are three gating units in LSTM network, which are the input gate, output gate and forgetting gate, in addition, many memory cells are added, and the combination of these memory cells and gating units can improve the processing ability of long sequence data.

i_t is the input gate that determines how much of the newly acquired information is updated in the LSTM cell at the t moment and saved in the memory cell as shown in Equation (5). 5 $i_{t} = σ (U_{i} h_{t - 1} + W_{i} e_{t} + b_{i})$

Where, σ is the sigmoid activation function, U_i is the parameter matrix of the input gate, h_t−1 is the output value of the hidden layer state in the previous moment, W_i is the weight matrix at the time of input, e_t is the input value of the current node of the LSTM at the moment t, and b_i is the bias vector at the time of input.

f_t is the forgetting gate, which decides the hidden layer cell state of the previous layer in the LSTM cell at the t moment with a certain probability of forgetting, as shown in Equation (6). 6 $f_{t} = σ (U_{f} h_{t - 1} + W_{f} e_{t} + b_{f})$

U_f, W_f and b_f are the parameter matrix, weight matrix and bias vector at the time of forgetting, respectively.

o_t is the output gate, which decides which information in the LSTM cell at moment t is used as output to generate the hidden layer state variables, as shown in Equation (7). 7 $o_{t} = σ (U_{o} h_{t - 1} + W_{o} e_{t} + b_{o})$

U_o, W_o and b_o are the parameter matrix, weight matrix and bias vector at output, respectively.

2.3

Element Recognition Model Based on Improved BiLSTM-CRF

Named entity recognition task plays a key role in natural language processing techniques. Case element recognition in this paper can also be categorized as named entity recognition. Named entity recognition, in its narrow sense, identifies three types of named entities, namely, human names, geographical names and organizational names [24]. The words are first converted into distributed representations, which are trained in the input to a deep neural network, and finally fed into a conditional random field to predict labels. The work in this section is to study and analyze the baseline model Character Vector Embedding-BiDirectional Long Short Memory Neural Network-Conditional Random Fields (CharEmbendding-BiLSTM-CRF) and improve some of the problems of this baseline model, so that the key elements can be identified more accurately.

2.3.1

Baseline model (BiLSTM-CRF)

The BiLSTM-CRF model consists of three layers, which are the embedding layer, the bidirectional long-short memory neural network, and the conditional random field. In the first embedding layer, there are two traditional choices of embedding layers: a word embedding layer and a word embedding layer. According to He, Liu, Li et al. made an explicit discussion on whether to use word-level or word-level sequence annotation in Chinese entity recognition, and the results show that the latter is empirically preferred. The word-level sequence annotation model has become the main method for Chinese entity recognition. So we chose the word embedding layer. The word embedding layer transforms each word into a distributed representation, also known as vectorization. It is worth mentioning that even punctuation marks are vectorized as words. The initialization of the word embedding vectors can be chosen randomly or trained using methods such as Glove, Word2 Vec, etc. We use Word2Vec trained word vectors, using with trained word vectors has the advantage of accelerating the training of the model and so on. Intuitively, reducing the word vectors to two dimensions using PCA will reveal that the distance between word vectors with similar meanings is also closer.

The CRF layer can add constraints to ensure that the final prediction is valid. These constraints can be automatically learned by the CRF layer when training the data. Possible constraints are that the sentence should start with “B-” or “O” instead of “I-”. “B-labell I-label2 I-label3...”, in this pattern, categories 1, 2, and 3 should be the same entity class. For example, “B-Person I-Person” is correct and “B-Person I-Organization” is wrong. “O I-label” is wrong, and the beginning of the named entity should be “B-” instead of “I-”. With these useful constraints, false prediction sequences will be greatly reduced. The loss function in the CRF layer consists of two types of fractions: the emission fraction, which is the output of our BiLSTM layer. Transfer scores: Scores from one label to the next, you can randomly initialize the scores of the transition matrix before training the model. These scores will be updated as the training iteration progresses. The CRF loss function consists of two parts, the score of the true path and the total score of all paths. The real path’s score should be the highest among all the paths. The loss function is as follows: 8 $L o s s F u n c t i o n = \frac{P_{R e a l p a t h}}{P_{1} + P_{2} + \dots + P_{N}}$

Since our training goal is usually to minimize the loss function, after adding the logarithmic and negative sign simplifications: 9 $\begin{array}{rcl} L o g L o s s F u n c t i o n & = & - \log \frac{P_{Re a l p u t h}}{P_{1} + P_{2} + \dots + P_{N}} \\ = & - \log \frac{e^{s_{g e s l l i t h}}}{e^{e_{1} + e^{s_{2}} + \dots + e^{s N}}} \\ = & - (\log (e^{s_{Re c l i p a n}}) - \log (e^{s_{1}} + e^{S_{2}} + \dots + e^{s_{N}})) \\ = & - (S_{Re a l p a t h} - \log (e^{S_{1}} + e^{S_{2}} + \dots + e^{S_{N}})) \\ = & - (\sum_{i = 1}^{N} x_{i y_{i}} + \sum_{i = 1}^{N - 1} t_{y, y_{i + 1}} - \log (e^{S_{1}} + e^{S_{2}} + \dots + e^{S_{N}})) \\ = & - (\sum_{i = 1}^{N} x_{i i_{i}} + \sum_{i = 1}^{N - 1} t_{y, y_{i + 1}} - \log (e^{s_{1}} + e^{s_{2}} + \dots + e^{s_{N}})) \end{array}$

2.3.2

Model improvements and innovations

In this paper, we propose four four improvements which have a greater increase in precision, recall, and F₁ value than Baseline, which are: improved BERT design based on Attention mechanism, improved design based on local position information, improved design based on Attention, and improved design based on LSTMDecoder. And better results can be obtained after mutual cross-fusion and overall fusion.The neural network structure of AttBERT-Pos-BiLSTM-LSTMD-Attention-CRF is shown in Fig. 2:

1)

Improved BERT design based on Attention mechanism (AttBERT)

The model structure of BERT is a bidirectional Transformer, but only the encoder part of Transformer is used.Transformer is divided into encoder and decoder parts, the encoder part is stacked by 6 encoders, and the decoder part is stacked by 6 decoders, which have the same structure but don’t share the weights. There is a residual connection in the middle, which makes up for the disadvantage that LSTM can not be computed in parallel with the multi-head self-attention mechanism. In this paper, we use the pre-trained Chinese model released by Google, which has 12 layers of encoder, 12 multi-heads of self-attention, and the hidden_size is 768. Considering that not only the highest layer of the Transformer contains the semantic information of the words, but also the 11 layers of the intermediate layer, which may encode different abstraction granularity of the input sentence feature information. So we consider stitching the last n layers together, but we do not know which layer contains more information. So we used the Attention mechanism of the forward neural network to let the model learn by itself how much weight each layer should take. Where n is also a hyperparameter. The structure of Attbert is shown in Figure 3:

The details of the operation are as follows: 10 $e_{i j} = a (h_{i j})$ 11 $α_{i j} = \frac{\exp (e_{i j})}{\sum_{i = 1}^{n} \exp (e_{i j})}$ 12 $w_{j} = \sum_{i = 1}^{n} α_{i j} h_{i j}$

Where $a ()$ represents the forward neural network, n represents the last n layers for Attention, h_ij represents the output of the jth position of the ith layer encoder, a_ij represents the weight of the ith position of the ith layer encoder, and w_j is the result of the weighted sum of the ith position, which is also the word vector we want to input into the model.

$X_{n} = [x_{i}, x_{2} \dots x_{n}]$ is the vector of the corresponding word onehot form, there is a corresponding dictionary in BERT to find the corresponding word vector through one-hot encoding and then through the bidirectional Transformer, that is, 12 encoders, the last n layers of each word encoder results are input into the forward neural network for computation to get the weights corresponding to each layer and then do the weighted summation to get the The final result is w_j. After arithmetic we can obtain more comprehensive word vector features, thus improving the overall element recognition. 2)

Improved design based on local position information

In order to utilize the structural information efficiently, we sum each word vector with the local positional information corresponding to it, giving each word the corresponding structural information. Window size n is the hyperparameter.

The window size is set to n and the sentence length is L. L + 2n vectors are randomly generated as position vectors h_t, and we utilize the Attention mechanism to generate a weight for each vector a. The structure of the neural network based on the local position information is shown in Fig. 4.

The details of the operation are as follows: 13 $e_{t} = a (h_{t})$ 14 $α_{t} = \frac{\exp (e_{t})}{\sum_{t = t - n / 2}^{t + n / 2} \exp (e_{t})}$ 15 $p_{t} = \sum_{t = t - n / 2}^{t + n / 2} α_{t} h_{t}$ 16 $x_{t} = p_{t} + w_{t}$ 3)

Improved design based on LSTMDecoder

For longer and complex entities, this paper proposes to decode them with a combination of LSTMDecoder and CRF, with bidirectional long and short memory neural networks to form an encoder-decoder model.LSTMDecoder was proposed by zheng et al. in the joint extraction model of entities and relations to solve the long term dependency between relation labels. The biggest defect of LSTMDecoder is that it cannot utilize the constraint relationship between labels well. Probably due to different scenarios, zheng et al. also did not try to perform overlay experiments. In this paper, we tried this approach and achieved good results.

The LSTMDecoder in this paper employs an LSTM structure to generate the state scores of the labels to be provided to the CRF layer.The inputs to the LSTMDecoder h_t are obtained from the BiLSTM layer, T_t−1 are the outputs of the previous LSTMDecoder sequence, c_t−1 are the cellular state of the previous sequence, $h_{t - 1}^{(2)}$ are the hidden layer state of the previous sequence, and the overall LSTMDecoder structure is shown in Fig. 5:

The specific details of the operation are defined below: 17 $i_{t} = δ (W_{w i} h_{t} + W_{h i} h_{t - 1}^{(2)} + W_{t i} T_{t - 1} + b_{i})$ 18 $f_{t} = δ (W_{w j} h_{t} + W_{h f} h_{t - 1}^{(2)} + W_{t j} T_{t - 1} + b_{f})$ 19 $z_{t} = \tanh (W_{w c} h_{t} + W_{h c} h_{t - 1}^{(2)} + W_{t c} T_{t - 1} + b_{c})$ 20 $c_{t} = f_{t} c_{t - 1} + i_{t} z_{t}$ 21 $o_{t} = δ (W_{w o} h_{t} + W_{h o} h_{t - 1}^{(2)} + W_{c o} c_{t} + b_{o})$ 22 $h_{t}^{(2)} = o_{t} \tanh (c_{t})$ 23 $T_{t} = W_{t s} h_{t}^{(2)} + b_{t s}$

4)

Improved design based on global attention mechanism

The method proposed by Luo et al. uses a conventional attention mechanism, which calculates the weights of the current word and all other words separately, and then weights the average, due to the long number of words in our data, the amount of attention calculations will be very large to the extent that it is difficult to calculate. We use the forward neural network attention mechanism as a good solution to this problem. The specific operational details are as follows: 24 $e_{t} = a (h_{t})$ 25 $α_{t} = \frac{\exp (e_{t})}{\sum_{k = 1}^{T} \exp (e_{k})}$ 26 $c = \sum_{i = 1}^{T} α_{t} h_{t}$ 27 $m_{t} = \tanh [h_{t}, c]$ 28 $n_{t} = \tanh (W_{m} m_{t} + b_{m})$ 29 $z_{t} = \tanh (W_{z} n_{t} + b_{z})$

3

Identification and analysis of key elements for modernizing the governance of vocational education

3.1

Indicators for experimental evaluation

The problem studied in this section belongs to classification problems, in which common evaluation metrics include precision rate, recall rate, and F1 value. Precision rate is the number of key elements correctly identified by the model as a proportion of all identified key elements, while FP is the number of key elements that were incorrectly identified. Recall rate is the number of key to be correctly identified by the model as a proportion of all real key elements, where FN is the number of non-key elements that are incorrectly identified.

3.2

Analysis of the recognition effect of different model basic elements

This section of the experiment will verify the effectiveness of the model proposed in this paper in identifying the key elements of modernizing vocational education governance. The experiment was designed with six control groups using the HMM model, CRF model, BiLSTM model, BiLSTM-CRF model, and BERT-CRF model.

In order to improve the recognition effect, in recent years various methods have achieved good results in specific tasks by building a lexicon and adding the lexicon information to the model training. In this section, the Lattice-LSTM model and SoftLexicon model are added as experimental controls. Both methods use character-level vector representations, Lattice LSTM introduces a unitary structure of a word, and for the current character, fuses all lexical information ending with that character.SoftLexicon simplifies the sequence modeling by integrating the lexicon into the character representation based on Lattice-LSTM, considering the role of the word in the word. The experimental comparison results are shown in Table 2. From the table of experimental comparison results, it can be seen that the recognition model of key elements of modernization of vocational education governance based on deep learning is higher than the HMM and CRF models under the two indicators of accuracy rate and F1. The CRF-based vocational education governance modernization key element extraction model is improved in all indicators than the HMM model.In the HMM model, the current state depends on the state of the previous moment, whereas the CRF can define more feature functions and use the feature function composed of the label of the current moment and its corresponding word plus the weight to represent the final output probability, which makes full use of contextual features and information. Information. The algorithmic models based on deep learning BiLSTM and BERT are superior to the models based on statistical learning methods in all aspects.CRF model training requires a larger dataset to ensure the validity of the results, and in the process of recognizing the basic elements, it needs to manually design the feature templates. The feature template is to establish the feature rules.LSTM usually encodes along the text sequence, and finally obtains the contextual semantic information of the sentence, but the previous text does not fit into the context after the target word, so this chapter uses the BiLSTM model, which combines two unidirectional, but opposite directions of the LSTM, and can be seen from the experimental results have a certain degree of improvement. The addition of BERT model improves the evaluation indexes to a certain extent, and character-level text vectors can be generated by the pre-training language model of BERT, and the BERT model realizes bidirectional learning and can deal with the polysemy problem of some texts more efficiently. In this chapter, in addition to vector representation of characters with the dynamic representation ability of BERT model, BiLSTM is also added to obtain the contextual semantic information of the text, which is combined with the label relevance analysis of CRF to recognize the basic elements in judicial documents. The analysis of the experimental results shows that the method of this paper achieves good results in recognizing the key elements, with an F1 value reaching 87.26%.

Table 2.

Experimental comparison

Model	Accuracy rate	Recall rate	F1 value
HMM	70.18%	80.27%	76.28%
CRF	81.73%	80.03%	79.92%
BiLSTM	85.08%	82.27%	82.64%
BiLSTM-CRF	86.21%	84.21%	85.65%
SoftLexicon(LSTM)	77.05%	82.63%	85.04%
Lattice-LSTM	88.48%	88%	81.2%
BERT-CRF	87.86%	81.89%	83.64%
Ours	87.65%	85.88%	87.26%

3.3

Comparison of the Recognition Effect of Different Basic Elements

This section conducts experiments in the dataset related to the modernization of vocational education governance. The results of the F1 value assessment for each entity are shown in Table 3. Comparison of the experimental results reveals that there is a decrease in the recognition effect in the task of identifying the basic elements of law. From the table, it can be seen that the identification model of basic elements proposed in this section achieved the best results in labeling sentencing results (R), with an indicator F1 value of 100%.

Table 3.

The results of each entity F1 assessment

Model	Basic type
Model	D	T	N	P	K	R
HMM	95.96%	80.13%	48.87%	31.03%	53.36%	83.56%
CRF	96.52%	92.72%	47.64%	37.39%	57.95%	96.24%
BiLSTM	95.27%	94.32%	49.32%	62.25%	90.57%	95.7%
BiLSTM-CRF	98.06%	93.36%	47.37%	55.04%	93.36%	95.9%
Lattice-LSTM	96.84%	85.97%	48.14%	37.45%	93.21%	92.37%
SoftLexicon(LSTM)	96.64%	88.52%	58.23%	48.61%	95.48%	98.83%
BERT-CRF	96.21%	94.51%	51.45%	57.55%	91.7%	97.52%
Ours	98.82%	94.16%	71.32%	70.44%	93.98%	100%

The model performance comparison is shown in Figure 6 (the left side is the analysis of the results of identifying the “education” element, and the right side is the analysis of the results of identifying the “governance” element, and 1~6 in the figure are the model of this paper, BERT-CRF, BiLSTM-CRF, BiLSTM, CRF, BiLSTM-CRF, BiLSTM, CRF and HMM, respectively). From the figure, it can be seen that the basic element recognition model proposed in this paper performs the best in recognizing the two types of elements, with an average recognition rate of 94% in terms of accuracy, FI value and recall. However, the performance of recognizing education has a small improvement relative to other models, while the additive criminalizable factors have a significant improvement relative to other models. This is precisely because the BiLSTM model is capable of learning the contextual information of text sequences, and the basic element recognition annotates the text sequences sequentially, and BiLSTM learns sequence features sequentially to help sequence annotation. In addition, the BERT model can learn word vectors with rich semantic information, and this dynamic characterization ability can solve the problem of multiple meanings of a word, which is of great help to the problem of recognition difficulty caused by the diversity of descriptions of additive and subtractive sentencing factors.

3.4

Model performance on publicly available datasets

In order to validate the feasibility of the basic element recognition model proposed in this paper, this paper has been validated on a publicly available dataset on modernization of vocational education governance. The dataset has been segmented and labeled with information such as person name, place name, and institution name using BIO’s annotation method. Among them, the person’s name is labeled as PER. place name is labeled as LOC. institution name is labeled as ORG. the evaluation metrics used in this experiment consists of accuracy, recall and F1 value, and the results of the dataset validation are shown in Table 4. From the table, it can be seen that BiLSTM and BERT fused with CRF layer can effectively improve the recognition effect.BiLSTM with the addition of CRF improves the accuracy, recall and F1 value by 0.8%, 2.27%, and 1.91%, respectively.CRF with the addition of BERT improves the three metrics by 4.62%, 2.41%, and 2.96%, respectively. It indicates that BiLSTM learns the contextual information of text sequences well. BERT incorporates global semantic information into character vectors, which contributes to the model’s performance. In addition, the basic element recognition model proposed in this chapter performs the best among all algorithms with an F1 value of 95.22%.

Table 4.

Data set validation results

Model	Accuracy rate	Recall rate	F1 value
HMM	85.76%	86.03%	83.99%
CRF	86.11%	86%	86.64%
BiLSTM	86.87%	87.94%	86.61%
BiLSTM-CRF	86.91%	88.27%	88.55%
BERT-CRF	90.73%	88.41%	89.6%
Lattice-LSTM	90.7%	90.12%	91.38%
SoftLexicon(LSTM)	90.74%	90.24%	90.13%
Ours	96.65%	96.3%	95.22%

4

Path design for modernizing the governance capacity of vocational education

4.1

The concept of vocational education governance as a pioneering element

“Education in the new era is a grand topic that needs to be discussed systematically, and the concept of educational governance is the first important issue with a navigational role to be discussed.” Idea is the precursor of action, advanced education governance concept for the modernization of education governance capacity to open up new ideas and methods. At present, the reality of the hot and difficult problems of education reflects the concept of education governance is still lagging behind the status quo of education governance practice, the concept of vocational education governance must keep abreast of the times and constantly innovate. With the in-depth promotion of the modernization of national governance, the modernization of the concept of governance is constantly integrated into the process of educational development, through a series of processes such as summarizing experience, broadening horizons, changing concepts, the concept of governance of vocational education, scientific, advanced, political, ideological and epoch-making to be continuously improved. The concept of education governance centered on “human modernization” takes the people’s sense of access to education and happiness in education as the foothold of modernizing education governance capacity, and guides education governance in the direction of modernity, democracy and the rule of law. In the new era, the in-depth implementation of the Party’s education policy should be led by its direction, and the innovation and change of the concept of vocational education governance should be taken as the overall leader, so as to make practical and useful guidance for the healthy development of vocational education and the steady improvement of the governance capacity of vocational education with the advanced governance concept.

4.2

Vocational education governance system as a prerequisite element

The modernization of the governance system of vocational education is to continuously improve the institutional mechanisms and institutional norms of education, and to promote high-quality development and high-quality enhancement of education governance under the new development pattern by means of complete and scientific institutional arrangements and an operationally coordinated organizational system. The modernization of vocational education governance capacity is intended to transform potential advantages of the education system into actual effectiveness of education governance within the norms and guidance of the education governance system. From the perspective of vocational education governance practice, the construction of governance system is often preceded by the construction of governance capacity, and a systematic and perfect education governance system is a prerequisite for improving education governance capacity. Similarly, the enhancement of vocational education’s governance capacity lay the foundation for the better functioning of the education governance system. The benign interaction and common role of the two promote the sustained and healthy development of education, and ultimately achieve modernization of education. Through the scientific construction and continuous optimization of the vocational education governance system, with a reasonable and clear division of responsibilities and a coordinated working mechanism, and giving full play to the unique advantages of the education system and the guiding function of education policies, the modernization of education governance capacity will certainly be able to provide stable and lasting protection.

4.3

Educational governance resources as foundational elements

The modernization of vocational education governance capacity must consolidate the fundamental status of resources and enhance the efficiency of the supply, allocation, integration, and use of vocational education governance resources in all aspects. “Resource allocation is the key to determining whether the modernization of education can be achieved on schedule and whether modern education governance can be realized.” Along with the development of the practice of governance of vocational education, the connotation of educational resources is constantly expanding and enriching, and it is generally believed that “educational resources are the sum of all human resources, material resources and financial resources in the field of education obtained through the allocation of total resources in the society”. Scientific integration and rational allocation of educational resources can provide space for the quality and balanced development of education, and also lay the foundation for consultation and common governance across multiple disciplines. From optimizing the structure of education to deepening the comprehensive reform of education, and from improving the quality of education to safeguarding the fairness of education, the whole process of modernization of educational governance capacity involves the integration and allocation of educational resources, and “the level of allocation of educational resources has become a measure of the The level of distribution of educational resources has become a key indicator of the level of educational governance of the main bodies of governance at all levels.”

4.4

Vocational education governance tools as supporting elements

The diversity, compatibility and effectiveness of the tools chosen in the process of vocational education governance are often the key indicators for measuring the level of education governance and an important basis for evaluating the strength of education governance capacity. The selection of traditional vocational education governance tools is mainly based on short-term governance goals and governance costs, and cannot be separated from the limitations of “education management”, which can easily lead to waste of resources, inefficiency, information occlusion and many other problems. Although this top-down administrative command-based governance tool is fast and efficient, it lacks systematic combination and optimal configuration, and is prone to suppress the function and utility of other tools, thus restricting the modernization of education governance capacity. Modernized governance tools are both an effective supplement to traditional governance tools and a key means of implementing education policies, laying a solid instrumental foundation for the change of education governance paradigms and the formation of education governance patterns. The enhancement of educational governance capacity needs to rely on effective governance tools, while innovative and efficient technological tools can provide strong technical support for the modernization of educational governance capacity by giving full play to the advantages of optimizing resource allocation, unimpeded information exchange and improved governance efficiency.

4.5

Vocational education governance mechanisms as a safeguarding element

Compared with the education governance system as a top-level design, the governance mechanism of vocational education is, in a certain sense, the use of bottom-level logic in the framework of the governance structure to ensure the orderly and standardized development of the education cause, and to guide and regulate the specific governance behaviors of the main body of education governance. The innovation of education governance system and mechanism is essentially the redesign and arrangement of education laws and regulations, institutional norms and code of conduct, the operation and internal mechanism formed in the process of education governance, and the perfect governance mechanism provides guidance and guarantee for the modernization of education governance capacity in terms of measures and programs. A modernized education governance mechanism requires sound and complete systems and regulations, a simple and efficient governance structure, coordinated procedures and processes, and a clear division of powers and responsibilities, so as to further enhance the compatibility and coordination of the education governance system and the work of education governance, and to ensure that the modernization of the capacity for education governance is steadily advanced.

5

Analysis of the configuration conditions for modernizing the governance of vocational education

5.1

Analysis of necessary conditions

In order to elucidate the specific explanatory effect of individual condition variables on the innovation outcomes, the study chose to take the consistency and coverage indicators to measure the degree of necessity of each condition variable on the outcome variable. The individual condition necessity analysis is shown in Table 5 (text indicates that the condition variable exists, ~ text indicates that the condition variable does not exist). In the analysis of influencing factors of innovation results, the consistency of all six types of condition variables is more than 0.5, indicating that the condition variables have a certain explanatory power for the innovation results, and the variable setting is basically reasonable. However, the consistency of individual conditional variables does not exceed 0.9, indicating that there is no single sufficient condition that influences the innovation results. The consistency of organizational leadership support is greater than 0.9, indicating that organizational leadership support is a sufficient condition that constitutes sustainability innovation. Similarly, in the analysis of conditional necessity of unsustainable innovation, the consistency of insufficient investment of educational resources is greater than 0.8, indicating that the lack of necessary educational resources is a sufficient condition that leads to unsustainable innovation. The pathways of educational resource input and organizational leadership support as important variables explaining innovation outcomes need to be verified through further group analysis.

Table 5.

Analysis of individual conditions necessity

Conditional variable	Variable code	Sustainable innovation		Unsustainable innovation
Conditional variable	Variable code	consistency	coverage	consistency	coverage
Creative task properties	T1	0.77687	0.84732	0.67322	0.59689
~ Creative task properties	~ T1	0.53751	0.70848	0.79554	0.70786
Innovative target type	T2	0.69851	0.76125	0.64441	0.57634
~ Innovative target type	~ T2	0.5184	0.68147	0.67747	0.70406
Multiple governance	O1	0.77828	0.82918	0.73316	0.73032
~ Multiple governance	~ O1	0.55959	0.75368	0.76369	0.80121
Organizational leadership support	O2	0.97876	0.96141	0.70312	0.56932
~ Organizational leadership support	~ O2	0.5581	0.69304	0.7925	0.81729
Economic development level	I1	0.74276	0.78653	0.67482	0.58544
~ Economic development level	~ I1	0.51757	0.69851	0.70474	0.74822
Innovation resource input	I2	0.73818	0.86072	0.61456	0.5851
~ Innovation resource input	~ I2	0.5571	0.68009	0.82442	0.78237

5.2

Conditional portfolio pathways for sustainable innovation

Group analysis forms a combined path solution by analyzing the influence of the combination effect among condition variables on the outcome variable. The common QCA analysis provides three combined paths of complex solution, intermediate solution and parsimonious solution, considering the need for comprehensive examination of variables and the universality of intermediate solution, the study refers to the intermediate solution and parsimonious solution results for in-depth interpretation of the data, and the paths of the sustained condition of local education governance innovation are shown in Table 6 (● indicates the presence of core conditions, _● indicates the presence of marginal conditions, and ⊗ indicates the absence of core conditions, _⊗ indicates the absence of marginal conditions, and blank conditions may or may not be present and are irrelevant to the results). First, organization-driven (H1). This grouping has a consistency of 0.8232, a unique coverage of 0.0889, and a raw coverage of 0.1692, covering four typical cases. Under this pathway model, the sustained promotion of innovative actions in education governance around the world relies mainly on the sustained promotion of pluralistic organizations, and less on the support and guarantee of the resource environment. Construct H1 can be expressed as follows: even if a place does not have a high level of economic development, as long as the place is able to clearly define the innovation goal, and through increasing the organizational support and strengthening the input of existing innovation resources, the place is able to achieve the continuous deepening of the innovation action, and thus guarantee the sustainability of innovation. Organizational support, type of innovation goal, and non-high economic development level are the core conditions.

Table 6.

Local education governance innovation continued path

Conditioned configuration	Sustainable innovation			Unsustainable innovation
Conditioned configuration	Combination H1	Combination H2	Combination H3	Combination W1	Combination W2
Creative task properties	●		•		⊗
Innovative target type		•		⊗	⊗
Multiple governance		⊗		⊗
Organizational leadership support	●	●	•		⊗
Economic development level	⊗	●	●	⊗
Innovation resource input	•		⊗	•	●
Primary coverage	0.1692	0.3645	0.2436	0.2599	0.3507
Unique coverage	0.0889	0.2925	0.069	0.2577	0.6494
Consistency	0.8232	0.9176	0.8009	0.8518	0.8331
The coverage of the solution	0.87216			0.72686
Consistency of solutions	0.90985			0.86125

Second, internal and external coupling type (H2). This grouping has a consistency of 0.9176, a unique coverage of 0.2925, and an original coverage of 0.3645, covering six typical cases. In this path model, the local area usually has a good economic environment and organizational capacity as a guarantee, and at the same time, it selects an innovative program that fits the local educational development, and many internal and external resource elements work together as antecedent conditions. Construct H2 can be expressed as follows: if the localities have a good economic foundation, pay close attention to the local educational needs when formulating governance innovation programs, and have strong organizational leadership support, the localities can achieve sustainable innovation in educational governance. Among them, the level of economic development and support for organizational leadership are the core conditions for sustainable innovation. Therefore, this path can be a good reference for some regions that have a high level of economic development and strong organizational capacity.

Third, resource empowerment (H3). This grouping has a consistency of 0.8009, a unique coverage of 0.069, and an original coverage of 0.2436, covering four typical cases. In this path mode, local economic development conditions and innovation resources are more adequate, and in terms of resource stock, it is slightly richer than the “organization-driven” and “internal and external coupling” path modes, but the organizational leadership support is weaker. The core conditions of configuration H3 are: the level of economic development, the input of innovation resources, the support of organizational leadership, and the type of innovation goals. The path can be expressed as follows: with strong economic security and continuous investment in innovation resources, localities can achieve sustainable development of innovation actions under the preconditions of clear innovation development tasks and strengthened organizational leadership.

5.3

Combined analysis of conditions for unsustainable innovation

Social phenomena do not exactly follow the law of causal symmetry, i.e., the emergence or non-emergence of a particular outcome requires different combinations of antecedent conditions to explain it separately. Therefore, in order to explore the causes of the demise of local education governance innovations in more depth, the study further analyzed the conditional groupings of non-sustainable innovations using “unsustainable innovations” as the outcome variable. The data results show that the model has an overall coverage of 0.7269, an overall consistency of 0.8613, and the existence of two combined paths of constructs W1-W2 that lead to the demise of local education governance innovations. Based on the intergroup comparison of the two shared core conditions, the study categorizes the mechanism leading to the demise of governance innovation as an internal and external constraint path, whereby unsustainable innovation is constrained by both the type of innovation goal and the input of innovation resources.

There are three core conditions of the internal and external constrained path, namely: vague innovation objectives, inadequate implementation of governance bodies, and insufficient investment of innovation resources. Among them, construct W1 indicates that when the regional economic development level is general and lacks innovation design, regardless of whether the innovation organization guarantee is fully constructed or not, as long as the regional innovation resources are insufficiently invested and the innovation goal is too macroscopic, the innovation action will be inhibited. Configuration W2 shows that regardless of the level of regional economic development, once the innovation design lacks standardization and the innovation goal is ambiguous, even if the local government invests a lot of resources in innovation, it cannot guarantee the sustainable development of innovation. The representative examples of this configuration are Nanjing and Zhengzhou. Taken together, the core conditions leading to unsustainable innovation configuration are mainly insufficient investment of innovation resources and ambiguous innovation goals.

6

Conclusion

Vocational education governance modernization is the basic content and important foundation of education modernization. The article combines deep learning technology to construct the identification method of key elements of vocational education governance modernization, and carries out the design of vocational education governance modernization path through five dimensions to explore its group effect.

In the experimental test on the public dataset, the recognition model in this paper performed best among all algorithms, with an F1 value of 95.22%. Overall, the proposed model in this paper performs well in the identification process of multiple elements, which is a good guide for the specific application of vocational education governance modernization.

Among the key elements of vocational education governance modernization, vocational education governance concept is the leading element, vocational education governance system is the prerequisite element, vocational education governance resources are the basic element, vocational education governance tools are the supporting element, and vocational education governance mechanism is the guarantee element. Reshaping the concept of vocational education governance, improving the vocational education governance system, optimizing the allocation of vocational education resources, innovating vocational education governance tools, and improving the vocational education governance mechanism are the realistic choices to promote the modernization of vocational education governance capacity.

Idioma:: Inglés

Calendario de la edición:: 1 veces al año
Temas de la revista:: Ciencias de la vida, Ciencias de la vida, otros, Matemáticas, Matemáticas aplicadas, Matemáticas generales, Física, Física, otros

RSS Feed de revista

Identification of Key Elements and Path Design of Vocational Education Governance Modernization Supported by Deep Learning

Fan Li

Publicado en línea: 21 mar 2025

Recibido: 30 oct 2024

Aceptado: 11 feb 2025

DOI: https://doi.org/10.2478/amns-2025-0567

Palabras clave<kwd>BiLSTM-CRF</kwd>, <kwd>Element identification</kwd>, <kwd>Path design</kwd>, <kwd>Vocational education governance modernization</kwd>

© 2025 Fan Li, published by Sciendo

This work is licensed under the Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Palabras clave
<kwd>BiLSTM-CRF</kwd>, <kwd>Element identification</kwd>, <kwd>Path design</kwd>, <kwd>Vocational education governance modernization</kwd>