A hybrid extraction model for semantic knowledge discovery of water conservancy big data

Data preprocessing and entity definition

Using web crawlers for data extraction is simple and convenient, but the quality of the data is often low. Therefore, it is necessary to first preprocess the obtained data to remove meaningless parts such as spaces or blank lines. During preprocessing, the text is segmented by sentences, ensuring that no sentence exceeds the maximum input limit of the network. Then, through data annotation, the relationships between the required information and the data are clarified, generating a usable training set.

In the WIEM-DL model, the multilayer perceptron (MLP) plays a crucial role in processing and extracting feature representations from unclassified data. The MLP utilizes layer-by-layer nonlinear transformations to convert input unclassified feature data into higher-level, abstract feature representations. Each layer in the MLP is designed to capture semantic information and patterns at varying levels, gradually extracting more expressive and meaningful features. The final layer of the MLP is a fully connected layer with a SoftMax activation function, which is used for classification and prediction tasks. This layer maps the feature representations to the corresponding sentiment categories or prediction outcomes, outputting a probability distribution via the SoftMax function.

In the WIEM-DL model, the Transformer is primarily utilized for semantic modeling and relationship identification of classified sentiment data. By leveraging the self-attention mechanism, the Transformer models relationships and dependencies within the input sequence, effectively capturing long-range dependencies. Through the computation of self-attention weights, the Transformer determines the level of focus on different positions within the input sequence, enabling a more nuanced understanding of semantic relationships.

The identification of node information is the first and a critical step in text information extraction. This process involves detecting the boundaries of phrases corresponding to nodes and determining their node types. For a given sentence composed of words $W:w_1,w_2,...,w_n$ node information recognition assigns a series of labels $N:N_1,N_2,...,N_n$ from a predefined set of categories $N_i\in \varphi ,\left|\varphi \right|=k$.

In this study, the Begin, Middle, Other, End, Single (BMOES) labeling scheme is utilized: sentences are input through word embedding, while words are input through character embedding. Here, B-X and M-X indicate that a character is at the beginning or middle of an entity X, O represents a non-entity, E-X indicates the end of entity X, and S-X represents a word that itself forms an entity. The specific definitions of node information labels are detailed in Table 1.

The word segmentation method based on word formation was proposed alongside the development of artificial neural networks. This approach treats Chinese word segmentation as a character labeling problem, utilizing machine learning methods for training to achieve segmentation results. For the sequence decoding method used in node information recognition, the WIEM-DL model employs a localized node learning approach. Its overall framework is illustrated in Fig. 4.

The first step is to map each input x_i to a low-dimensional vector h_i through a feedforward neural network (FNN), $h_i=f\left(x_i\right)$, where f is the mapping function of the FNN. Next, calculate the weight ${\omega }_i$ for each input x_i, using the following formula:

(1) $${\omega }_i={\displaystyle \frac{\exp (e_i)}{{\sum }_{j=1}^n\exp (e_j)}.}$$

In which e_i is a score that measures the correlation between h_i and the entire input sequence. It can be computed using methods such as dot product or a multilayer perceptron (MLP). This article adopts the dot product scoring method, expressed as:

(2) $$e_i=h_i^T\cdot u$$

In which u is a learnable vector.

Finally, the output y is calculated by taking the weighted sum of the inputs x_i, as follows:

(3) $$y=\sum _{i=1}^n{\omega }_i\cdot h_i.$$

In which ${\omega }_i$ is the attention weight assigned to each input $x_i$, and $h_i$ is the transformed low-dimensional vector representation of x_i.

In the BERT Embedding layer with integrated sentiment analysis, the attention mechanism can be applied at different levels of public opinion text representation, such as word, sentence, or document level, to capture various levels of sentiment information. The structure of the BERT Embedding layer with integrated sentiment analysis is shown in Fig. 5.

The innovation of the BERT Embedding layer integrated with sentiment analysis lies in its ability to not only use the BERT model as a text representation tool but also introduce elements of sentiment analysis to enhance the model’s understanding and processing of emotions. Specifically, this layer builds on BERT by incorporating additional sentiment tags, which strengthens the model’s ability to capture emotional information. As a result, the text representation generated by BERT becomes more comprehensive by including sentiment-related features, allowing the model to better understand and classify the emotional tone within the text. This integration helps the model achieve a deeper understanding of both the semantic meaning and the emotional nuances of the text, making it particularly effective for sentiment-sensitive tasks such as public opinion extraction, sentiment classification, and emotion-aware information retrieval. WIEM‑DL prepends or appends tokens such as [SENT_POS] and [SENT_NEG] to each paragraph input, guiding BERT’s self‑attention layers to treat sentiment polarity as an explicit context feature. Sentiment scores from lexicons are concatenated with BERT’s wordpiece embeddings. This fusion of data‑driven and lexicon‑driven features supplies predefined polarity signals directly into the embedding space. WIEM‑DL adds a parallel sentiment classification head alongside the primary NER/extraction head. During fine‑tuning, the combined loss $\mathcal{L}=\alpha {\mathcal{L}}_{extraction}+\left(1-\alpha \right){\mathcal{L}}_{sentiment}$ encourages the Transformer layers to internalize both semantic and emotional representations. Inspired by syntax‑aware models, WIEM‑DL supervises intermediate phrase representations with sentiment labels (positive/negative/neutral). This ensures that embeddings capture compositional sentiment semantics, such as negation or contrast clauses.

BiLSTM layer of WIEM-DL model

In the WIEM-DL model, the BiLSTM layer is primarily used to process the textual information of each DOM node. Specifically, this layer takes the text representation of each node as input and maps it to a fixed-length vector. The BiLSTM layer is a type of recurrent neural network that processes sequential data by repeatedly applying the same layer at each time step (Wang et al., 2019b). The key feature of BiLSTM is its ability to capture both past and future context, as it processes the input sequence in two directions: forward and backward. This allows the model to better understand the semantics of a node by leveraging the context of surrounding nodes, rather than just focusing on the node’s own textual information. By incorporating this bidirectional context, the BiLSTM layer enables the model to generate more accurate and comprehensive semantic representations of the text, which is essential for tasks such as information extraction and sentiment analysis in the context of web pages.

In the WIEM-DL model, the BiLSTM layer consists of two LSTM layers: one is the forward LSTM, and the other is the backward LSTM. Both LSTM layers share the same parameters, but they process the input sequence in opposite directions. The forward LSTM processes the sequence from left to right, while the backward LSTM processes the sequence from right to left.

This bidirectional structure allows the model to simultaneously consider both the forward and backward context of each node, capturing the information from both past and future nodes in the sequence. By integrating these two LSTM layers, the BiLSTM layer can generate more comprehensive and context-aware representations, which is crucial for understanding the full semantic meaning of the node’s text and improving the performance of tasks such as information extraction and sentiment analysis.

The output of the BiLSTM layer is a sequence of hidden states for each node, where each hidden state is a vector representing the semantic information of the node. These vectors can then be input into the attention mechanism to compute the importance weight of each node, helping the model better learn the semantic information of key nodes.

The model’s structure includes components such as memory cells, input gates, forget gates, and output gates. These components form a control mechanism that significantly enhances the model’s ability to capture long-range historical information. In the WIEM-DL model, the BiLSTM layer processes the input vectors generated by the BERT Embedding layer. The vector sequence is split into forward and backward LSTMs, and the outputs from these two LSTMs are concatenated in the hidden layer to form a BiLSTM. The final output layer consists of a fully connected layer, and the length of the output vector corresponds to the number of labels in the task.

This design allows the BiLSTM layer to effectively capture both past and future context of the input nodes, enabling more accurate modeling of the semantic information needed for tasks such as sentiment analysis and information extraction. Let the sequence X = (x₁, x₂, . . ., x_s) represent the input sequence, where s denotes the sequence length. In this article, for each word in the sentence, an embedding vector i = [w; c_h; c_a] is constructed, consisting of three parts: the semantic-level features w, the character-level features c_h, and the case-sensitive features c_a. The dimensions of these three parts are denoted as d_w, d_o and d_c.

In which d_w represents the dimension of the semantic-level features w, d_o refers to the dimension of the output-level features c_h , d_c is the dimension of the character-level features c_a. Each of these components contributes a specific aspect to the word’s embedding, allowing the model to learn richer and more nuanced representations of the input sequence. In order to obtain comprehensive and complete feature information, character-level features c_h are introduced. Specifically, the BERT Embedding model is used to input the character embeddings, weighted by emotional attention, into an LSTM. The n output vector of the LSTM represents the character-level feature C_h. This approach allows the model to capture finer-grained information at the character level, which is particularly useful for tasks such as handling out-of-vocabulary words or understanding the morphological structure of words. By integrating this character-level information with the semantic-level features from BERT, the model can create a more robust and nuanced representation of the input sequence. Where the character length of a word is denoted by n, so the total length of the input vector is given by d_i = d_w + d_o + d_c. The detailed formulas for the LSTM are as follows:

(4) $$i_t=\sigma \left(w_{xi}{x}_t+W_{hi}{h}_{t-1}+W_{ci}{c}_{t-1}+b_i\right)$$

(5) $$f_t=\sigma \left(w_{fi}{x}_t+W_{hf}{h}_{t-1}+W_{cf}{c}_{t-1}+b_f\right)$$

(6) $$o_t=f_t{c}_{t-1}+i_t\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{h}\left(W_{xc}{x}_t+W_{hc}{h}_{t-1}+b_c\right)$$

(7) $$c_t=\sigma \left(w_{xo}{x}_t+W_{ho}{h}_{t-1}+W_{co}{c}_{t-1}+b_o\right)$$

(8) $$h_t=o_t\tanh h\left(c_t\right).$$

In which i_t, f_t, o_t, c_t represent the input gate, forget gate, output gate, and memory cell at time t, respectively. The computations for these gates involve weight matrices W_x, W_h, W_c, and bias terms b_i, b_f, b_c, b_o. The activation function is denoted by σ. All these parameters are updated through network training to determine their final values. To simplify the notation, the LSTM model will be represented as LSTM (i_t, h_t−1) in the subsequent text.

The BiLSTM layer is a deep learning model commonly used for processing sequential data, achieving excellent results in natural language processing tasks. Compared to traditional unidirectional LSTM, BiLSTM considers the contextual information for each word, utilizing the words on both the left and right sides of the current word to predict its label. Specifically, BiLSTM consists of a forward LSTM and a backward LSTM, which compute the hidden states of each word from left to right and from right to left, respectively. The two hidden states are then concatenated to form the vector representation of the current word. This approach enables the model to capture contextual information more effectively (Wang, Xue & Zhang, 2021). In named entity recognition (NER) tasks, a BiLSTM can pass the vector output of each word as input to a conditional random field (CRF) layer to achieve more accurate sequence labeling. The CRF layer models all possible labeling sequences by calculating scores for each potential sequence based on the given input. It then outputs the labeling sequence with the highest probability. During training, the CRF layer uses the backpropagation algorithm to update model parameters, minimizing the gap between the model output and the ground truth labels.

The BiLSTM layers in the WIEM-DL model can be stacked to further enhance the model’s performance. Specifically, in the WIEM-DL model, the BiLSTM consists of two stacked layers, each containing 128 hidden units. These hidden units capture the relationships between the current word and its surrounding context, generating vector representations of the words for subsequent use by the CRF layer (Zhang, Xue & Zhang, 2019). The primary purpose of the LSTM model is to extract latent features from sequences. The extraction model can obtain embeddings for all words in the embedding layer, and these embeddings are ordered to form a sequence I = (i₁, . . ., i_t, . . ., i_s), which serves as the input to the LSTM, where $i_t\in {\mathbb{R}}^{dt}$. In the LSTM model, there are two separate LSTM modules: one operates forward and the other backward, extracting hidden features $\overrightarrow{h_t}$ and $\stackrel{\leftharpoonup }{h_t}$, respectively. The bidirectional LSTM can be described using the following equations:

(9) $$\overrightarrow{h_t},\overrightarrow{c_t}=LSTM(i_{t,}{\overrightarrow{h}}_{t-1},{\overrightarrow{c}}_{t-1}),\stackrel{\leftharpoonup }{h_t},\stackrel{\leftharpoonup }{c_t}=LSTM(i_{t,}\stackrel{\leftharpoonup }{h_{t-1}},\stackrel{\leftharpoonup }{c_{t-1}}).$$

The output of the LSTM is a concatenation of the forward and backward hidden layer features $h_t=\left[\overrightarrow{h_t},\stackrel{\leftharpoonup }{h_t}\right]$. Subsequently, an activation function is applied to extract features from $\overrightarrow{h_t}$ and $\stackrel{\leftharpoonup }{h_t}$, and a linear layer maps the hidden features to $d_l$ dimensional space. Where $d_l$ represents the number of entity types to be recognized. W_l and W_t are the feature matrices, and b_t and b_l are the bias terms. Since the BiLSTM layer can capture the contextual information of words in a sentence, it performs exceptionally well in sentiment analysis. The innovation of the BiLSTM layer in the WIEM-DL model lies primarily in its use of a bidirectional LSTM structure, coupled with the integration of an attention mechanism to enhance the precision of sentiment classification.

By leveraging the bidirectional flow of information, the BiLSTM layer ensures that both preceding and succeeding contexts are considered. The attention mechanism further refines this process by focusing on the most relevant parts of the input sequence, dynamically weighting the importance of words based on their contribution to the sentiment classification task. This combination allows the WIEM-DL model to achieve a deeper understanding of the emotional and semantic nuances present in the text. Traditional LSTM considers only unidirectional information flow, focusing on the influence of preceding words on the current word. In contrast, BiLSTM incorporates bidirectional information flow, capturing contextual information from both preceding and succeeding words. This bidirectional structure enables the model to more accurately grasp key semantic information critical to sentiment analysis.

The attention mechanism enhances this capability by learning the significance of different parts of the text and applying weighted calculations, improving the precision of sentiment classification. It dynamically assigns weights to various positions in the input sequence, allowing the model to focus on information most relevant to the classification task. This approach enhances the model’s robustness and generalization ability. Additionally, the BiLSTM layer in the WIEM-DL model integrates batch normalization, which accelerates training, mitigates issues such as vanishing and exploding gradients, and contributes to the model’s overall stability and performance.

CRF layer in WIEM-DL model

Under the condition of a given set of input random variables, the CRF layer models the conditional probability distribution of the output random variables. In the WIEM-DL model, the CRF layer is used for sequence labeling of sentiment labels in text sequences. This layer takes the output of the BiLSTM layer as input and uses a series of parameterized functions to map the feature vectors at each position in the input sequence to corresponding label scores. The CRF layer then performs global normalization to calculate the probability of each label sequence and optimizes the label sequence based on these probabilities, ensuring the best match with the observed sequence. Additionally, the CRF layer incorporates an attention mechanism, which automatically learns the important information in the text, thereby improving the accuracy of sentiment analysis and event extraction. Specifically, the CRF layer effectively captures the contextual information within a sequence because it considers the entire label sequence simultaneously. Compared to traditional classifiers, the CRF layer can model the dependencies between labels while ensuring the relevance of the output labels, thereby improving the accuracy of sentiment classification. In the WIEM-DL model, the combination of the CRF layer and the BiLSTM layer achieves better performance in sentiment classification tasks compared to traditional methods (Zhang, Xue & Zhang, 2018). The linear-chain CRF is a commonly used model in sequence labeling problems, as shown in the brief structural diagram in Fig. 6. In sequence labeling tasks, the labeled sequence and the sequence to be labeled correspond to the state sequence variables x and y, respectively.

Given a known training dataset, the model can derive a conditional probability model for the training data using the maximum likelihood estimation (MLE) algorithm. When given an input sequence y to label new data, the model determines the output x* by maximizing the conditional probability P(x|y), where x* represents the model output that maximizes the conditional probability. CRF outperform other models in capturing the dependencies between labels, which improves the accuracy of entity recognition. Specifically, in the model, the output of the LSTM layer is combined into a sequence L = [l₁, l₂, . . ., l_s−1, l_s] ^T, which serves as the input to the CRF model. Where L has dimensions of s × d_l, L_i,j represents the probability that the i word in the sentence is predicted to be labeled as the j tag. For a predicted sequence y = (y₁, y₂, ..., y_s−1, y_s), the score function of the CRF model is defined as follows:

(10) $$f\left(X,y\right)=\sum _{i=0}^{s+1}{T}_{yi,yi+1}+\sum _{i=1}^s{L}_{i,yi}.$$

The transition matrix T is used to represent the transition probabilities between labels, where T_i,j denotes the probability of transitioning from the i label to the j label. The labels for the start and end positions are represented by y_₀ and y_s+1, respectively. The final output sequence y probability is computed by the SoftMax normalization function in the joint extraction model.

(11) $$p(y|x)={\displaystyle \frac{1}{Z\left(x\right)}\exp \left(\sum _{i=1}^T\sum _{k=1}^K{w}_k{f}_k\left(y_{t-1},y_t,x,t\right)\right).}$$

In which Z(x) is the normalization factor, which ensures that the sum of the conditional probabilities of all possible label sequences equals 1. w_k is the weight of the feature function f_k, where $f_k\left(y_{t-1},y_t,x,t\right)$ represents the combination of features at time step t, involving the label y_t, its previous label y_t-1, and the input x. The model parameters w_k can be optimized using algorithms such as stochastic gradient descent. The final trained model can then be used to predict new label sequences by solving the following:

(12) $$\hat{y}=argma{x}_y(p(x|X)).$$

In which $\hat{y}$ represents the label sequence that maximizes the conditional probability $p(x|X)$, where xxx is the input sequence and x represents the model parameters.

The CRF layer in the WIEM-DL model is applied in the field of online public opinion analysis. It performs sentiment analysis and event extraction on online texts, enabling real-time monitoring and analysis of large-scale data from platforms such as social media. This helps users stay informed about the public’s attitudes and emotional tendencies toward events and topics. The detailed description of the WIEM-DL algorithm (Algorithm 1) is as follows.

Evaluation indicators

The text classification extraction model uses accuracy (Precision), recall (Recall), and F-measure (F1) to evaluate the performance of webpage content extraction. The evaluation formulas are as follows:

(13) $$Pre={\displaystyle \frac{\left|S_e\cap S_t\right|}{\left|S_e\right|}}$$

(14) $$Rec={\displaystyle \frac{\left|S_e\cap S_t\right|}{\left|S_t\right|}}$$

(15) $$F1={\displaystyle \frac{2\ast Pre\ast \mathrm{Re}c}{Pre+\mathrm{Re}c}.}$$

In which $S_e$ represents the extracted results, and $S_t$ represents the manually annotated results.

Model implementation and training parameter settings

In the Python programming environment, the algorithm was implemented using the Torch framework. The dataset consists of 5,000 sentences, with the training and validation sets split at a ratio of dev_split_size = 0.1. The maximum vocabulary size is set to max_vocab_size = 1,000,000, and the maximum sentence length is limited to max_len = 500. The learning rate is set to learning_rate = 1e−5, with a weight decay of weight_decay = 0.01. Gradient clipping is applied with a value of clip_grad = 5. The batch size is batch_size = 4, and the model is trained for epoch_num = 20 epochs, with an early stopping patience of patience_num = 4. The Adam optimizer was chosen, and the model was trained on the CPU for 20 epochs. After completing the model training, the model can be applied to web information extraction tasks. By using this model, we can extract target data from web pages and generate corresponding entity information. The application of these entity information can be referenced in Fig. 2, which demonstrates an example of entity information generated by the extraction model. In this way, unstructured web information can be transformed into structured entity data, achieving the integration between structured and unstructured data. Below is a detailed justification for each hyperparameter choice:

• (1)

  A learning rate of 1e−5 is commonly recommended for fine-tuning large pre-trained models like BERT, especially when dealing with smaller datasets. This conservative rate helps prevent catastrophic forgetting and ensures stable convergence during training.

• (2)

  Applying a weight decay of 0.01 serves as a regularization technique to prevent overfitting by penalizing large weights. This value is standard in fine-tuning scenarios and aligns with practices observed in training Transformer models.

• (3)

  Gradient clipping at a norm of 5.0 is employed to mitigate the risk of exploding gradients, which can destabilize training. This technique is particularly beneficial when training deep networks or when using small batch sizes.

• (4)

  A batch size of 4 is chosen to accommodate memory limitations inherent in CPU-only training environments. While smaller batch sizes can introduce more noise into the training process, they are often necessary when computational resources are limited.

• (5)

  Setting the maximum sentence length to 500 tokens allows the model to capture sufficient context from web page paragraphs, which often contain longer sequences of text. This length balances the need for contextual information with the computational constraints of the training environment.

• (6)

  A large vocabulary size of 1,000,000 ensures comprehensive coverage of diverse terms found in web data, including domain-specific jargon and colloquial expressions. This extensive vocabulary aids in accurate tokenization and representation of input text.

• (7)

  Allocating 10% of the dataset for validation is a standard practice that provides a sufficient sample to evaluate model performance and generalization without significantly reducing the training data size.

• (8)

  Training the model for up to 20 epochs with an early stopping patience of four epochs helps prevent overfitting. Early stopping monitors validation performance and halts training when improvements plateau, which is especially important when working with limited data.

• (9)

  The Adam optimizer is selected for its adaptive learning rate capabilities and efficiency in handling sparse gradients, making it well-suited for training deep learning models like BERT.

The hyperparameter choices for the WIEM-DL model are informed by best practices in training Transformer-based models under resource constraints. These settings aim to balance model performance with computational efficiency, ensuring effective learning from the available data.

Comparative analysis of experimental results

The comparison methods in this study are BERT-CRF (Ma & Tian, 2020) and BiLSTM-CRF (Chen et al., 2019) models. The experimental performance of information extraction is evaluated using three metrics: precision (P), recall (R), and F1 score (F1), to verify the effectiveness of the proposed model. In the experiment, 10 manually labeled news pages, sourced from three different media platforms—Sina Weibo, Sina News, and Toutiao—are used as the training set. The other 20,000 pages from the same websites are used as the test set to validate the model’s performance. The overall results of the experiment are shown in Tables 2–4. Table 2 is experimental training dataset,this table outlines the composition of the training dataset used for the WIEM-DL model. Table 3 is training results of information extraction models for three types of websites, this table presents the performance metrics of the WIEM-DL model compared to baseline models across the three platforms. Table 4 is transfer training results of information extraction models for two types of websites, this table evaluates the transfer learning capabilities of the WIEM-DL model when trained on one platform and tested on another.

While the WIEM-DL model demonstrates superior performance over baseline models like BiLSTM-CRF and BERT-CRF, it’s crucial to ascertain whether these improvements are statistically significant. Incorporating confidence intervals (CIs) for evaluation metrics such as precision, recall, and F1 score can provide insights into the reliability of the results. For instance, employing bootstrapping methods to compute 95% CIs can offer a clearer picture of the model’s performance variability. This approach involves resampling the test data multiple times and calculating the metrics for each sample, thus estimating the range within which the true performance metrics lie with 95% confidence.

From the analysis of the experimental results, we can conclude that the WIEM-DL model achieves an overall extraction accuracy of over 92% for web pages, outperforming both BiLSTM-CRF and BERT-CRF. Specifically, the F1 score for BiLSTM-CRF is 90.64%, while that for BERT-CRF is 89.69%. Except for a few pages with special text formats, the model successfully extracts information from the majority of pages. The detailed performance can be seen in Table 3.

• (1)

  In terms of the F1 evaluation metric, it is evident that the proposed WIEM-DL model achieves the best performance on most datasets. Notably, the model attains the highest F1 score of 95% on the dataset sourced from Sina News. The average F1 score across all datasets exceeds 93%. These experimental results demonstrate that the comprehensive capability of WIEM-DL significantly surpasses that of the other comparative methods.

• (2)

  By comparing the Weibo and news datasets, it can be observed that datasets with longer content tend to achieve higher F1 scores. Unlike the news dataset, the Weibo dataset primarily consists of short texts with relatively higher noise levels. Training the model on such datasets makes it challenging for the model to accurately pinpoint the relevant content, which subsequently reduces its performance.

• (3)

  In comparison with the BERT-CRF and BiLSTM-CRF models, the WIEM-DL model demonstrates superior performance in completing information extraction tasks regardless of text length. It consistently retrieves accurate content information. These comparative results validate the effectiveness of the proposed model and highlight that leveraging title-based content localization can significantly enhance the accuracy and robustness of extraction models.

• (4)

  Finally, analyzing the experimental results from the perspective of accuracy, the WIEM-DL model achieved the highest accuracy compared to the BERT-CRF and BiLSTM-CRF models. Specifically, it achieved an accuracy of 93% on the Sina Weibo and Sina News datasets. This high level of accuracy demonstrates the proposed model’s strong adaptability and its potential for superior performance in open information extraction tasks.

While WIEM-DL integrates components like BERT embeddings, BiLSTM, and CRF layers, similar architectures have been explored in prior research. Emphasize any novel aspects of the model, such as innovative integration techniques, unique training strategies, or application to new domains. Compare WIEM-DL’s performance with a wider array of models, including recent state-of-the-art approaches, to contextualize its effectiveness. Conducting ablation studies is vital to understand the contribution of each component within the WIEM-DL architecture. By systematically removing or altering components like the BERT embeddings, BiLSTM layer, or CRF layer, and observing the impact on performance, one can identify which elements are most critical to the model’s success.

To isolate the impact of WIEM‑DL’s components—BERT embeddings, sentiment‑aware tagging, BiLSTM, attention, CRF decoding, and ontology integration—we propose systematic ablations. Each study variant removes or replaces one element, measuring resultant drops in precision, recall, and F1. Paired bootstrap tests and 95% confidence intervals will assess statistical significance. This approach follows established ablation methodologies in NLP and information extraction research. Experimental setup as:

• (1)

  Use the established 20,000‑page test set, retaining the original train/validation split.

• (2)

  For each variant, compute precision, recall, and F1, employ paired bootstrap (10,000 resamples) to derive 95% confidence intervals for F1, Apply McNemar’s test for pairwise significance (α = 0.05) between Full WIEM‑DL and each ablation.

• (3)

  Ensure comparable training regimes (same hy vperparameters, epochs, and early stopping), Report mean ± CI for each metric.

A large F1 drop when removing sentiment tags indicates high reliance on emotional cues. If the 95% CIs for Full WIEM‑DL’s F1 do not overlap with those of a variant, the component’s effect is statistically significant. Qualitatively examine extraction failures per variant to understand error patterns.

To facilitate data analysis, the statistical results shown in Tables 2–4 were visualized using graphical representations. The final results are illustrated in Figs. 7 to 10. These figures provide a clear comparison of performance metrics, including precision, recall, and F1 score, across different datasets and models, making it easier to evaluate the effectiveness of the WIEM-DL model. Figure 7 presents a comparative analysis of the WIEM-DL model against baseline models such as BERT-CRF and BiLSTM-CRF, focusing on key performance metrics: precision, recall, and F1 score. Figure 8 delves into the performance of the WIEM-DL model across different datasets, such as Sina News, Weibo, and Toutiao. Figure 9 assesses the WIEM-DL model’s ability to generalize across different websites, evaluating its performance when trained on one platform and tested on another. Figure 10 presents the results of ablation studies, where specific components of the WIEM-DL model are systematically removed to assess their individual contributions to overall performance.

Analysis of cross website information extraction results

In this section, we will analyses the transfer capability of the WIEM-DL model in the task of cross-site web page information extraction through specific experimental design. This experiment uses a small amount of initial labelled data (such as Weibo, Sina News, and Toutiao) to train a cross-site information extraction model that can handle data from different websites. Although these websites contain similar target information (e.g., public opinion topics, sentiment analysis, etc.), their web structures differ significantly. Therefore, the model needs to have strong transfer learning ability to apply the knowledge learned from one website to others. Implementing bootstrapping methods to compute 95% CIs for evaluation metrics (precision, recall, F1 score) can provide insights into the reliability of the results. This involves resampling the test data multiple times and calculating the metrics for each sample, thus estimating the range within which the true performance metrics lie with 95% confidence. While WIEM-DL integrates components like BERT embeddings, BiLSTM, and CRF layers, similar architectures have been explored in prior research. Compare WIEM-DL’s performance with a wider array of models, including recent state-of-the-art approaches, to contextualize its effectiveness.