Knowledge-driven intelligent recommendation method for emergency plans in water diversion projects

Abbreviations

Expansion

PTMs

pre-trained language models

PTM–ICNN

fusion of PTM and ICNN

BILSTM

bidirectional long short-term memory

CNN

convolutional neural network

GPTs

generative pre-trained transformers

BERT–CRF

fusion of BERT and CRF

ICNN

interactive convolutional neural network

RNNs

recurrent neural networks

CRF

conditional random field

BERT

bidirectional encoder representations from transformers

BERT–BILSTM–CRF

fusion of BERT, BILSTM, and CRF

BERT–CNN

fusion of BERT and CNN

Variables

Meaning

⁠, K, and

word vector matrices

number of attention heads

relation embedding vector

one-dimensional convolutional kernel

feature map obtained after the one-dimensional convolutional calculation

feature alternating matrix

feature map obtained after the two-dimensional convolution calculation

masked keyword sequence

Vector concatenation

trainable weight matrix

count of entity relationships

entity to be retrieved

number of correctly predicted entities

scaling factor

entity embedding vector

convolution operator

vector alternating matrix

two-dimensional convolution kernel

text sequence

extracted text features

embedding matrix for entity relationships

Softmax function

input entity

joint probability distribution of the two random variables

number of predicted entities that are not actual entities

number of entities that were not predicted

Emergency plans serve as critical foundations for engineering emergency response and precise decision-making. Constructing emergency plans as knowledge graphs and enabling intelligent recommendations are the keys to advancing the level of intelligence in emergency management (Kreibich et al. 2022). Currently, emergency plans are predominantly in document form, making it difficult to achieve effective knowledge integration and updates. They suffer from weak knowledge association, limited timeliness, and insufficient support for intelligent decision-making (Liu & Chen 2019; Tang et al. 2022). Hence, investigating the methods for constructing an emergency knowledge graph in water diversion projects and achieving intelligent recommendations of emergency plans is a crucial issue that urgently needs to be addressed in the field of water diversion emergency management.

Traditional methods for constructing knowledge graphs primarily rely on rule-based matching techniques, where textual information is extracted and knowledge is built using logical rules devised by experts (Popovski et al. 2019; Zhu & Zheng 2020). Mykowiecka et al. (2009) combined semantic rules and dictionaries to build a relational knowledge graph between patients and drugs, providing support for medical decision-making and knowledge discovery. Chiticariu et al. (2013) utilized regular expressions and rule templates to construct a biomedical knowledge graph, facilitating automatic recommendation of biomedical knowledge. Milosevic et al. (2019) integrated rule matching and template matching techniques to extract event information from biomedical literature, enabling support for biomedical research. Rule-based methods necessitate frequent rule definition and adjustment, which limits the scalability and expansiveness of knowledge graphs (Bouraoui & Schockaert 2019; Lan et al. 2020).

The knowledge graph construction approach based on deep learning utilizes deep learning models to automatically extract entities and relationships from textual data, offering advantages such as strong scalability and high automation (Ma et al. 2021a, 2021b; Li et al. 2022). Deep learning models can be categorized into individual models and hybrid models (Banan et al. 2020; Fan et al. 2020; Chen et al. 2022). Individual models, such as the one presented by Zeng & Wang (2022), employ recurrent neural networks (RNNs) to establish an intelligent recommendation knowledge graph, thereby augmenting the precision and dependability of recommendation systems. Zhou et al. (2020) have put forth a knowledge graph construction methodology, focusing on the industrial domain, which leverages the power of bidirectional long short-term memory (BILSTM) neural networks. This approach has resulted in notable enhancements in both the efficiency and quality of industrial manufacturing. However, individual models often suffer from poor generalization and low precision in knowledge extraction (Afan et al. 2021; Xiao & Zhang 2021; Fan & Wang 2022).

Hybrid models can effectively address the shortcomings and limitations of individual models. Chang et al. (2022) enhanced the accuracy and reliability of knowledge graph construction by combining conditional random fields (CRFs) with BILSTM. Zhou et al. (2020) addressed the issue of BILSTM's inability to capture spatial dependencies in text by proposing a convolutional neural network (CNN)-based BILSTM model for constructing an industrial process knowledge graph. With the emergence of advanced pre-trained models such as a bidirectional encoder representation from transformers (BERTs) and generative pre-trained transformers (GPTs), scholars have sought to leverage these models to enhance the limitations of deep learning models (Lee & Toutanova 2018; Radford et al. 2018). Meng et al. (2022) developed a knowledge graph for power equipment failures based on BERT–BILSTM–CRF, offering more accurate and reliable support for fault information in the field of electrical power. Zhang et al. (2021) employed BERT–CRF to construct a knowledge graph for the management of chemical hazardous materials. Currently, hybrid models based on pre-training still face issues such as fixed pre-training patterns and low accuracy in domain knowledge extraction.

Building upon the aforementioned discussions, this research proposes an intelligent recommendation approach for emergency plans in water diversion projects, based on knowledge graph, to address issues such as weak knowledge associations, poor timeliness, and insufficient intelligent decision support. The main contributions are as follows:

• (1)

  By employing a pre-trained language model (PTM) with entity masking, the model's ability to identify domain-specific entities is enhanced.

• (2)

  By leveraging matrix-based two-dimensional transformations and feature recombination, the interactive CNN (ICNN) is constructed to enhance the processing capability of complex relationships.

• (3)

  By combining the PTM with ICNN, we construct a PTM–ICNN-based joint extraction method for emergency entity relationships.

• (4)

  Utilizing the mutual information criterion, we achieve intelligent retrieval and recommendation of emergency plans.

The remaining sections of this paper are organized as follows: Section 2 introduces the relevant work of PTM–ICNN. Sections 3 and 4 provide detailed explanations of the construction process of PTM–ICNN and the process of intelligent recommendation of emergency plans. Section 5 presents the experimental design, results, and discussions. Finally, in Section 6, the paper concludes and highlights future research directions.

PTMs aim to acquire extensive linguistic knowledge through self-supervised learning on text corpora (Zhang et al. 2021). Early pre-training language models primarily relied on n-gram and rule-based methods. Cichosz (2018) introduced the Global Vectors model, which accomplishes pre-training by leveraging the statistical co-occurrence information of words and global semantic information. Ma et al. (2021a, 2021b) proposed the Word Vector model, which achieves pre-training by representing words as continuous vectors through self-supervised learning.

With the advancement of deep learning, Han et al. (2021) introduced the transformer model, which employs self-attention mechanism for sequence modeling. The transformer model effectively captures long-range dependencies in text through its self-attention mechanism and has demonstrated significant performance improvements in machine translation tasks. Subsequently, Lee & Toutanova (2018) proposed the BERT model, building upon the foundation of the transformer. The BERT model employs a strategy of pre-training and fine-tuning, engaging in large-scale unsupervised training on vast amounts of data, followed by fine-tuning on downstream tasks. The groundbreaking innovations of the BERT model have yielded outstanding results across multiple natural language processing tasks, sparking widespread scholarly interest in PTMs.

A CNN is a feedforward neural network renowned for its formidable feature extraction capabilities (Yamashita et al. 2018; Tan & Le 2019). The origins of CNNs can be traced back to the 1980s. LeCun et al. (1989) introduced the LeNet model, which was developed for the task of handwritten digit recognition. This model incorporates the structural components of convolution and pooling and incorporates the use of the backpropagation algorithm for training. With the emergence and development of big data, researchers have embarked on exploring deep CNN models. Krizhevsky et al. (2017) introduced the AlexNet model, which achieved a significant breakthrough in the ImageNet image classification competition by employing multiple convolutional layers and fully connected layers.

The initial application of the CNN in natural language processing tasks was primarily focused on text classification. Chen (2015) proposed a text classification model based on the CNN, which utilizes convolution operations on sentences to extract local features. The model has achieved favorable performance on multiple text classification tasks. Subsequently, the CNN has been employed in sequence modeling tasks, such as named entity recognition and part-of-speech tagging. Gehring et al. (2017) proposed a sequence labeling model based on the CNN, which performs convolution and pooling operations on the input sequence to make label predictions.

The emergency plans for water diversion projects encounter challenges involving a plethora of technical terminologies and complexities in entity naming rules. Leveraging PTMs allows the transfer of linguistic knowledge acquired during the pre-training phase to the task of extracting emergency knowledge, thereby significantly enhancing entity recognition accuracy. Conventional PTMs, such as BERT, typically employ a character-level masking strategy during pre-training, where a solitary [MASK] character is utilized to conceal individual characters within the text. The character-level masking strategy employed by BERT falls short in fully concealing the entire entity, leading to biases and erroneous predictions when the model attempts to forecast the masked entity. To facilitate the model's effective extraction of contextual features from the input sequence, this study employs the pre-training strategy of entity masking. By randomly masking certain entities within the input sequence, the model can leverage context to infer the original meaning of the masked positions, thus enhancing its understanding of domain-specific entities and consequently improving the accuracy of entity recognition tasks. The PTM proposed in this study is a self-supervised deep language model that undergoes pre-training on multi-layer transformer structures through a random masking mechanism. It comprises input, feature extraction, and output layers, with the fundamental implementation as follows.

Among them, W represents the weight matrix, signifies the number of attention heads, and ⁠, ⁠, ⁠, respectively, denote the weight parameter matrices for Q, K, and V. To expedite the convergence of the model, this study incorporates residual networks and applies normalization techniques.

Output layer: The output layer of PTM comprises the results of feature extraction performed by 24 layers of transformer encoders and decoders, presented in the form of word vectors. These word vectors encapsulate the rich semantic information of the input text.

Among them, and ⁠, respectively, denote the reorganized matrices after vector alternation and feature alternation.

Among them, represents the dimensionality of the two-dimensional convolutional filter, and denotes the convolution operator. represents the feature map obtained after the two-dimensional convolution computation. Then, utilizing matrix addition, the feature map is fused, and subsequently, through inner product operations, the model achieves its final output. The ICNN model's output consists of the relational features between entities.

Pre-training module: Initialize the text sequence, referred to as ⁠, and apply entity masking to the text sequence using the entity masking mechanism. The masked input sequence is constructed using character embedding vectors, word embedding vectors, and positional embedding vectors. By utilizing vector addition, we integrate the character embedding vectors, word embedding vectors, and positional embedding vectors together as inputs to the model. After multiple layers of bidirectional language encoding, the output of the pre-training phase consists of extracted textual features, denoted as ⁠.

Among them, N represents the count of entity relationships.

Among them, represents the input entity, represents the entity to be retrieved, and represents the joint probability distribution of the two random variables.

This section validates the effectiveness of the proposed method through joint extraction of emergency entity relationships, hyperparametric sensitivity analysis, ablation experiments and case study experiments. Firstly, an explanation is provided regarding the datasets, baseline models, and evaluation indexes utilized in the experiment. Furthermore, the experiment is divided into four groups to achieve various research objectives.

• (1)

  In the experiment of entity relationship joint extraction on emergency plan data for water diversion projects, the PTM–ICNN model was utilized, and the experimental results were compared and analyzed with previous state-of-the-art models to evaluate the effectiveness of the PTM–ICNN model;

• (2)

  Conducting a sensitivity analysis of the hyperparameters aims to validate the model's performance sensitivity to different parameter settings;

• (3)

  By conducting ablation experiments, the impact of entity masking-based pre-training tasks and interactive convolution on the performance of PTM–ICNN can be analyzed;

• (4)

  Conducting a case study experiment on intelligent recommendations for emergency plans, the experiment aims to intricately demonstrate how our proposed method ensures the reliability of plan recommendations.

This study utilizes the comprehensive emergency plans released by provincial and municipal levels, as well as 17 specialized emergency plans from the South-to-North Water Diversion Project in China between 2014 and 2021, as experimental data. The comprehensive emergency plans at the provincial and municipal levels, as well as the specific contingency plans, all originate from the 47 subordinate management offices of the South-to-North Water Diversion Project in China. The provincial and municipal-level emergency plans delineate the unified emergency response procedures for significant risk events within the respective province or city. The process comprises four stages: forecasting and early warning, graded response, emergency handling, and post-event support. There are 17 specialized emergency plans, encompassing major risk events such as floods, fires, traffic accidents, earthquakes, water pollution, and others. These plans outline the emergency response procedures for each respective event. After data pre-processing, a total of 733,041 records were obtained, which included provincial and municipal emergency plans as well as specialized emergency plans. Employing a 5-fold cross-validation method, the emergency plan data was divided into five sets, distributed according to predetermined proportions. Among these, four sets comprising a total of 586,432 records were utilized as the training set, while the remaining set of 146,608 records served as the validation set. The aforementioned steps were repeated five times, with each iteration using a distinct validation set. Ultimately, the average of the five validation results was taken as the model's performance level, thereby assessing the overall effectiveness of the model.

This study selects two widely used models, the individual model, and the hybrid model, as the baseline models. By comparing the two types of models, a more comprehensive understanding of the applicability and reliability of different models in entity relation extraction tasks can be achieved. The individual models comprise FastText (Oral et al. 2020), TextCNN (Bu et al. 2022), and BILSTM (Zhou et al. 2020; Lin et al. 2022). The hybrid models consist of BERT–CNN (Qi et al. 2021) and BERT–BILSTM–CRF (Wang et al. 2021; Meng et al. 2022). Regarding the model's parameter configuration, excessively high or low learning rates can lead to an increase in the model's loss or a decrease in convergence speed. An excessively high or low dropout rate can lead to model overfitting or underfitting. In this study, we ensure the fundamental performance of the model while comprehensively considering the fairness and reliability of the experiments, thereby eliminating the impact of parameter variations on the experimental results. Overall, the parameter configurations of the PTM–ICNN model and the baseline model are presented in Table 1.

Among them, denotes the number of correctly predicted entities, denotes the number of predicted entities that are not actual entities, and denotes the number of entities that were not predicted.

Figure 7(a)–7(c) illustrates the relationship between dropout rate, number of iteration rounds, and evaluation indexes. When the dropout rate is less than 0.2, the model exhibits excellent performance and remains relatively stable. When the dropout rate exceeds 0.2, the model's performance deteriorates and its stability decreases. At a dropout rate of 0.9, the model performs the worst. The experimental results indicate that the dropout rate is a sensitive parameter. Setting a dropout rate too high implies that more neurons are randomly discarded during the training process, leading to a weakening of the network's expressive capability. This could potentially lead to the model encountering the issue of underfitting, where the model fails to adequately fit the data, resulting in a lower accuracy level. Figure 7(d)–7(f) illustrates the relationship between entity relation embedding dimension, number of iteration rounds, and evaluation indexes. The model performs well when the embedding dimension is in the range of 80–200. When the embedding dimension of the model exceeds 200, overfitting occurs, leading to a decline in overall performance. The experimental results also demonstrate that the embedding dimension for entity relations is a sensitive hyperparameter. When the model's embedding dimension is set too large, it implies that each entity feature is represented as a lengthy vector. Embedding with such high dimensionality will escalate the quantity of model parameters, resulting in a more intricate model structure, thereby inducing the issue of overfitting.

To analyze the impact of entity-masked training tasks and interactive convolution on model performance. In this study, two variant models, namely PTM–ICNN–CUT and PTM–ICNN–UNI, were designed based on the PTM–ICNN model. PTM–ICNN–CUT denotes the exclusion of the entity masking pre-training task from the PTM–ICNN model. PTM–ICNN–UNI represents the exclusion of interactive convolution from the model. Table 3 displays the average values of accuracy, recall, and F1 score for the PTM–ICNN model and its variant models in the entity relation extraction task.

According to Table 3, the PTM–ICNN–CUT model demonstrates accuracy, recall, and F1 scores of 0.749, 0.679, and 0.712, respectively. The PTM–ICNN–UNI model achieved accuracy, recall, and F1 scores of 0.896, 0.898, and 0.896, respectively. The PTM–ICNN model achieved accuracy, recall, and F1 scores of 0.909, 0.922, and 0.913, respectively. Compared to the PTM–ICNN–CUT model, the PTM–ICNN model exhibited an improvement of 21.36% in accuracy, 35.79% in recall, and 28.23% in F1 score. The removal of the entity masking pre-training task in the PTM–ICNN–CUT model has resulted in a diminished ability to extract domain entities from emergency plans, leading to a decline in model performance. Compared to PTM–ICNN–UNI, PTM–ICNN demonstrated improvements of 1.50% in accuracy, 2.63% in recall, and 1.93% in F1 score. By removing the interactive convolution from PTM–ICNN–UNI, the model's capability to extract complex relationships has been weakened, resulting in a decline in overall performance.

This study evaluates the reliability of the proposed method by measuring whether the recommended solution set includes the correct emergency plans. Table 4 presents the accuracy, recall, and F1 values for the top 1, 3, and 5 recommended solutions as correct solutions. The table reveals that the accuracy of the correct solution being ranked first in the recommended solution set is 72.32%, while it is 74.30% for the top three and 75.36% for the top five. Considering the current recommendation outcomes in conjunction with the actual conditions of the South-to-North Water Diversion Project, they largely meet the demand for accurate recommendations of emergency plans. For the minority of correct solutions that are not included in the recommended set, the primary causes leading to the failure of recommendations are as follows: (1) Due to potential annotation errors during manual supervision of data labeling, the entity recognition algorithm may inaccurately identify certain entities. (2) The insufficient amount of training data for the entity recognition model resulted in inaccurate identification of relevant entities. (3) The excessive presence of similar entities within certain emergency knowledge results in the correct solution being located beyond the top five recommended options.

In response to the issues of weak knowledge association, poor timeliness, and insufficient intelligent decision support in the emergency plans for water diversion projects. Based on the knowledge graph technology, this paper proposes a knowledge-driven intelligent recommendation method for emergency plans in water diversion projects, yielding the following main conclusions:

• (1)

  Taking into consideration the abundant domain knowledge present in the text of emergency plans for water diversion projects, we propose a pre-training task based on entity masking. Compared to BERT–BILSTM–CRF, PTM–ICNN showed improvements of 2.27% in accuracy, 3.73% in recall, and 3.90% in F1 score for entity extraction, effectively meeting the extraction needs of emergency entities in water diversion projects.

• (2)

  To address the issue of complex relationships in emergency plan texts, an ICNN has been proposed. The experimental results demonstrate that PTM–ICNN achieved accuracy, recall, and F1 score of 0.872, 0.871, and 0.868, respectively, making it the top-performing model. This model exhibits stronger capabilities in handling complex relationships.

• (3)

  The entity relationship triplets are stored and the emergency knowledge graph of the water diversion project is constructed using the Neo4j graph database.

• (4)

  The proposed work is an emergency plan retrieval and recommendation method based on the mutual information criterion, with a recommendation accuracy of 75.36% for the top-5. Considering the actual operational conditions of the South-to-North Water Diversion Project, it can effectively meet the requirements for accurately recommending emergency plans.

However, the approach presented in this paper has not taken into account the multimodal information in emergency plan data, leading to a gap between knowledge graph-based emergency plan recommendations and practical engineering applications. Future work will focus on addressing the challenge of extracting multimodal knowledge and investigating the construction of a multimodal knowledge graph to enhance the effectiveness of our approach.

L.W., X.L., and Y.L. conceptualized the study; L.W. and H.L. did data curation; L.W. and J.L. did formal analysis; Y.L. acquired funds; L.W. and X.L. performed the methodology; L.W., X.L. and Y.L. wrote the original draft; and L.W. and X.L. wrote, reviewed, and edited the article.

This work was supported in part by grants from the National Natural Science Foundation of China (72271091), Projects of Open Cooperation of Henan Academy of Sciences (220901008), and Major Science and Technology Projects of the Ministry of Water Resources (SKS-2022029).

This data was provided by the Ministry of Water Resources of China and the South-to-North Water Diversion Central Line Administration and is available on the official website of the Ministry of Water Resources of China: http://www.mwr.gov.cn/.

The authors declare there is no conflict.