Research on Agricultural Meteorological Disaster Event Extraction Method Based on Character–Word Fusion

Abstract

Meteorological disasters are significant factors that impact agricultural production. Given the vast volume of online agrometeorological disaster information, accurately and automatically extracting essential details—such as the time, location, and extent of damage—is crucial for understanding disaster mechanisms and evolution, as well as for enhancing disaster prevention capabilities. This paper constructs a comprehensive dataset of agrometeorological disasters in China, providing a robust data foundation and strong support for event extraction tasks. Additionally, we propose a novel model named character and word embedding fusion-based GCN network (CWEF-GCN). This integration of character- and word-level information enhances the model’s ability to better understand and represent text, effectively addressing the challenges of multi-events and argument overlaps in the event extraction process. The experimental results on the agrometeorological disaster dataset indicate that the F1 score of the proposed model is 81.66% for trigger classification and 63.31% for argument classification. Following the extraction of batch agricultural meteorological disaster events, this study analyzes the triggering mechanisms, damage patterns, and disaster response strategies across various disaster types using the extracted event. The findings offer actionable decision-making support for research on agricultural disaster prevention and mitigation.

1. Introduction

An event refers to a specific occurrence involving participants, characterized by changes in state, and represents an action or process that takes place within a defined context [1]. Event extraction aims to identify event types and their corresponding arguments in the text, encompassing two primary subtasks: event detection and event argument extraction [2]. Event detection aims to identify triggers and classify them as event types. Event triggers are specific words or phrases—typically verbs or nouns—that clearly indicate the occurrence of an event and intuitively reflect the associated state changes. Event arguments are the participants or attributes involved in the event [3,4].

Agrometeorological disaster event extraction enables the timely acquisition of critical information during disasters, enhances the development of response strategies, improves the efficiency of information retrieval, and supports precise and informed decision-making for agricultural disaster prevention and mitigation [5]. The rapid evolution of internet technologies has made online resources a major source of information. Consequently, a large amount of agrometeorological disaster information is now presented in unstructured text form, containing valuable insights with significant potential for application.

As illustrated in Figure 1, the event triggers “hit,” “affected,” and “collapsed” are extracted. Based on these triggers and the surrounding context, it is determined that “hit” signals a “Torrential Flood/Disaster-causing event,” while “affected” and “collapsed” indicate a “Torrential Flood/Disaster-impact event.” For the “Torrential Flood/Disaster-causing event,” the argument extraction module identifies the arguments “From May 29 to 30” (duration), “Guizhou Province” (location), and “torrential or heavy rain” (severity). This process transforms unstructured text into a structured format, which facilitates rapid retrieval and deeper analysis, ultimately maximizing resource utilization and enhancing scientific management of agrometeorological disasters. This approach also plays a critical role in downstream tasks such as information retrieval and knowledge graph construction.

However, there are two challenges in extracting agrometeorological disaster events. First, as shown in Figure 1, a single sentence can contain multiple events (e.g., “Torrential Flood/Disaster-causing event” and “Torrential Flood/Disaster-impact event”) that may share common arguments such as “duration” and “location.” Second, argument overlap is evident: as illustrated in Figure 2, the argument “public” serves as the “Recovery target” in the “Torrential Floods/Production recovery” event and as the “Responsible Party” in another event. Additionally, the “Responsible party” argument, “Agricultural technicians from the Luo Ding Municipal Bureau of Agriculture and Rural Affairs”, contains a nested “Location” argument, “Luo Ding”. These complexities increase the difficulty of event extraction, which requires advanced techniques to accurately identify interconnected triggers and arguments within the text.

To address these issues, this study proposes a Chinese event extraction model named character and word embedding fusion-based GCN network (CWEF-GCN). Methods based on Graph Neural Networks (GNNs) have achieved remarkable results in event extraction tasks by modeling structural relationships between words. These approaches primarily rely on dependency parsing to construct sentence graph structures, which reveal dependency relations between words and help capture long-range contextual dependencies. However, current GNN-based event extraction methods suffer from high complexity in graph construction. To capture more diverse relations, complex heterogeneous graphs incorporating multiple types of relational edges are often constructed, which inevitably introduce substantial noise. Moreover, the characteristics of the Chinese language pose specific challenges: using only character-level features leads to word segmentation ambiguity and high complexity, whereas relying solely on word-level features risks error propagation due to segmentation mistakes. To address these issues, this paper proposes a fusion of character- and word-level features, leveraging character features to mitigate segmentation errors and word features to disambiguate character-level meanings. To efficiently align with the aforementioned graph structure and accurately identify triggers and arguments, this paper employs a pointer network for decoding. Traditional sequence labeling-based decoding methods often underperform when handling long sequences and nested arguments. In contrast, pointer decoding effectively addresses issues such as argument overlap and nesting by identifying the start and end positions of triggers and arguments. Furthermore, in the absence of a dataset for agricultural meteorological disaster event extraction, we constructed and annotated a dataset from two sources: the China Meteorological Disaster Yearbook and China News Network. This dataset defines 16 event-type labels and 13 argument roles, providing a valuable resource for evaluating the effectiveness of our algorithm for agricultural meteorological disaster event extraction tasks. The main contributions of this study are as follows:

(1)

We propose a novel graph construction method. Compared to the complex graph structures commonly used in existing studies, this strategy simplifies the graph construction process and reduces model complexity. Furthermore, by leveraging the fusion of character-level and word-level information, the graph convolutional network can learn semantic associations between triggers and arguments more efficiently and accurately.

(2)

A comprehensive agrometeorological disaster dataset is constructed to validate the proposed model. The experimental results show that our model outperforms competitive models.

(3)

We analyze the characteristics of disasters by examining the extracted event triggers and arguments, thereby offering decision-making support for the prevention and mitigation of agricultural disasters.

2. Related Work

Event extraction research has evolved through the eras of pattern matching, machine learning, and deep learning. While early pattern-based and statistical machine learning methods (e.g., Conditional Random Fields, CRFs) laid the groundwork, they relied heavily on manual feature engineering and demonstrated limited generalization. Since then, the field has been dominated by deep learning approaches, which can be categorized into several paradigms. Early deep learning models, such as Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (RNNs), initially approached event extraction through pipeline models or joint learning frameworks. CNN-based models and RNN-based models—including Long Short-Term Memory networks (LSTMs) and Gated Recurrent Units (GRUs) [6,7]—demonstrated the ability to automatically learn features from raw text, overcoming previous limitations. However, a fundamental weakness of these sequence-based architectures is their inability to effectively capture long-range dependencies between event triggers and arguments, which are crucial for accurate event extraction.

Graph-based methods were subsequently introduced to explicitly model the syntactic structure of sentences, effectively addressing the long-distance dependency problem. By constructing graphs over dependency trees, Graph Convolutional Networks (GCNs) [8,9] can propagate information along syntactic paths, leading to significant performance gains. More recent studies have formulated event extraction entirely as a graph parsing problem [10,11]. While powerful, a key drawback of these methods is their tendency to construct overly complex and noisy heterogeneous graphs—incorporating multiple relation types—which can increase computational costs and obscure the most critical connections.

Transformer-based pre-trained language models (PLMs) have revolutionized the field by providing deeply contextualized token representations. Models such as BERT have become the standard encoder, substantially boosting performance [3]. A recent paradigm shift reformulates the task using generative PLMs (e.g., T5, GPT) through prompt-based learning [12,13], demonstrating great promise in low-resource scenarios and in handling complex cases such as argument overlap. Despite their strengths, purely sequence-based PLMs often lack explicit inductive biases for syntactic structure, which can be vital for accurate relational reasoning in event extraction. This limitation has spurred interest in hybrid approaches that integrate PLMs with graph neural networks to synergize contextualized representations with structured reasoning.

Task-specific reformulations include framing event extraction as a question-answering task [14] to address argument overlap. However, these methods often require designing numerous question–answer templates, which can be cumbersome and suboptimal.

Finally, Chinese event extraction presents unique challenges due to the absence of clear word delimiters, which makes trigger identification and word segmentation error-prone. Research in this area often focuses on effectively integrating character-level and word-level information [15,16,17,18] or incorporating external lexical knowledge [19,20] to mitigate these issues. A common challenge remains designing a model architecture that is both effective in capturing the nuances of Chinese and computationally efficient, avoiding the complexity introduced by many graph-based methods.

Our work is positioned at the intersection of these recent advancements. We propose a novel graph construction method that simplifies the complex graphs commonly used in GCN approaches, thereby reducing noise and computational overhead. Additionally, our model integrates a pre-trained Transformer (BERT) as a backbone encoder to generate powerful contextualized features. We specifically design a character–word fusion mechanism to address the unique challenges of Chinese text. This approach enables us to leverage the strengths of PLMs while incorporating a more efficient and focused structural inductive bias through our simplified graph.

3. Materials and Methods

In this section, we propose a novel event extraction model named character and word embedding fusion-based GCN network. As shown in Figure 3, the proposed CWEF-GCN model comprises four components: a sentence encoding layer, dependency feature extraction layer, feature fusion layer, and event decoding layer. First, the sentence encoding layer utilizes the robust encoding capabilities of BERT to capture contextual nuances and generate character-level embeddings from the input text. Subsequently, the dependency feature extraction layer implements a novel graph construction approach based on character and word embedding fusion, which is then processed using a graph convolutional network. This method effectively captures long-range dependencies between event triggers and arguments, uncovers complex semantic relationships, and addresses the challenge of disambiguating argument roles. The feature fusion layer employs a Bidirectional Long Short-Term Memory (BiLSTM) network to capture sequential dependencies within a sentence, facilitating the integration of trigger information with other contextual features. Finally, the event decoding layer utilizes a pointer mechanism to decode event triggers, arguments, event types, and their corresponding roles, thereby ensuring accurate event extraction and role classification.

3.1. Sentence Encoding Layer

In this study, we employ BERT for sentence encoding owing to its ability to capture rich and complex semantic relationships using a bidirectional training approach. BERT encodes each character by simultaneously considering both the preceding and following contextual information, resulting in dynamic word representation. This approach allows BERT to generate distinct vector representations for the same word in different contexts, thereby enhancing the model’s ability to accurately interpret polysemous words and contextually related terms. Given a sentence $S=\{c_1,c_2,c_3,\dots ,c_n\}$, where $c_i$ denotes the ith character, we obtain the corresponding character encoding: token embedding, segment embedding, and position embedding. Token embedding encodes each word’s vocabulary index, segment embedding distinguishes between sentences, and position embedding indicates the relative position of each word in the sentence, guiding the model’s focus during processing. These character encodings are fed into the pre-trained BERT model, which produces semantically enriched vector representations $E=\{e_1,e_2,e_3,\dots ,e_n\}$, as illustrated in Figure 4.

3.2. Dependency Feature Extraction Layer

To comprehensively explore both the semantic and syntactic features of a sentence, we employ GCNs to extract deeper semantic insights. Our approach utilizes dependency trees to propagate information across graphs, thereby enhancing the node representations. Previous approaches to event extraction using graph structures include research by Wu et al. [9], who proposed a word–character graph method. That model involved two sets of nodes derived from dependency parsing: character nodes, $V_c$, and word nodes, $V_w$, derived from dependency parsing. If a dependency existed between nodes $V_{w_{ij}}$ and $V_{w_{kl}}$, edges were constructed between the corresponding character nodes $V_{c_i}$ and $V_{c_j}$ and word nodes $V_{w_{kl}}$. In addition, to integrate sequential features, the characters between nodes $V_{c_i}$ and $V_{c_j}$ were connected. Similarly, Xu et al. [21] employed word-level modeling for Chinese event extraction by constructing word-level graphs using dependency parsing. Liu et al. [22] further developed a character-level dependency tree that linked the internal characters of word nodes via dependency relationships to form edges. While word-based graphs provide a clear structure, they are vulnerable to errors arising from Chinese word segmentation. Conversely, character-based graphs reduce segmentation errors but often introduce substantial noise and create complex graph structures, which increase computational complexity.

In contrast to existing methods, we propose a novel graph construction approach that preserves essential syntactic structures while significantly reducing complexity and avoiding word segmentation errors. Specifically, this approach defines a dependency graph G = (V, E), where $V=\{v_1,v_2,v_3,\dots ,v_n\}$ represents n nodes corresponding to the words in the sentence, and E represents the set of edges between the nodes. We capture the relationships between characters using the embedding vector $E=\{e_1,e_2,e_3,\dots ,e_n\}$ generated from the sentence encoding layer. We also employ the spaCy tool to segment the sentence and derive the word set $W=\{w_1,w_2,w_3,\dots ,w_m\}$. SpaCy is a widely validated, high-performance industrial-strength natural language processing library. Its tokenization and dependency parsing pipelines demonstrate robust and consistent performance across multiple standard benchmark datasets. A word-length mask matrix M, represented as an $m\times n$ matrix, is constructed, where $M_{ij}=1$ indicates that character $c_j$ belongs to word $w_i$, and $M_{ij}=0$ otherwise. The BERT character vector is then multiplied by the mask matrix to obtain the word vector W as follows:

$$W=M\times E$$

(1)

The word vectors are then averaged, resulting in averaged word vectors $\overline{w_i}$, which are used as graph nodes.

$$\overline{w_i}={\displaystyle \frac{1}{l_i}}\sum _{j=1}^n{M}_{ij}{e}_j$$

(2)

where $l_i$ denotes the length of each word. In spaCy, the dependency parser analyzes the syntactic structure of a sentence, which is used to generate a dependency graph for sentences by extracting pairs of words based on specific syntactic dependency paths. If a dependency relationship exists between nodes $v_i$ and $v_j$, a forward syntactic edge $(v_i,v_j)$ is created. To allow bidirectional information flow, reverse syntactic edges $(v_j,v_i)$ are added and self-loop edges $(v_i,v_i)$ are included for each node to capture their own information. For example, in the sentence “Drought occurred in Hebei Province,” dependency parsing reveals that “occurred” is the core word, with a subject–predicate (SBV) relationship to “Hebei Province” and a verb–object (VOB) relationship to “drought.” Consequently, the dependency graph includes forward edges (occurred, Hebei Province) and (occurring, drought) along with the corresponding reverse edges and self-loops. This syntactic information is then encoded into an $n\times n$ adjacency matrix A, where n is the number of nodes in the dependency syntax tree. Finally, the nodes $\overline{W}$ and edges are input into the GCN to obtain word vector $W_1$, which incorporates the dependency features. The calculation formula is as follows:

$$H^{(l+1)}=\sigma \left({\tilde{D}}^{{\displaystyle \frac{1}{2}}}\tilde{A}{\tilde{D}}^{-{\displaystyle \frac{1}{2}}}{H}^l{W}^l\right)$$

(3)

Here, $\tilde{A}=A+I_N$ represents the adjacency matrix of the graph, which includes self-loops for the nodes. $I_N$ denotes the identity matrix, ${\tilde{D}}_{ii}={\sum }_j{\tilde{A}}_{ij}$ the degree matrix, $W^l$ the weight matrix for layer l, and $\sigma$ the activation function. $H^{l+1}$ represents the node feature matrix for layer l + 1, and $H^l$ denotes the node feature matrix for layer l. This graph construction method simplifies the process, while allowing word vectors to capture more semantic information. Because some character-level information contains too much detail, which can distract the model with irrelevant data, this approach minimizes the noise introduced by character-level dependencies. It preserves the relationships between words and maintains the structural integrity of the text, thereby improving the effectiveness of the event extraction. The dependency feature extraction layer is shown in Figure 5.

3.3. Feature Fusion Layer

The GCN model first generates word embeddings $W_1$, which are then transformed into character embeddings using a word-length mask matrix. The transformed character embeddings $C_1$ are then concatenated with the BERT output E along the last dimension, thereby enriching the feature representation. This process produces a concatenated feature vector denoted by $E_1$. The formula for this concatenation is as follows:

$$E_1=[C_1,E]$$

(4)

Subsequently, the fused sequence output $E_1$ is fed into the BiLSTM network. This network consists of two LSTMs: a forward LSTM, which processes the sequence from beginning to end, and a backward LSTM, which processes the sequence from end to beginning. Together, these LSTMs capture bidirectional contextual dependencies within the sentences. The encoding of character $e_i$ by the BiLSTM network is given by:

$${\overrightarrow{p}}_t=\mathrm{LSTM}\left({\overrightarrow{p}}_{t-1},e_t\right)$$

(5)

$$\overleftarrow{{\mathbf{p}}_t}=\mathrm{LSTM}\left(\overleftarrow{{\mathbf{p}}_{t-1}},{\mathbf{e}}_t\right)$$

(6)

After BiLSTM processing, character $e_i$ is represented as a concatenated vector ${\overline{e}}_i=\left[{\overrightarrow{p}}_t,\overleftarrow{p_t}\right]$. This means the forward and backward encodings, ${\overrightarrow{p}}_t\mathrm{and}\overleftarrow{p_t}$, are combined to form a new encoded vector for the sequence, denoted as $\overline{E}=\left({\overline{e}}_1,{\overline{e}}_2,{\overline{e}}_3,\dots ,{\overline{e}}_n\right)$. This process is illustrated in Figure 6.

3.4. Dependency Feature Extraction Layer

3.4.1. Trigger Classification Layer

The BiLSTM model processes a sequence to produce a hidden state sequence, denoted by H. That sequence is then fed into a fully connected layer, which maps it onto the event feature space, generating the event feature. The event features are then averaged along the second dimension to derive sentence-level embedding $S_e$. To enrich the feature representation, we concatenate the generated sentence embedding $S_e$, original event features H, and part-of-speech (POS) embedding along the last dimension to form a comprehensive feature vector. This method enables the model to capture finer sequential information and improve performance. This process is described by the following equations:

$$H=\mathrm{FC}\left(L\right)$$

(7)

$$S_e={\displaystyle \frac{1}{n}}\sum _{i=1}^n{H}_i$$

(8)

$$C=[S_e,H,P]$$

(9)

where FC denotes a fully connected layer. The concatenated feature vector C is then passed through a linear layer to map the input features into the output label space, followed by a sigmoid activation layer to perform the trigger classification. In this study, we adopted pointer decoding rather than the commonly used Conditional Random Field structure to identify the trigger labels. Traditional CRF methods have limitations in handling nested or overlapping arguments, often necessitating complex label designs while still yielding suboptimal performance. In contrast, the pointer network effectively addresses argument overlap by directly and independently predicting the start and end positions of each argument. Additionally, the node representations produced by the GCN module, enriched with structural information, offer high-quality feature inputs for the pointer network. Specifically, the goal of trigger classification is to predict whether a token marks the start or end position of a trigger for a given event type i. By inputting the concatenated feature vector C into a fully connected layer, followed by a sigmoid activation function, the model predicts the start and end positions of the event trigger with probabilities calculated as follows:

$$P_{T_s}^{E_i}\left(c_i\right)=\sigma \left(W_{T_s}^{E_i}{c}_i+b_{T_s}^{E_i}\right)$$

(10)

$$P_{T_e}^{E_i}\left(c_i\right)=\sigma \left(W_{T_e}^{E_i}{c}_i+b_{T_e}^{E_i}\right)$$

(11)

Here, the subscripts s and e represent the start and end of the trigger, $E_i$ denotes the ith event type; $W_{T_s}$ and $W_{T_e}$ are the trainable weights of the binary classifiers for detecting the start and end of the trigger, respectively; and $b_{T_s}$ and $b_{T_e}$ represent the corresponding biases. Thresholds ${\delta }_s$ and ${\delta }_e$ are established for the start and end positions of the trigger, respectively. If $P_{T_s}^{E_i}\left(c_i\right)>{\delta }_s$, the token is identified as the start of a candidate event-type label. Similarly, if $P_{T_e}^{E_i}\left(c_i\right)>{\delta }_e$, the token is identified as the end of the candidate event-type label.

3.4.2. Argument Classification Layer

Once the trigger is identified, a trigger mask matrix T_m is generated, based on the positional information of the trigger. This mask is then applied to the BiLSTM output sequence L to generate the trigger embedding T. The embedding T encapsulates the contextual information of the trigger as follows:

$$T=T_m·L$$

(12)

Trigger embedding T is added to the corresponding elements of the BiLSTM output L to construct the event argument embedding, integrating the trigger information into each token’s embedding:

$$\mathrm{Arg}=T+L$$

(13)

Sentence-level embedding S_a is obtained by averaging the argument embeddings across the first dimension. The sentence-level embedding S_a is then concatenated with the original argument embedding along the last dimension to form a richer representation. Finally, the argument embeddings are passed through linear layers and classifiers to generate an argument classification result for each token.

$$S_a={\displaystyle \frac{1}{n}}\sum _{i=1}^n{\mathrm{Arg}}_i$$

(14)

$$C^a=[S_a,\mathrm{Arg}]$$

(15)

Like trigger classification, argument extraction seeks to predict whether a token represents the start or end of argument role. The model predicts the start and end positions of the argument by passing the concatenated vector C^a through a fully connected layer, followed by a sigmoid activation function, resulting in the following probabilities:

$$P_{A_s}^{R_i}\left(c_i^a\right)=\sigma \left(W_{A_s}^{R_i}{c}_i^a+b_{A_s}^{R_i}\right)$$

(16)

$$P_{A_e}^{R_i}\left(c_i^a\right)=\sigma \left(W_{A_e}^{R_i}{c}_i^a+b_{A_e}^{R_i}\right)$$

(17)

In this case, the subscripts s and e represent the argument’s start and end positions, respectively, R_i refers to the ith argument role, and $W_{A_s}$ and $W_{A_e}$ are the trainable weights for detecting the start and end of the argument, respectively. The bias terms, $b_{A_s}$ and $b_{A_e}$, are also included. Thresholds ${\delta }_s^A$ and ${\delta }_e^A$ are set for the argument’s start and end positions, respectively. If $P_{A_s}^{R_i}\left(c_i^a\right)>{\delta }_s^A$, then the token is identified as the start of the candidate argument role. Similarly, if $P_{A_e}^{R_i}\left(c_i^a\right)>{\delta }_e^A$, then the token is identified as the end of the candidate argument role.

4. Results

4.1. Datasets

To validate the effectiveness of the proposed model, experiments were conducted on two datasets: the publicly available ACE2005 Chinese dataset, and an agrometeorological disaster dataset developed as part of this study. The ACE2005 dataset is a widely used benchmark for information extraction tasks and contains eight major categories and 33 subcategories of event types.

We developed a new agrometeorological disaster dataset by classifying events into three main types—disaster causing, disaster impact, and disaster relief—based on the evolution of agrometeorological disasters, referencing the framework proposed by Ning et al. [23]. The disaster-causing event type primarily captured agrometeorological disasters. This study focused on the most common agrometeorological disasters, identifying four major categories: typhoons, droughts, torrential floods, and low-temperature, frost, and snow disasters. The disaster-impact event type recorded the direct impact and damage caused by disasters on agricultural production. These events integrated agricultural disaster indicators from the national standard, “Natural Disaster Loss Statistics—Part 2: Extended Indicators,” [24] and primarily covered three categories: crop damage, infrastructure damage, and economic loss. The disaster-relief event type pertained to the measures taken to mitigate and respond to agricultural meteorological disasters, drawing from the Ministry of Agriculture’s emergency response plan for major agricultural natural disasters. These events were divided into two main categories: production recovery and post-disaster claims. The production recovery category included material allocation and farmland management, whereas the post-disaster claim category focused on providing financial support and protection to affected farmers. Each event type was associated with specific event argument roles for extraction, covering 4 event types and 16 subtypes, as shown in Appendix A.

The agrometeorological disaster dataset in this study was compiled from two main sources: the China Meteorological Disaster Yearbook, which offers comprehensive records of meteorological disaster events in China, and the China News Network, which added data on disaster relief. The event annotation task involved, for a given sentence, first determining whether it was an event sentence. For event sentences, the process included annotating event triggers and arguments and assigning each argument to its corresponding trigger. Generally, an event sentence must contain a trigger—typically a noun or verb—which can be identified based on the sentence’s core word. Non-event sentences can be categorized by core word types such as declarative, predictive, directive, and descriptive. Declarative words represent facts expressed by others and require further evaluation. Predictive words refer to events that have not yet occurred; sentences containing such words were not annotated. Directive words indicate instructions, guidance, or suggestions and were excluded from the annotation process. Descriptive words objectively state certain situations; although they may express facts, further analysis is needed to determine whether subordinate clauses contain event sentences.

Reports from the China Meteorological Disaster Yearbook and China News Network often encompass all impacts of disasters, including effects on transportation, society, communications, agriculture, economic losses, and more. Since this study focused specifically on agrometeorological disasters—that is, the impact of meteorological disasters on agricultural production—the annotation process prioritized the event arguments defined in this paper. Furthermore, it was stipulated that an event must have a trigger but does not need to include all types of arguments, allowing for overlapping and nested argument structures.

We used the Doccano tool to annotate the data in this study. Doccano is an open-source text-annotation tool. Based on the defined categories of agrometeorological disaster events, we designed sequence annotation tasks using Doccano. Sixteen labels were created for event types and thirteen labels for argument roles. Event types were identified by labeling the triggers in the format “trigger-event type.” Argument role labels were used to identify specific elements associated with each event type. Because sentences often contain multiple events, relational labeling was applied to distinguish between them during the annotation process. The annotated data were exported in the JSON format.

This study successfully annotated 6919 sentences, comprising 4022 events and 9095 argument roles. Among the disaster types, rainstorm accounted for the largest proportion at 61.5%, followed by typhoon (14.1%), low-temperature, frost, and snow disasters (13.7%), and drought (10.7%). Regarding the distribution of argument roles, it was notably uneven due to the fact that not every event contained all argument types and the frequent omission of elements in Chinese expression. Among them, the categories of “affected entity” and “damage” constituted the majority.

4.2. Evaluation Metrics

The performance of the event extraction model was evaluated using the standard metrics precision (P), recall (R), and F1 score, defined as follows:

$$P={\displaystyle \frac{TP}{TP+FP}}\times 100\%$$

(18)

$$R={\displaystyle \frac{TP}{TP+FN}}\times 100\%$$

(19)

$$F1={\displaystyle \frac{2·P·R}{P+R}}\times 100\%$$

(20)

where true positive (TP) denotes the number of instances in which the model correctly identifies positive cases. False positive (FP) refers to number of instances where negative cases are incorrectly classified as positive, whereas false negative (FN) indicates the number of instances where the model fails to recognize actual positive cases. The F1 score represents the harmonic mean of the precision and recall and provides a comprehensive measure of the model’s performance. In the context of event extraction, the evaluation comprised two distinct subtasks, trigger classification and argument classification, each with specific criteria for successful identification. For trigger classification, a trigger was classified correctly if it was correctly identified and assigned to the correct type. In argument classification, an argument was considered correctly classified if it was correctly identified and the predicted role and event subtype matched the ground-truth labels.

4.3. Experimental Parameter Settings

The experiments in this study were conducted using the PyTorch 2.1.1 deep learning framework within the PyCharm Python 3.11 development environment. For encoding, the model utilized BERT, which comprises 12 layers of transformer encoders. The architecture also included a two-layer GCN with an output dimension of 768, followed by a single-layer BiLSTM network with an output dimension of 768. An AdamW optimizer was employed, incorporating weight decay as a regularization technique to reduce overfitting. The detailed parameter settings are listed in

Table 1. Parameter Settings.

Parameter	Parameter Value	Description
Bert_hidden_size	768	Hidden layer dimension of the BERT model
max_seq_len	256	The maximum sequence length is set, with longer sequences truncated and shorter ones padded
Pos_embedding	200	POS encoding dimension
Bert_learning_rate	5 × 10−5	Learning rate of the BERT model
Other_learning_rate	1 × 10−3	Learning rate in non-BERT models
Batch_size	32	Number of samples per gradient descent update

.

4.4. Experimental Results and Analysis

To evaluate the effectiveness of the proposed model, we compared it to several established baseline models.

CAEE: Wu et al. utilized a graph attention network-based framework for Chinese event extraction, enhancing word representations by integrating character-level embeddings into the vocabulary [9].

ATCEE: Liu et al. employed graph attention and table pointer networks to construct character-level syntactic dependency trees for Chinese event extraction [22].

JMCEE: Xu et al. introduced a joint event extraction framework that addressed the issue of argument overlap by defining event relation triples to capture the interdependencies among event triggers, arguments, and argument roles [25].

CasEE: Sheng et al. utilized a cascade decoding joint-learning framework for overlapping event extraction, leveraging dependencies between subtasks through joint learning [26].

PLMEE: Yang. addressed the overlapping argument issue by extracting role-specific arguments based on event triggers [27].

Table 2. Comparison of the overall performance of the models on the ACE2005 dataset.

Models	Trigger Classification (%)			Argument Classification (%)
	P	R	F1	P	R	F1
CAEE	76.11	68.98	72.37	54.12	45.49	49.43
ATCEE	73.40	79.10	76.20	56.50	54.40	55.50
JMCEE	70.90	65.90	68.30	61.60	47.30	53.50
CasEE	71.80	66.70	69.20	52.00	47.30	49.50
PLMEE	70.30	75.20	72.70	54.00	50.70	52.30
CWEF–GCN	75.36	78.79	77.04	50.12	55.03	52.46

and

Table 3. Comparison of the overall performance of the models on agrometeorological disaster dataset.

Models	Trigger Classification (%)			Argument Classification (%)
	P	R	F1	P	R	F1
CasEE	50.20	70.20	58.50	49.30	63.00	55.30
JMCEE	35.60	80.30	49.40	53.30	50.60	51.90
ATCEE	70.35	68.13	72.73	56.97	61.74	52.88
CWEF–GCN	83.60	79.82	81.66	57.16	70.95	63.31

display the performance comparison of the different methods on the ACE2005 and agrometeorological disaster datasets, with the bolded numbers representing the best results. Although our proposed model achieved good F1 scores in trigger classification tasks, it fell short in argument classification compared to ATCEE and JMCEE on the ACE2005 dataset. However, on the agrometeorological disaster dataset, our model outperformed the other methods across most metrics, except for recall, which trailed behind JMCEE. Notably, both CasEE and JMCEE exhibited significantly higher recall than precision. The main reasons for these discrepancies are as follows: CAEE leverages lexical embeddings, providing richer semantic information compared to the tokenization performed by SpaCy in our model. JMCEE employs multiple classifiers to accurately predict trigger start and end positions, leading to fewer false positives (FPs) and higher precision. CWEF-GCN has relatively high recall but low precision on argument classification, probably due to the data imbalance on the ACE2005 dataset. ATCEE introduces a table-filling strategy that enhances information interactions both within and between events, improving the model’s extraction performance for event categories with fewer samples. This approach ultimately leads to higher precision.

In agrometeorological disaster datasets, which often contain multiple events per sentence, CasEE and JMCEE generate samples for each event, thereby expanding the dataset and enhancing semantic feature learning. This approach results in more true positives (TPs) and higher recall but may also lead to increased false positives and lower precision. Compared to CasEE and JMCEE, our model’s precision and recall remained more stable, indicating that GNN-based multi-event extraction effectively facilitates event information interaction and enhances extraction performance.

4.5. Ablation Experiments

To assess the contributions of key model components, we conducted ablation experiments. The results are detailed in

Table 4. Results of ablation experiments with different modules.

Models	Trigger Classification (%)			Argument Classification (%)
	P	R	F1	P	R	F1
w/o character-level fusion	81.00	78.31	79.63	53.43	69.01	60.23
CWEF–GCN	83.60	79.82	81.66	57.16	70.95	63.31

and

Table 5. Results of ablation experiments with different layers.

Layer	Trigger Classification (%)			Argument Classification (%)
	P	R	F1	P	R	F1
1	83.60	79.82	81.66	57.16	70.95	63.31
2	82.84	82.53	82.28	52.10	75.39	61.62
3	78.86	83.13	80.94	48.04	72.93	57.92

, with bold figures indicating the optimal outcomes. The experiments evaluated the impact of excluding character-level information fusion (“w/o character level fusion”) and the different BiLSTM layers.

• Impact of character-level fusion: We explored the impact of combining word vectors from a GCN with BERT character embeddings on performance improvement. Disabling the BERT splicing module and using only GCN-encoded embeddings resulted in a 2.03% decrease in the F1 score for trigger classification and a 3.08% decrease in argument classification. This drop is attributed to BERT’s focus on local semantic and syntactic information, whereas GCN emphasizes global semantic and structural information. The concatenation of these embeddings integrates the features from multiple perspectives, thereby underscoring the importance of character-embedding integration.
• Effect of the different BiLSTM layers: To evaluate the effect of the different BiLSTM layers, we increased the number of layers in the BiLSTM from one to three. The experiments demonstrated that the one-layer BiLSTM network achieved the highest F1 score for argument classification. However, its performance in trigger classification was slightly inferior to that of the two-layer BiLSTM network, although the difference remained within an acceptable range. Additionally, the lower complexity of the one-layer model enhances training efficiency, reduces the risk of overfitting, and enables faster inference speeds in practical applications.

4.6. Analysis of Event Extraction Effect

We applied the proposed model to large-scale event extraction, processing 5653 sentences with an average length of 76.59 characters. The total processing time was 425.48 s, resulting in an average speed of 13.3 sentences per second.

(1)

Analysis of event trigger extraction performance: The model successfully extracted 75% of events. Errors in event trigger recognition were categorized into two main types: recognition errors and boundary errors, with their distribution illustrated in Figure 7.

Recognition errors can be categorized into general recognition errors and event-type recognition errors. General recognition errors occur when the model identifies triggers in sentences that do not contain any events. Event-type recognition errors arise when the model misclassifies the event type, leading to incorrect trigger identification. This issue is particularly challenging in the field of agricultural meteorological disasters, where different event types often share overlapping contexts. The model, which relied on local context windows, struggled to capture global document-level cues, resulting in misclassification of event types. Furthermore, the long-tail distribution of disaster event types contributed to poorer prediction performance for low-frequency events. Boundary errors can be categorized into two types: general boundary errors and multi-event boundary errors. General boundary errors occur when the model identifies only a portion of a trigger word or mistakenly classifies a phrase as a trigger. Multi-event boundary errors arise when the model incorrectly identifies a phrase as a trigger, as illustrated in

Table 6. Examples of event trigger recognition errors.

Error Type	Example
General Recognition Error	Sentence: “Affected by the above four typhoons, the cumulative rainfall during the process in most areas of Hainan is over 250 mm. Specifically, Qiongzhong County recorded 925 mm, Tongshi City 888 mm, and Baoting County 830 mm.”
	Incorrect Identification: Event Type: Typhoon/Disaster-causing; Trigger: Bao
	Correct Identification: Event Type: None
Event-Type Recognition Error	Sentence: “The area of crops affected by the frost disaster in Ningxia Autonomous Region is 344,000 hectares”
	Incorrect Identification: Event Type: Low temperatures, Frosts, Snow/Disaster-causing; Trigger: suffer
	Correct Identification: Event Type: Low temperatures, Frosts, Snow/Disaster-impact; Trigger: suffer from the frost disaster
General Boundary Error	Sentence: “ In September, the precipitation in Inner Mongolia Autonomous Region ranged from 2 to 118 mm, which was less than the same period of previous years. Meteorological drought occurred in most areas. Some areas in Xing’an League suffered from severe drought, and there were local extreme drought conditions”
	Incorrect Identification: Event Type: Drought/Disaster-causing; Trigger: occur
	Correct Identification: Event Type: Drought/Disaster-causing; Trigger: occur
Multi-event Boundary Error	Sentence: “In 21 counties in Anhui Province, including Suixi, Su County, Huaiyuan, Laian, and Quanjiao, disasters were caused by wind and rain. The waterlogged area reached 513,300 hectares, among which 400,000 hectares of crops were damaged. 17,400 houses collapsed, 33,900 houses were damaged, and there were over 140 casualties.”
	Incorrect Identification: Event Type: Torrential Floods/Disaster-impact; Trigger: 17,400 houses collapsed and…were damaged
	Correct Identification: Event Type: Rainstorm Flood/Disaster-impact; Trigger: collapse

, where the correct trigger (“collapse”) was misidentified due to overlapping event boundaries. The pointer decoding mechanism independently predicts the start and end positions, selecting characters with logit values above a specified threshold for the current event type, without considering the joint constraints between the start and end positions. In agricultural meteorological disaster texts, the presence of multiple events increases the probability of boundary overlaps, and shared contexts between adjacent event triggers contribute to feature confusion, resulting in boundary errors. Examples of event recognition errors are presented in Table 6.

To address these issues, this study removed incorrectly recognized events, reprocessed triggers with boundary errors, completed partially recognized triggers, and truncated overly long triggers. A trigger word dictionary and a set of regular expression rules were constructed to facilitate matching, completion, and truncation, ultimately yielding 4685 events.

(2)

Analysis of event argument extraction: The model successfully identified 70% of arguments, with errors categorized into recognition errors, classification errors, and boundary errors. Boundary errors accounted for only 2% of the total errors. This analysis focused on the three argument roles with the highest classification accuracy and the three with the lowest. The model achieved precision rates of 92.23%, 89.38%, and 86.67% for time, location, and disaster severity arguments, respectively. Despite the varied expressions of time, such as “early May to mid-July,” “late July to early August,” “summer,” and “autumn,” the model effectively captured core temporal elements (e.g., year, month, day, hour, minute, second, early/mid/late month, season) and inferred logical relationships to maintain high precision. Location arguments, being clear and semantically explicit, were recognized with even higher accuracy. Disaster severity expressions, which are relatively fixed, facilitated feature extraction by the model. However, the precision rates for damaged subjects, damage conditions, and typhoon names were only 64.96%, 44.11%, and 31.39%, respectively. The recognition precision—defined as the accurate identification of boundaries and classification of argument roles—for damaged subjects and damage conditions reached 97.94% and 95.87%, respectively. The lower classification precision for these roles was primarily due to misclassification, where the model correctly identified the arguments but assigned them to incorrect roles. In sentences containing multiple events, the model may inappropriately assign all arguments to all triggers, particularly in complex sentences with multiple clauses, leading to inaccurate classifications. The low precision for typhoon names was attributed to recognition errors, as the model lacked comprehensive training on diverse expressions of typhoon names and had limited data for this category. To address these issues, misclassified data were removed, and boundary errors were corrected through completion and truncation, ultimately yielding 10,768 event arguments that provided valuable data support for downstream event knowledge graph construction and data analysis.

5. Analysis of Characteristics of Agricultural Meteorological Disaster Events

Based on the previously mentioned event extraction results, this study analyzed the characteristics of agricultural meteorological disasters. Event triggers were utilized to identify the occurrence or changes in the state of events. By examining the high-frequency triggers associated with various types of agricultural meteorological disaster events, the core characteristics of these events could be revealed. This analysis can reflect the damage patterns experienced by affected entities, providing a foundation for developing targeted disaster mitigation measures. Furthermore, it can aid in uncovering semantic patterns related to disasters, offering data support for optimizing models. To achieve this, the study investigated the high-frequency triggers for each event type, with the specific distribution detailed in

Table 7. High-frequency triggers for each event type.

Event Type	High-Frequency Triggers
Torrential floods/disaster causing	Occur, fall, widespread fall, suffer, happen
Droughts/disaster causing	Occur, happen, develop, drought
Typhoons/disaster causing	Landfall, occur
Low-temperature, frost, snow/disaster causing	Occur, suffer, happen, continuous fall
Torrential floods/disaster impact	Affected, collapse, loss, damage, inundated
Droughts/disaster impact	Drought-affected, affected, dry up, loss, run dry
Typhoons/disaster impact	Affected, loss, collapse, damage
Low temperature, frost, snow/disaster impact	Affected, loss, death, frost-damaged, freeze to death
Torrential floods/production recovery	Guide, carry out, repair, issue, rush to repair
Droughts/production recovery	Dispatch, carry out, invest, issue, implement
Typhoons/production recovery	Guide, issue, participate, carry out
Low-temperature, frost, snow/production recovery	Guide, arrange, dispatch
Torrential floods/post-disaster claims	Pay, estimate loss, report loss
Droughts/post-disaster claims	Compensate
Typhoons/post-disaster claims	Pay, compensation
Low-temperature, frost, snow/post-disaster claims	Estimate

.

In the disaster phase, high-frequency words for rainstorms and floods focus on precipitation descriptions (fall/widespread fall) and disaster occurrence (occur/happen), reflecting the direct triggering mechanism of heavy rainfall. For droughts, the high-frequency word “develop” indicates the progressive nature of droughts, highlighting the need to monitor long-term meteorological cumulative effects. The high-frequency word “landfall” for typhoons marks the transition of disaster impact from the ocean to land. For low-temperature, frost, and snow disasters, the high-frequency word “continuous fall” emphasizes the persistent low-temperature process, while “suffer” reflects insufficient preparation for cold weather. In the affected phase, the damage patterns of rainstorms and floods are reflected in infrastructure damage (collapse/damage) and short-term agricultural losses (inundated). Drought damage patterns primarily involve the failure of water infrastructure (dry up/run dry) and direct yield reduction (drought-affected). Typhoon damage patterns mainly include infrastructure damage (collapse/damage), while low-temperature, frost, and snow disasters primarily cause livestock deaths (death/freeze to death) and frost damage to cash crops (frost-damaged). In the recovery phase, post-disaster recovery strategies vary. For engineering recovery, such as “repair” and “rush to repair” for rainstorms and floods, the focus is on infrastructure. For typhoons, “guide” and “participate” emphasize social mobilization capabilities. In terms of resource allocation, “issue” for rainstorms, droughts, and typhoons reflects government directive-driven emergency responses, while “arrange” and “dispatch” for low-temperature disasters indicate localized management. In terms of compensation response speed, “pay” and “compensate” for rainstorms and typhoons reflect rapid economic compensation.

The arguments of agricultural meteorological disaster events can elucidate the entities affected by such disasters. For instance, by examining the primary entities impacted by various disasters, we can gain insights into the key elements of agricultural production that are influenced. This study explored the main affected entities under the categories of rainstorm and flood, drought, typhoon, and low-temperature, frost, and snow events. The entities affected by rainstorms, floods, and typhoons were relatively concentrated. Specifically, the consequences of rainstorms and floods primarily included house collapses, crop damage, direct economic losses, and farmland inundations. Similarly, the damage caused by typhoons predominantly involved house collapses, crop damage, and direct economic losses. In contrast, the entities affected by droughts and low-temperature, frost, and snow events were more dispersed. The entities affected by droughts primarily included crop damage, farmland desiccation, reservoir depletion, direct economic losses, reduced grain yields, and the drying up of rivers. The damage caused by low-temperature, frost, and snow-related disasters typically resulted in crop loss, livestock fatalities due to freezing, direct economic losses, structural collapses, and frost damage to wheat. The distribution of affected entities highlights the varying impacts of different disasters on various sectors during the affected phase, providing clear evidence for targeted rescue and recovery strategies. For instance, in the case of rainstorms, floods, and typhoons, rescue efforts should prioritize repairing damaged homes, reinforcing structures, and salvaging crops. In contrast, drought responses must focus on addressing water shortages and ensuring adequate agricultural irrigation. During low-temperature, frost, and snow events, special attention should be given to protecting livestock from the cold and implementing frost protection measures for crops.

In the future, a knowledge graph can be constructed based on event extraction to visually represent the evolution of disasters, analyze the mechanisms behind their occurrence, and contribute to research on the prevention and mitigation of agricultural meteorological disasters.

6. Discussion

Although the proposed character–word fusion-based GCN event extraction model achieved baseline performance on the agrometeorological disaster task, it still faces challenges such as limited generalization due to data sparsity and insufficient modeling of complex event semantic relationships. Future improvements to the model’s performance may proceed in the following directions: First, to address the data sparsity issue in low-frequency event categories, data augmentation strategies can be employed. Beyond traditional methods such as synonym replacement and syntactic transformation, generative techniques using large language models can be leveraged to design domain-specific prompt templates for generating high-quality training samples that align with the characteristics of agrometeorological disasters. Second, we can integrate prompt engineering with the proposed model. Specifically, the strong semantic understanding capabilities of large language models can be utilized to perform deep semantic encoding of the text, generating representation vectors rich in global semantic information as input to the GCN model, thereby effectively enhancing the ability to model complex events such as nested multi-events and cross-sentence event associations. Third, introducing an external knowledge enhancement mechanism that incorporates entity relations from agricultural and meteorological knowledge graphs can be combined with attention mechanisms to strengthen the model’s comprehension of domain-specific semantics and improve the accuracy and generalization of event extraction.

The current agrometeorological disaster dataset is relatively limited in scope. First, the data sources are fairly homogeneous, relying exclusively on news networks and the China Statistical Yearbook, without supplementation from diverse, multi-source heterogeneous data. This limitation may restrict the model’s generalization ability and robustness. Second, there is a noticeable imbalance in the distribution of event types, which is likely to degrade recognition performance for minority classes. Additionally, the dataset covers a limited range of disaster types and does not encompass all categories of agrometeorological disasters. The dataset boundaries require further clarification, and future work should focus on incorporating additional disaster types to develop a more comprehensive agrometeorological disaster dataset.

7. Conclusions

This paper explored the extraction of agrometeorological disaster events, a technology that efficiently and accurately identifies key disaster information from textual data. It offered valuable data support for real-time disaster response, assessment, and subsequent relief efforts. Additionally, it advanced research on agrometeorological disasters, opening new avenues for investigating disaster impact mechanisms and enhancing capabilities for disaster prevention and mitigation. We addressed the challenges of multi-events and argument overlap in agrometeorological disaster events by proposing a novel joint extraction method based on character and word embedding fusion. We introduced a new graph construction approach that utilized a word-length mask matrix to transform the character embeddings generated by the BERT language model into word embeddings. These embeddings were processed through a graph convolutional network, which facilitated efficient character-to-word transitions and reduced the graph construction complexity. The graph convolutional network effectively modeled complex relationships and semantic information. Our experiments demonstrated the significant performance of CWEF-GCN on both the public ACE2005 dataset and a constructed agricultural meteorological disaster dataset. In the future, the powerful language understanding and generation capabilities of large models will be leveraged for data annotation through prompt learning. By designing prompt templates that integrate extensive knowledge and comprehension abilities acquired during pre-training, these models can generate annotation results. This approach can reduce the manual annotation workload while enhancing both efficiency and quality. Furthermore, future work will concentrate on constructing an event knowledge graph utilizing structured event information obtained from the extraction step. This effort aims to uncover the mechanisms underlying hazard causation, such as spatiotemporal patterns and cascading effects, thereby advancing research on the prevention and mitigation of agricultural meteorological disasters.