Method and device for extracting argument of rainstorm disaster event based on event text enhancement

Abstract

The invention discloses a rainstorm disaster event argument extraction method and device based on event text enhancement, and the method comprises the steps: constructing a rainstorm disaster event template, and defining an event type and an argument role; extracting data related to the rainstorm disaster from the microblog platform; preprocessing the data; labeling the preprocessed data to obtain a data set; generating a problem template; performing event text enhancement; integrating the problem template and the enhanced event text as model input; encoding the integrated text by using a pre-trained Chine-BERT-WWM model, and obtaining a hidden layer representation; deep features are extracted through an intermediate linear layer, and starting and ending positions of an input sequence are predicted by using a full connection layer; according to the method, the rich MRC data set can be directly utilized to enhance event extraction, the problem of data scarcity is relieved, and possibility is provided for zero sample event extraction of invisible event types.

Description

Technical Field

[0001] The present invention belongs to the technical field of natural language processing, and in particular relates to a method and device for extracting arguments of a rainstorm disaster event based on event text enhancement. Background Art

[0002] As the frequency and severity of rainstorm events continue to increase, the urgency of effective disaster management becomes more prominent. Traditional methods face huge challenges in information timeliness, rapid response, and resource coordination. Traditional data sources often fail to provide real-time and detailed insights in emergency situations, which makes it particularly important to extract relevant information from unstructured rainstorm disaster event texts.

[0003] Weibo has become a key source of data on rainstorm disaster events due to its wide coverage and real-time update capabilities. Compared with official reports, Weibo data can instantly reflect the latest situation of disaster victims. Its immediacy and extensiveness provide unique advantages for extracting structured event information from massive unstructured text. Event extraction can automatically identify and organize event information in unstructured text, convert it into actionable structured data, and provide decision support for emergency management.

[0004] Event extraction usually consists of two subtasks: event detection and event argument extraction. Event detection aims to identify trigger words in text and classify them into event types, while event argument extraction identifies event arguments and divides argument roles. Despite a lot of progress, classification-based methods still have problems with data scarcity and inability to handle new event types. Summary of the invention

[0005] In order to solve the above technical problems, the present invention provides a method and device for extracting rainstorm disaster event arguments based on event text enhancement, introduces a new event extraction learning paradigm, and regards event extraction as a machine reading comprehension (MRC) problem. This method uses the latest progress of MRC to significantly enhance the reasoning ability of the model, and alleviates the problem of data scarcity through cross-domain data enhancement, opening up a new situation for zero-sample event extraction of invisible event types, and can directly identify relevant event information from unseen data.

[0006] To achieve the above purpose, the technical solution adopted by the present invention is as follows: A method for extracting arguments of rainstorm disaster events based on event text enhancement, the method comprising: Construct event templates to define rainstorm disaster event types and corresponding argument roles; Obtain and pre-process rainstorm disaster event data from the Internet to obtain valid rainstorm disaster event text; According to the constructed event template, the event type and argument role of the valid rainstorm disaster event text are annotated to build a standardized data set; Structuring the event types and argument roles in the standardized data set to generate question templates; Performing an event text enhancement operation on the valid rainstorm disaster event text to obtain an enhanced event text; Integrate the question template and the enhanced event text into a unified input format, encode the integrated data after obtaining it, and convert the integrated data into a high-dimensional representation vector; The high-dimensional representation vector is linearly transformed to extract deep features and convert text data into structured event information.

[0007] Furthermore, the rainstorm disaster event type includes at least one of waterlogging, vegetation loss, traffic problems, rescue operations, personal safety, landslides, and building collapses, and the argument roles include at least one of time and place.

[0008] Furthermore, the obtaining and preprocessing of rainstorm disaster event data from the Internet includes applying a network crawler framework to obtain the rainstorm disaster event data, and using duplicate data deletion technology, a keyword-based template filtering method and a dependency syntax analysis technology to preprocess the rainstorm disaster event data.

[0009] Furthermore, the question template is generated using a prompt word template extracted from prior knowledge based on the event type and corresponding event pattern determined by the event detection subtask.

[0010] Furthermore, the event text enhancement operation includes adding special delimiters before and after event trigger words to improve the model's understanding of the event text.

[0011] Furthermore, the encoding after obtaining the integrated data includes encoding the integrated data through a pre-trained BERT model to convert the integrated data into a high-dimensional representation vector.

[0012] Furthermore, the linear transformation of the high-dimensional representation vector, extraction of deep features, and conversion of text data into structured event information includes linear transformation of the high-dimensional representation vector through an intermediate linear layer to obtain deep features, inputting the deep features into a fully connected layer, and obtaining predictions of the starting position and the ending position of the input sequence.

[0013] On the other hand, the present invention provides a rainstorm disaster event argument extraction device based on event text enhancement, comprising: The event template construction module is used to construct event templates and define the types of rainstorm disaster events and the corresponding argument roles; The data acquisition module is used to obtain and pre-process the rainstorm disaster event data from the Internet to obtain valid rainstorm disaster event text; The dataset construction module is used to label the event type and argument role of the valid rainstorm disaster event text according to the constructed event template, and construct a standardized dataset; The question template generation module is used to structure the event types and argument roles in the standardized data set to generate question templates; An event text enhancement module, used for performing an event text enhancement operation on the valid rainstorm disaster event text to obtain an enhanced event text; A feature extraction module, used to integrate the question template and the enhanced event text into a unified input format, encode the integrated data after obtaining it, and convert the integrated data into a high-dimensional representation vector; The location prediction module is used to perform linear transformation on the high-dimensional representation vector, extract deep features, and convert text data into structured event information.

[0014] In a third aspect, the present invention provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method for extracting arguments of rainstorm disaster events based on event text enhancement.

[0015] In a fourth aspect, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enables the processor to implement the aforementioned method for extracting arguments of rainstorm disaster events based on event text enhancement.

[0016] The beneficial effects of the present invention are: (1) This application proposes a trigger-based machine reading comprehension (MRC) framework ETEN_BERT_QA, which enhances the semantic understanding of event text by combining event trigger words, thereby improving the accuracy and completeness of event argument extraction. Compared with the traditional BERT-based model, the method of this application has achieved significant improvements in event extraction tasks, especially in handling the complexity and noise of social media data.

[0017] (2) This application overcomes the limitations of traditional methods in generating questions describing event scenarios through an unsupervised question generation process. This process can generate context-related questions, which not only improves the model's ability to recognize event details, but also better copes with insufficient training data. Unlike supervised learning methods that rely on a large amount of labeled data, the method of this application can directly use rich MRC datasets to enhance event extraction, alleviate the problem of data scarcity, and provide the possibility for zero-shot event extraction of invisible event types.

[0018] (3) This application constructs a special dataset RainEE based on the characteristics of rainstorm disaster events, and designs event extraction templates that are suitable for specific event types and argument roles. These templates not only help to more clearly present the dynamic changes of event argument roles, but also provide targeted analysis basis for subsequent emergency response and disaster relief plans for different types of disasters. By using the RainEE dataset for evaluation, this application demonstrates the superiority of its method in processing event information in specific fields, especially when dealing with complex and unstructured data posted on social media platforms such as Weibo. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 It is a logic diagram of the rainstorm disaster event argument extraction method based on event text enhancement of the present invention; Figure 2 The flowchart of constructing a data set in an embodiment of the present invention is shown in FIG. DETAILED DESCRIPTION

[0020] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0021] As mentioned above, the present invention provides a method for extracting arguments of rainstorm disaster events based on event text enhancement, and generates a rainstorm event extraction (RainEE) dataset based on rainstorm disaster texts on Weibo; in order to enhance the perception of trigger words in event texts and improve the accuracy and completeness of event argument extraction, a BERT-based question-answering network (ETEN_BERT_QA) is proposed, and an event text enhancement module network is added to the encoding part, which avoids the poor event extraction effect and the problem of missing annotated corpus in the field of rainstorm disasters.

[0022] Specifically, Figure 1As shown, the present invention first defines seven types of rainstorm disaster events and their corresponding argument roles by constructing event templates. These definitions provide a standard event classification framework for the system and also provide a structural basis for data annotation and information extraction in subsequent steps. Then, based on the event types and keywords in the event template, a data acquisition process is designed. A large amount of text data related to rainstorm disasters is captured on the social media platform through web crawlers to provide rich text samples for model training and verification. The collected text data undergoes a data preprocessing step to remove redundant and irrelevant information and filter out content that meets the requirements of the event template, thereby ensuring the accuracy and relevance of the data; then, based on the preprocessed data, the text is carefully annotated with event types and argument roles according to the event template to construct a standardized data set to ensure that the data structure is consistent with the event template, laying the foundation for model learning and verification. On this basis, in order to further simplify model processing and improve efficiency, a "question template" based on the event template is generated to structure the event type and argument roles so that the model can directly locate key information through event trigger words without natural language understanding. Next, through the event text enhancement step, the recognition of the argument roles in the text is improved, so that the model can better capture event trigger words. After that, the question template and the enhanced event text are integrated into a unified input format for the model to perform question and answer processing. The model encodes the integrated input through the pre-trained BERT model and converts it into a high-dimensional representation vector to capture the semantic information in the text. The encoded text is then feature processed through the intermediate linear layer and the fully connected layer, and finally the start and end positions in the event are identified, and the text data is successfully converted into structured event information. Through such an interlocking step design, the present invention realizes the automated, standardized and efficient processing of information related to rainstorm disasters from raw text to structured output.

[0023] Example The present invention is further described below according to specific embodiments, and the present invention is performed according to the following steps: Step 1: Construct an event template to define the rainstorm disaster event type and the corresponding argument role; The event types in the dataset are defined. According to the quantity and type of available data, the data are classified into event categories. The final dataset event template includes 7 event types and 28 argument roles, such as waterlogging, vegetation loss, traffic problems, rescue operations, personal safety, landslides, building collapse / damage event types and argument roles such as time and location. The event template is shown in Table 1.

[0024] Table 1

[0025]

[0026] Step 2: Obtain rainstorm disaster data from the Internet based on the event template; like Figure 2 As shown in the figure, the specific process of dataset construction is given. First, a customized web crawler framework is designed and implemented to extract information related to rainstorm disasters from online platforms such as Weibo based on keywords and event time. First, based on the correlation analysis of rainstorm events, a set of accurate keyword sets is constructed. These keywords are highly correlated with rainstorm events, ensuring the effectiveness of data collection. On this basis, from July 15, 2021 to August 3, 2021, keywords such as "flood in a certain place", "rainstorm in a certain place" and other related phrases were used to successfully collect more than 70,000 Weibo contents related to rainstorm events in a certain place in 2021, and they were parsed to obtain relevant rainstorm disaster texts.

[0027] Step 3: Preprocess the acquired rainstorm disaster data to obtain valid rainstorm disaster event text; To ensure the high quality and relevance of the data, the data was preprocessed. First, deduplication technology was applied to identify and remove redundant posts to avoid data duplication affecting the analysis results. Secondly, a keyword-based template filtering method was used to screen out posts that met the predetermined content standards to further improve the accuracy of the data. Finally, the posts were evaluated through dependency syntax analysis technology to obtain valid rainstorm disaster event texts to ensure that their grammatical structure and content were highly relevant to the research topic.

[0028] Step 4: Label the event type and argument role of valid rainstorm disaster event texts according to the constructed event template to build a standardized dataset.

[0029] During the annotation process, a detailed text analysis was conducted on the text related to the rainstorm disaster, and the event trigger words were identified and annotated, and classified into different event types. At the same time, the event arguments were annotated and divided into different argument roles. These annotation information were recorded in a predefined event template to ensure the standardization of the data. Finally, a dataset containing 1,432 annotated texts was constructed, of which 70% were used for the training set, a total of 1,021 data, and 30% were used for the validation set, a total of 411 data. In order to ensure the consistency and accuracy of data annotation, the previously completed annotation work was regularly reviewed, and consistency checks were completed to ensure that the data structure was consistent with the event template, laying the foundation for model learning and verification. By taking strict review and control measures, the quality and reliability of the RainEE dataset during the annotation process were ensured.

[0030] Step 5: Structure the event type and argument roles to generate a question template; In order to ensure efficient processing and avoid long texts affecting model performance, prompt word templates extracted from prior knowledge are used instead of traditional natural language templates. The format of the question template is as follows: ; in, Indicates the event type. Indicates the argument role corresponding to the event type, m indicates the length of the event type, and n indicates the length of the argument role. Determined by the event detection subtask, From Obtained from the corresponding event mode. and Respectively represent and Length.

[0031] Step 6: Perform event text enhancement on the valid rainstorm disaster event text; in order to improve the model's understanding of the event text, the representation of trigger words is enhanced. This includes adding trigger words to the event text. The specific process is as follows: Constructing event text: For the original event text , the text is vectorized according to the event content, and its form is: ; in, Indicates the length after word segmentation.

[0032] Trigger word representation: The trigger word form obtained by the event detection subtask is: ; Enhanced representation construction: In order to improve the model's ability to understand event text, the event trigger words are used The strategy of adding special separators "[SEP]" before and after. Specifically, the enhanced text constructed The form is as follows: ; in, Special tokens representing sequence segmentation, Indicates event trigger words, Indicates the length of the event trigger word, Indicates the length of the enhanced text.

[0033] Step 7: Integrate the question template and the enhanced valid rainstorm disaster event text into a unified input format to obtain integrated data.

[0034] The question template and event text enhancement are integrated as the input C of the machine question answering model. The integration form is as follows: ; in, A special token indicating the beginning of a sequence, Represents a filling token, which is used to supplement the insufficient length of the fragment or sequence in the input sequence. Here, Mark the beginning, Separate the ingredients, Ensure uniform input length.

[0035] Step 8: Encode the integrated data through the pre-trained BERT model and convert the integrated data into a high-dimensional representation vector.

[0036] The pre-trained Chinese-BERT-WWM model will be used to train the integrated question-answer text Encode and get its hidden layer representation .

[0037] ; in, Represents input Length, Represents the hidden layer representation The dimension of Represents a pre-trained language representation model.

[0038] Step 9: Perform linear transformation on the high-dimensional representation vector through the intermediate linear layer to extract deep features; The role of this intermediate linear layer is to perform a linear transformation on the input data, mapping it from the original dimension to the intermediate feature space. Through this linear layer, the model can extract deeper features and lay the foundation for subsequent nonlinear processing and feature selection, thereby improving the expression ability and classification effect of the neural network.

[0039] Linear layer: This linear layer receives input data and performs a linear transformation through a weight matrix to generate an output signal of intermediate dimensions. The input dimension of the linear layer is out_dims and the output dimension is mid_linear_dims.

[0040] Activation function: After the linear layer, a nonlinear activation function is applied to introduce nonlinear characteristics. Preferably, the activation function is ReLU (rectifying linear unit), which can zero out the negative part of the input signal and keep the positive part unchanged.

[0041] Dropout layer: To prevent overfitting of the model, a Dropout layer is introduced after the ReLU activation function. The Dropout layer randomly sets the input units to zero with a certain probability (0.3 in this embodiment), thereby enhancing the generalization ability of the model during training.

[0042] Step 10: Input the deep features into the fully connected layer to obtain the prediction of the starting and ending positions of the input sequence, and convert the text data into structured event information; The prediction of the starting and ending positions of the input sequence is achieved through the fully connected layer.

[0043] Starting position prediction: Use a fully connected layer to linearly map the input features and output a label related to the starting position.

[0044] End position prediction: Similarly, another fully connected layer is used to linearly transform the same input features to output a label related to the end position and predict the end position of the input sequence. The input and output dimensions of this layer are the same as those of the starting position prediction layer.

[0045] On the other hand, the present invention provides a rainstorm disaster event argument extraction device based on event text enhancement, comprising: The event template construction module is used to construct event templates and define the types of rainstorm disaster events and the corresponding argument roles; The data acquisition module is used to obtain and pre-process the rainstorm disaster event data from the Internet to obtain valid rainstorm disaster event text; The dataset construction module is used to label the event type and argument role of the valid rainstorm disaster event text according to the constructed event template, and construct a standardized dataset; The question template generation module is used to structure the event types and argument roles in the standardized data set to generate question templates; An event text enhancement module, used for performing an event text enhancement operation on the valid rainstorm disaster event text to obtain an enhanced event text; A feature extraction module, used to integrate the question template and the enhanced event text into a unified input format, encode the integrated data after obtaining it, and convert the integrated data into a high-dimensional representation vector; The location prediction module is used to perform linear transformation on the high-dimensional representation vector, extract deep features, and convert text data into structured event information.

[0046] In a third aspect, the present invention provides an electronic device, comprising: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method for extracting arguments of rainstorm disaster events based on event text enhancement.

[0047] In a fourth aspect, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enables the processor to implement the aforementioned method for extracting arguments of rainstorm disaster events based on event text enhancement.

[0048] The specific embodiments described above further illustrate the objectives, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A rainstorm disaster event argument extraction method based on event text enhancement, characterized in that: The method comprises: Construct event templates to define rainstorm disaster event types and corresponding argument roles; Obtain and pre-process rainstorm disaster event data from the Internet to obtain valid rainstorm disaster event text; According to the constructed event template, the event type and argument role of the valid rainstorm disaster event text are annotated to build a standardized data set; Structuring the event types and argument roles in the standardized data set to generate question templates; Performing an event text enhancement operation on the valid rainstorm disaster event text to obtain an enhanced event text; Integrate the question template and the enhanced event text into a unified input format, encode the integrated data after obtaining it, and convert the integrated data into a high-dimensional representation vector; The high-dimensional representation vector is linearly transformed to extract deep features and convert text data into structured event information.

2. According to the method for extracting rainstorm disaster event arguments based on event text enhancement according to claim 1, it is characterized in that: The rainstorm disaster event type includes at least one of waterlogging, vegetation loss, traffic problems, rescue operations, personal safety, landslides, and building collapses, and the argument roles include at least one of time and place.

3. The method for extracting rainstorm disaster event arguments based on event text enhancement according to claim 1 is characterized in that: The method of obtaining and preprocessing the rainstorm disaster event data from the network includes: applying a network crawler framework to obtain the rainstorm disaster event data, and using duplicate data deletion technology, a keyword-based template filtering method and a dependency syntax analysis technology to preprocess the rainstorm disaster event data.

4. The method for extracting arguments of rainstorm disaster events based on event text enhancement according to claim 1 is characterized in that: The question template uses a prompt word template extracted from prior knowledge and is generated based on the event type and corresponding event pattern determined by the event detection subtask.

5. The method for extracting arguments of rainstorm disaster events based on event text enhancement according to claim 1 is characterized in that: The event text enhancement operation includes adding special delimiters before and after event trigger words to improve the model's understanding of the event text.

6. The method for extracting arguments of rainstorm disaster events based on event text enhancement according to claim 1 is characterized in that: The encoding after obtaining the integrated data includes encoding the integrated data through a pre-trained BERT model to convert the integrated data into a high-dimensional representation vector.

7. The method for extracting arguments of rainstorm disaster events based on event text enhancement according to claim 1 is characterized in that: The method of performing a linear transformation on the high-dimensional representation vector, extracting deep features, and converting text data into structured event information includes performing a linear transformation on the high-dimensional representation vector through an intermediate linear layer to obtain deep features, inputting the deep features into a fully connected layer, and obtaining a prediction of the starting position and the ending position of the input sequence.

8. A rainstorm disaster event argument extraction device based on event text enhancement, characterized in that: include: The event template construction module is used to construct event templates and define the types of rainstorm disaster events and the corresponding argument roles; The data acquisition module is used to obtain and pre-process the rainstorm disaster event data from the Internet to obtain valid rainstorm disaster event text; The dataset construction module is used to label the event type and argument role of the valid rainstorm disaster event text according to the constructed event template, and construct a standardized dataset; The question template generation module is used to structure the event types and argument roles in the standardized data set to generate question templates; An event text enhancement module, used for performing an event text enhancement operation on the valid rainstorm disaster event text to obtain an enhanced event text; A feature extraction module, used to integrate the question template and the enhanced event text into a unified input format, encode the integrated data after obtaining it, and convert the integrated data into a high-dimensional representation vector; The location prediction module is used to perform linear transformation on the high-dimensional representation vector, extract deep features, and convert text data into structured event information.

9. An electronic device, characterized in that: include: one or more processors; A memory for storing one or more programs; Among them, when one or more programs are executed by the one or more processors, the one or more processors implement the method for extracting rainstorm disaster event arguments based on event text enhancement as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: Executable instructions are stored thereon, which, when executed by a processor, enable the processor to implement the method for extracting rainstorm disaster event arguments based on event text enhancement as described in any one of claims 1 to 7.

Images