Improved disaster news three-element extraction method

Abstract

The application discloses an improved disaster news three-element extraction method. The method comprises the following steps: constructing and training a BILSTM-based element classification model, a disaster theme classification model and a named entity recognition model; performing data preprocessing on disaster news texts; determining the disaster theme of the preprocessed disaster news texts by using the disaster theme classification model; determining the three elements of the preprocessed disaster news texts by using the element classification model; and extracting the time and place entities of the region by using the element entity recognition model, so as to obtain the time and place of disaster occurrence.

Description

Technical Field

[0001] This invention relates to an improved method for extracting the three elements of disaster news. Background Technology

[0002] According to the "Basic Situation of Natural Disasters in China in 2021" released by the Ministry of Emergency Management of the People's Republic of China, the natural disaster situation in my country in 2021 was complex and severe, with frequent extreme weather and climate events. Various natural disasters affected 107 million people and caused direct economic losses of 334.02 billion yuan throughout the year. Faced with the ever-emerging threat of disasters, obtaining effective disaster information extensively and studying the patterns of occurrence and development of various disaster events are of great significance for researching disaster prevention and mitigation and reducing disaster risks. [1] .

[0003] With the widespread use of the internet, information and news reports about disasters can be continuously obtained online within minutes to days after they occur. Disaster news releases are multi-layered, comprehensive, and three-dimensional, and their content reflects the authenticity and accuracy of the disaster, making them suitable for spatiotemporal analysis of disaster information. Extracting valuable information from massive amounts of disaster news text data is a crucial step in the study of the spatiotemporal patterns of disaster events. (An Lu et al.) [5] This study divides breaking news data from Weibo into four lifecycles, uses the Word2vec model and K-means clustering algorithm to extract the themes of each part, and then uses the H-index to filter out the themes of the breaking news, thereby determining the derivative events of the themes. Alberto Téllez Valero et al. applied Support Vector Machine (SVM) and Naive Bayes (Naive Bayes) methods. Three algorithms—Bayes (NB), decision tree (C4.5)—are used to classify the features of disaster-related documents in news reports to obtain the disaster category of the news.

[0004] In the era of big data, deep learning has significant advantages in multiple fields such as computer vision, natural language processing, and information retrieval, and its application in disaster emergency management is becoming increasingly widespread. Liu Shuhan et al. used a convolutional neural network (CNN) to classify topics related to rainstorm disaster information on Weibo. Wu Jianhua [9]A BILSTM-CRF model was built to automate the identification of spatiotemporal information of emergencies in Weibo, improving the accuracy of such identification. Given the linguistic characteristics of news texts and the domain-specific issues of disaster news, how to better extract basic elements from disaster news is a current hot research topic. Song Dunjiang et al. applied a pre-trained language model with Bidirectional Encoder Representation (BERT) to extract triplet information from geological disasters, optimizing the "low-rank bottleneck" problem caused by the underlying multi-head attention mechanism. This significantly improved the error tolerance and accuracy of extracting key information such as the type, time, and location of geological disasters from news reports and other texts.

[0005] With the continuous updates to deep learning network models, current research focuses more on optimizing and innovating these models, enabling them to learn more feature information from complex input data. Summary of the Invention

[0006] To overcome the above-mentioned shortcomings, this invention provides an improved method for extracting the three elements of disaster news.

[0007] To achieve the above objectives, the present invention provides an improved method for extracting the three elements of disaster news, comprising the following steps:

[0008] Construct and train a BILSTM-based feature classification model, a disaster theme classification model, and a feature entity recognition model;

[0009] Data preprocessing of disaster news texts;

[0010] The disaster theme is determined by using a disaster theme classification model to identify the disaster theme in the preprocessed disaster news text;

[0011] The scope of the three elements of the preprocessed disaster news text is determined using an element classification model;

[0012] The element entity recognition model is used to extract the time and location entities within the area, thus obtaining the time and location of the disaster.

[0013] Furthermore, the step of determining the disaster theme of the preprocessed disaster news text using a disaster theme classification model includes:

[0014] A disaster topic classification model based on BILSTM is constructed, and a 1×M sliding window is used to carry out regional identification in news texts, where M is the length of the sliding window.

[0015] The classification network performs sliding predictions in the text with a step size l; where, in the i-th window, it returns the predicted class X. i And prediction accuracy p iAfter filtering by setting a threshold α, the predicted values ​​X that meet the conditions are... i Add it to the topic prediction set S;

[0016] After all windows have been processed, a voting decision is made on the set of topic predictions, and the prediction category with the most predictions is taken as the disaster topic for the news.

[0017] Furthermore, the step of determining the three elements of the preprocessed disaster news text using the element classification model includes:

[0018] Based on the BILSTM feature classification model, a sliding window with an initial size of 1×N is constructed and equipped with a feature classification network to slide through the news text. The classification result of each sliding window is saved to the feature prediction set S.

[0019] The prediction set S is corrected for misjudgments and similar items are merged to show the order in which the three elements of location, time, and event appear.

[0020] A 1×3 sliding window is constructed, sliding within the set with a step size of 1 to filter for the three elements based on their nearest neighbor distance. The event element must appear at the end of the sliding window, and the window must contain all three elements. After filtering, a preliminary set S of all time and location ranges containing disaster events is obtained. l .

[0021] Furthermore, it also includes the step of using a disaster news topic identification model to analyze the set S. l The similarity of themes is judged, and the similarity of each range with the disaster theme in the news is output. The time and location range with the highest similarity with the disaster theme is taken as the region to be extracted as the spatiotemporal elements of the disaster. The named entity recognition model based on BILSTM is used to extract the time and location entities in this region to obtain the time and location of the disaster.

[0022] This invention addresses the extraction process of the three key elements (LOE, ABO, and B) in disaster news. Based on a Chinese disaster classification title dataset and a Chinese emergency corpus, it proposes an improved method for extracting these elements. A sliding window classification and detection model based on BILSTM (Bidirectional Long Short-Term Memory) combined with a named entity recognition model is constructed to identify and extract the three LEE. Extraction tests were conducted on disaster and accident information texts published by the Ministry of Emergency Management of the People's Republic of China, and the recognition performance of four different models was compared. The proposed method helps improve the efficiency and accuracy of disaster news information extraction and provides new insights for disaster information extraction research. Attached Figure Description

[0023] Figure 1This is a flowchart of the data annotation and model training process for this invention.

[0024] Figure 2 This is a flowchart of the three elements extraction process for disaster news in this invention.

[0025] Figure 3 LSTM architecture diagram

[0026] Figure 4 BILSTM structure diagram

[0027] Figure 5 Example flowchart of the time and location extraction optimization algorithm Detailed Implementation

[0028] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0029] In the description of this invention, it should be understood that the terms "center", "upper", "lower", "front", "rear", "left", "right", "vertical", "horizontal", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0030] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, unless otherwise stated, "a plurality of" means two or more.

[0031] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "joining" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0032] Design of a technical framework for disaster news topic extraction in this invention

[0033] Overall technical framework

[0034] The proposed method for extracting the three elements of disaster news is based on natural language processing analysis using a deep neural network model. It consists of two stages: data annotation and model training, and the identification and extraction of the three elements of disaster news. The overall technical framework is as follows: Figure 1 , Figure 2 As shown.

[0035] Figure 1 This section covers data annotation and model training. It mainly involves two aspects: 1) This experiment uses two datasets as the primary training data: a Chinese emergency events corpus dataset and a Chinese disaster classification title dataset. Due to the small amount of data, external data was added to expand the data during the category annotation process. The external data comes from the Small-Chinese-Corpus dataset on GitHub. 2) A BILSTM-based multi-classification model and an element entity recognition model were built on the PyTorch deep learning framework to achieve the recognition and extraction of elements from news text.

[0036] Figure 2 For the identification and extraction of the three elements of disaster news, this method comprehensively uses different models combined with sliding window detection to identify and extract the disaster theme and the time and location of the disaster from the news text. The extraction process is mainly divided into two parts. On the left, a sliding window + multi-classification model + voting decision mechanism is used to identify and extract the disaster news theme. On the right, a sliding window with an element classification model is first used to determine the time and location range of the news content. Then, the nearest neighbor distance filtering and named entity recognition model are used to complete the precise extraction of the range and the time and location of the disaster.

[0037] 1.2 Disaster News Theme Identification and Extraction Model

[0038] News theme extraction refers to finding or extracting words or phrases from news texts that embody the core content of the news. For disaster news, the type of disaster event described in the news text often represents the disaster theme of the news. However, common disaster event types are not abundant. Therefore, this invention considers transforming the task of extracting disaster themes from news into a task of classifying and recognizing disaster event types in news.

[0039] In 1997, Hochreiter proposed an improved Temporal Recurrent Transformer (LSTM) network, which effectively addresses the gradient explosion problem caused by long-term dependencies in general Recurrent Neural Networks (RNNs). The LSTM network contains three different special gate structures to perform forgetting, updating, and output operations on the incoming information, thereby controlling the cell state. Figure 3 This is a diagram of the LSTM structure.

[0040] However, using LSTM for sentence feature extraction presents a problem: it cannot encode information from the end to the beginning. This prevents LSTM from fully utilizing contextual information for learning, resulting in unsatisfactory performance on some fine-grained classification tasks. BILSTM consists of a forward LSTM and a backward LSTM. The forward LSTM extracts features from the preceding context at the current time step, while the backward LSTM extracts features from the following context. Finally, the outputs of the forward and backward LSTMs at the current time step are concatenated to obtain an output vector that incorporates contextual semantic information. Compared to LSTM, the BILSTM model learns richer features, thus achieving better results in text classification or named entity recognition tasks. Figure 4 This is a structural diagram of the BILSTM model.

[0041] This invention employs a BILSTM-based disaster topic classification model in its disaster news topic identification and extraction model. A 1×M sliding window is constructed to carry out regional identification within the news text using a disaster topic classification network, where M is the length of the sliding window. The classification network makes sliding predictions in the text with a step size l. In the i-th window, the predicted category X is returned. i And prediction accuracy p i After filtering by setting a threshold α, the predicted values ​​X that meet the conditions are... i Add it to the topic prediction set S, as shown in equation (1):

[0042] S = {X i |p i >α} (1)

[0043] After all windows have been processed, a voting decision is made on the set of topic predictions, and the prediction category with the most predictions is taken as the disaster topic for the news.

[0044] Time and location extraction optimization algorithm

[0045] Currently, extracting elements such as time and location from text can be considered a named entity recognition task, i.e., a sequence labeling task. This invention intends to use text classification as an experimental method to study the feasibility of applying text classification methods to entity recognition. The optimized algorithm flow for time and location extraction is as follows: Figure 5 As shown, it is mainly divided into the following parts.

[0046] ① Data preprocessing. Stop words are removed from the text to eliminate phrases such as "rescue," "reporting," and "reporting" that are irrelevant to the timing of the disaster, preventing misjudgments during subsequent sliding window classification.

[0047] ② Determining the scope of time and location elements of disaster occurrence in news texts. A BILSTM-based element classification model is built, and a sliding window of initial size 1×N carrying the element classification network is constructed and slid across the news text. The classification result of each sliding window is saved to the element prediction set S.

[0048] ③ Correct misjudgments and merge similar categories in the prediction set S. For example, the predicted category X of the i-th window (where i is not at the beginning or end). i If the predicted category is different from that of the surrounding window, then X will be... i It was determined to be a misjudgment, so let X... i =X i-1 The corrected prediction set S is obtained by merging consecutive predictions of the same type X in the original prediction set S. m This makes the order of the three elements clearer.

[0049] ④ Perform nearest neighbor distance filtering for the three elements. After merging similar elements, the prediction set will show the order in which the three elements—location, time, and event—appear. Construct a 1×3 sliding window with a step size of 1, sliding it through the set to filter for nearest neighbor distance among the three elements. The event element must appear at the end of the sliding window, and the window must contain all three elements. After filtering, a preliminary set S of all time and location ranges containing disaster events will be obtained. l .

[0050] ⑤ Accurately define the scope of spatiotemporal elements and extract spatiotemporal elements. Use a disaster news topic recognition model to analyze set S. l The algorithm performs topic similarity assessment, outputting the similarity score of each range to the disaster topic in the news. The range with the highest similarity to the disaster topic in terms of time and location is selected as the region to be extracted as the spatiotemporal element of the disaster. A BILSTM-based named entity recognition model is then built to extract time and location entities from this region, thus obtaining the time and location of the disaster.

[0051] Example

[0052] Disaster News Data Acquisition and Processing

[0053] This invention uses the Chinese Disaster Classification Title Dataset and the Chinese Event Corpus (CEC) as the main datasets for the research. The Chinese Disaster Classification Title Dataset was developed by an open-source website and the Whale Community. https: / / www.heywhale.comThe dataset, collected to extract the main content of disaster news, contains 48,529 news headlines across 51 disaster categories. The categories and corresponding data volumes are shown in Table 1. A Chinese emergency corpus is used for training to extract the time and location elements of disasters from the news reports. Considering the correlation between spatiotemporal information and trigger words for emergencies, the distance relationship between these three elements in the text can be used to assist in determining the range of spatiotemporal elements. Because the dataset contains many location elements and numerous duplicate texts for time elements, the dataset is noisy and this can negatively impact the learning of the classification model. Therefore, this invention removes duplicate data and expands the dataset with external data to increase the number of time and event elements.

[0054] Table 1 Disaster News Theme Dataset

[0055]

[0056] Dataset labeling

[0057] For the Chinese disaster classification title dataset, there are 51 disaster categories. The dataset is labeled with categories numbered 0-50 in ascending order, as shown in Table 2. Examples of labeled data are shown in Table 3. The Chinese emergency event corpus is labeled using the same steps. Furthermore, this invention uses the BIO (B: beginning of a noun phrase, I: middle of a noun phrase, O: not a noun phrase) labeling system to perform sequence labeling on the Chinese emergency event corpus, dividing absolute time and relative time into time entities. Detailed information on sequence labeling is shown in Table 4. Finally, 332 labeled entries are obtained. All three datasets are divided into training and validation sets in a 7:3 ratio.

[0058] Table 2 Examples of Category Labeling for Disaster News Theme Dataset

[0059]

[0060] Table 3 Examples of Category Annotations for Chinese Emergency Event Corpus

[0061]

[0062] Table 4 Examples of Sequence Annotation for Chinese Emergency Event Corpus

[0063]

[0064] Experimental design and evaluation indicators

[0065] The computer used in this experiment was configured with an Intel(R) Core(TM) i5-9300H processor, 16GB of RAM, and an NVIDIA GeForce GTX 1650 graphics card. The experimental environment consisted of PyTorch deep learning framework version 1.9 and Python version 3.7.0.

[0066] To verify the effectiveness of the improved method for extracting the three elements of disaster news, this invention conducted comparative experiments. For the extraction of spatiotemporal elements of disaster occurrence, LSTM, BILSTM, and BILSTM-CRF models were compared with the time-location optimization extraction method of this invention. A three-element classification model was built to classify the three elements of disaster time, location, and event. The relevant parameters of the model are shown in Table 6. An element entity recognition model was built to accurately extract the disaster occurrence time and location of the area to be extracted. The relevant parameters of the model are shown in Table 7. Regarding disaster news topic recognition, since this experiment focuses on extracting disaster topics from disaster news, and the topics are clear and the number of topics is small, supervised classification methods would yield greater benefits than unsupervised models such as LDA. Therefore, this invention did not include comparative experiments in disaster news topic extraction. The parameters of the disaster news topic recognition model built in the experiment are shown in Table 5.

[0067] Table 5 Parameters of the Disaster News Theme Recognition Model

[0068]

[0069] Table 6 Parameters of the Element Classification Model

[0070]

[0071]

[0072] Table 7 Training parameters for the element entity recognition model

[0073]

[0074] Evaluation metrics: This invention uses precision (P), recall (R), and F1 score to evaluate the recognition accuracy of the above methods. The formulas for the three metrics are shown below:

[0075]

[0076]

[0077]

[0078] Where TP is the number of correct classes predicted as correct, FP is the number of incorrect classes predicted as correct, FN is the number of correct classes predicted as incorrect, and F1 is the harmonic mean of precision and recall.

[0079] Experimental Results and Analysis

[0080] (1) Disaster News Theme Identification

[0081] This invention crawled various types of disaster reports from CCTV's website using disaster names as keywords. After filtering out irrelevant news and removing small sample sizes, a total of 571 news items in five categories—hail, earthquake, tsunami, landslide, and fire—were obtained as test data for the disaster theme recognition model. This data included some early warning and post-disaster reports as noise data to test the model's robustness. The model recognition results are shown in Table 8.

[0082] Table 8. Disaster Theme Model Identification Results

[0083]

[0084] As shown in Table 8, the model of this invention achieves good recognition results for the three themes of earthquake, landslide, and fire, with F1 scores all above 90%. However, the precision and recall differ significantly for the themes of hail and tsunami, with the recall being lower. Examination of the test text reveals that most news reports of tsunamis are caused by undersea earthquakes, and news reports of hail disasters are often accompanied by reports of heavy rain, lightning, and convective weather. Since the model of this invention focuses on single-theme recognition of disaster news, it identifies many other categories during the sliding window recognition process, interfering with the final voting judgment. Future improvements could consider enhancing the model's voting judgment by changing single-theme disaster recognition to multi-theme disaster recognition.

[0085] (2) Extraction of spatiotemporal elements of disaster news occurrence

[0086]

[0087]

[0088] For testing the spatiotemporal element extraction model, this invention crawled news articles about fires and earthquakes from the disaster and accident information publicly available from the Ministry of Emergency Management of the People's Republic of China. Other disaster events were not included in the test data due to their small data volume. After crawling, irrelevant data was filtered and formatted, resulting in a total of 134 news articles about fires and earthquakes, including 45 fire news articles, 44 earthquake news articles, and 35 landslide news articles. An example of the processed news data is shown in Table 9.

[0089] Table 9 Example of Test Data

[0090] The results of each method on the test data are shown in Table 10.

[0091] Table 10 Comparison of extraction effects of each model

[0092]

[0093] Table 10 shows that the extraction performance of the four methods varies. LSTM network did not achieve ideal results in extracting the time and location elements of disasters from disaster news texts, with low precision and recall. Although BILSTM has higher precision, its recall is insufficient compared to the latter two methods, at only 78.8%, which also leads to a lower F1 score for BILSTM. BILSTM-CRF replaces the softmax layer with a CRF layer in the last layer compared to BILSTM. The CRF layer uses the probabilities of each label obtained from the BILSTM layer to calculate the transition probabilities between labels, thereby controlling the output order of each label. Therefore, compared to BILSTM, BILSTM-CRF avoids situations where there are two consecutive B labels or a string of labels all being I labels, improving recall and F1 score by 13.6% and 7.2% respectively. The method of this invention, compared to BILSTM-CRF, achieves a slight improvement in precision while maintaining recall.

[0094] Further analysis of the performance of several methods on three types of test data is presented. Tables 11, 12, and 13 show the test results of the four methods on earthquake, fire, and landslide datasets, respectively. All models achieved high accuracy in extracting earthquake news, but LSTM and BILSTM performed poorly in extracting fire and landslide news, with F1 scores below 80%. BILSTM-CRF achieved 84.4% precision and 83.3% recall in extracting the time and location of fire news, and 71.4% precision and recall in extracting landslide news. The time and location extraction method proposed in this invention achieved the best results, with an overall improvement of 1%–2% over BILSTM-CRF in fire news and approximately 3% in landslide news. The results indicate that the precise time and location range method in this invention has better performance than current named entity recognition algorithms, correctly extracting the time and location range of disasters. This allows subsequent named entity recognition to extract information only within a defined range, reducing errors in time periods or locations that occur when directly recognizing the entire text.

[0095] Overall, the method proposed in this invention achieves good results in extracting the three elements of disaster news: topic, time, and location. Compared with LSTM and BILSTM, the method proposed in this invention has a significant improvement in F1 score. Compared with BILSTM-CRF, the method proposed in this invention has slight improvements in precision and recall. In the task of extracting the time and location of disasters, because the method proposed in this invention can initially determine the time and location range in the news text, it narrows down the target of entity recognition to a specific range, thereby increasing the accuracy of recognition.

[0096] Table 11 Comparison of Earthquake News Extraction Results by Different Models

[0097]

[0098] Table 12 Comparison of Extraction Results of Different Fire News Models

[0099]

[0100] Table 13 Comparison of extraction effects of different models for landslide news.

[0101]

[0102]

[0103] In summary, the present invention has the following characteristics:

[0104] 1) This invention proposes a method for identifying and extracting disaster news topics. The sliding window detection model constructed in this invention, which combines a sliding window, a multi-classification model, and a voting decision-making mechanism, can effectively reduce the identification error caused by the accuracy problem of the classification model.

[0105] 2) This invention optimizes the time and location extraction method for news by adding a precise step of determining the range of time and location elements of disasters, which improves the accuracy of the recognition range for the named entity recognition model. The effectiveness of this method is demonstrated by comparing the recognition results with those of LSTM, BILSTM, and BILSTM-CRF models on different news items.

[0106] In the description of this specification, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0107] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. An improved method for extracting the three elements of disaster news, characterized in that, The method includes the following steps: Construct and train a BILSTM-based feature classification model, a disaster theme classification model, and a feature entity recognition model; Data preprocessing of disaster news texts; The disaster theme is determined by using a disaster theme classification model to identify the disaster theme in the preprocessed disaster news text; The scope of the three elements of the preprocessed disaster news text is determined using an element classification model; The element entity recognition model is used to extract the time and location entities in this area to obtain the time and location of the disaster. The steps for determining the three elements of preprocessed disaster news text using an element classification model include: Based on the BILSTM feature classification model, a sliding window with an initial size of 1×N is constructed and equipped with a feature classification network to slide through the news text. The classification result of each sliding window is saved to the feature prediction set S. The prediction set S is corrected for misjudgments and similar items are merged to show the order in which the three elements of location, time, and event appear. A 1×3 sliding window is constructed, sliding within the set with a step size of 1 to filter for the three elements based on their nearest neighbor distance. The event element must appear at the end of the sliding window, and the window must contain all three elements. After filtering, a preliminary set S of all time and location ranges containing disaster events is obtained. l ; It also includes the step of: using a disaster news topic identification model to analyze the set S l The similarity of themes is judged, and the similarity of each range with the disaster theme in the news is output. The time and location range with the highest similarity with the disaster theme is taken as the region to be extracted as the spatiotemporal elements of the disaster. The element entity recognition model based on BILSTM uses the named entity recognition method to extract the time and location entities in the region, so as to obtain the time and location of the disaster.

2. The improved method for extracting the three elements of disaster news as described in claim 1, characterized in that, The steps for determining the disaster theme of preprocessed disaster news text using a disaster theme classification model include: A disaster topic classification model based on BILSTM is constructed, and a 1×M sliding window is used to carry out regional identification in news texts, where M is the length of the sliding window. The classification network performs sliding predictions in the text with a step size l; where, in the i-th window, it returns the predicted class X. i And prediction accuracy p i After filtering by setting a threshold α, the predicted values ​​X that meet the conditions are... i Add it to the topic prediction set S; After all windows have been processed, a voting decision is made on the set of topic predictions, and the prediction category with the most predictions is taken as the disaster topic for the news.

Images