A document-level event factuality detection method and system

Abstract

This invention discloses a text-level event factuality detection method and system. The method includes: acquiring a corpus containing an event description title and corresponding body text; translating the corpus and extracting information on the main events and key elements; encoding the extracted bilingual information and using an improved graph attention network-based model to connect the encoded bilingual information as global nodes with the body text; adjusting the weights of the Chinese and English texts through grid search; and using a data-augmented dataset to determine the event factuality detection result in the title. This invention improves the accuracy and efficiency of the model by obtaining key information through data augmentation before data preprocessing. Multilingual representation can obtain different levels of sensitivity for the same text in multiple languages. Furthermore, this invention applies the GAT model in graph models, combining the extracted key information as global nodes with the entire text, resulting in better model performance and stronger robustness.

Description

Technical Field

[0001] This invention relates to the field of event fact detection technology, and in particular to a method and system for detecting event facts at the chapter level. Background Technology

[0002] Existing methods for document-level event fact detection rarely employ data augmentation before preprocessing, and those that do mostly involve multilingual representations. Information extraction, another important task in natural language processing, can also be applied to this work. Previous works have attempted to build upon this by introducing the standard-ner package for information extraction, but due to accuracy limitations, the improvement is not significant. This means that key information that could be obtained through data augmentation before data preprocessing needs to be discovered by spending considerable computational resources in the model to find semantic relationships between texts. Furthermore, the accuracy of this approach is far lower than direct extraction, making it even more difficult to find global structural features and latent textual features.

[0003] Most existing document-level event fact detection methods focus on building models to extract more suitable semantic features and improve classification performance, often proceeding to the word segmentation module with little or no data processing. However, in real-world scenarios, data augmentation is feasible. This allows the model to acquire more effective semantic features from the outset, reduces the probability of the model acquiring incorrect features, minimizes noise input, and improves model effectiveness. Summary of the Invention

[0004] In view of the aforementioned existing problems, the present invention is proposed.

[0005] Therefore, this invention provides a method and system for detecting factual information about events at the chapter level, which solves the problem that existing methods for detecting factual information about events at the chapter level lack effective data augmentation methods in the preprocessing stage, resulting in the model needing to consume a lot of computing power to extract key information and achieving poor results.

[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0007] In a first aspect, the present invention provides a text-level event factual detection method, comprising: acquiring a corpus containing an event description title and corresponding body text; translating the corpus and extracting information on the main events and key elements; encoding the extracted bilingual information and using an improved model based on graph attention networks to connect the encoded bilingual information as global nodes to the body text; adjusting the weights of the Chinese and English texts through grid search to obtain a data-enhanced dataset; and using the data-enhanced dataset to determine the event factual detection result in the title.

[0008] As a preferred embodiment of the chapter-level event fact detection method of the present invention, the step of encoding the extracted bilingual information includes:

[0009] The main events and key elements are extracted from English and Chinese corpora, where the key elements are the subject, verb, and object extracted from the main events.

[0010] A segmentation strategy is used to segment texts whose length exceeds a first threshold.

[0011] The Longformer encoder, suitable for long text processing, is selected to encode the pre-extracted and segmented bilingual text, generating a fixed-length vector representation, which is then input into the improved model based on the graph attention network.

[0012] As a preferred embodiment of the chapter-level event fact detection method of the present invention, the step of using an improved model based on graph attention networks to connect the encoded bilingual information as global nodes with the main text includes:

[0013] Construct a relationship graph containing the main event node, sentence node, word node, and global nodes that the key elements become, and define the connections between the nodes;

[0014] An improved model based on graph attention network is used to update the node representation. When updating the representation of each node, the weight is reassigned according to the strength of the relationship between it and its neighboring nodes. That is, sentence nodes that are closely related to the global node will be given higher weights. The updated global node is used as the first node to add sentences in the text in turn, thereby completing the connection with the main text.

[0015] As a preferred embodiment of the document-level event fact detection method described in this invention, the step of adjusting the weights of Chinese and English texts through grid search to obtain a data-enhanced dataset includes:

[0016] Set weight ratios to represent the importance of Chinese text relative to English text, and create a grid based on the weight ratios, wherein each cell in the grid corresponds to a specific weight combination;

[0017] For each pair of weights in the grid, the text features are reweighted using the corresponding proportions and the model is trained. After each training iteration, the model's performance is measured using a predetermined evaluation metric on the validation set.

[0018] Select the weight ratio that optimizes the model performance from all weight combinations, and adjust the weight ratio of Chinese and English texts according to the optimized weight ratio to retain the information of the original English text to the greatest extent.

[0019] As a preferred embodiment of the article-level event factuality detection method of the present invention, the method for determining the event factuality detection result in the title using the data-enhanced dataset includes:

[0020] By utilizing nodes and edges in a graph structure to capture complex dependencies in text, and by strengthening the influence of key information nodes through an attention mechanism, a new vector representation that integrates contextual information is obtained.

[0021] The new vector representation that incorporates the contextual information is input into the classification layer for prediction. The classification layer maps each node to one of five classification labels to obtain the detection result of the factuality of the event in the title. The five classification labels are: definitely happened, possibly happened, definitely not happened, highly unlikely to happen, and unknown.

[0022] As a preferred embodiment of the article-level event factuality detection method of the present invention, the method further includes: using the data-enhanced dataset to determine the event factuality detection result in the title.

[0023] The cross-entropy loss function is used to measure the gap between the model's predicted values ​​and the true labels, and the AdamW optimizer is used to minimize the loss. The model parameters are adjusted through continuous iterative training to classify the data-augmented dataset and thus determine the detection results of the factual nature of the events in the title.

[0024] As a preferred embodiment of the document-level event fact detection method of the present invention, the extraction of information on major events and key elements after translating the corpus includes:

[0025] The English corpus containing event descriptions and corresponding text is translated to form a Chinese corpus. The main events and key elements are extracted from the English and Chinese corpora. The key elements are the subject, verb, and object extracted from the main events.

[0026] Secondly, the present invention provides a chapter-level event fact detection system, comprising:

[0027] The data acquisition module is used to acquire a corpus containing event descriptions and corresponding body text.

[0028] The information extraction module is used to extract information about the main events and key elements after translating the corpus.

[0029] The multidimensional data augmentation module is used to encode the extracted bilingual information and use an improved model based on graph attention network to connect the encoded bilingual information as global nodes with the main text. The weights of the Chinese and English texts are adjusted through grid search to obtain a data-augmented dataset.

[0030] An event authenticity identification module is used to determine the detection results of the factuality of the event in the title using the data-enhanced dataset.

[0031] Thirdly, the present invention provides a computing device, comprising:

[0032] Memory and processor;

[0033] The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions, which, when executed by the processor, implement the steps of the chapter-level event fact detection method.

[0034] Fourthly, the present invention provides a computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the chapter-level event fact detection method.

[0035] Compared with existing technologies, the beneficial effects of this invention are as follows: ① This invention obtains key information through data augmentation before data preprocessing to improve the accuracy and efficiency of the model. This is the first time that multidimensional data augmentation has been applied to the identification of factual events at the document level. Compared with other one-dimensional or no data augmentation methods, it has a stronger grasp of global structural features and text sensitivity. ② This invention uses GPT as the source of multidimensional data augmentation. With the continuous upgrading of technology, it will gain greater advantages in the accuracy and various indicators of data augmentation in the future. ③ This invention uses a graph model to connect the extracted key nodes as global nodes to the entire article, which not only improves the generalization and accuracy of the model, but also reduces the model's insensitivity to specific keywords in a certain language by combining Chinese and English. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 This is a schematic diagram of the overall process logic of the chapter-level event fact detection method according to an embodiment of the present invention;

[0038] Figure 2 This is a flowchart illustrating the overall process logic of a chapter-level event fact detection method according to an embodiment of the present invention. Detailed Implementation

[0039] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0040] Example 1

[0041] Reference Figures 1-2 As one embodiment of the present invention, a method for detecting factual information of events at the chapter level is provided, such as... Figure 1 The specific steps shown are as follows:

[0042] S100: Obtain a corpus containing titles and corresponding body texts of events, translate the corpus, and extract information on the main events and key elements;

[0043] S200: The extracted bilingual information is encoded, and an improved model based on graph attention network is used to connect the encoded bilingual information as global nodes with the main text. The weights of the Chinese and English texts are adjusted through grid search to obtain a data-enhanced dataset.

[0044] S300: Utilize data-enhanced datasets to determine the factual nature of events in the captions;

[0045] It should be noted that this invention obtains key information through data augmentation before data preprocessing to improve the accuracy and efficiency of the model. This is the first time that multidimensional data augmentation has been applied to document-level event fact recognition. Compared with other one-dimensional or no data augmentation methods, it has a stronger grasp of global structural features and text sensitivity. Moreover, this invention uses GPT as the source of multidimensional data augmentation. With continuous technological upgrades, it will gain greater advantages in the accuracy and various indicators of data augmentation in the future. In addition, this invention uses a graph model to connect the extracted key nodes as global nodes to the entire article, which not only improves the generalization and accuracy of the model, but also reduces the model's insensitivity to specific keywords in a certain language by combining Chinese and English.

[0046] In this embodiment of the application, step S100 includes the following sub-steps A1-A2;

[0047] In A1: The English corpus containing event descriptions and corresponding text is translated to form a Chinese corpus;

[0048] In A2: extract the main events and key elements from the English and Chinese corpora. The key elements are the subject, verb, and object extracted from the main events.

[0049] In an optional embodiment, a Chinese corpus is formed by translating the acquired English corpus containing event descriptions and corresponding body text, and the main events and key elements from both the English and Chinese corpora are extracted, such as... Figure 2 The original two sets of data are expanded into six sets, as shown below:

[0050] The original data is:

[0051] 1) Key event: Voters support for the prime minister;

[0052] 2) English document text;

[0053] The data augmented is as follows:

[0054] 1) Key event: Voters support for the prime minister;

[0055] 2) English document text;

[0056] 3) Key elements: [Voters], [support], [prime minister];

[0057] 4) Key Event: Do voters support the Prime Minister?

[0058] 5) Chinese document text;

[0059] 6) Key elements: [Voter], [Support], [Prime Minister];

[0060] It should be noted that bilingual translation utilizes the original English corpus to form a Chinese corpus through GPT, thus expanding the size of the corpus. Key element extraction extracts the most important elements from the events: subject, verb, and object. These elements help the model obtain more effective semantic features from the beginning, while reducing the probability of the model acquiring incorrect features. This makes the text more sensitive to the events in the model and provides important information for subsequent modules.

[0061] In this embodiment of the application, step S200 includes the following sub-steps B1-B3;

[0062] In B1: The extracted bilingual information is encoded;

[0063] In B2: An improved model based on graph attention network is used to connect the encoded bilingual information as global nodes to the main text;

[0064] In B3: Adjust the weights of Chinese and English texts through grid search to obtain a data-enhanced dataset;

[0065] Specifically, encoding the extracted bilingual information includes:

[0066] Extract the main events and key elements from English and Chinese corpora;

[0067] A segmentation strategy is used to segment texts whose length exceeds a first threshold.

[0068] The Longformer encoder, suitable for long text processing, is selected to encode the pre-extracted and segmented bilingual text, generating a fixed-length vector representation, which is then input into an improved model based on a graph attention network.

[0069] It should be noted that the first threshold is 512. This is because encoding long text requires calculating the encoded length of the generated text and whether it meets the maximum length limit of the selected encoder. Since the most widely used BERT has a maximum encoding length of 512, and half of the corpus lengths already exceed this value, the model performance will degrade significantly if this value is exceeded. In the experiments using BERT segmentation and changing the encoder, this embodiment selected the Longformer encoder, which performs well in handling long texts.

[0070] Specifically, an improved model based on graph attention networks is used to connect the encoded bilingual information as global nodes to the main text, including:

[0071] Construct a relationship graph containing main event nodes, sentence nodes, word nodes, and global nodes that become key elements, and define the connections between each node;

[0072] An improved model based on graph attention network is used to update node representations. When updating the representation of each node, the weights are reassigned according to the strength of the relationship between it and its neighboring nodes. That is, sentence nodes that are closely related to the global node are given higher weights. The updated global node is used as the first node to add sentences in the text in turn, thereby completing the connection with the main text.

[0073] It's important to note that constructing a relational graph containing nodes such as text paragraphs, events, sentences, and key elements means extracting information units at different levels from the encoded vector representations, such as entire text segments, specific event descriptions, individual sentences, and key elements extracted from sentences (e.g., subject, verb, object). Next, the connections between nodes are defined; for example, the relationship between sentence nodes and event nodes reflects the sentence's description of the event; the connection between sentence nodes and key element nodes indicates that the sentence contains this key information. This constructs a multi-layered graph structure that includes various logical connections within the text. Node representations are then updated using a graph attention mechanism. Graph attention allows each node to assign different weights based on the strength of its relationships with neighboring nodes when updating its representation. This means that in the graph, each node recalculates its feature representation based on information from its surrounding nodes. During this calculation, sentence nodes closely related to key information nodes (such as event descriptions and key elements) are given higher weights, thus enhancing the model's focus on this key information.

[0074] It should be noted that, in terms of graph-based models, this approach extends the widely used GAT model. Nodes are categorized into main event nodes, sentence nodes, word nodes, and global nodes that represent key elements. Compared to the standard GAT method, this embodiment adds global nodes and constructs edges between them and all sentence nodes. This makes each sentence more sensitive to keywords in the main event, reduces the impact of irrelevant noise, increases the importance of sentences related to keywords, and improves model performance.

[0075] Specifically, datasets that utilize grid search to adjust the weights of Chinese and English text to obtain data augmentation include:

[0076] Set weight ratios to represent the importance of Chinese text relative to English text, and create a grid based on the weight ratios, where each cell in the grid corresponds to a specific weight combination;

[0077] For each pair of weights in the grid, the text features are reweighted using the corresponding proportions and the model is trained. After each training iteration, the model's performance is measured using a predetermined evaluation metric on the validation set.

[0078] Select the weight ratio that makes the model perform best from all weight combinations, and adjust the weight ratio of Chinese and English texts according to the weight ratio that makes the model perform best, so as to retain the information of the original English text to the greatest extent.

[0079] It's important to note that a reasonable weight range is determined, such as from 0.1 to 0.9, representing the relative importance of Chinese text to English text. A grid is then created based on these weight values, with each cell corresponding to a specific weight combination. For each pair of weights in the grid, the text features are reweighted using the appropriate ratio, and the model is trained. After each training iteration, a predetermined evaluation metric on the validation set is used to measure the model's performance, such as accuracy or F1 score. Finally, the optimal weight ratio is selected from all weight combinations to maximize the retention of information from the original English text. Once the optimal weights are determined, the model's generalization ability is validated on an independent test set to ensure its reliability and effectiveness.

[0080] In an optional embodiment, the main event is: Xiaoming went fishing today; the key elements are: [Xiaoming], [fishing], [fish]; the text description is: ... Xiaohong did not go fishing today. ...

[0081] In the text description, if [Xiaoming] is not concatenated with 'Xiaohong is not going fishing today' as a node, the simplest BERT model would consider the probability that Xiaoming is not going fishing today to be 70%, because Xiaohong and Xiaoming differ by only one character.

[0082] By using a graph model to concatenate [Xiaoming] as a node with the sentence, the model can fully grasp the information about Xiaoming, while considering the event 'Xiaohong isn't going fishing today' as a useless event. This is just one example. For the entire document, [Xiaoming], [fishing], and [fish] are the three global key elements, concatenated into each node by the graph model. Each node is tightly connected to these key elements. Therefore, these three nodes, as global structural features, are directly obtained before data preprocessing, thus improving model performance.

[0083] In this embodiment of the application, step S300 includes the following sub-steps C1-C3;

[0084] In C1: Nodes and edges in the graph structure are used to capture complex dependencies in the text, and the influence of key information nodes is enhanced through an attention mechanism to obtain a new vector representation that incorporates contextual information.

[0085] In C2: The new vector representation that incorporates contextual information is input into the classification layer for prediction. The classification layer maps each node to one of five classification labels to obtain the detection result of the factuality of the event in the title. The five classification labels are: definitely happened (CT+), possibly happened (PS+), definitely not happened (CT-), highly unlikely to happen (PS-), and unknown (Uu).

[0086] In C3: The cross-entropy loss function is used to measure the gap between the model's predicted values ​​and the true labels, and the AdamW optimizer is used to minimize the loss. The model parameters are adjusted through continuous iterative training to classify the data-augmented dataset and thus determine the detection results of the factuality of the events in the title.

[0087] It should be noted that in this embodiment, multidimensional data augmentation is applied to this task by referencing GPT. By performing extensive preliminary work before data preprocessing, the data features are made more sensitive. Multilingual representation can obtain expressions of the same text with different levels of sensitivity in multiple languages, and the subject, verb, and object, as the most important factors of the main events, are extracted as the main elements of the entire text. Furthermore, the GAT model from the graph model is applied to the method provided in this embodiment, allowing the extracted key information to be combined with the entire text as global nodes, resulting in better model performance and stronger robustness.

[0088] The above is an illustrative scheme of a chapter-level event fact detection method according to this embodiment. It should be noted that the technical solution of this chapter-level event fact detection system belongs to the same concept as the technical solution of the chapter-level event fact detection method described above. For details not described in detail in the technical solution of the chapter-level event fact detection system in this embodiment, please refer to the description of the technical solution of the chapter-level event fact detection method described above.

[0089] This embodiment also provides a chapter-level event fact detection system, including:

[0090] The data acquisition module is used to acquire a corpus containing event descriptions and corresponding body text.

[0091] The information extraction module is used to extract information about major events and key elements from the corpus after translation.

[0092] The multidimensional data augmentation module encodes the extracted bilingual information and uses an improved model based on graph attention networks to connect the encoded bilingual information as global nodes with the main text. The weights of the Chinese and English texts are adjusted through grid search to obtain a data-augmented dataset.

[0093] The event authenticity identification module is used to determine the factual nature of events in the title using data-enhanced datasets.

[0094] The above-mentioned unit modules can be embedded in the processor of the computer device in hardware form or independent of it, or they can be stored in the memory of the computer device in software form, so that the processor can call and execute the corresponding operations of the above modules.

[0095] This embodiment also provides a computing device, which includes a processor, memory, communication interface, display screen, and input device connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. When the computer program is executed by the processor, it implements a chapter-level event fact detection method. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input device can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the device's casing, or an external keyboard, touchpad, or mouse.

[0096] This embodiment also provides a storage medium on which a computer program is stored, which, when executed by a processor, implements the method proposed in the above embodiments.

[0097] The storage medium proposed in this embodiment belongs to the same inventive concept as the method proposed in the above embodiments. Technical details not described in detail in this embodiment can be found in the above embodiments, and this embodiment has the same beneficial effects as the above embodiments.

[0098] Based on the above description of the implementation methods, those skilled in the art can clearly understand that the present invention can be implemented using software and necessary general-purpose hardware, and of course, it can also be implemented using hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as a computer floppy disk, read-only memory (ROM), random access memory (RAM), flash memory, hard disk, or optical disk, etc., including several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods of the various embodiments of the present invention.

[0099] Example 2

[0100] Based on the previous embodiment, this embodiment provides an application example of a chapter-level event fact detection method and system to verify and illustrate the technical effects of the method.

[0101] In this embodiment, the acquired English corpus containing event descriptions, titles, and corresponding body text is first translated to form a Chinese corpus. Major events and key elements are extracted from both the English and Chinese corpora, where key elements are the subject, verbs, and objects extracted from the major events. The extracted bilingual information is then encoded. An improved model based on graph attention networks is used to connect the encoded bilingual information as global nodes to the body text. The weights of the English and Chinese texts are adjusted using grid search to obtain a data-enhanced dataset. After completing the above sub-modules, the cross-entropy loss function is used as the classification loss function, and AdamW is selected as the optimizer. The model is trained and five classification results are obtained on the test set: definitely happening (CT+), possibly happening (PS+), definitely not happening (CT-), highly unlikely not happening (PS-), and unknown (Uu).

[0102] The prediction results were evaluated using the prediction accuracy of the three main categories (definitely CT+, definitely not CT-, and possibly PS+) and five commonly used indicators, including macro F1 and micro F1. The control group consisted of baseline models with good current performance and some basic models. The evaluation data is shown in Table 1. The results show that the method provided in this embodiment has the best effect.

[0103] Table 1: Evaluation data for five indicators;

[0104]

[0105] As described above, this invention obtains key information through data augmentation before data preprocessing to improve the accuracy and efficiency of the model. This is the first time that multidimensional data augmentation has been applied to document-level event fact recognition. Compared with other one-dimensional or no data augmentation methods, it has a stronger grasp of global structural features and text sensitivity. Moreover, this invention uses GPT as the source of multidimensional data augmentation. With continuous technological upgrades, it will gain greater advantages in the accuracy and various indicators of data augmentation in the future. In addition, this invention uses a graph model to connect the extracted key nodes as global nodes to the entire article, which not only improves the generalization and accuracy of the model, but also reduces the model's insensitivity to specific keywords in a certain language by combining Chinese and English.

[0106] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

[0107] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. The solutions in the embodiments of this application can be implemented in various computer languages, such as the object-oriented programming language Java and the interpreted scripting language JavaScript.

[0108] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0109] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0110] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1The steps of the function specified in one or more boxes.

[0111] Although preferred embodiments of this application have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including the preferred embodiments as well as all changes and modifications falling within the scope of this application.

[0112] Obviously, those skilled in the art can make various modifications and variations to this application without departing from the spirit and scope of this application. Therefore, if such modifications and variations fall within the scope of the claims of this application and their equivalents, this application also intends to include such modifications and variations.

Claims

1. A method for detecting factual information about events at the chapter level, characterized in that, include: Obtain a corpus containing event descriptions, titles, and corresponding body text; After translating the corpus, information on major events and key elements is extracted. The extracted bilingual information is encoded, and an improved model based on graph attention network is used to connect the encoded bilingual information as global nodes to the main text. Adjusting the weights of Chinese and English texts using grid search to obtain a data-enhanced dataset; The data-enhanced dataset is used to determine the detection results of the factuality of events in the title; The process of encoding the extracted bilingual information includes: The main events and key elements are extracted from English and Chinese corpora, where the key elements are the subject, verb, and object extracted from the main events. A segmentation strategy is used to segment texts whose length exceeds a first threshold. The Longformer encoder, suitable for long text processing, is selected to encode the pre-extracted and segmented bilingual text, generating a fixed-length vector representation, which is then input into the improved model based on the graph attention network. The improved model based on graph attention networks, which connects the encoded bilingual information as global nodes to the main text, includes: Construct a relationship graph containing the main event node, sentence node, word node, and global nodes that the key elements become, and define the connections between the nodes; An improved model based on graph attention network is used to update the node representation. When updating the representation of each node, the weight is reassigned according to the strength of the relationship between it and its neighboring nodes. That is, sentence nodes that are closely related to the global node will be given higher weights. The updated global node is used as the first node to add sentences in the text in turn, thereby completing the connection with the main text. The dataset used to adjust the weights of Chinese and English text through grid search to obtain data augmentation includes: Set weight ratios to represent the importance of Chinese text relative to English text, and create a grid based on the weight ratios, wherein each cell in the grid corresponds to a specific weight combination; For each pair of weights in the grid, the text features are reweighted using the corresponding proportions and the model is trained. After each training iteration, the model's performance is measured using a predetermined evaluation metric on the validation set. Select the weight ratio that makes the model perform optimal from all weight combinations, and adjust the weight ratio of Chinese and English texts according to the weight ratio that makes the model perform optimal, so as to retain the information of the original English text to the greatest extent. Information extracted from the corpus after translation, including key events and elements, includes: The English corpus containing event descriptions and corresponding text is translated to form a Chinese corpus. The main events and key elements are extracted from the English and Chinese corpora. The key elements are the subject, verb, and object extracted from the main events.

2. The method for detecting factual information of events at the chapter level as described in claim 1, characterized in that, The detection results for determining the factuality of events in the title using the data-enhanced dataset include: By utilizing nodes and edges in a graph structure to capture complex dependencies in text, and by strengthening the influence of key information nodes through an attention mechanism, a new vector representation that integrates contextual information is obtained. The new vector representation that incorporates the contextual information is input into the classification layer for prediction. The classification layer maps each node to one of five classification labels to obtain the detection result of the factuality of the event in the title. The five classification labels are: definitely happened, possibly happened, definitely not happened, highly unlikely to happen, and unknown.

3. The method for detecting factual information of events at the chapter level as described in claim 2, characterized in that, The detection results for determining the factuality of events in the title using the data-enhanced dataset also include: The cross-entropy loss function is used to measure the gap between the model's predicted values ​​and the true labels, and the AdamW optimizer is used to minimize the loss. The model parameters are adjusted through continuous iterative training to classify the data-augmented dataset and thus determine the detection results of the factual nature of the events in the title.

4. A system applying the document-level event fact detection method as described in any one of claims 1 to 3, characterized in that, include: The data acquisition module is used to acquire a corpus containing event descriptions and corresponding body text. The information extraction module is used to extract information about the main events and key elements after translating the corpus. The multidimensional data augmentation module is used to encode the extracted bilingual information and use an improved model based on graph attention network to connect the encoded bilingual information as global nodes with the main text. The weights of the Chinese and English texts are adjusted through grid search to obtain a data-augmented dataset. An event authenticity identification module is used to determine the detection results of the factuality of the event in the title using the data-enhanced dataset.

5. An electronic device, comprising: Memory and processor; The memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions. When the computer-executable instructions are executed by the processor, they implement the steps of the chapter-level event fact detection method according to any one of claims 1 to 3.

6. A computer-readable storage medium storing computer-executable instructions that, when executed by a processor, implement the steps of the chapter-level event fact detection method according to any one of claims 1 to 3.

Images