Event extraction method and device and electronic equipment

Abstract

The invention discloses an event extraction method and device and electronic equipment. The method comprises the following steps: acquiring a communication fault processing text to be subjected to event extraction; an event extraction model is adopted to determine a first semantic vector and a second semantic vector of the communication fault processing text, the first semantic vector is used for representing a word vector corresponding to each vocabulary in the communication fault processing text, and the second semantic vector is used for representing a word vector corresponding to each vocabulary in the communication fault processing text; the second semantic vector is used for representing a semantic vector fusing the first semantic vector and a dependency relationship between vocabularies in the communication fault processing text; and fusing the first semantic vector and the second semantic vector to obtain a third semantic vector, and generating an event extraction result corresponding to the communication fault processing text according to the third semantic vector. According to the event extraction method and device, the technical problem that the event extraction effect is poor due to the fact that an event extraction method based on deep learning is less applied to the communication field and syntactic structures and dependency relationships between texts are not fully considered in the related technologies is solved.

Description

Technical Field

[0001] The present application relates to the field of communication fault processing, and in particular, to an event extraction method, device and electronic equipment. Background Art

[0002] As communication technology becomes increasingly complex and the scale of networks continues to expand, it is particularly important to ensure the stable operation of communication systems. As a result of practical experience and accumulated skills, communication fault handling knowledge covers key links such as equipment installation, network debugging, troubleshooting and repair, and is an important basis for operation and maintenance personnel to quickly locate and solve faults. However, this knowledge often exists in the form of unstructured text, such as technical manuals, operation logs, fault reports, etc., which makes it difficult to efficiently retrieve and use the valuable information contained therein. Therefore, automatically extracting communication fault handling events and converting them into structured data is of great significance for building an intelligent operation and maintenance system and improving fault handling efficiency.

[0003] Event extraction technology plays an important role in the field of natural language processing, aiming to extract key event information from text, including event types, trigger words, and argument roles. Historically, event extraction methods have evolved from rule-based, machine learning-based, and deep learning-based. Rule-based methods, such as the AutoSlog system, match information through predefined dictionaries and templates. Although they can achieve high accuracy, they rely on expert knowledge and have poor flexibility and scalability. Subsequently, machine learning models, such as CRF (Conditional Random Field) learning text features to overcome the limitations of rule-based methods, are still insufficient in the depth and breadth of feature extraction, resulting in limited accuracy in event extraction.

[0004] In recent years, the rise of deep learning, especially pre-trained language models, has provided a new solution for event extraction. The introduction of the BERT (Bidirectional Encoder Representations from Transformers) model has significantly enhanced the semantic expression ability of word vectors, enabling event extraction to understand the meaning of text more accurately. However, in the vertical field of communication fault handling, the application of deep learning methods is still in its infancy, and related research often ignores the rich syntactic structure and inter-lexical dependencies in fault handling texts, which are crucial for understanding the context of events.

[0005] Therefore, although the event extraction technology in related technologies has made significant progress in general scenarios, it still has problems such as insufficient adaptability and insufficient utilization of syntactic structure information in the vertical field of communication fault handling.

[0006] To address the above-mentioned problems, no effective solution has been proposed yet. Summary of the invention

[0007] The embodiments of the present application provide an event extraction method, device and electronic device to at least solve the technical problem that the event extraction method based on deep learning in the related technology is rarely used in the communication field and does not fully consider the syntactic structure and dependency relationship between texts, resulting in poor event extraction effect.

[0008] According to one aspect of an embodiment of the present application, an event extraction method is provided, including: obtaining a communication fault processing text for event extraction; using an event extraction model to determine a first semantic vector and a second semantic vector of the communication fault processing text, wherein the first semantic vector is used to represent a word vector corresponding to each vocabulary in the communication fault processing text, and the second semantic vector is used to represent a semantic vector that integrates the dependency relationship between the first semantic vector and each vocabulary in the communication fault processing text; integrating the first semantic vector and the second semantic vector to obtain a third semantic vector, and generating an event extraction result corresponding to the communication fault processing text based on the third semantic vector.

[0009] Optionally, during the training phase of the event extraction model, data preprocessing is performed on the historical communication fault processing text, including: obtaining text information that is manually annotated on the historical communication fault processing text, wherein the text information includes the communication fault processing text and annotation information, and the annotation information includes the trigger word position and argument position that are manually annotated on the communication fault processing text; comparing the text information with the communication fault processing text to obtain a comparison result; when the comparison result indicates that the annotation information in the text information is inconsistent with the initial position information in the communication fault processing text, determining that there is an offset in the trigger word position or argument position in the annotation information, and correcting the trigger word position or argument position in the label information, wherein the initial position information is used to indicate the trigger word position and argument position in the communication fault processing text corresponding to the manually annotated area, the trigger word is used to indicate the key vocabulary for the occurrence of an event in the historical communication fault processing text, and the argument is used to indicate the entity or concept involved when the event in the historical communication fault processing text occurs.

[0010] Optionally, an event extraction model is used to determine the first semantic vector of the communication fault processing text, including: processing the communication fault processing text through the first model in the event extraction model to obtain an initial word vector corresponding to each word in the communication fault processing text, wherein the initial word vector includes position information and fragment information of each word in the communication fault processing text; generating a query vector, a key vector and a value vector corresponding to the initial word vector through a multi-head attention mechanism; and determining the first semantic vector based on the query vector, the key vector and the value vector.

[0011] Optionally, an event extraction model is used to determine the second semantic vector of the communication fault processing text, including: performing dependency syntactic analysis on the communication fault processing text through the event extraction model to obtain the dependency relationship between each word in the communication fault processing text; determining the dependency matrix of the communication fault processing text based on the dependency relationship; and fusing the dependency matrix and the first semantic vector through a graph convolutional neural network to obtain a second semantic vector.

[0012] Optionally, determining a dependency matrix of a communication fault handling text based on dependencies includes: determining a first word vector and a second word vector of the communication fault handling text, wherein the first word vector and the second word vector are arbitrary vectors in a first semantic vector; determining a target dependency between the first word vector and the second word vector; determining a dependency weight between the first word vector and the second word vector based on the target dependency and a preset dependency embedding vector; and determining a dependency matrix based on the dependency weight.

[0013] Optionally, the first semantic vector and the second semantic vector are fused to obtain a third semantic vector, including: processing the first semantic vector and the second semantic vector through a bidirectional long short-term memory network to obtain a forward output vector and a reverse output vector corresponding to the communication fault processing text; and splicing the forward output vector and the reverse output vector through a splicing function to obtain the third semantic vector.

[0014] Optionally, an event extraction result corresponding to the communication fault processing text is generated based on the third semantic vector, including: determining the label score of the third semantic vector through the second model in the event extraction model, wherein the label score is used to represent the degree of match between each word in the communication fault processing text and the corresponding label information; determining the label sequence of the communication fault processing text based on the label score and the transition probability between the label scores; determining the target label sequence based on the label sequence and the maximum likelihood function, wherein the target label sequence is the optimal label sequence in the label sequence; determining the event extraction result based on the target label sequence, wherein the event extraction result is used to represent the structured event description in the target label sequence.

[0015] According to another aspect of an embodiment of the present application, an event extraction device is also provided, including: an acquisition module, used to acquire a communication fault processing text to be subjected to event extraction; a determination module, used to determine a first semantic vector and a second semantic vector of the communication fault processing text using an event extraction model, wherein the first semantic vector is used to represent a word vector corresponding to each vocabulary in the communication fault processing text, and the second semantic vector is used to represent a semantic vector that integrates the dependency relationship between the first semantic vector and each vocabulary in the communication fault processing text; a generation module, used to integrate the first semantic vector and the second semantic vector to obtain a third semantic vector, and generate an event extraction result corresponding to the communication fault processing text based on the third semantic vector.

[0016] According to another aspect of the embodiment of the present application, there is also provided an electronic device, including: a memory and a processor, wherein the memory is used to store program instructions; the processor is connected to the memory and is used to execute the above-mentioned event extraction method.

[0017] According to another aspect of the embodiments of the present application, a non-volatile storage medium is provided, the non-volatile storage medium includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the above-mentioned event extraction method by running the computer program.

[0018] According to another aspect of the embodiments of the present application, a computer program product is provided, including computer instructions, which implement the above event extraction method when executed by a processor.

[0019] In an embodiment of the present application, a communication fault processing text for event extraction is obtained; an event extraction model is used to determine a first semantic vector and a second semantic vector of the communication fault processing text, wherein the first semantic vector is used to represent a word vector corresponding to each vocabulary in the communication fault processing text, and the second semantic vector is used to represent a semantic vector that integrates the first semantic vector and the dependency relationship between each vocabulary in the communication fault processing text; the first semantic vector and the second semantic vector are integrated to obtain a third semantic vector, and an event extraction result corresponding to the communication fault processing text is generated based on the third semantic vector, thereby achieving the purpose of structuring the key information in the communication fault processing text, thereby achieving the technical effect of improving the accuracy and efficiency of event extraction, and further solving the technical problem that the event extraction method based on deep learning in the related technology is less used in the communication field and does not fully consider the syntactic structure and dependency relationship between texts, resulting in poor event extraction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0021] Figure 1 is a hardware structure diagram of a computer terminal for implementing an event extraction method according to an embodiment of the present application;

[0022] Figure 2 is a flow chart of an event extraction method according to an embodiment of the present application;

[0023] Figure 3 is a schematic diagram of the overall architecture of an event extraction model according to an embodiment of the present application;

[0024] Figure 4is a schematic diagram of a data preprocessing process according to an embodiment of the present application;

[0025] Figure 5 It is a structural diagram of an event extraction device according to an embodiment of the present application. DETAILED DESCRIPTION

[0026] In order to enable those skilled in the art to better understand the solution of the present application, the technical solution in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present application.

[0027] It should be noted that the terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable where appropriate, so that the embodiments of the present application described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0028] First, some nouns or terms that appear in the process of explaining the embodiments of the present application are subject to the following explanations:

[0029] Communication fault handling: In a communication system, when situations such as network interruption, signal interference, hardware or software failure occur that lead to abnormal communication services, communication fault handling is a set of systems or processes used to diagnose, repair and restore the normal operation of communication services to ensure the stability and reliability of the communication network.

[0030] Event extraction: A subtask in natural language processing that aims to automatically identify and extract specific types of event information from unstructured text data, including event types, trigger words (sign words of events), entities involved in events (i.e. arguments) and their roles in events. Event extraction is widely used in information extraction, monitoring systems, automatic summarization and other fields.

[0031] Dependency Parsing: A syntactic analysis technique used to identify dependencies between words in a sentence, that is, how words depend on other words in terms of grammar and meaning. Dependency parsing is usually represented in the form of a directed graph, where each word is a node of the graph and the dependency relationships are expressed as directed edges connecting these nodes, which is used to reveal the deep semantic structure of the sentence.

[0032] BERT (Bidirectional Encoder Representations from Transformers): A pre-trained deep learning model proposed by Google in 2018 based on the Transformer architecture. BERT is trained in a bidirectional manner and is able to consider all words in the context to generate high-quality word vectors. This model performs well in a variety of natural language processing tasks, including question answering, text classification, and event extraction.

[0033] GCNs (Graph Convolutional Networks): A deep learning model specifically designed to process graph-structured data, such as social networks, molecular structure graphs, etc. In natural language processing tasks, GCNs can be used to analyze the dependency structure of sentences and update the representation of each node by aggregating the information of neighboring nodes, thereby capturing the dependency relationship between words and deeper semantic information.

[0034] CRF (Conditional Random Field): A statistical modeling method commonly used in sequence labeling and structured prediction tasks. CRF can consider the dependencies between labels and predict the labels of the entire sequence through global optimization, avoiding errors caused by local optimal solutions, and is very suitable for tasks such as event extraction.

[0035] Bi-LSTM (Bidirectional Long Short-Term Memory): It can process sequence data from both the forward and backward directions, thereby capturing long-distance dependencies in the sequence. In event extraction, Bi-LSTM can enhance the contextual semantic representation of the text, allowing the model to understand the meaning of the sentence more accurately.

[0036] BIOSE label construction method: a label system for sequence annotation, where B represents the start label, I represents the internal label, O represents the non-event label, S represents the event trigger word or argument label composed of a single character, and E represents the end label. Compared with the traditional BIO system, BIOSE can more accurately locate the boundaries of trigger words and arguments and reduce misidentification.

[0037] In order to solve the problem of low event extraction efficiency in the related art, the present application embodiment provides an event extraction method, which can be run on Figure 1 In the computer terminal shown, the computer terminal is described below.

[0038] The event extraction method embodiment provided in the embodiment of the present application can be executed in a mobile terminal, a computer terminal or a similar computing device. Figure 1 FIG. 1 shows a hardware structure block diagram of a computer terminal for implementing an event extraction method. Figure 1 As shown, the computer terminal 10 may include one or more (102a, 102b, ..., 102n are used to illustrate) processors (the processor may include but is not limited to a processing device such as a microprocessor MCU or a programmable logic device FPGA), a memory 104 for storing data, and a transmission module 106 for communication functions connected via a wired and / or wireless network. In addition, it may also include: a display, a keyboard, a cursor control device, an input / output interface (I / O interface), a universal serial bus (USB) port (which may be included as one of the ports of the I / O interface), a network interface, and a BUS bus. Those skilled in the art can understand that Figure 1 The structure shown is only for illustration and does not limit the structure of the above electronic device. Figure 1 More or fewer components as shown, or with Figure 1 Different configurations shown.

[0039] It should be noted that the one or more processors and / or other data processing circuits described above may generally be referred to herein as "data processing circuits". The data processing circuits may be embodied in whole or in part as software, hardware, firmware, or any other combination thereof. In addition, the data processing circuit may be a single independent processing module, or may be incorporated in whole or in part into any of the other components in the computer terminal 10. As described in the embodiments of the present application, the data processing circuit acts as a processor control (e.g., selection of a variable resistor terminal path connected to an interface).

[0040] The memory 104 can be used to store software programs and modules of application software, such as the program instructions / data storage device corresponding to the event extraction method in the embodiment of the present application. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory 104, that is, realizing the above-mentioned event extraction method. The memory 104 may include a high-speed random access memory, and may also include a non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some instances, the memory 104 may further include a memory remotely arranged relative to the processor, and these remote memories may be connected to the computer terminal 10 via a network. Examples of the above-mentioned network include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0041] The transmission module 106 is used to receive or send data via a network. The specific example of the above network may include a wireless network provided by a communication provider of the computer terminal 10. In one example, the transmission module 106 includes a network adapter (Network Interface Controller, NIC), which can be connected to other network devices through a base station so as to communicate with the Internet. In one example, the transmission module 106 can be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0042] The display may be, for example, a touch screen liquid crystal display (LCD) that enables a user to interact with a user interface of the computer terminal 10 .

[0043] It should be noted that, in some optional embodiments, the above Figure 1 The computer terminal shown may include hardware elements (including circuits), software elements (including computer code stored on a computer-readable medium), or a combination of hardware elements and software elements. It should be noted that Figure 1 This is merely one example of a particular embodiment and is intended to illustrate the types of components that may be present in the computer terminal described above.

[0044] In the above-mentioned operating environment, an embodiment of the present application provides an event extraction method embodiment. It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a set of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described can be executed in an order different from that shown here.

[0045] Figure 2 is a flow chart of an event extraction method according to an embodiment of the present application. Figure 2 As shown, the method comprises the following steps:

[0046] Step S202, obtaining the communication fault processing text for event extraction.

[0047] In the above step S202, the acquisition of the communication fault processing text is the basis of the entire event extraction process. The communication fault processing text may come from other files such as operation and maintenance logs, fault reports and technical manuals.

[0048] Step S204, using an event extraction model to determine a first semantic vector and a second semantic vector of the communication fault processing text, wherein the first semantic vector is used to represent a word vector corresponding to each vocabulary in the communication fault processing text, and the second semantic vector is used to represent a semantic vector that integrates the dependency relationship between the first semantic vector and each vocabulary in the communication fault processing text.

[0049] In the above step S204, the text is encoded by the BERT model pre-trained in the event extraction model DPEE, and the first semantic vector that incorporates the contextual semantics can be generated for each word. Among them, through the multi-layer multi-head self-attention mechanism, BERT can capture the complex semantic structure of the words in the communication fault processing text, providing a basis for the subsequent fusion of dependency syntactic information. After obtaining the word vector, the DPEE model further performs dependency syntactic analysis on the text to construct a dependency weight matrix between words. By defining multiple dependency types and using a trainable parameter matrix to dynamically adjust the weights, the model can learn deeper inter-word relationships. Then, the graph convolutional neural network (GCNs) is used to fuse these dependency information to generate a second semantic vector, wherein the second semantic vector not only contains the word vector information, but also integrates the syntactic structure and dependency between words, which is the key to understanding the contextual relationship of communication fault processing events.

[0050] It should be noted that the above event extraction model (DPEE) comprehensively uses advanced technologies and deep learning models such as BIOSE sequence annotation label construction method, BERT model, dependency syntax analysis, graph convolutional neural network (GCNs), bidirectional long short-term memory network (Bi-LSTM) and conditional random field (CRF) to process communication fault processing text in a multi-stage fusion manner. The overall architecture is as follows: Figure 3As shown. In the overall event extraction method (taking the event of abnormal power supply of the equipment caused by the power supply module failure as an example), data preprocessing and label construction are first performed, among which label construction is only performed in the model training stage; secondly, the BERT model provides the initial word vector to capture the contextual semantics of the word; dependency syntactic analysis reveals the structural relationship between words, and further integrates the dependency information through GCNs to enhance the semantic representation of the word vector; then, the Bi-LSTM network strengthens the interaction of contextual information and generates a more context-rich semantic vector; finally, CRF ensures the global optimal solution of the event label sequence, thereby obtaining accurate event extraction results (optimal solution label sequence). The entire model effectively improves the accuracy and efficiency of event extraction in the communication field through meticulous data preprocessing, multi-dimensional semantic information fusion and accurate label prediction.

[0051] Step S206: The first semantic vector and the second semantic vector are fused to obtain a third semantic vector, and an event extraction result corresponding to the communication fault processing text is generated according to the third semantic vector.

[0052] In the above step S206, the DPEE model uses a bidirectional long short-term memory network (Bi-LSTM) to re-encode the first semantic vector and the second semantic vector. Through the forget gate, input gate and output gate mechanism of Bi-LSTM, the model can more effectively integrate contextual information and generate a third semantic vector that combines word vectors and dependency syntactic information. Furthermore, based on the generated third semantic vector, the DPEE model uses conditional random fields (CRF) to predict event label sequences. Among them, CRF can capture the dependencies between event labels, thereby obtaining a globally optimal label sequence during the decoding process. Finally, based on the predicted label sequence, the model can generate event extraction results corresponding to the communication fault processing text, including trigger words, trigger word types, arguments and their role types, etc.

[0053] Through the above steps S202 to S206, the purpose of structuring the key information in the communication fault processing text is achieved, thereby achieving the technical effect of improving the accuracy and efficiency of event extraction, and further solving the technical problem that the event extraction method based on deep learning in the related technology is rarely used in the communication field and does not fully consider the syntactic structure and dependency relationship between texts, resulting in poor event extraction effect. The following is a detailed description.

[0054] Optionally, the above method also includes performing data preprocessing on the historical communication fault processing text during the training phase of the event extraction model, including: obtaining text information that is manually annotated on the historical communication fault processing text, wherein the text information includes the communication fault processing text and the annotation information, and the annotation information includes the trigger word position and argument position that are manually annotated on the communication fault processing text; comparing the text information with the communication fault processing text to obtain a comparison result; when the comparison result indicates that the annotation information in the text information is inconsistent with the initial position information in the communication fault processing text, determining that there is an offset in the trigger word position or argument position in the annotation information, and correcting the trigger word position or argument position in the label information, wherein the initial position information is used to indicate the trigger word position and argument position in the communication fault processing text corresponding to the manually annotated area, the trigger word is used to indicate the key vocabulary of the event occurrence in the historical communication fault processing text, and the argument is used to indicate the entity or concept involved when the event occurs in the historical communication fault processing text.

[0055] In the embodiment of the present application, during the training phase of the DPEE model, especially for the data preprocessing of the communication fault handling text, ensuring the quality of the input data is crucial to the accuracy and reliability of the model. Figure 4 As shown in the figure, it mainly includes text cleaning, position offset detection and correction, English component unification, and special character replacement of spaces and line breaks. First, when the DPEE model is in the training stage, the position offset detection and correction algorithm is used to automatically locate and correct the deviation between the trigger word or argument position in the training data and the manually marked position, ensuring that the model training is based on accurate position information. Secondly, the case of English words in the text is unified to enhance the consistency and generalization ability of the model in processing English components. It should be noted that if the DPEE model is not in the model training stage at this time, the English letter format of the sample can be directly unified, and spaces and line breaks can be replaced and data missing can be deleted. Furthermore, special characters are used to replace spaces and line breaks in the text to avoid the secondary position offset problem that may be caused by directly deleting these characters. This series of preprocessing measures provides high-quality input data for the subsequent core steps of the DPEE model, such as context representation learning, dynamic dependency syntactic information construction, and event label prediction, and is a key step to improve the accuracy of communication fault processing event extraction and model performance.

[0056] Among them, through the position offset detection and correction algorithm, the deviation of the trigger word or argument position in the training data can be automatically located and corrected, which can ensure that the model can learn from high-quality labeled data during the training phase and avoid training deviations caused by position offsets. Through continuous correction and learning, the model can gradually improve the accuracy of trigger word and argument position recognition, so that it can perform more stably and reliably in event extraction tasks. This method is not only applicable to communication fault processing texts, but also has reference significance for event extraction tasks in other fields, which can improve the generalization ability and extraction effect of the model. The specific process can be as follows:

[0057] First, before model training, a set of manually annotated communication fault processing text datasets is required. These datasets contain real event information of the text, such as trigger words, argument information, trigger words, argument types and role information, and their exact positions in the text. This real event information is the benchmark for model training and is used to guide the model to learn to correctly identify and locate event information.

[0058] During the training phase of the model, the historical communication fault handling text will be input into the DPEE model. The DPEE model will match the text fragment corresponding to the manually annotated area in the original text based on the position offset and correction algorithm combined with the manually annotated label information. The text fragment contains the initial position information of the trigger words and arguments in the historical communication fault handling text.

[0059] Secondly, the manually annotated trigger words and arguments are compared with the matching text fragments in the original text to obtain the comparison results. The comparison results will indicate whether the manually annotated trigger words and arguments are accurate, as well as possible position deviations. If the comparison results show that the manually annotated trigger words or arguments do not match the corresponding position fragments of the original text, this usually means that there is a deviation in the manually annotated trigger words and arguments position information.

[0060] Finally, if the manually annotated trigger words and arguments are confirmed to have offsets, the trigger words or arguments are corrected. Specifically, a position offset detection and correction algorithm is used to retrieve text segments that are consistent with the trigger words and arguments in the original text and record the position information of the segments. The position information with the smallest distance between the manually annotated position and the retrieved position is used as the trigger word and argument position information to be corrected. Finally, the algorithm will update the manually annotated trigger words or arguments to correct the position offset.

[0061] In addition, in the embodiment of the present application, the DPEE model uses the BIOSE label construction method to construct event extraction labels. After pre-processing, each training sample is annotated using the BIOSE system to form sequence labels. These sequence labels will serve as guiding signals for the model, and can help the model learn how to identify and locate trigger words and their types, arguments and their roles. The model optimizes its own parameters by minimizing the gap between the predicted label and the actual BIOSE label, with the goal of most accurately predicting the correct label for each word. During the model iterative training process, the BIOSE label system ensures that the model can make more accurate predictions from the beginning to the end of the vocabulary, thereby improving the integrity and accuracy of communication fault handling event extraction.

[0062] Specifically, BIOSE is an extended sequence annotation system, in which B (Begin) indicates the beginning of an element, I (Inside) indicates the inside of an element, O (Outside) indicates other text that is not an event or entity, S (Single) is used to identify trigger words or arguments consisting of a single character, and E (End) is used to clearly identify the end of an element. Compared with the traditional BIO system, BIOSE adds two labels, S and E, which help to more accurately locate the boundaries of trigger words and arguments and reduce the boundary misrecognition rate.

[0063] For the label construction process, for example, for each independent event, the positions of its trigger words and arguments in the text are first determined, where the trigger words are the iconic words of the event and the arguments are the entities involved in the event. In the process of extracting trigger words and arguments, the trigger word event type and argument role are classified at the same time. The event type and argument role are shown in Table 1.

[0064] For example, for a trigger word, if it consists of multiple characters, its first character is marked as B-softhardwarefault (or other event types), the inner characters are I-softhardwarefault, and the last character is E-softhardwarefault.

[0065] For trigger words consisting of a single character, S-softhardwarefault is used for marking.

[0066] The labeling of arguments follows the same principle, using B, I, E, or S, suffixed with the argument role, such as B-subject, I-subject, E-subject, or S-subject.

[0067] Parts of the text that do not belong to any trigger word or argument are marked with O.

[0068] Table 1 Summary of communication failure event types and argument roles

[0069]

[0070]

[0071] In the above step S204, an event extraction model is used to determine the first semantic vector of the communication fault processing text, including: processing the communication fault processing text through the first model in the event extraction model to obtain an initial word vector corresponding to each word in the communication fault processing text, wherein the initial word vector includes the position information and fragment information of each word in the communication fault processing text; generating a query vector, a key vector and a value vector corresponding to the initial word vector through a multi-head attention mechanism; and determining the first semantic vector based on the query vector, the key vector and the value vector.

[0072] In the embodiment of the present application, determining the word vector corresponding to each word in the communication fault text (i.e., the first semantic vector mentioned above) is crucial for the DPEE model, because it ensures that the model can extract a representation that incorporates rich contextual information from the communication fault handling text, and provides high-quality semantic features for the construction of dynamic dependency syntactic information and event extraction in subsequent steps. By combining BERT pre-training knowledge with the contextual information of specific tasks, the DPEE model can more accurately identify trigger words and arguments, thereby improving the extraction effect of fault handling events. The specific process can be as follows:

[0073] First, the communication fault handling text is processed using the pre-trained BERT model. The BERT model is a deep learning model based on the Transformer architecture. It can handle long-distance dependencies through a multi-layer multi-head attention mechanism to better understand the context. Before entering the BERT model, each word in the communication fault handling text includes not only its own embedding vector, but also position information and fragment information. The position information is used to indicate the relative position of the word in the sentence, and the fragment information helps the model distinguish different parts of the text. These two pieces of information are added to the embedding vector of the word to form the initial word vector, which is used as the input of the BERT encoding and is recorded as

[0074] Secondly, combined with the multi-head attention mechanism in the BERT model, the initial word vector is mapped into a query vector, a key vector, and a value vector, which are respectively denoted as Q = X emb W Q , K = X emb W K 、V=X emb W V , and W Q , W K , Among them, the multi-head attention mechanism allows the model to pay attention to information at different positions at the same time, so as to consider the relationship between words more comprehensively; at the same time, the BERT model can accurately calculate the attention weights between words, which reflect the relative importance of words in the sentence and the potential connection between them. Specifically, the query vector is used to represent the "question" information, the key vector is used to represent the "response" information, and the value vector contains the substantive information of the vocabulary. The multi-head attention mechanism decomposes these vectors into different heads, each head calculates the attention weight independently, and finally splices the results of all heads together, and further maps them through a feedforward neural network to obtain a richer word vector representation.

[0075] Finally, after the encoding process of the 12-layer multi-head self-attention mechanism, the BERT model can generate word vectors that incorporate contextual semantic information. These word vectors are the first semantic vectors. Then the corresponding word vector sequence can be expressed as E = {e 1 ,e 2 …e n-1 ,e n The first semantic vector not only contains the original information of each word, but also takes into account its relative position in the sentence, fragment information, and dependency information with other words. It is an important input feature for event extraction.

[0076] The specific calculation formula is as follows:

[0077] E=BERT(X)

[0078] head i =Attention(QW i Q ,KW i K ,VW i V )

[0079]

[0080] MutiHead(Q,K,V)=concat(head 1 ,…,head 12 )W O

[0081] Among them, the key steps of calculating the word vector E in the first formula can be decomposed into the following three formulas.

[0082] In the formula, head i Denotes the output of the i-th self-attention head by combining Q, K, V with the corresponding head weight matrix W i Q , W iK , W i V After multiplication, attention calculation is performed. Represents the calculation formula of the self-attention mechanism, where is the query, key, and value vector after linear mapping. The softmax function is used to convert the dot product result into a probability distribution, and finally multiply it with the value vector to get the attention output; d k =d mod el ÷N, N represents the number of self-attention heads, here it is 12. MutiHead(Q,K,V) represents the output of multi-head self-attention. The outputs of 12 self-attention heads are concatenated through the concatenation function concat, and then the output weight matrix W is used. O Perform a linear transformation.

[0083] Furthermore, an event extraction model is used to determine the second semantic vector of the communication fault processing text, including: performing dependency syntactic analysis on the communication fault processing text through the event extraction model to obtain the dependency relationship between each word in the communication fault processing text; determining the dependency matrix of the communication fault processing text based on the dependency relationship; and fusing the dependency matrix and the first semantic vector through a graph convolutional neural network to obtain a second semantic vector.

[0084] Optionally, determining a dependency matrix of a communication fault handling text based on dependencies includes: determining a first word vector and a second word vector of the communication fault handling text, wherein the first word vector and the second word vector are arbitrary vectors in a first semantic vector; determining a target dependency between the first word vector and the second word vector; determining a dependency weight between the first word vector and the second word vector based on the target dependency and a preset dependency embedding vector; and determining a dependency matrix based on the dependency weight.

[0085] In an embodiment of the present application, in the DPEE model, the conversion from the first semantic vector to the second semantic vector is essentially to enhance the vocabulary embedding representation of the communication fault handling text from a vector containing only contextual semantic information to a richer representation that integrates syntactic structure and dependencies between words. Through dependency syntactic analysis and dynamic dependency syntactic information fusion of GCNs, the DPEE model can better understand the grammatical structure of sentences and the interaction between words, thereby providing more accurate and comprehensive semantic features for subsequent event extraction tasks. This enhanced vocabulary representation helps the model to more effectively utilize syntactic information when extracting event trigger words and arguments, and improve the accuracy and coverage of event extraction, especially in the specific field of communication fault handling, where syntax and dependencies are crucial to understanding event logic and key elements. The specific process can be as follows:

[0086] First, the communication fault handling text is subjected to dependency syntactic analysis, which can be done using tools such as spacy, with the goal of identifying the role of each word in the sentence structure and the dependency relationships between them. Dependency relationships can be direct objects (dobj), nominal subjects (nsubj), etc. Each dependency relationship corresponds to a specific type, as shown in Table 2, which will provide a basis for the subsequent construction of dynamic dependency syntactic information.

[0087] Table 2 Inter-word dependency relationship type table

[0088]

[0089] Second, for each pair of words in the communication failure processing text, we embed them according to their syntactic dependencies and their corresponding dependency embedding vectors (e.g., pre-defined trainable parameter matrices As the initial embedding vector of the dependency relationship), calculate the dependency weights between them. Among them, each element in the weight matrix represents the dependency strength between words. By adaptively learning the parameter matrix of the neural network, the noise component in the dependency relationship can be reduced. The obtained weight matrix needs to be further normalized to ensure the correct distribution of weights, so as to obtain the final dynamic dependency matrix. The specific calculation method can be as follows:

[0090] For the word μ in the communication fault processing text, i … w j} and word ν={w k … w m}, the corresponding word embedding representations obtained through the BERT model are the first word vector E μ ={e i … e j} and the second word vector E ν ={e k … e m}, if the word μ There is a syntactic dependency relationship between word ν and word ν, namely the target dependency relationship mentioned above, and the corresponding dependency embedding vector is The word μ The dependency weight α between α and word ν is calculated as follows:

[0091]

[0092] In the formula, W 2 ∈R 256×1 , b 1 ∈R 1×256 , b 2 ∈R 1×1are all learnable parameters; i, j, f, k are all different position indexes, e p Represents the embedding vector of the pth word in the word vector; Sigmoid and Gelu are activation functions, and the Sigmoid activation function limits the dependency weight α to between 0 and 1.

[0093] For each word in word μ and word ν, the shared dependency weight α is calculated. Then, after all words in the sentence are analyzed by dependency syntactic analysis, the dynamic dependency matrix A∈R can be obtained. n×n , then the dependency matrix A is normalized, and the specific processing formula can be as follows:

[0094]

[0095] In the formula, I represents the unit matrix, which is used to increase the self-loop of the word itself; D is a diagonal matrix, which is used to keep the influence of each node on the neighboring nodes at the same order of magnitude, so as to obtain the final dependency matrix The calculation formula of D is as follows:

[0096]

[0097] In the formula, Represents the sum of all elements in the i-th row of the dependency matrix.

[0098] Finally, based on the obtained dynamic dependency matrix Multi-layer graph convolutional neural networks (GCNs) can be used to fuse the word vectors (i.e., the first semantic vector) and dependency information generated by BERT in the first stage. GCNs can propagate information in the graph structure and integrate the dependency weights in the adjacency matrix by iteratively updating the features of the nodes. Each layer of GCNs updates the word vector so that it not only contains the original context information but also integrates the dependencies between words. This fusion process generates a new word vector representation, i.e., the second semantic vector. The specific expression is as follows:

[0099]

[0100] In the formula, H l represents the output of the lth layer of the graph convolutional neural network, H l-1 represents the output of the l-1th layer of the graph convolutional neural network, W l-1 represents the weight matrix of the l-1th layer of the graph convolutional neural network, b l-1 Represents the bias of the l-1th layer of the graph convolutional neural network.

[0101] In the above step S206, the first semantic vector and the second semantic vector are fused to obtain a third semantic vector, including: processing the first semantic vector and the second semantic vector through a bidirectional long short-term memory network to obtain a forward output vector and a reverse output vector corresponding to the communication fault processing text; and splicing the forward output vector and the reverse output vector through a splicing function to obtain the third semantic vector.

[0102] In the embodiment of the present application, through the bidirectional processing of the Bi-LSTM network and the splicing of the positive and negative output vectors, the generated third semantic vector can better capture the actual semantic and syntactic roles of the vocabulary in the communication fault processing text, providing a higher quality word vector representation for the event extraction task, thereby improving the event extraction accuracy and effect of the model. The specific process can be as follows:

[0103] First, the Bi-LSTM network receives word vectors from two sources and concatenates them as input, namely the first semantic vector and the second semantic vector mentioned above. Among them, the first semantic vector is obtained by processing the communication fault processing text through the BERT model. This vector carries the semantic information of the word in the context and can reflect the meaning of the word and its relationship with the surrounding words. The second semantic vector is a word vector that integrates dynamic dependency syntactic information through graph convolutional neural networks (GCNs), capturing the dependency relationship and syntactic structure between words. These two vectors describe the semantic features of vocabulary from different perspectives, providing a basis for fusion.

[0104] Secondly, the forward output vector and the reverse output vector are determined by a bidirectional long short-term memory network (Bi-LSTM). Among them, Bi-LSTM is a special type of LSTM, which can not only process sequences from front to back (forward), but also process sequences from back to front (reverse), thereby capturing the bidirectional contextual information of words in the text sequence. In the DPEE model, the Bi-LSTM network receives the first semantic vector and the second semantic vector as input. Through bidirectional processing, it can more comprehensively understand the role of vocabulary in the text. The forward LSTM reads the text from left to right, updates its internal state word by word, and integrates the future contextual information of each word into the word vector; while the reverse LSTM reads the text from right to left and integrates the past contextual information of each word into the word vector. Such bidirectional processing enables the model to make full use of the previous and subsequent context information of the word to generate a more complete and accurate word vector representation. That is, after being processed by the Bi-LSTM network, each word will get two output vectors: one is the output vector of the forward LSTM, and the other is the output vector of the reverse LSTM. These two vectors contain the semantic information of the words in the positive and negative directions, as well as the dependency syntactic information fused through the first and second semantic vectors. Such an output vector has stronger context dependency and syntactic structure, and can more accurately reflect the role and meaning of words in the communication fault handling text.

[0105] It should be noted that the Bi-LSTM network controls the selection and output of semantic information through the forget gate, input gate, and output gate, thereby enhancing the ability of contextual semantic information interaction. The forget gate determines which historical information should be discarded; the input gate determines which information should be stored in the cell state; and the output gate determines which information should be output from the cell state. Through these gating mechanisms, the Bi-LSTM can effectively process long sequence dependencies, thereby more finely encoding and integrating the contextual information of the communication fault processing text.

[0106] Finally, the DPEE model uses a concatenation function to concatenate the forward output vector and the reverse output vector in the word vector dimension to generate a third semantic vector. The third semantic vector integrates the direct context semantics, dependency, and bidirectional context information of the word, and is a complete semantic representation of the word in the communication fault processing text. This enhanced word vector representation is crucial for subsequent event extraction tasks because it not only considers the meaning of the word, but also the grammatical structure of the word in the sentence and the relationship with other words, so that it can more accurately identify event trigger words and arguments and improve the overall performance of event extraction.

[0107] In the above process, the encoding formula involved is as follows:

[0108] f t =Sigmoid(W f·[h t-1 ,x t ]+b f )

[0109] i t =Sigmoid(W i ·[h t-1 ,x t ]+b i )

[0110] c t =f t ·c t-1 +i t Tanh (W c ·[h t-1 ,x t ]+b c )

[0111] o t =Sigmoid(W o ·[h t-1 ,x t ]+b o )

[0112] h t =o t ·Tanh (c t )

[0113]

[0114] In the formula, h t-1 、x t They represent the output of the Bi-LSTM network at time t-1 and the input at time t respectively; W f , W i , W c , is the trainable parameter matrix, b f , b i , b c , b o is the bias; f t is the calculation result of the forget gate, i t Calculate the result for the input gate, o t Calculate the result for the output gate, c t is the cell state at time t, c t-1 is the cell state at time t-1, h t Represents the output of the Bi-LSTM network at time t; Represents the concatenated forward output vector and the reverse output vector The third semantic vector obtained is

[0115] Furthermore, an event extraction result corresponding to the communication fault processing text is generated based on the third semantic vector, including: determining the label score of the third semantic vector through the second model in the event extraction model, wherein the label score is used to represent the degree of matching between each word and each category label information in the communication fault processing text; determining the label sequence of the communication fault processing text based on the label score and the transition probability between the label scores; determining the target label sequence based on the label sequence and the maximum likelihood function, wherein the target label sequence is the optimal label sequence in the label sequence; determining the event extraction result based on the target label sequence, wherein the event extraction result is used to represent the structured event description in the target label sequence.

[0116] In the embodiment of the present application, the DPEE model can utilize the rich semantic and syntactic information contained in the third semantic vector, and through the label score and transition probability calculation of the CRF layer, combined with the maximum likelihood function to determine the optimal label sequence, and finally generate high-quality event extraction results. This result not only reflects the key event information of the communication fault processing text, but also takes into account the grammatical structure and the dependency relationship between words when the event occurs, making the extracted event description more accurate and complete. The specific process can be as follows:

[0117] First, in the DPEE model, the second model usually refers to the conditional random field (CRF) layer. The role of the CRF layer is to model the relationship between labels. It can capture the global characteristics of the label sequence, thereby avoiding the label prediction bias caused by the local optimal solution. Based on the third semantic vector, a label score is determined for each word in the communication fault processing text through the CRF layer. This score represents the degree of match between the vocabulary and the specific event label (including the trigger word type and the argument role), which provides a basis for the subsequent label sequence prediction.

[0118] It should be noted that the CRF layer not only considers the label score of a single word, but also the transition probability between labels. Among them, the transition probability represents the possibility of one label being transformed into another label, which is particularly important when dealing with continuous event labels. By combining the label score and the transition probability, the CRF layer can predict the entire label sequence, not just the label of a single word. The specific calculation formula can be as follows:

[0119]

[0120] In the formula, P represents the label score, P∈R n×|class| ; is the weight parameter, b 3 For bias.

[0121] For a text sequence X = {w 1 ,w 2 …wn-1 ,w n}, corresponding to the output label sequence Y = {y 1 ,y 2 …y n-1 ,y n The calculation formula for the probability of} is as follows:

[0122]

[0123] In the formula, represents the score of the i-th word belonging to the j-th tag, represents the transfer score from the ith label to the i+1th label, Score(X,Y) represents the total score of the entire sequence X and the predicted label sequence Y; P(Y|X) is the probability of predicting the label sequence Y after normalizing the score, Y X represents all possible predicted label sequences.

[0124] Secondly, the maximum likelihood function is used to determine the optimal label sequence, that is, the target label sequence. The goal of the maximum likelihood function is to find a sequence among all possible label sequences so that the probability of occurrence of the sequence is the highest. It can consider the label sequence with the highest joint probability from a global perspective, thereby solving the common local optimal solution problem in event extraction. The specific calculation formula is as follows:

[0125]

[0126]

[0127] Where, ln[P(Y|X)] represents the logarithmic result, which is used to maximize the likelihood function; Y * Represents the target label sequence.

[0128] Once the target label sequence is determined, the DPEE model can generate event extraction results based on this sequence. The event extraction result is a structured description of key event information in the communication fault handling text, including the type of event, trigger words, and related arguments and their roles. Based on the target label sequence, the model can accurately identify which words are event trigger words, which are arguments, and the specific roles of the arguments, thereby generating structured data describing the communication fault handling event. These structured data are of great value for building fault handling knowledge graphs, semantic models, etc., and can assist operation and maintenance personnel to quickly locate and resolve communication faults, and improve the operating efficiency and stability of the system.

[0129] The specific troubleshooting example is as follows:

[0130] Step 1: For the training samples in Table 3, firstly, the trigger words and arguments are detected and corrected for position shift. Algorithm (1) is an implementation example of the position shift detection and correction method. Then, check whether there are English characters in the sample. If so, change the English letters to lowercase. Finally, check whether there are spaces or line breaks in the sample. If so, replace them with “@” to avoid secondary position shift caused by directly deleting spaces or line breaks.

[0131] Table 3 Communication fault handling event extraction sample table

[0132]

[0133] Algorithm 1 Trigger word and argument position offset detection and correction algorithm example table

[0134]

[0135]

[0136] Step 2: After the data preprocessing in step 1, the training samples are annotated using the BIOSE sequence annotation label construction method. For multiple events in the text, each independent event is annotated separately, and the suffix event type and argument role are added.

[0137] Step 3: Segment the training samples and input them into the BERT model. Encode them through a multi-layer, multi-head self-attention mechanism to obtain the word embedding representation of the fault handling text.

[0138] Step 4: Perform dependency syntactic analysis on the sample to obtain the dependency relationship and dependency type between sample words. Here, the spacy dependency syntactic analysis tool is used. For the sample text in Table 3, the dependency syntactic analysis results are shown in Table 4. Then, dynamic dependency syntactic information is constructed, and the noise component in the dependency relationship can be reduced through the adaptive learning parameter matrix of the neural network.

[0139] Table 4 Dependency parsing example table

[0140] text Dependencies Parent Node Parent node start index position nsubj cause 9 for case Fault 7 powered by Compound:nn Fault 7 Modules Compound:nn Fault 7 Fault nmod:prep cause 9 cause dep abnormal 15 equipment Compound:nn powered by 13 powered by dobj cause 9 abnormal Root Root 0

[0141] Step 5: After obtaining the dynamic dependency syntactic information between words through step 4, a two-layer graph neural network is used to aggregate the information. After aggregation, the semantic information of the words will be integrated with the dependency words and dependency relationship information. Then, Bi-LSTM is used to strengthen the interaction of contextual semantic information, which is conducive to improving the effect of event extraction.

[0142] Step 6: In the communication fault processing event extraction label prediction stage, DPEE uses conditional random fields to output the global optimal solution label sequence. Table 5 shows the output results of the sample, which is consistent with the expected output results.

[0143] Table 5. Sample table of DPEE model event extraction results

[0144]

[0145]

[0146] By extracting events from the communication fault handling process through the DPEE model, key information can be structured and data preparation can be provided for the construction of the communication fault handling knowledge graph and semantic big model, thereby assisting operation and maintenance personnel to locate and handle communication faults in a timely manner and improve system operation stability.

[0147] In an embodiment of the present application, in order to verify the performance of the DPEE model, the effect of the DPEE model is tested on the CCKS2021: Knowledge event extraction dataset of fault handling process for the communication field. Specifically, under the evaluation indicators of Macro-F1 (macro average F1 value), Micro-F1 (micro average F1 value), and F1-Average (average of macro and micro average F1), the results of DPEE and the comparison model in the extraction of trigger words and their types, arguments and their roles are shown in Tables 6 and 7. It can be seen from the table that the DPEE model has achieved better results than the comparison model, especially in the Macro-F1 value, which is 8.44% and 4.31% higher than the optimal comparison model, respectively.

[0148] Table 6 DPEE model trigger words and their types extraction results (unit: %)

[0149]

[0150] Table 7 DPEE model argument and role extraction results (unit: %)

[0151]

[0152] In order to demonstrate the effectiveness of the dynamic dependency syntactic information constructed by DPEE, further ablation experiments were conducted, and the experimental results are shown in Tables 8 and 9. The ablation experiments were carried out around the importance of different modules of dynamic dependency syntactic information, namely, removing dependency types, removing dependency words, and removing dependency types and dependency words at the same time. From the experimental results in the table, it can be seen that when different modules are removed, the event extraction effect is reduced, which proves the rationality and effectiveness of the construction of DPEE dynamic dependency syntactic information.

[0153] Table 8 DPEE model trigger words and their types extraction and cancellation experiment (unit: %)

[0154]

[0155] Table 9 DPEE model argument and its role extraction and elimination experiment (unit: %)

[0156]

[0157] In summary, this application designs a communication fault handling event extraction model that integrates dependency syntactic information. This model significantly enhances the understanding of communication fault handling text by dynamically constructing dependency syntactic information and integrating this information using graph convolutional neural networks (GCNs). DPEE not only uses the pre-trained language model BERT for initial context representation learning, but also further re-encodes the word vectors that incorporate dependency syntactic information through a bidirectional long short-term memory network (Bi-LSTM), and finally uses the conditional random field (CRF) to predict the optimal label sequence, thereby improving the accuracy of event extraction. This method has shown excellent results in the extraction of fault handling events in the communication field. Through experimental verification, DPEE has significantly improved in the extraction of trigger words and types, arguments and roles, especially in the Macro-F1 value, compared with the existing technology. This is beneficial to building a more accurate communication fault handling knowledge graph and semantic model, providing operation and maintenance personnel with a fast and effective fault location and processing solution, thereby improving the operating efficiency and stability of the communication system.

[0158] According to an embodiment of the present application, an event extraction device is provided. It should be noted that the event extraction device of the embodiment of the present application can be used to execute the event extraction method provided by the embodiment of the present application. The event extraction device provided by the embodiment of the present application is introduced below.

[0159] Figure 5 is a structural diagram of an event extraction device provided according to an embodiment of the present application. Figure 5 As shown, the device comprises:

[0160] An acquisition module 50 is used to acquire a communication fault processing text for event extraction;

[0161] A determination module 52 is used to determine a first semantic vector and a second semantic vector of the communication fault processing text by using an event extraction model, wherein the first semantic vector is used to represent a word vector corresponding to each word in the communication fault processing text, and the second semantic vector is used to represent a semantic vector that integrates the dependency relationship between the first semantic vector and each word in the communication fault processing text;

[0162] The generating module 54 is used to fuse the first semantic vector and the second semantic vector to obtain a third semantic vector, and generate an event extraction result corresponding to the communication fault processing text according to the third semantic vector.

[0163] Through the acquisition module 50, determination module 52 and generation module 54 in the above-mentioned event extraction device, the purpose of structuring the key information in the communication fault processing text is achieved, thereby realizing the technical effect of improving the accuracy and efficiency of event extraction, and further solving the technical problem that the event extraction method based on deep learning in the related technology is less used in the communication field and does not fully consider the syntactic structure and dependency relationship between texts, resulting in poor event extraction effect.

[0164] The event extraction device provided in the embodiment of the present application also includes a training module 56, which is used to perform data preprocessing on the historical communication fault processing text in the training stage of the event extraction model, including: obtaining text information that is manually annotated on the historical communication fault processing text, wherein the text information includes the communication fault processing text and the annotation information, and the annotation information includes the trigger word position and argument position that are manually annotated on the communication fault processing text; comparing the text information with the communication fault processing text to obtain a comparison result; when the comparison result indicates that the annotation information in the text information is inconsistent with the initial position information in the communication fault processing text, determining that there is an offset in the trigger word position or argument position in the annotation information, and correcting the trigger word position or argument position in the label information, wherein the initial position information is used to indicate the trigger word position and argument position corresponding to the manually annotated area in the communication fault processing text, the trigger word is used to indicate the key vocabulary of the event occurrence in the historical communication fault processing text, and the argument is used to indicate the entity or concept involved when the event in the historical communication fault processing text occurs.

[0165] In the event extraction device provided in the embodiment of the present application, the determination module is also used to process the communication fault processing text through the first model in the event extraction model to obtain an initial word vector corresponding to each word in the communication fault processing text, wherein the initial word vector includes the position information and fragment information of each word in the communication fault processing text; generate a query vector, a key vector and a value vector corresponding to the initial word vector through a multi-head attention mechanism; and determine the first semantic vector based on the query vector, the key vector and the value vector.

[0166] In the event extraction device provided in the embodiment of the present application, the determination module is also used to perform dependency syntactic analysis on the communication fault processing text through the event extraction model to obtain the dependency relationship between each word in the communication fault processing text; determine the dependency matrix of the communication fault processing text based on the dependency relationship; and fuse the dependency matrix and the first semantic vector through a graph convolutional neural network to obtain a second semantic vector.

[0167] In the event extraction device provided in the embodiment of the present application, the determination module is also used to determine the first word vector and the second word vector of the communication fault processing text, wherein the first word vector and the second word vector are arbitrary vectors in the first semantic vector; determine the target dependency between the first word vector and the second word vector; determine the dependency weight of the first word vector and the second word vector based on the target dependency and the preset dependency embedding vector; and determine the dependency matrix based on the dependency weight.

[0168] In the event extraction device provided in the embodiment of the present application, the generation module is also used to process the first semantic vector and the second semantic vector through a bidirectional long short-term memory network to obtain a forward output vector and a reverse output vector corresponding to the communication fault processing text; and to splice the forward output vector and the reverse output vector through a splicing function to obtain a third semantic vector.

[0169] In the event extraction device provided in the embodiment of the present application, the generation module is also used to determine the label score of the third semantic vector through the second model in the event extraction model, wherein the label score is used to indicate the degree of match between each word in the communication fault processing text and the corresponding label information; determine the label sequence of the communication fault processing text based on the label score and the transition probability between the label scores; determine the target label sequence based on the label sequence and the maximum likelihood function, wherein the target label sequence is the optimal label sequence in the label sequence; determine the event extraction result based on the target label sequence, wherein the event extraction result is used to represent the structured event description in the target label sequence.

[0170] An embodiment of the present application also provides an electronic device, including: a memory and a processor, wherein the memory is used to store program instructions; the processor is connected to the memory and is used to execute the above-mentioned event extraction method.

[0171] It should be noted that the above electronic equipment is used to execute Figure 2 The event extraction method shown, therefore the relevant explanations and instructions in the above event extraction method are also applicable to the electronic device and will not be repeated here.

[0172] An embodiment of the present application also provides a non-volatile storage medium, which includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the above-mentioned event extraction method by running the computer program.

[0173] It should be noted that the above non-volatile storage medium is used to execute Figure 2 The event extraction method shown, therefore the relevant explanations in the above event extraction method are also applicable to the non-volatile storage medium and will not be repeated here.

[0174] An embodiment of the present application also provides a computer program product, including computer instructions, which implement the above-mentioned event extraction method when executed by a processor.

[0175] It should be noted that the above-mentioned computer program product is used to execute Figure 2 The event extraction method shown, therefore the relevant explanations and instructions in the above event extraction method are also applicable to the computer program product and will not be repeated here.

[0176] The serial numbers of the above-mentioned embodiments of the present application are for description only and do not represent the advantages or disadvantages of the embodiments.

[0177] In the above embodiments of the present application, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, please refer to the relevant description of other embodiments.

[0178] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. Among them, the device embodiments described above are only schematic. For example, the division of the units can be a logical function division. There may be other division methods in actual implementation. For example, multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of units or modules, which can be electrical or other forms.

[0179] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple units. Some or all of the units may be selected according to actual needs to achieve the purpose of the present embodiment.

[0180] In addition, each functional unit in each embodiment of the present application may be integrated into one processing unit, or each unit may exist physically separately, or two or more units may be integrated into one unit. The above-mentioned integrated unit may be implemented in the form of hardware or in the form of software functional units.

[0181] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present application, in essence, or the part that contributes to the prior art or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, disk or optical disk and other media that can store program codes.

[0182] The above is only a preferred implementation of the present application. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present application. These improvements and modifications should also be regarded as the scope of protection of the present application.

Claims

1. An event extraction method, characterized in that: include: Obtain the communication fault processing text for event extraction; An event extraction model is used to determine a first semantic vector and a second semantic vector of the communication fault processing text, wherein the first semantic vector is used to represent a word vector corresponding to each vocabulary in the communication fault processing text, and the second semantic vector is used to represent a semantic vector that integrates the dependency relationship between the first semantic vector and each vocabulary in the communication fault processing text; The first semantic vector and the second semantic vector are fused to obtain a third semantic vector, and an event extraction result corresponding to the communication fault processing text is generated according to the third semantic vector.

2. The method according to claim 1, characterized in that The method further includes, during the training phase of the event extraction model, performing data preprocessing on historical communication fault processing texts, including: Acquire text information manually annotated for the historical communication fault processing text, wherein the text information includes the communication fault processing text and annotation information, and the annotation information includes trigger word positions and argument positions manually annotated for the communication fault processing text; Comparing the text information with the communication fault processing text to obtain a comparison result; When the comparison result indicates that the annotation information in the text information is inconsistent with the initial position information in the communication fault processing text, it is determined that there is an offset in the trigger word position or the argument position in the annotation information, and the trigger word position or the argument position in the annotation information is corrected, wherein the initial position information is used to represent the trigger word position and the argument position corresponding to the manually annotated area in the communication fault processing text, the trigger word is used to represent the key vocabulary of the event occurrence in the historical communication fault processing text, and the argument is used to represent the entity or concept involved when the event in the historical communication fault processing text occurs.

3. The method according to claim 1, characterized in that Adopting an event extraction model to determine a first semantic vector of the communication fault processing text includes: Processing the communication fault processing text by the first model in the event extraction model to obtain an initial word vector corresponding to each word in the communication fault processing text, wherein the initial word vector includes position information and segment information of each word in the communication fault processing text; Generate a query vector, a key vector, and a value vector corresponding to the initial word vector through a multi-head attention mechanism; The first semantic vector is determined according to the query vector, the key vector, and the value vector.

4. The method according to claim 1, characterized in that: Adopting an event extraction model to determine a second semantic vector of the communication fault processing text includes: Performing dependency syntactic analysis on the communication fault processing text through the event extraction model to obtain dependency relations between each word in the communication fault processing text; Determining a dependency matrix of the communication fault processing text according to the dependency relationship; The dependency matrix and the first semantic vector are fused through a graph convolutional neural network to obtain the second semantic vector.

5. The method according to claim 4, characterized in that Determining a dependency matrix of the communication fault processing text according to the dependency relationship includes: Determine a first word vector and a second word vector of the communication fault processing text, wherein the first word vector and the second word vector are any vectors in the first semantic vector; Determining a target dependency relationship between the first word vector and the second word vector; Determining dependency weights of the first word vector and the second word vector according to the target dependency and a preset dependency embedding vector; The dependency matrix is ​​determined according to the dependency weights.

6. The method according to claim 1, characterized in that The first semantic vector and the second semantic vector are fused to obtain a third semantic vector, including: Processing the first semantic vector and the second semantic vector through a bidirectional long short-term memory network to obtain a forward output vector and a reverse output vector corresponding to the communication fault processing text; The forward output vector and the reverse output vector are concatenated by a concatenation function to obtain the third semantic vector.

7. The method according to claim 1, characterized in that Generating an event extraction result corresponding to the communication fault processing text according to the third semantic vector includes: Determining a label score of the third semantic vector by using the second model in the event extraction model, wherein the label score is used to indicate a degree of matching between each word in the communication fault handling text and corresponding label information; Determining a label sequence of the communication fault processing text according to the label scores and the transition probabilities between the label scores; Determining a target tag sequence according to the tag sequence and a maximum likelihood function, wherein the target tag sequence is an optimal tag sequence among the tag sequences; The event extraction result is determined according to the target tag sequence, wherein the event extraction result is used to represent the structured event description in the target tag sequence.

8. An event extraction device, characterized in that: include: An acquisition module is used to acquire the communication fault processing text to be extracted; A determination module, configured to determine a first semantic vector and a second semantic vector of the communication fault processing text by using an event extraction model, wherein the first semantic vector is used to represent a word vector corresponding to each vocabulary in the communication fault processing text, and the second semantic vector is used to represent a semantic vector that integrates a dependency relationship between the first semantic vector and each vocabulary in the communication fault processing text; A generation module is used to fuse the first semantic vector and the second semantic vector to obtain a third semantic vector, and generate an event extraction result corresponding to the communication fault processing text based on the third semantic vector.

9. An electronic device, characterized in that: include: A memory and a processor, wherein the memory is used to store program instructions; The processor is connected to the memory and is used to execute the event extraction method described in any one of claims 1 to 7.

10. A non-volatile storage medium, characterized in that: The non-volatile storage medium includes a stored computer program, wherein the device where the non-volatile storage medium is located executes the event extraction method described in any one of claims 1 to 7 by running the computer program.

11. A computer program product comprising computer instructions, characterized in that: When the computer instructions are executed by a processor, the event extraction method described in any one of claims 1 to 7 is implemented.

Images