Text event extraction method and system based on frame semantic mapping and type awareness

Abstract

The present invention relates to the fields of knowledge graphs and natural language processing, and discloses a method and system for extracting chapter events based on frame semantic mapping and type perception. The system includes an encoder module, an event detection module, and an argument recognition module. In the method, frame semantic mapping is to map the Chinese FrameNet (CFN) with chapter events in the event detection module, and at the same time, the chapter is segmented using a sliding window mechanism and the trigger word interpretation information is incorporated into the text input stage, thereby improving the data sparsity problem in chapter event extraction; while type perception is to separate multiple chapter events through event type labels in the argument recognition module, converting multiple chapter events into multiple single events, thereby alleviating the problem of multi-valued argument coupling. Finally, the present invention integrates the output labels of the two modules of event detection and argument recognition to form a complete event representation, and has achieved good results on relevant data sets for Chinese chapter event extraction.

Description

Technical Field

[0001] The present invention relates to the fields of knowledge graphs and natural language processing, and specifically to a method and system for extracting chapter events based on framework semantic mapping and type perception. Background Art

[0002] Document Event Extraction (DEE) is a key research area in information extraction, aiming to extract structured event types and event elements from unstructured text. DEE is a key task in public opinion monitoring and finance. Events are also a key component of knowledge graphs, making event extraction crucial for graph reasoning and event analysis.

[0003] DEE consists of two parts: event detection and argument recognition. The main challenges of this task lie in the complexity of the text and the complexity of the event structure. The complexity of the text is reflected in the fact that the input text is a paragraph, which requires models with limited input text to perform data preprocessing while considering the global information of the paragraph. The complexity of the event structure is reflected in the fact that a paragraph contains multiple events, and different events overlap, making it more difficult to extract a single event.

[0004] Analysis of mainstream text event extraction datasets reveals that the DEE task faces two main challenges: 1) Text data is sparse. Manual annotation of existing datasets is difficult, resulting in a long-tail distribution of event types within a text; 2) Texts contain multiple events. The same argument is distributed across different events, leading to multi-valued argument coupling, or argument overlap. Therefore, this paper explores the introduction of the Chinese FrameNet (CFN) to establish a mapping between events to address data sparsity. This integration of event type information and global text information mitigates the multi-valued argument coupling problem.

[0005] Deep learning-based event extraction methods avoid the tedious task of manually designing features and use word vectors with rich linguistic features as input to reduce errors caused by underlying NLP tools. While the introduction of pre-trained models enriches text semantic representations to a certain extent, the current approach to improving the performance of Chinese text event detection by introducing external knowledge is not considered. Event detection and argument recognition are performed in isolation, lacking information interaction. Most existing text event extraction methods use deep learning and joint learning for feature interaction, capturing the interaction between event types and arguments through graph construction to avoid error propagation. However, they do not effectively address the problem of multi-valued argument coupling in text events.

[0006] In reality, the contextual representations in the event detection and argument identification stages essentially capture different information. However, a joint extraction approach shares the model structure and parameters of both, thus affecting the overall performance of event extraction. This paper adopts a pipeline extraction approach that, compared to most pipeline models, places greater emphasis on the interaction between event types and arguments. In the event detection stage, the introduction of frame semantic mapping and the use of CFN-related event sentences as external data partially alleviate the event sparsity problem. In the argument identification stage, a type-aware multi-event separation strategy is designed to alleviate the problem of multi-valued argument coupling to a certain extent. Summary of the Invention

[0007] To address the above-mentioned issues, the present invention provides a method and system for extracting chapter events based on framework semantic mapping and type awareness. This method and system utilize framework semantic mapping and type awareness to extract chapter events. Framework semantic mapping, during the event detection phase, establishes a mapping between CFN and chapter events, incorporating trigger word interpretation information and text fragments processed by a sliding window, thereby improving the data sparsity problem in chapter event extraction. Type-aware labeling, on the other hand, separates multiple chapter events using event type labels during the argument recognition phase, thereby converting multiple chapter events into multiple single events and further alleviating the multi-valued argument coupling problem. Finally, the present invention jointly optimizes the event detection and argument recognition stages, achieving good results on datasets related to Chinese chapter event extraction.

[0008] In order to achieve the above object, the present invention adopts the following technical solutions:

[0009] In a first aspect, the present invention provides a method for extracting chapter events based on framework semantic mapping and type awareness, comprising the following steps:

[0010] Step 1: The encoder module encodes the text of the event detection module and the argument identification module respectively to obtain the vector representation of each character or word in the passage segment, trigger word interpretation information, and event type label;

[0011] Step 2: In the event detection module, to address the sparsity of text event data, a two-level mapping between the CFN framework and Chinese events is established. Then, the pre-defined event types are predicted by integrating the interpretation information of text trigger words and CFN-related event sentences.

[0012] Step 3: In the argument recognition module, a multi-event separator based on type label perception is established to address the multi-valued argument coupling problem. The arguments in the text are predicted based on the predefined event argument roles.

[0013] Step 4: Concatenate all event type labels predicted by the event detection module with all argument labels predicted by the argument recognition module to complete the final chapter event extraction.

[0014] Furthermore, the specific operation of step 1 is: using the pre-trained language model BERT-Fin as an encoder to encode the text of the event detection module and the argument recognition module respectively;

[0015] The encoding of the event detection module text is as follows: the input text is preprocessed using a sliding window mechanism, and the text is divided into multiple segments, each segment s={c1,c2,c3,...,c i} as an input for the event detection task, c i Indicates the i-th character in the text segment; at the same time, in order to enhance the semantic understanding of the trigger word to the event type, its Chinese interpretation information t={t1,t2,...,t j} is spliced ​​to the trigger word of the text fragment and used as the input of the text. j Represents the jth character in the trigger word interpretation, and the calculation formula is as follows:

[0016] s1=[CLS]+s+[SEP]+t+[SEP]

[0017] s′1=Bert-FinEncoder(s1)

[0018] Where s1 represents the input of the event detection model, [CLS] represents the starting position, and [SEP] represents the separator; s′1={c1,c2,...,c len}, s′1∈R len×d , len represents the total length of the text segment and the interpretation information, and d is the hidden layer dimension;

[0019] The encoding of the argument recognition module text is as follows: the input text is preprocessed using a sliding window mechanism, and the text is divided into multiple segments, each segment s={c1,c2,c3,...,c i} as an input for the argument recognition task, c i Represents the i-th character in the text segment; each event type label e={e1,e2,...,e j} are concatenated with the input text fragments as the model input features, e j Indicates the j-th character of the current event type. The calculation formula is as follows:

[0020] s2=[CLS]+s+[SEP]+e+[SEP]

[0021] s′2=Bert-FinEncoder(s2)

[0022] Among them, s2 represents the input of the argument recognition model, s′2={c1,c2,...,c len}, s′2∈R len×d , len represents the total length of the text segment and type label, and d is the hidden layer dimension.

[0023] Furthermore, step 2 includes two parts: CFN framework and Chinese event mapping and sequence modeling; the specific operations are:

[0024] (1) Mapping between CFN framework and Chinese events: The relevant event sentences, trigger word interpretations, and text segments of the sliding window after mapping the chapter events to the CFN framework are passed through the encoder module to obtain the vector s′1 of the segments and trigger word interpretations. In order to further enhance the representation of long entity information in the chapter, bi_lstm is added after the pre-training model to further enhance its semantic representation. The specific calculation process is shown in the following formula:

[0025] h=bi_lstm(s′1)

[0026]

[0027] h represents the hidden layer vector, including the forward Vector and Backward Vector, after vector splicing, serves as the input I of the next layer;

[0028] (2) Sequence modeling: In the decoding stage, the hidden layer output is used as the CRF emission probability, and the CRF layer transition probability matrix A i,j Indicates that from the i-th tag tag i To the jth tag j The transfer score of s′1={c1,c2,...,c len The corresponding output label result is y e =(y1,y2,...,y n ), the calculation of the current sequence score is defined as follows:

[0029]

[0030] Among them, I i,yi Indicates that the softmax output of the i-th position is y i The probability of the complete sequence score score(s′1,y e) is equal to the sum of the scores of each position. The score of each position is composed of the hidden layer output vector I and the CRF transfer matrix A. Finally, softmax is used to calculate the normalized probability of the event type label, as shown in the following formula:

[0031]

[0032] Where y′ e Represents the normalized value of the output label. The maximum log-likelihood function is used to optimize the objective function. The log-likelihood calculation is shown in the following formula:

[0033]

[0034] Use the dynamic programming Viterbi algorithm to find the optimal label sequence, s1′ represents the input sequence, y e * represents the true label, y e Represents the predicted label, which is calculated as follows:

[0035] y e * =argmax(score(y e |s1′))

[0036] The event detection model after integrating the semantic mapping of the CFN framework can effectively solve the data sparsity problem in chapter events and improve the accuracy of event type recognition.

[0037] Furthermore, step 3 includes two parts: a multi-type perception layer and a multi-label decoding layer; the specific operations are:

[0038] (1) Multi-type perception layer: A type perception module is set up in this stage. Its core is to convert each event type label e = {e1, e2, ..., e j} respectively with the input text fragment s={c1,c2,c3,.. i .c, splicing together as the model input features, according to different type labels, multiple events in the text are converted into multiple single events. In order to further enhance the interaction between event detection and argument recognition stages, the global information of the chapter in event detection is obtained by maxpooling to obtain a vector Then it is fused with the text segment encoding vector s′2 to further enhance the text information representation. The hidden layer calculation is shown in the following formula:

[0039]

[0040]

[0041] Among them, I nrepresents the hidden layer encoding of the nth segment after the sliding window segmentation in the article, and s″2 represents the hidden layer representation that integrates the global information of the article;

[0042] (2) Multi-label decoding layer: After the type labels are separated, the single events are passed through their respective CRF decoding layers, and the original classification of multiple event argument roles is converted into the classification of multiple single event argument roles; the CRF decoding structure is consistent with the event detection stage, and the CRF decoding sequence calculation formula is as follows:

[0043]

[0044]

[0045] in, represents the label sequence of the i-th event after CRF decoding, Represents the set of argument label sequences for all events. To address the imbalanced distribution of positive and negative samples, the Focal loss function is used. The calculation process is shown in the following formula:

[0046]

[0047]

[0048] Among them, s′2 represents the input sequence, y′ a Represents the normalized value of the output label, P represents the normalized probability value of the current sequence score after softmax, p * represents the probability value of the true label of the sample, α is the positive and negative sample ratio weight balance factor, and the initial value is set to 1.0; γ is the sample difficulty weight balance factor, and the initial value is set to 2.0;

[0049] Use the dynamic programming Viterbi algorithm to find the optimal label sequence. represents the true label, y a Represents the predicted label, which is calculated as follows:

[0050]

[0051]

[0052] All event type labels predicted by event detection All argument labels predicted by argument recognition Splicing, y * It represents a complete event and finally completes the overall chapter event extraction.

[0053] In a second aspect, the present invention provides a text event extraction system based on frame semantic mapping and type awareness, comprising an encoder module, an event detection module, and an argument recognition module; wherein:

[0054] The encoder module encodes the text fragments and trigger word interpretation information of the event detection input, and encodes the text fragments and event type labels of the argument recognition input to obtain the vector representation of each character or word in the input feature;

[0055] The event detection module is to judge the event type of the vector after the event detection encoding of the encoder module through the sequence modeling layer, and then use the sequence labeling method to obtain the optimal event type sequence label;

[0056] The argument recognition module judges the vector after the argument recognition encoding of the encoder module through the multi-type perception layer and the multi-label decoding layer, and then obtains the optimal argument sequence label by using the sequence labeling method.

[0057] Furthermore, the encoder module includes two parts: 1) Chapter text encoding in the event detection stage. By encoding the chapter segments processed by the sliding window, the trigger word interpretation information, and the event sentences after the CFN and the chapter event mapping, the vector representation of each character or word in the chapter segments and the trigger word interpretation information is obtained. 2) Chapter text encoding in the argument recognition stage. By encoding the chapter segments and event type labels processed by the sliding window, the vector representation of each character or word in the chapter segments and event type labels is obtained; the event detection module is a two-level mapping based on the similarity between the Chinese frame network and the Chinese event structure, which improves the data sparsity problem in event detection; the argument recognition module is a multi-event separation method based on event type label perception, which separates multiple events in the chapter into multiple single events, alleviating the problem of multi-valued argument coupling in the chapter.

[0058] In a third aspect, the present invention provides an electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for extracting chapter events based on framework semantic mapping and type perception is implemented.

[0059] In a fourth aspect, the present invention provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is used to implement the above-mentioned method for extracting chapter events based on framework semantic mapping and type perception when executed by a processor.

[0060] Compared with the prior art, the present invention has the following advantages:

[0061] (1) This paper combines CFN with the Chinese text event extraction task to explore the relationship between frameworks and events, which helps to improve the data sparsity problem in text event extraction.

[0062] (2) In the event detection stage, the present invention introduces external knowledge such as CFN, constructs a secondary mapping between the CFN framework and events, and uses the relevant event sentences after CFN mapping as external data to improve the data sparsity problem in chapter event extraction.

[0063] (3) In the argument recognition stage, by fusing event type labels and integrating global context information, multiple events in the passage are converted into multiple single events, which alleviates the problem of multi-valued argument coupling to a certain extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0064] Figure 1 The overall architecture diagram of a chapter event extraction method based on framework semantic mapping and type awareness provided by an embodiment of the present invention;

[0065] Figure 2 Data samples used in this invention;

[0066] Figure 3 This is a structural diagram of the event detection module of the present invention;

[0067] Figure 4 This is a structural diagram of the argument recognition module of the present invention;

[0068] Figure 5 This is a diagram of the multi-tag integration structure of the present invention;

[0069] Figure 6 A schematic diagram of the structure of a chapter event extraction system based on framework semantic mapping and type awareness provided by an embodiment of the present invention; DETAILED DESCRIPTION

[0070] The present invention is further described in detail below with reference to the embodiments and drawings. The embodiments are only used to explain the present invention and are not used to limit the scope of protection of the present invention.

[0071] Example 1

[0072] Figure 1 The overall architecture diagram of a passage event extraction method based on frame semantic mapping and type perception provided in an embodiment of the present invention mainly includes an encoder module, an event detection module and an argument recognition module.

[0073] The specific contents of each module are as follows:

[0074] The encoder module encodes event detection and argument recognition separately to obtain the vector representation of each character or word in the passage segment, trigger word interpretation information, and event type label;

[0075] The event detection module utilizes a two-level mapping between CFN and Chinese events, a sliding window mechanism, and trigger word interpretation information to address data sparsity during the event detection phase. Through two-level framework semantic mapping, relevant event sentences from CFN are used for event detection, improving performance in event detection tasks.

[0076] The argument recognition module uses type-aware tags and integrates global information from the text to alleviate the problem of multi-valued argument coupling in the text. The type-aware module separates multiple events in the text into multiple single events for recognition.

[0077] Label integration: All event type labels predicted by the event detection module are spliced ​​with all argument labels predicted by the argument recognition module to complete the final chapter event extraction.

[0078] Example 2

[0079] Figure 2 This is an example from Baidu's newly released 2021 Financial Chapter-Level Event Extraction Dataset (DuEE-fin). It contains 11,700 chapters across 13 event types, including some non-target chapters as negative examples. The event types are derived from common financial events, and the chapters in the dataset are from news and announcements in the financial sector, covering many difficult problems in real-world application scenarios. Chapter-Level Event Extraction input: one or more consecutive, complete chapters containing event information; output: event arguments with predefined event types and argument roles.

[0080] 1. Use the pre-trained language model BERT-Fin as the encoder to encode the text of the event detection module and the argument recognition module respectively.

[0081] The encoding of the event detection module text is as follows: the sliding window mechanism is used to preprocess the input text data, and the text is divided into multiple segments, each segment s={c1,c2,c3,...,c i} as an input for the event detection task, c i Indicates the i-th character in the text segment; at the same time, in order to enhance the semantic understanding of the trigger word to the event type, its Chinese interpretation information t={t1,t2,...,t j} is spliced ​​to the trigger word of the text fragment and used as the input of the text. j Represents the jth character in the trigger word interpretation, and the calculation formula is as follows:

[0082] s1=[CLS]+s+[SEP]+t+[SEP]

[0083] s′1=Bert-FinEncoder(s1)

[0084] Where s1 represents the input of the event detection model, [CLS] represents the starting position, and [SEP] represents the separator; s′1={c1,c2,...,c len}, s′1∈R len×d , len represents the total length of the text segment and the interpretation information, and d is the hidden layer dimension;

[0085] The encoding of the argument recognition module text is as follows: the sliding window mechanism is used to preprocess the input text data, and the text is divided into multiple segments, each segment s={c1,c2,c3,...,c i} as an input for the argument recognition task, c i Represents the i-th character in the text segment; each event type label e={e1,e2,...,e j} are concatenated with the input text fragments as the model input features, e j Indicates the j-th character of the current event type. The calculation formula is as follows:

[0086] s2=[CLS]+s+[SEP]+e+[SEP]

[0087] s′2=Bert-FinEncoder(s2)

[0088] Among them, s2 represents the input of the argument recognition model, s′2={c1,c2,...,c len}, s′2∈R len×d , len represents the total length of the text segment and type label, and d is the hidden layer dimension.

[0089] 2. In the event detection module, in order to solve the problem of sparse event data in the text, the pre-defined event types are predicted by integrating the interpretation information of the text trigger words and the CFN related event sentences, including the CFN framework and Chinese event mapping and sequence modeling. The event detection module structure is shown in the figure below. Figure 3 shown.

[0090] (1) Mapping between CFN framework and Chinese events: The relevant event sentences, trigger word interpretations, and text segments of the sliding window after mapping the chapter events to the CFN framework are passed through the encoder module to obtain the vector s′1 of the segments and trigger word interpretations. In order to further enhance the representation of long entity information in the chapter, bi_lstm is added after the pre-training model to further enhance its semantic representation. The specific calculation process is shown in the following formula:

[0091] h=bi_lstm(s′1)

[0092]

[0093] h represents the hidden layer vector, including the forward Vector and Backward Vector, after vector splicing, serves as the input I of the next layer;

[0094] (2) Sequence modeling: In the decoding stage, the hidden layer output is used as the CRF emission probability, and the CRF layer transition probability matrix A i,j Indicates that from the i-th tag tag i To the jth tag j The transfer score of s′1={c1,c2,...,c len The corresponding output label result is y e =(y1,y2,...,y n ), the calculation of the current sequence score is defined as follows:

[0095]

[0096] Among them, I i,yi Indicates that the softmax output of the i-th position is y i The probability of the complete sequence score score(s′1,y e ) is equal to the sum of the scores of each position. The score of each position is composed of the hidden layer output vector I and the CRF transfer matrix A. Finally, softmax is used to calculate the normalized probability of the event type label, as shown in the following formula:

[0097]

[0098] Where y′ e Represents the normalized value of the output label. The maximum log-likelihood function is used to optimize the objective function. The log-likelihood calculation is shown in the following formula:

[0099]

[0100] Use the dynamic programming Viterbi algorithm to find the optimal label sequence, s1′ represents the input sequence, y e * represents the true label, y e Represents the predicted label, which is calculated as follows:

[0101] y e * =argmax(score(y e |s1′))

[0102] The event detection model after integrating the semantic mapping of the CFN framework can effectively solve the data sparsity problem in chapter events and improve the accuracy of event type recognition.

[0103] 3. In the argument recognition stage, in order to solve the problem of multi-valued argument coupling, the arguments in the text are predicted through the pre-defined event argument roles, which includes two parts: the multi-type perception layer and the multi-label decoding layer. The structure diagram of the argument recognition module is as follows: Figure 4 shown.

[0104] (1) Multi-type perception layer: A type perception module is set up in this stage. Its core is to convert each event type label e = {e1, e2, ..., e j} respectively with the input text fragment s={c1,c2,c3,.. i .c, splicing together as the model input features, according to different type labels, multiple events in the text are converted into multiple single events. In order to further enhance the interaction between event detection and argument recognition stages, the global information of the chapter in event detection is obtained by maxpooling to obtain a vector Then it is fused with the text segment encoding vector s′2 to further enhance the text information representation. The hidden layer calculation is shown in the following formula:

[0105]

[0106]

[0107] Among them, I n represents the hidden layer encoding of the nth segment after the sliding window segmentation in the article, and s″2 represents the hidden layer representation that integrates the global information of the article;

[0108] (2) Multi-label decoding layer: After the type labels are separated, the single events are passed through their respective CRF decoding layers, and the original classification of multiple event argument roles is converted into the classification of multiple single event argument roles; the CRF decoding structure is consistent with the event detection stage, and the CRF decoding sequence calculation formula is as follows:

[0109]

[0110]

[0111] in, represents the label sequence of the i-th event after CRF decoding, Represents the set of argument label sequences for all events. To address the imbalanced distribution of positive and negative samples, the Focal loss function is used. The calculation process is shown in the following formula:

[0112]

[0113]

[0114] Among them, s′2 represents the input sequence, y′ a Represents the normalized value of the output label, P represents the normalized probability value of the current sequence score after softmax, p * represents the probability value of the true label of the sample, α is the positive and negative sample ratio weight balance factor, and the initial value is set to 1.0; γ is the sample difficulty weight balance factor, and the initial value is set to 2.0;

[0115] Use the dynamic programming Viterbi algorithm to find the optimal label sequence. represents the true label, y a Represents the predicted label, which is calculated as follows:

[0116]

[0117]

[0118] All event type labels predicted by event detection All argument labels predicted by argument recognition Splicing, y * It represents a complete event and finally completes the overall chapter event extraction.

[0119] The multi-label integration structure diagram of the present invention in the final prediction stage is as follows Figure 5 shown.

[0120] Combine Figure 2 It can be seen that the method of the present invention has achieved good results on the Chinese text event extraction related data set.

[0121] Example 3

[0122] Figure 6 A structural diagram of a chapter event extraction system based on framework semantic mapping and type perception provided by an embodiment of the present invention is shown in FIG. Figure 6 As shown in Figure 1, the chapter event extraction system includes: encoder module, event detection module, and argument recognition module, among which:

[0123] Encoder module: Encodes the text fragments and trigger word interpretation information of the event detection input, and encodes the text fragments and event type labels of the argument recognition input, obtaining the vector representation of each character or word in the input feature;

[0124] Event detection module: The vector encoded by event detection is passed through the sequence modeling layer to determine the event type, and then the optimal event type sequence label is obtained through sequence labeling.

[0125] Argument recognition module: The vector after argument recognition encoding is judged by the multi-type perception layer and the multi-label decoding layer, and then the optimal argument sequence label is obtained by sequence labeling;

[0126] An embodiment of the present invention provides a chapter event extraction system based on framework semantic mapping and type awareness, which specifically executes the above-mentioned embodiment process of a chapter event extraction method based on framework semantic mapping and type awareness. For details, please refer to the content of the above-mentioned embodiment of a chapter event extraction method based on framework semantic mapping and type awareness, which will not be repeated here.

[0127] Example 4

[0128] This embodiment provides an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the above-mentioned method for extracting chapter events based on framework semantic mapping and type perception is implemented.

[0129] Example 5

[0130] This embodiment provides a non-transitory computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the computer program is used to implement the above-mentioned chapter event extraction method based on framework semantic mapping and type awareness.

[0131] The system embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units. They may be located in one place or distributed across multiple network units. Some or all of the modules may be selected based on actual needs to achieve the objectives of this embodiment. Persons of ordinary skill in the art will be able to understand and implement the present invention without inventive effort.

[0132] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the same. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be included in the scope of the technical solutions of the present invention.

Claims

1. A method for extracting chapter events based on frame semantic mapping and type awareness, characterized by: The following steps are involved: Step 1: Encode the text of the event detection module and argument identification module separately through the encoder module to obtain the vector representation of each character or word in the passage segment, trigger word interpretation information, and event type label; Step 2: In the event detection module, to address the sparsity of text event data, a two-level mapping between the CFN framework and Chinese events is established. Then, the pre-defined event types are predicted by integrating the interpretation information of text trigger words and CFN-related event sentences. Step 3: In the argument recognition module, a multi-event separator based on type label perception is established to address the multi-valued argument coupling problem. The arguments in the text are predicted based on the predefined event argument roles. Step 4: Concatenate all event type labels predicted by the event detection module with all argument labels predicted by the argument recognition module to complete the final chapter event extraction. Step 2 includes two parts: CFN framework and Chinese event mapping and sequence modeling; the specific operations are: (1) CFN framework and Chinese event mapping: The relevant event sentences, trigger word interpretations and text fragments of the sliding window after the chapter event is mapped to the CFN framework are obtained through the encoder module to obtain the vectors of the fragments and trigger word interpretations. In order to further enhance the representation of long entity information in the article, we add To further enhance its semantic representation, the specific calculation process is shown in the following formula: ； ； Represents the hidden layer vector, including the forward Vector and Backward Vector, after vector splicing as the input of the next layer ; (2) Sequence modeling: In the decoding stage, the hidden layer output is used as the CRF emission probability, and the CRF layer transfer probability matrix Indicates that from Tags To Tags The transfer score of The corresponding output label result is , the calculation of the current sequence score is defined as follows: ； in, Indicates the Locations The output is The probability of the complete sequence Equal to the sum of the scores of each position, each position score is output by the hidden layer vector and Transfer Matrix Together constitute; finally use Calculate the normalized probability of the event type label as shown in the following formula: ； In the formula Represents the normalized value of the output label. The maximum log-likelihood function is used to optimize the objective function. The log-likelihood calculation is shown in the following formula: ； Using dynamic programming The algorithm obtains the optimal label sequence. represents the input sequence, represents the true label, Represents the predicted label, which is calculated as follows: ； The event detection model after integrating the semantic mapping of the CFN framework can effectively solve the data sparsity problem in chapter events and improve the accuracy of event type recognition.

2. The method for extracting chapter events based on frame semantic mapping and type awareness according to claim 1, characterized in that: The specific operation of step 1 is: using the pre-trained language model BERT-Fin as an encoder to encode the text of the event detection module and the argument recognition module respectively; The encoding of the event detection module text is as follows: the input text is preprocessed using a sliding window mechanism, and the text is divided into multiple segments. As an input to the event detection task, Indicates the first characters; at the same time, in order to enhance the semantic understanding of the trigger word to the event type, its Chinese interpretation information After being spliced ​​into the trigger words of the text fragment, they serve as the input of the text. Indicates the first characters, the calculation formula is as follows: ； ； in, Represents the input of the event detection model, [CLS] represents the start position, and [SEP] represents the separator; , , Indicates the total length of the text segment and interpretation information, is the hidden layer dimension; The encoding of the argument recognition module text is as follows: the input text is preprocessed using a sliding window mechanism, and the text is divided into multiple segments. As an input to the argument recognition task, Indicates the first characters; each event type label predicted in the event detection phase They are concatenated with the input text fragments and used as model input features. Indicates the current event type. characters, the calculation formula is as follows: ； ； in, represents the input of the argument recognition model, , , Indicates the total length of the text segment and type tag, is the hidden layer dimension.

3. The method for extracting chapter events based on frame semantic mapping and type awareness according to claim 1, characterized in that: Step 3 includes two parts: a multi-type perception layer and a multi-label decoding layer; the specific operations are: (1) Multi-type perception layer: A type perception module is set up in this stage. Its core is to label each event type perceived in the event detection stage. Respectively with the input text fragment The splicing is used as the model input feature, and multiple events in the text are converted into multiple single events according to different type labels. In order to further enhance the interaction between event detection and argument recognition, the global information of the chapter in event detection is transformed into Get vector , and then the text fragment encoding vector Fusion is performed to further enhance the text information representation. The hidden layer calculation is shown in the following formula: ； ； in, Indicates the first The hidden layer code of the fragment, The hidden layer representation that represents the fusion of the global information of the passage; (2) Multi-label decoding layer: After the type labels are separated, the single events are passed through their respective CRF decoding layers, and the original classification of multiple event argument roles is converted into the classification of multiple single event argument roles; the CRF decoding structure is consistent with the event detection stage, and the CRF decoding sequence calculation formula is as follows: ； ； in, Indicates the The label sequence of each event after CRF decoding, Represents the set of argument label sequences for all events; To solve the imbalanced distribution of positive and negative samples, The loss function is calculated as follows: ； ； in, represents the input sequence, represents the normalized value of the output label, Indicates the current sequence score The normalized probability value after represents the probability value of the true label of the sample, is the positive and negative sample ratio weight balance factor, and its initial value is set to 1.0; is the sample difficulty weight balancing factor, with an initial value of 2.0; Using dynamic programming The algorithm obtains the optimal label sequence. represents the true label, Represents the predicted label, which is calculated as follows: ； ； All event type labels predicted by event detection All argument labels predicted by argument recognition To splice, It represents a complete event and finally completes the overall chapter event extraction.

4. A text event extraction system based on frame semantic mapping and type awareness, characterized by: A method for extracting text events based on frame semantic mapping and type awareness according to any one of claims 1 to 3, comprising an encoder module, an event detection module, and an argument recognition module; wherein: The encoder module encodes the text fragments and trigger word interpretation information of the event detection input, and encodes the text fragments and event type labels of the argument recognition input to obtain the vector representation of each character or word in the input feature; The event detection module is to judge the event type of the vector after the event detection encoding of the encoder module through the sequence modeling layer, and then use the sequence labeling method to obtain the optimal event type sequence label; The argument recognition module judges the vector after the argument recognition encoding of the encoder module through the multi-type perception layer and the multi-label decoding layer, and then obtains the optimal argument sequence label by using the sequence labeling method.

5. An electronic device, characterized in that: The method comprises a memory, a processor and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for extracting chapter events based on framework semantic mapping and type perception as described in any one of claims 1 to 3 is implemented.

6. A non-transitory computer-readable storage medium, characterized in that: The medium stores a computer program, which is used to implement the chapter event extraction method based on framework semantic mapping and type awareness as described in any one of claims 1 to 3 when executed by a processor.

Images