Text data reverse generation method and device based on event pattern

Abstract

The present invention provides a method and device for reverse generation of text data based on event patterns, the method comprising: extracting event trigger words and event argument roles corresponding to the event type from event trigger words and argument candidate vocabulary based on a preset event type; generating corresponding temporal event quintuples based on the preset event type, the extracted corresponding event trigger words and event argument roles, and the corresponding event argument roles, and forming structured text data; inputting the structured text data into a large language model so that the large language model outputs event text data corresponding to the preset event type, and adding annotation information to the output event text data. The present invention can reversely generate accurate and effective text data containing event pattern annotation information at low cost and high efficiency, and ensure the rationality and naturalness of the generated text data with respect to the event pattern and event structure.

Description

Technical Field

[0001] The present invention relates to the technical field of natural language processing, and in particular to a method and device for reversely generating text data based on event patterns. Background Art

[0002] The goal of event extraction is to use algorithmic models to extract structured information about events mentioned in natural language texts. It plays a vital role in tasks such as natural language understanding and structured knowledge extraction from knowledge graphs. In order to train an effective event extraction model, a large corpus containing annotated event patterns is usually required. However, collecting and annotating such corpora is not only difficult but also costly. The industry generally uses rule-based methods or language model-based methods to generate corpora.

[0003] Rule-based methods use easy-to-compute predefined rules to generate new data, including random operations on word units, expanding vocabulary diversity, grammatical variations, etc. This type of method is relatively simple, but the data generated is relatively limited, and the gain is limited when used for model training. The method of using language models to train data includes small language models (such as BERT) and large language models (LLM). Small language models themselves have limitations, and the data generated is also relatively limited. Existing methods for generating data based on large language models generally rely mainly on the self-reflexive ability of LLM itself, randomly combining the trigger words and event arguments of the sampled events, and do not fully utilize the structural knowledge of event patterns, and cannot guarantee the rationality and naturalness of the generated data with respect to event patterns and event structures. Summary of the invention

[0004] In view of this, an embodiment of the present invention provides a method and device for reverse generation of text data based on event patterns to eliminate or improve one or more defects existing in the prior art.

[0005] One aspect of the present invention provides a method for reverse generation of text data based on event patterns, the method comprising the following steps:

[0006] Based on a preset event type, event trigger words and event argument role instantiations corresponding to the event type are jointly extracted from an event trigger word and argument candidate vocabulary library, wherein the event trigger word and argument candidate vocabulary library is pre-formed by instantiations of a plurality of event trigger words and a plurality of event argument roles corresponding to a plurality of event types, and the instantiations of the plurality of event trigger words and a plurality of event argument roles corresponding to the plurality of event types are respectively obtained by inputting a plurality of determined event patterns including event types, event trigger words and event argument roles into a large language model, and then outputted by the large language model;

[0007] Based on the preset event type, the extracted instantiation representation of the corresponding event trigger word and event argument role, and the corresponding event argument role, a corresponding time series event quintuple is generated to form structured text data;

[0008] The structured text data is input into a large language model so that the large language model outputs event text data corresponding to the preset event type, and adds annotation information to the output event text data.

[0009] In some embodiments of the present invention, before the step of adding annotation information to the output event text data, the method further includes: performing quality assessment and proofreading on the output event text data based on the generated temporal event quintuple.

[0010] In some embodiments of the present invention, quality assessment and proofreading of output event text data based on the generated time series event quintuple includes:

[0011] Perform string matching on the output event text data and the corresponding time series event quintuple to obtain the string matching result;

[0012] The output event text data and the corresponding time series event quintuple are subjected to semantic similarity calculation to obtain a semantic similarity calculation result;

[0013] A first score corresponding to the output event text data is determined based on the string matching result and the semantic similarity calculation result; if the first score is less than a first threshold, the step of inputting the structured text data into a large language model is repeated to optimize the large language model until the first score finally obtained is greater than or equal to the first threshold.

[0014] In some embodiments of the present invention, the output event text data and the corresponding time series event quintuple are subjected to semantic similarity calculation to obtain a semantic similarity calculation result, including:

[0015] Segment the output event text data to obtain multiple segmented words;

[0016] Inputting the multiple word segments and all words of the time sequence event quintuple into a pre-trained language model respectively, so that the pre-trained language model outputs multiple semantic vectors corresponding to the multiple word segments and all words of the time sequence event quintuple respectively;

[0017] Perform pairwise similarity calculation on the multiple semantic vectors corresponding to the multiple word segments and the multiple semantic vectors corresponding to all words of the time series event quintuple to obtain multiple similarity calculation results;

[0018] A portion of similarity calculation results greater than a second threshold is screened out from the plurality of similarity calculation results and averaged to obtain a semantic similarity calculation result.

[0019] In some embodiments of the present invention, based on a preset event type, an event trigger word and an instantiation representation of an event argument role corresponding to the event type are jointly extracted from an event trigger word and an argument candidate vocabulary library, including:

[0020] Encoding the instantiation representations of multiple event trigger words and multiple event argument roles corresponding to multiple event types in the event trigger word and argument candidate vocabulary library respectively, and obtaining multiple text vector representations accordingly;

[0021] A named entity recognition pre-trained model based on global normalization is used to decode multiple text vector representations, and the corresponding relationships between the instantiation representations of event trigger words and event argument roles and event types are extracted jointly;

[0022] According to the correspondence between the instantiation representations of event trigger words and event argument roles and the event types, respectively, event trigger words and event argument roles corresponding to the event type are jointly extracted from the event trigger word and argument candidate vocabulary library based on the preset event type.

[0023] In some embodiments of the present invention, before forming the structured text data, the method further includes: constructing a corresponding event graph based on the generated temporal event quintuple.

[0024] In some embodiments of the present invention, when there is a labeled data set, the event pattern is determined by extracting the event pattern in the labeled data set; when there is no labeled data set, the event pattern is determined by customizing the event pattern.

[0025] Another aspect of the present invention provides a text data reverse generation device based on event patterns, the device comprising: a computer device, the computer device comprising a processor and a memory, the memory storing computer instructions, the processor being used to execute the computer instructions stored in the memory, and the device implementing the steps of the aforementioned method when the computer instructions are executed by the processor.

[0026] Another aspect of the present invention provides a computer-readable storage medium having a computer program stored thereon, and the computer program is executed by a processor to implement the steps of the aforementioned method.

[0027] Another aspect of the present invention provides a computer program product, comprising computer instructions, which implement the steps of the above method when executed by a processor.

[0028] The method and device for reverse generation of text data based on event patterns of the present invention can make full use of the structural knowledge of event patterns to ensure the rationality and naturalness of the generated text data with respect to event patterns and event structures, and can reversely generate accurate and effective text data containing event pattern annotation information at low cost and high efficiency, fully exploring the advantages and capabilities of large language models in generating event text data based on the structured information of event patterns, and can well support low-resource scenarios with less and insufficient training data.

[0029] Additional advantages, purposes, and features of the present invention will be described in part in the following description, and will become apparent to those skilled in the art after studying the following, or may be learned from the practice of the present invention. The purposes and other advantages of the present invention may be achieved and obtained by the structures specifically indicated in the specification and the accompanying drawings.

[0030] Those skilled in the art will appreciate that the objectives and advantages that can be achieved with the present invention are not limited to the above specific description, and the above and other objectives that can be achieved by the present invention will be more clearly understood from the following detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The drawings described herein are used to provide a further understanding of the present invention, constitute a part of this application, and do not constitute a limitation of the present invention.

[0032] Figure 1 Schematic diagram of a flow chart of a method for reversely generating text data based on an event pattern in one embodiment of the present invention;

[0033] Figure 2 It is a specific flow chart of a method for reversely generating text data based on an event pattern in one embodiment of the present invention;

[0034] Figure 3 It is a schematic diagram of a specific example of a method for reversely generating text data based on an event pattern in another embodiment of the present invention. DETAILED DESCRIPTION

[0035] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the embodiments and the accompanying drawings. Here, the illustrative embodiments of the present invention and their descriptions are used to explain the present invention, but are not intended to limit the present invention.

[0036] It should also be noted that, in order to avoid obscuring the present invention due to unnecessary details, only structures and / or processing steps closely related to the solutions according to the present invention are shown in the accompanying drawings, while other details that are not closely related to the present invention are omitted.

[0037] It should be emphasized that the term “include / comprises” when used herein refers to the presence of features, elements, steps or components, but does not exclude the presence or addition of one or more other features, elements, steps or components.

[0038] It should also be noted that, unless otherwise specified, the term “connection” herein may refer not only to a direct connection but also to an indirect connection involving an intermediate.

[0039] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the accompanying drawings, the same reference numerals represent the same or similar components, or the same or similar steps.

[0040] In low-resource scenarios where text data containing annotated event patterns is scarce, the annotated data used to train event extraction models is seriously insufficient, and the cost of annotated data is high. An event pattern refers to a predefined event type and an event structure corresponding to the event type in an event extraction task, wherein the event structure includes an event trigger (Event Trigger) and an event argument role (Event Argument). In order to make full use of the structural knowledge of event patterns, ensure the rationality and naturalness of the generated text data about event patterns and event structures, and be able to reversely generate accurate and effective text data containing event pattern annotation information at low cost and high efficiency, an embodiment of the present invention provides a method and device for reverse generation of text data based on event patterns, which fully explores and utilizes the advantages and capabilities of large language models to generate event text data based on the structured information of event patterns, and can well support low-resource scenarios with less and insufficient training data.

[0041] Figure 1 FIG. 1 is a flow chart of a method for reversely generating text data based on an event pattern in one embodiment of the present invention. Figure 1 As shown, the method comprises the following steps:

[0042] Step S110, based on a preset event type, jointly extracting instantiation representations of event trigger words and event argument roles corresponding to the event type from an event trigger word and argument candidate vocabulary library, wherein the event trigger word and argument candidate vocabulary library is pre-formed by instantiation representations of multiple event trigger words and multiple event argument roles corresponding to multiple event types, and the instantiation representations of multiple event trigger words and multiple event argument roles corresponding to the multiple event types are obtained by inputting a plurality of determined event patterns containing event types, event trigger words and event argument roles into a large language model and then outputting the large language model.

[0043] Event type refers to the category of the event that occurred, which is used to describe the nature of the event, such as attack events, conflict events, birth events or marriage events; event trigger words refer to the key actions that lead to the occurrence of the event, such as attack / throwing, birth / birth or marriage; event argument role refers to the relationship between the event argument and the events it participates in, such as the time, place, attacker, means, and target of the event; event argument refers to the main attribute of the event, that is, the instantiation of the event argument role, such as September 30, 2024 (time of occurrence), the entrance to the town (place), children (attackers), stones and empty bottles (means), bad guys who bully children (target), etc.

[0044] Figure 2 FIG. 1 is a schematic diagram of a specific flow chart of a method for reversely generating text data based on an event pattern in an embodiment of the present invention. Figure 2 As shown, in the process of forming a candidate vocabulary library of event trigger words and arguments, first determine whether there is a labeled data set for training the event extraction model. If there is a labeled data set for training the event extraction model (for example, it can be the classic ACE 2005 data set and Baidu DuEE Chinese event extraction data set, etc.) and the data in the data set is relatively small, the event pattern (event type, event trigger words, and event argument roles) of the events in the labeled data set is identified and extracted based on the annotation information in the labeled data set, and the event pattern is determined by the event pattern extracted from the labeled data set; if there is no labeled data set for training the event extraction model, customize the event type, event trigger words, and event argument roles, and determine the event pattern by customizing the event pattern. For example, Figure 3 FIG. 1 is a schematic diagram of a specific example of a method for reversely generating text data based on an event pattern in another embodiment of the present invention. Figure 3 As shown in the figure, by extracting or customizing various event types such as attack events, the time (Time) and place (Place) of various events such as conflict (Conflict) type attack (Attack) subtype events, the attacker (Attacker), means (Instrument), target (Target) and other event argument roles in the event, and event trigger words (Trigger), and using the above content as the context prompt (Prompt) information of the large language model, the large language model is guided to generate specific event trigger words corresponding to each event type (attack, throw, and other verbs with similar semantics that can clearly indicate the corresponding event type), specific event arguments (multiple words such as stone and empty bottle indicating means, words indicating the specific object of the attacker, etc.), and then a large number of event trigger words and event argument roles generated by the large language model are used to form an event trigger word and argument candidate vocabulary library.

[0045] like Figure 3 As shown, the instantiation of the event argument role extracted in this step includes specific instances of the event occurrence time and the non-event occurrence time (i.e., all elements that constitute a complete event), that is, including September 30, 2024 (the specific time of the event), the town entrance (specific location), the child (specific attacker), the stone and the empty bottle (specific means), the bad guy who bullies the child (specific target), etc. corresponding to the conflict event.

[0046] Specifically, step S110, based on a preset event type, jointly extracts the event trigger word and event argument role instantiation representation corresponding to the event type from the event trigger word and argument candidate vocabulary library, including the following steps:

[0047] Step S111, respectively encoding instantiation representations of a plurality of event trigger words and a plurality of event argument roles corresponding to a plurality of event types in the event trigger word and argument candidate vocabulary library, and obtaining a plurality of corresponding text vector representations;

[0048] Step S112, using a globally normalized named entity recognition pre-trained model to decode multiple text vector representations, and jointly extract the corresponding relationships between the instantiation representations of event trigger words and event argument roles and the event types;

[0049] Step S113, according to the correspondence between the instantiation representations of event trigger words and event argument roles and the event types respectively, based on the preset event types, the event trigger words and event argument roles corresponding to the event types are jointly extracted from the event trigger word and argument candidate vocabulary library.

[0050] By using the idea of ​​global normalization to identify and semantically understand named entities, nested entities and non-nested entities of instantiated representations can be identified indiscriminately, with good recognition effects for both types of entities. It can also extract the correspondence between the instantiated representation of event trigger words and event types, as well as the correspondence between the instantiated representation of event argument roles and event types based on semantic understanding, thereby jointly extracting the instantiated representation of event trigger words (such as throwing, throwing, etc.) and the instantiated representation of event argument roles (such as September 30, 2024 and children, stones, empty bottles, etc.) corresponding to the preset event types (such as attack events, conflict events, etc.) based on these two types of correspondences.

[0051] Step S120, based on the preset event type, the extracted instantiation representation of the corresponding event trigger word and event argument role, and the corresponding event argument role, a corresponding temporal event quintuple is generated to form structured text data.

[0052] The specific form of the generated time series event quintuple is (event type, event occurrence time, event trigger word, event argument role, event argument), where the event argument role refers to other argument roles except the event occurrence time, and the event argument refers to other arguments except the event occurrence time. For example, Figure 3 As shown, for example, it can be: (attack / conflict, September 30, 2024, throwing, location, town entrance), (attack / conflict, September 30, 2024, throwing, attacker, child), (attack / conflict, September 30, 2024, throwing, means, stones and / or empty bottles), (attack / conflict, September 30, 2024, throwing, target, bad guy who bullies children), etc., where September 30, 2024 is the time when the attack / conflict incident occurs.

[0053] In some embodiments, before forming the structured text data in step S120, the method further includes: constructing a corresponding event graph based on the generated temporal event quintuple.

[0054] That is to say, the structured text data can be a five-tuple of time-series events, or an event graph constructed by five-tuples of time-series events. The event graph helps to mine more relationships and details in a certain type of event.

[0055] Step S130: input the structured text data into a large language model, so that the large language model outputs event text data corresponding to the preset event type, and adds annotation information to the output event text data.

[0056] In the specific implementation process, the large language model used in step S130 is not limited to which large language model is used, and the large language model used in the process of forming the event trigger word and argument candidate vocabulary library in step S110 is not limited to which large language model is used. The large language models used in step S110 and step S130 can be the same or different. The structured text data formed by the time series event quintuple or event graph is input into the large language model as the context prompt information of the large language model, guiding the large language model to generate the corresponding specific event text data, and finally automatically generating the event text data with annotation information by adding annotations, such as Figure 3 As shown, a data set or enhanced data set is formed for training the event extraction model in the corresponding task scenario. <time> <time>< / time> 、 <attacker> <attacker / > 、 <place> <place / > 、 <trigger> <trigger / > 、 <instrument> <instrument / > 、 <target> <target / > etc. are added to the corresponding event text data generated.

[0057] In some embodiments, before the step of adding annotation information to the output event text data, the method further includes: performing quality assessment and proofreading on the output event text data based on the generated temporal event quintuple.

[0058] In the actual implementation process, due to some limitations of some large language models themselves, such as the so-called hallucination, it cannot be fully guaranteed that the generated event text data can be generated according to the structured text data as contextual prompt information. Therefore, it is necessary to use the generated time series event quintuples to conduct quality assessment and proofreading of the output event text data.

[0059] Specifically, the quality assessment and proofreading of the output event text data is performed based on the generated time series event quintuples, including the following steps:

[0060] Perform string matching on the output event text data and the corresponding time series event quintuple to obtain the string matching result;

[0061] The output event text data and the corresponding time series event quintuple are subjected to semantic similarity calculation to obtain a semantic similarity calculation result;

[0062] A first score corresponding to the output event text data is determined based on the string matching result and the semantic similarity calculation result; if the first score is less than a first threshold, the step of inputting the structured text data into a large language model is repeated to optimize the large language model until the first score finally obtained is greater than or equal to the first threshold.

[0063] The output event text data and the corresponding time series event quintuple are subjected to semantic similarity calculation to obtain the semantic similarity calculation result, including the following steps:

[0064] Segment the output event text data to obtain multiple segmented words;

[0065] Inputting the multiple word segments and all words of the time sequence event quintuple into a pre-trained language model respectively, so that the pre-trained language model outputs multiple semantic vectors corresponding to the multiple word segments and all words of the time sequence event quintuple respectively;

[0066] Perform pairwise similarity calculation on the multiple semantic vectors corresponding to the multiple word segments and the multiple semantic vectors corresponding to all words of the time series event quintuple to obtain multiple similarity calculation results;

[0067] A portion of similarity calculation results greater than a second threshold is screened out from the plurality of similarity calculation results and averaged to obtain a semantic similarity calculation result.

[0068] After the above steps, the generated event text data is quality assessed and proofread to achieve higher accuracy.

[0069] In summary, the method for reverse generation of text data based on event patterns in the embodiment of the present invention utilizes the asymmetric characteristics of large language models to generate text data. This method fully exploits the advantages of large language models, fully exploits event structure information, and converts event structure information into time series event quintuples or event graphs, more effectively providing contextual prompt information for large language models, that is, fully exploiting the ability of large language models to generate event text data from the structured information of event patterns, making the generated event text data diverse in type, and making the text semantic information more fluent, reasonable and natural, thereby reversely generating a data set or enhanced data set for training event extraction models in specific scenario tasks. In addition, by automatically generating event text data with annotation information through annotation, annotation overhead and cost can be saved.

[0070] Corresponding to the above method, the present invention also provides a text data reverse generation device based on event pattern, which includes a computer device, the computer device includes a processor and a memory, the memory stores computer instructions, the processor is used to execute the computer instructions stored in the memory, and when the computer instructions are executed by the processor, the device implements the steps of the above method.

[0071] The embodiment of the present invention also provides a computer-readable storage medium on which a computer program is stored, and when the computer program is executed by a processor, the steps of the aforementioned method are implemented. The computer-readable storage medium can be a tangible storage medium, such as a random access memory (RAM), a memory, a read-only memory (ROM), an electrically programmable ROM, an electrically erasable programmable ROM, a register, a floppy disk, a hard disk, a removable storage disk, a CD-ROM, or any other form of storage medium known in the technical field.

[0072] An embodiment of the present invention further provides a computer program product, comprising computer instructions, which implement the steps of the aforementioned method when executed by a processor.

[0073] It should be understood by those skilled in the art that the exemplary components, systems and methods described in conjunction with the embodiments disclosed herein can be implemented in hardware, software or a combination of the two. Whether it is performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of the present invention. When implemented in hardware, it can be, for example, an electronic circuit, an application specific integrated circuit (ASIC), appropriate firmware, a plug-in, a function card, etc. When implemented in software, the elements of the present invention are programs or code segments used to perform the required tasks. The program or code segment can be stored in a machine-readable medium, or transmitted on a transmission medium or a communication link via a data signal carried in a carrier.

[0074] It should be clear that the present invention is not limited to the specific configuration and processing described above and shown in the figures. For the sake of simplicity, a detailed description of the known method is omitted here. In the above embodiments, several specific steps are described and shown as examples. However, the method process of the present invention is not limited to the specific steps described and shown, and those skilled in the art can make various changes, modifications and additions, or change the order between the steps after understanding the spirit of the present invention.

[0075] In the present invention, features described and / or illustrated for one embodiment may be used in the same or similar manner in one or more other embodiments, and / or combined with features of other embodiments or replace features of other embodiments.

[0076] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the embodiments of the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.< / target> < / instrument> < / trigger> < / place> < / attacker> < / time>

Claims

1. A method for reverse generation of text data based on event patterns, characterized in that: The method comprises: Based on a preset event type, event trigger words and event argument role instantiations corresponding to the event type are jointly extracted from an event trigger word and argument candidate vocabulary library, wherein the event trigger word and argument candidate vocabulary library is pre-formed by instantiations of a plurality of event trigger words and a plurality of event argument roles corresponding to a plurality of event types, and the instantiations of the plurality of event trigger words and a plurality of event argument roles corresponding to the plurality of event types are respectively obtained by inputting a plurality of determined event patterns including event types, event trigger words and event argument roles into a large language model, and then outputted by the large language model; Based on the preset event type, the extracted instantiation representation of the corresponding event trigger word and event argument role, and the corresponding event argument role, a corresponding time series event quintuple is generated to form structured text data; The structured text data is input into a large language model so that the large language model outputs event text data corresponding to the preset event type, and adds annotation information to the output event text data.

2. The method according to claim 1, characterized in that Before the step of adding annotation information to the output event text data, the method further includes: performing quality assessment and proofreading on the output event text data based on the generated time series event quintuple.

3. The method according to claim 2, characterized in that Based on the generated time series event quintuples, the output event text data is quality assessed and proofread, including: Perform string matching on the output event text data and the corresponding time series event quintuple to obtain the string matching result; The output event text data and the corresponding time series event quintuple are subjected to semantic similarity calculation to obtain a semantic similarity calculation result; A first score corresponding to the output event text data is determined based on the string matching result and the semantic similarity calculation result; if the first score is less than a first threshold, the step of inputting the structured text data into a large language model is repeated to optimize the large language model until the first score finally obtained is greater than or equal to the first threshold.

4. The method according to claim 3, characterized in that The output event text data and the corresponding time series event quintuple are used to calculate the semantic similarity, and the semantic similarity calculation results are obtained, including: Segment the output event text data to obtain multiple segmented words; Inputting the multiple word segments and all words of the time sequence event quintuple into a pre-trained language model respectively, so that the pre-trained language model outputs multiple semantic vectors corresponding to the multiple word segments and all words of the time sequence event quintuple respectively; Perform pairwise similarity calculation on the multiple semantic vectors corresponding to the multiple word segments and the multiple semantic vectors corresponding to all words of the time series event quintuple to obtain multiple similarity calculation results; A portion of similarity calculation results greater than a second threshold is screened out from the plurality of similarity calculation results and averaged to obtain a semantic similarity calculation result.

5. The method according to claim 1, characterized in that Based on the preset event type, the event trigger word and the instantiation representation of the event argument role corresponding to the event type are jointly extracted from the event trigger word and argument candidate vocabulary library, including: Encoding the instantiation representations of multiple event trigger words and multiple event argument roles corresponding to multiple event types in the event trigger word and argument candidate vocabulary library respectively, and obtaining multiple text vector representations accordingly; A named entity recognition pre-trained model based on global normalization is used to decode multiple text vector representations, and the corresponding relationships between the instantiation representations of event trigger words and event argument roles and event types are extracted jointly; According to the correspondence between the instantiation representations of event trigger words and event argument roles and the event types, respectively, event trigger words and event argument roles corresponding to the event type are jointly extracted from the event trigger word and argument candidate vocabulary library based on the preset event type.

6. The method according to claim 1, characterized in that Before forming the structured text data, the method further includes: constructing a corresponding event graph based on the generated time series event quintuple.

7. The method according to claim 1, characterized in that When there is a labeled data set, the event pattern is determined by extracting the event pattern in the labeled data set; when there is no labeled data set, the event pattern is determined by customizing the event pattern.

8. A text data reverse generation device based on event mode, comprising a processor, a memory and computer instructions stored in the memory, characterized in that: The processor is used to execute the computer instructions, and when the computer instructions are executed, the device implements the steps of the method according to any one of claims 1 to 7.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 7 are implemented.

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

Images