WO2024066041A1 - Electronic letter of guarantee automatic generation method and apparatus based on sequence adversary and priori reasoning - Google Patents

Abstract

The present invention relates to an electronic letter of guarantee automatic generation method and apparatus based on a sequence adversary and priori reasoning. In the method and the apparatus, acquired initial training data is inputted into a generator G in a letter of guarantee generation model; then, the generator G receives whole-sentence input text in the initial training data by means of a self-attention encoder of the generator G, outputs an encoder latent vector, and the latent vector is then inputted into a decoder in the generator G; finally, in a decoding stage, the decoder predicts the next word according to one word inputted each time, and finally outputs a complete sentence for an electronic letter of guarantee. In the present invention, on the basis of a sequence adversarial generative network and prior knowledge in a knowledge graph, automatic letter of guarantee generation can be achieved for insurance of specific types by means of learning the form, style, and structure of text content of electronic letters of guarantee written in an electronic natural language, thereby reducing the time costs of insurance managers.

Description

Method and device for automatically generating electronic letter of guarantee based on sequence confrontation and prior reasoning Technical Field

The present invention relates to the field of electronic letters of guarantee, and in particular to a method and device for automatically generating an electronic letter of guarantee based on sequence adversarial and priori reasoning.

Background technique

A guarantee letter refers to a written credit guarantee certificate issued by a bank, insurance company, guarantee company or individual to a third party at the request of an applicant. However, with the increase in the number of electronic guarantee letters, it is difficult to display the business information related to the guarantee letter in a centralized manner. The timeliness tracking and classified information query of the guarantee letter itself still require manual intervention. The traditional electronic guarantee letter only changes the display form of the guarantee letter. The management of the guarantee letter and related business information has not changed much from the original offline paper guarantee letter. The entire guarantee letter issuance process requires a lot of time and manpower costs, which is inefficient.

technical problem

The embodiments of the present invention provide a method and device for automatically generating an electronic letter of guarantee based on sequence adversarial and a priori reasoning, so as to at least solve the technical problem of low efficiency in issuing existing electronic letters of guarantee.

Technical Solutions

According to an embodiment of the present invention, a method for automatically generating an electronic letter of guarantee based on sequence adversarial and prior reasoning is provided, comprising the following steps:

S101: Input the acquired initial training data into the generator G in the letter of guarantee generation model;

S102: Generator G receives the entire sentence input text in the initial training data through its self-attention encoder, outputs the encoder latent vector, and the latent vector is then input into the decoder in generator G;

S103: During the decoding stage, the decoder predicts the next word based on each word input, and finally outputs a complete sentence for the electronic letter of guarantee.

Furthermore, after step S103, the method further includes:

S104: The generator G inputs the output result to the discriminator D in the guarantee generation model to train the guarantee generation model.

Furthermore, step S104 specifically includes:

The discriminator D adopts the EMLo model. The EMLo model receives the serialized text generated by the generator G and outputs a vector with the same dimension as the existing topic word type. The probability of each type of topic is calculated through Softmax, and the topic type with the highest probability is selected for comparison with the actual topic type label. The discriminator D and the generator G are updated simultaneously through the cross-entropy loss back-propagation method to train the guarantee generation model.

Furthermore, the discriminator D adds a penalty term when calculating the loss function, indicating whether the sentence to be judged includes a priori high-frequency words. If it does, it then determines whether they appear in order and imposes a penalty based on the difference in order, imposing a higher loss on generated sentences that do not conform to empirical rules.

Furthermore, the generator G uses the ALBERT model, and the input is a text sentence and its topic type label; the ALBERT model is configured as: a lightweight optimized version of the classic NLP model BERT, which reduces the number of model parameters by linearly decoupling the input matrix and splitting the input matrix of the original BERT model into the product of two small matrices.

Furthermore, step S102 specifically includes:

The generator G first uses the pre-trained word vector to convert the words in the text sentence into semantic numerical vectors according to the word segmentation results, and then inputs them into the ALBERT model. It receives the entire sentence input text through its self-attention encoder and outputs the encoder latent vector, which is then input into its decoder.

Furthermore, before step S101, the method further includes:

S100: Based on the experience of safety risk management of construction projects, sort out the subject types of current electronic guarantees for construction projects; collect existing electronic guarantee documents of different subject types, extract the text of the electronic guarantee documents and use open source Chinese natural language processing tools for pre-processing, and annotate the text according to the subject category to which it belongs as initial training data.

According to another embodiment of the present invention, there is provided an apparatus for automatically generating an electronic letter of guarantee based on sequence adversarial and priori reasoning, comprising:

A data input unit, used to input the acquired initial training data into the generator G in the letter of guarantee generation model;

A data receiving unit is used for the generator G to receive the whole sentence input text in the initial training data through its self-attention encoder, output the encoder latent vector, and the latent vector is then input into the decoder in the generator G;

The decoding unit is used for the decoder to predict the next word according to each word input during the decoding stage, and finally output a complete sentence for the electronic letter of guarantee.

Furthermore, the device also includes:

The training unit is used for the generator G to input the output result into the discriminator D in the guarantee generation model to train the guarantee generation model.

Furthermore, the device also includes:

The data acquisition unit is used to sort out the subject types of current electronic guarantees for construction projects based on the experience of safety risk management of construction projects; collect existing electronic guarantee documents of different subject types, extract the text of the electronic guarantee documents and use open source Chinese natural language processing tools for pre-processing, and annotate the text according to the subject category to which it belongs as initial training data.

A storage medium storing a program file capable of implementing any one of the above-mentioned methods for automatically generating an electronic letter of guarantee based on sequence adversarial and a priori reasoning.

A processor is used to run a program, wherein when the program is run, any one of the above-mentioned methods for automatically generating an electronic letter of guarantee based on sequence adversarial and priori reasoning is executed.

Beneficial Effects

The method and device for automatically generating an electronic letter of guarantee based on sequence adversarial and prior reasoning in the embodiments of the present invention are based on the prior knowledge in the sequence adversarial generation network and the knowledge graph. By learning the form, style, and structure of the electronic letter of guarantee text content written in electronic natural language, it can realize automatic generation of a letter of guarantee for a given specific insurance type, thereby saving the time cost of insurance managers.

BRIEF DESCRIPTION OF THE DRAWINGS

The drawings described herein are used to provide a further understanding of the present invention and constitute a part of this application. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

FIG1 is a flow chart of a method for automatically generating an electronic letter of guarantee based on sequence adversarial and prior reasoning according to the present invention;

FIG2 is a preferred flow chart of a method for automatically generating an electronic letter of guarantee based on sequence confrontation and a priori reasoning according to the present invention;

FIG3 is another preferred flow chart of the method for automatically generating an electronic letter of guarantee based on sequence confrontation and a priori reasoning according to the present invention;

FIG4 is a module diagram of a device for automatically generating an electronic letter of guarantee based on sequence adversarial and priori reasoning according to the present invention;

FIG5 is a preferred module diagram of a device for automatically generating an electronic letter of guarantee based on sequence confrontation and a priori reasoning according to the present invention;

FIG6 is another preferred module diagram of the device for automatically generating an electronic letter of guarantee based on sequence adversarial and a priori reasoning according to the present invention.

Embodiments of the present invention

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

It should be noted that the terms "first", "second", etc. in the specification and claims of the present invention 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 interchanged where appropriate, so that the embodiments of the present invention described herein can be implemented in an order other than those illustrated or described herein. In addition, the terms "including" and "having" and any variations thereof are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device that includes 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.

Example 1

According to an embodiment of the present invention, a method for automatically generating an electronic letter of guarantee based on sequence adversarial and prior reasoning is provided, as shown in FIG1 , and includes the following steps:

S101: Input the acquired initial training data into the generator G in the letter of guarantee generation model;

S102: Generator G receives the entire sentence input text in the initial training data through its self-attention encoder, outputs the encoder latent vector, and the latent vector is then input into the decoder in generator G;

S103: During the decoding stage, the decoder predicts the next word based on each word input, and finally outputs a complete sentence for the electronic letter of guarantee.

The method for automatically generating an electronic letter of guarantee based on sequence adversarial and prior reasoning in the embodiment of the present invention is based on the prior knowledge in the sequence adversarial generation network and the knowledge graph. By learning the form, style, and structure of the electronic letter of guarantee text content written in electronic natural language, it can realize automatic generation of letters of guarantee for a given specific insurance type, thereby saving the time cost of insurance managers.

After step S103, referring to FIG. 2 , the method further includes:

S104: The generator G inputs the output result to the discriminator D in the guarantee generation model to train the guarantee generation model.

Wherein, step S104 specifically includes:

The discriminator D adopts the EMLo model. The EMLo model receives the serialized text generated by the generator G and outputs a vector with the same dimension as the existing topic word type. The probability of each type of topic is calculated through Softmax, and the topic type with the highest probability is selected for comparison with the actual topic type label. The discriminator D and the generator G are updated simultaneously through the cross-entropy loss back-propagation method to train the guarantee generation model.

Among them, the discriminator D adds a penalty term when calculating the loss function, which indicates whether the sentence to be judged includes a priori high-frequency words. If it does, it then determines whether they appear in order and imposes a penalty based on the difference in order, imposing a higher loss on the generated sentences that do not conform to the empirical rules.

The generator G uses the ALBERT model, and the input is a text sentence and its topic type label; the ALBERT model is configured as a lightweight optimized version of the classic NLP model BERT, which reduces the number of model parameters by linearly decoupling the input matrix and splitting the input matrix of the original BERT model into the product of two small matrices.

Wherein, step S102 specifically includes:

The generator G first uses the pre-trained word vector to convert the words in the text sentence into semantic numerical vectors according to the word segmentation results, and then inputs them into the ALBERT model. It receives the entire sentence input text through its self-attention encoder and outputs the encoder latent vector, which is then input into its decoder.

Before step S101, referring to FIG3 , the method further includes:

S100: Based on the experience of safety risk management of construction projects, sort out the subject types of current electronic guarantees for construction projects; collect existing electronic guarantee documents of different subject types, extract the text of the electronic guarantee documents and use open source Chinese natural language processing tools for pre-processing, and annotate the text according to the subject category to which it belongs as initial training data.

The following is a detailed description of the method for automatically generating an electronic letter of guarantee based on sequence confrontation and prior reasoning according to the present invention using a specific embodiment:

The method for automatically generating electronic letters of guarantee for construction projects based on sequence adversarial and prior reasoning is based on the prior knowledge in the sequence adversarial generation network and the knowledge graph. By learning the form, style, and structure of the electronic letter of guarantee text written in electronic natural language, it can realize automatic letter of guarantee generation for a given specific insurance type (i.e., subject), saving the time cost of insurance managers. It specifically includes the following steps:

Based on the experience of safety risk management of construction projects, the subject types of current electronic guarantees for construction projects are sorted out. For example, according to the severity, they can be divided into death, serious injury, and minor injury; according to the type of accident, they can be divided into falling from height, being hit by objects, and electric shock, etc.; existing electronic guarantee documents of different subject types are collected, the text of the electronic guarantee documents is extracted, and pre-processed using open source Chinese natural language processing tools such as hanNLP and Jeba. The pre-processing mainly includes sentence segmentation, word segmentation, stop word removal, etc. The text is annotated according to the subject category to which it belongs as initial training data.

Given the initial training data, use it to train the generator G in the guarantee generation model; the generator G uses the ALBERT model, and the input is the text sentence and its topic type label. The ALBERT model is a lightweight optimized version of the classic NLP model BERT. It reduces the number of model parameters by linearly decoupling the input matrix, that is, splitting the input matrix of the original BERT model into the product of two smaller matrices. The degree of optimization depends on the total number of words that the BERT model needs to process. The higher the total number, the more obvious the acceleration effect, which can generally increase the calculation speed by 8-10 times.

First, use pre-trained word vectors (word2vec or glove) to convert the words in the text sentence into semantic numerical vectors according to the word segmentation results, and then input them into the ALBERT model. This model is the same as the BERT model except for decoupling and tuning at the input end. It receives the entire sentence input text through its self-attention encoder and outputs the encoder latent vector, which is then input into its decoder. In the decoding stage, one word is input each time, and the guarantee generation model calculates the probability of all words in the corpus in the current state for selection. Then the decoder predicts the next word, thereby realizing continuous text generation. The decoder finally outputs a complete sentence for the electronic guarantee part of the construction project. Generator G inputs the output result into discriminator D, whose purpose is to determine whether the generated sequence comes from the real data set, so as to guide the improvement and optimization of generator G; discriminator D adopts EMLo model, which contains a multi-layer bidirectional long short-term memory network Bi-LSTM, which can well capture the semantics of sequence text and will not lose the information of words at earlier positions in the sequence; EMLo model receives the serialized text generated by generator G, outputs a vector with the same dimension as the existing topic word type, calculates the probability of each type of topic through Softmax, selects the topic type with the highest probability and compares it with the actual topic type label, and simultaneously updates discriminator D and generator G through the cross-entropy loss back propagation method, thereby realizing the training of the guarantee generation model.

The innovative features and beneficial effects of the present invention are at least:

1) Different from the general text generation model based on the generative adversarial network, the guarantee generation model of the present invention takes into account the general structure of the engineering electronic guarantee, including various fill-in-the-blanks. For example, the blanks such as the name of the insured and the age are all included in the training data, so the model can learn and generate automatically.

2) Through prior knowledge, we can know that in the specific field of construction project insurance, certain high-frequency words appearing in a fixed order can indicate that a sentence is reasonable, and violating this rule is likely to indicate that the sentence does not conform to the empirical rules of this field. Therefore, the model of the present invention adds a penalty term when the discriminator D calculates the loss function, indicating whether the sentence to be judged includes the prior high-frequency words. If included, it is then determined whether they appear in order, and penalties are imposed based on the difference in order, imposing higher losses on generated sentences that do not conform to the empirical rules.

Example 2

According to another embodiment of the present invention, there is provided an apparatus for automatically generating an electronic letter of guarantee based on sequence confrontation and a priori reasoning, as shown in FIG4 , comprising:

The data input unit 201 is used to input the acquired initial training data into the generator G in the letter of guarantee generation model;

A data receiving unit 202 is used for the generator G to receive the whole sentence input text in the initial training data through its self-attention encoder, output the encoder latent vector, and the latent vector is then input to the decoder in the generator G;

The decoding unit 203 is used for the decoder to predict the next word according to each word input during the decoding stage, and finally output a complete sentence for the electronic guarantee part.

The automatic generation device for electronic letters of guarantee based on sequence adversarial and prior reasoning in the embodiment of the present invention is based on the prior knowledge in the sequence adversarial generation network and the knowledge graph. By learning the form, style and structure of the electronic letter of guarantee text content written in electronic natural language, it can realize automatic generation of letters of guarantee for a given specific insurance type, thereby saving the time cost of insurance managers.

Wherein, referring to FIG5, the device further includes:

The training unit 204 is used for the generator G to input the output result to the discriminator D in the guarantee generation model to train the guarantee generation model.

Wherein, referring to FIG6 , the device further includes:

The data acquisition unit 200 is used to sort out the subject types of current electronic guarantees for construction projects based on the experience of safety risk management of construction projects; collect existing electronic guarantee documents of different subject types, extract the text of the electronic guarantee documents and use open source Chinese natural language processing tools for pre-processing, and annotate the text according to the subject category to which it belongs as initial training data.

The following is a detailed description of the electronic guarantee automatic generation device based on sequence confrontation and prior reasoning of the present invention with a specific embodiment:

The automatic generation device of electronic guarantee for construction projects based on sequence adversarial and prior reasoning is based on the prior knowledge in the sequence adversarial generation network and the knowledge graph. By learning the form, style and structure of the electronic guarantee text content written in electronic natural language, it can realize the automatic generation of guarantee for a given specific insurance type (i.e., subject), saving the time cost of insurance managers. It specifically includes the following steps:

Data acquisition unit 200: Based on the experience of construction project safety risk management, sort out the subject types of current electronic guarantees for construction projects, such as death, serious injury, and minor injury according to severity, and fall from height, object impact, electric shock, etc. according to accident type; collect existing electronic guarantee documents of different subject types, extract the text of the electronic guarantee documents and use open source Chinese natural language processing tools such as hanNLP and Jeba for preprocessing. The preprocessing mainly includes sentence segmentation, word segmentation, stop word removal, etc. The text is annotated according to the subject category to which it belongs as initial training data.

Data input unit 201: Given initial training data, use it to train the generator G in the guarantee generation model; the generator G uses the ALBERT model, and the input is a text sentence and its topic type label. The ALBERT model is a lightweight optimized version of the classic NLP model BERT. It reduces the number of model parameters by linearly decoupling the input matrix, that is, splitting the input matrix of the original BERT model into the product of two smaller matrices. The degree of optimization depends on the total amount of words that the BERT model needs to process. The higher the total amount, the more obvious the acceleration effect, which can generally increase the calculation speed by 8-10 times.

Data receiving unit 202: First, use the pre-trained word vector (word2vec or glove) to convert the words in the text sentence into semantic numerical vectors according to the word segmentation results, and then input them into the ALBERT model. This model is the same as the BERT model except for the decoupling tuning at the input end. It receives the entire sentence input text through its self-attention encoder and outputs the encoder latent vector, which is then input into its decoder.

Decoding unit 203: Each time a word is input in the decoding stage, the guarantee generation model calculates the probability of all words in the corpus in the current state for selection, and then the decoder predicts the next word, thereby realizing continuous text generation, and the decoder finally outputs a complete sentence for the electronic guarantee part of the construction project. Training unit 204: Generator G inputs the output result into the discriminator D, the purpose of which is to determine whether the generated sequence comes from a real data set, thereby guiding the improvement and optimization of generator G; discriminator D adopts the EMLo model, which contains a multi-layer bidirectional long short-term memory network Bi-LSTM, which can well capture the semantics of the sequence text without losing the information of the words at the earlier position of the sequence; the EMLo model receives the serialized text generated by generator G, outputs a vector with the same dimension as the existing topic word type, calculates the probability of each type of topic through Softmax, selects the topic type with the highest probability and compares it with the actual topic type label, and updates the discriminator D and generator G simultaneously through the cross entropy loss back propagation method, thereby realizing the training of the guarantee generation model.

The innovative features and beneficial effects of the present invention are at least:

1) Different from the general text generation model based on the generative adversarial network, the guarantee generation model of the present invention takes into account the general structure of the engineering electronic guarantee, including various fill-in-the-blanks. For example, the blanks such as the name of the insured and the age are all included in the training data, so the model can learn and generate automatically.

2) Through prior knowledge, we can know that in the specific field of construction project insurance, certain high-frequency words appearing in a fixed order can indicate that a sentence is reasonable, and violating this rule is likely to indicate that the sentence does not conform to the empirical rules of this field. Therefore, the model of the present invention adds a penalty term when the discriminator D calculates the loss function, indicating whether the sentence to be judged includes the prior high-frequency words. If included, it is then determined whether they appear in order, and penalties are imposed based on the difference in order, imposing higher losses on generated sentences that do not conform to the empirical rules.

Example 3

A storage medium storing a program file capable of implementing any one of the above-mentioned methods for automatically generating an electronic letter of guarantee based on sequence adversarial and a priori reasoning.

Example 4

A processor is used to run a program, wherein when the program is run, any one of the above-mentioned methods for automatically generating an electronic letter of guarantee based on sequence adversarial and priori reasoning is executed.

The serial numbers of the above embodiments of the present invention are only for description and do not represent the advantages or disadvantages of the embodiments.

In the above embodiments of the present invention, the description of each embodiment has its own emphasis. For parts that are not described in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

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 system embodiments described above are only schematic. For example, the division of 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.

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.

In addition, each functional unit in each embodiment of the present invention 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.

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 invention, 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 for 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 methods of each embodiment of the present invention. The aforementioned storage medium includes: U disk, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), mobile hard disk, magnetic disk or optical disk and other media that can store program codes.

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

Claims (10)

A method for automatically generating an electronic letter of guarantee based on sequence adversarial and prior reasoning, characterized in that it comprises the following steps: S101: Input the acquired initial training data into the generator G in the letter of guarantee generation model; S102: Generator G receives the entire sentence input text in the initial training data through its self-attention encoder, outputs the encoder latent vector, and the latent vector is then input into the decoder in generator G; S103: During the decoding stage, the decoder predicts the next word based on each word input, and finally outputs a complete sentence for the electronic letter of guarantee. The method for automatically generating an electronic letter of guarantee based on sequence adversarial and priori reasoning according to claim 1, characterized in that after step S103, the method further comprises: S104: The generator G inputs the output result to the discriminator D in the guarantee generation model to train the guarantee generation model. The method for automatically generating an electronic letter of guarantee based on sequence adversarial and priori reasoning according to claim 2 is characterized in that step S104 specifically comprises: The discriminator D adopts the EMLo model. The EMLo model receives the serialized text generated by the generator G and outputs a vector with the same dimension as the existing topic word type. The probability of each type of topic is calculated through Softmax, and the topic type with the highest probability is selected for comparison with the actual topic type label. The discriminator D and the generator G are updated simultaneously through the cross-entropy loss back-propagation method to train the guarantee generation model. The method for automatically generating an electronic letter of guarantee based on sequence adversarial and prior reasoning according to claim 3 is characterized in that a penalty term is added when the discriminator D calculates the loss function, indicating whether the sentence to be judged includes a priori high-frequency words. If included, it is then judged whether they appear in sequence, and a penalty is imposed based on the difference in sequence, imposing a higher loss on the generated sentences that do not conform to the empirical rules. According to claim 1, the method for automatically generating an electronic letter of guarantee based on sequence adversarial and prior reasoning is characterized in that the generator G uses the ALBERT model, and the input is a text sentence and its topic type label; the ALBERT model is configured as: a lightweight optimized version of the classic NLP model BERT, which reduces the number of model parameters by linearly decoupling the input matrix and splitting the input matrix of the original BERT model into the product of two small matrices. The method for automatically generating an electronic letter of guarantee based on sequence adversarial and priori reasoning according to claim 5 is characterized in that step S102 specifically comprises: The generator G first uses the pre-trained word vector to convert the words in the text sentence into semantic numerical vectors according to the word segmentation results, and then inputs them into the ALBERT model. It receives the entire sentence input text through its self-attention encoder and outputs the encoder latent vector, which is then input into its decoder. The method for automatically generating an electronic letter of guarantee based on sequence adversarial and a priori reasoning according to claim 1 is characterized in that, before step S101, the method further comprises: S100: Based on the experience of safety risk management of construction projects, sort out the subject types of current electronic guarantees for construction projects; collect existing electronic guarantee documents of different subject types, extract the text of the electronic guarantee documents and use open source Chinese natural language processing tools for pre-processing, and annotate the text according to the subject category to which it belongs as initial training data. An automatic generation device for an electronic letter of guarantee based on sequence confrontation and a priori reasoning, characterized by comprising: A data input unit, used to input the acquired initial training data into the generator G in the letter of guarantee generation model; A data receiving unit is used for the generator G to receive the whole sentence input text in the initial training data through its self-attention encoder, output the encoder latent vector, and the latent vector is then input into the decoder in the generator G; The decoding unit is used for the decoder to predict the next word according to each word input during the decoding stage, and finally output a complete sentence for the electronic letter of guarantee. The device for automatically generating an electronic letter of guarantee based on sequence confrontation and a priori reasoning according to claim 8, characterized in that the device also includes: The training unit is used for the generator G to input the output result into the discriminator D in the guarantee generation model to train the guarantee generation model. The device for automatically generating an electronic letter of guarantee based on sequence confrontation and a priori reasoning according to claim 8, characterized in that the device also includes: The data acquisition unit is used to sort out the subject types of current electronic guarantees for construction projects based on the experience of safety risk management of construction projects; collect existing electronic guarantee documents of different subject types, extract the text of the electronic guarantee documents and use open source Chinese natural language processing tools for pre-processing, and annotate the text according to the subject category to which it belongs as initial training data.