WO2021174827A1

WO2021174827A1 - Text generation method and appartus, computer device and readable storage medium

Info

Publication number: WO2021174827A1
Application number: PCT/CN2020/118456
Authority: WO
Inventors: 邓悦; 金戈; 徐亮
Original assignee: 平安科技（深圳）有限公司
Priority date: 2020-03-02
Filing date: 2020-09-28
Publication date: 2021-09-10
Also published as: CN111428448B; CN111428448A

Abstract

A text generation method and apparatus, a computer device and a readable storage medium, which belong to the field of text processing. The text generation method and apparatus, the computer device and the readable storage medium generate data of a target text according to target guide data by using a pretrained text generative adversarial network model, thus solving the problem of being unable to update a discrete output, thereby achieving the objective of being able to generate a text statement according to sentence header data by using a text generative adversarial network model.

Description

Text generation method, device, computer equipment and readable storage medium

This application claims the priority of a Chinese patent application filed on March 2, 2020 with the application number 202010136551.6 and titled "text generation method, device, computer equipment and readable storage medium", the entire content of which is incorporated by reference In this application.

Technical field

This application relates to the field of text processing, in particular to text generation methods, devices, computer equipment and readable storage media.

Background technique

In the intelligent interview scenario, artificial intelligence (Artificial Intelligence, referred to as AI), in addition to asking the candidates according to the pre-set questions, should also ask the candidates open questions based on the actual situation to test the candidates’ Actual response ability. Open questions require AI to use a generative model to generate question text.

The current generative model mainly adopts the generative confrontation network (GAN). Since the generative confrontation network needs to update parameter variables based on continuous output data, it is mainly used in image processing. Various image generation tasks include unsupervised generation, labeled generation, Super-resolution restoration, automatic coloring, street view generation, etc., the quality of the generated pictures is so real that it is difficult for the human eye to distinguish the authenticity.

The inventor realized that when the generative confrontation network is applied to the text generation task, because in the text generation process, the generative confrontation network needs to output the probability distribution of the next word in the vocabulary based on the generated text sequence, and then select the word , The output result is discrete data, and discrete data cannot achieve network training update. Therefore, the current generative confrontation network cannot be applied to text generation tasks.

Summary of the invention

Aiming at the problem that the existing generative confrontation network only supports continuous output, a text generation method, device, computer equipment and readable storage medium based on a text generation confrontation network that can be updated based on discrete data are now provided.

In order to achieve the above purpose, this application provides a text generation method, which includes the following steps:

Collect the answer data generated by the business object in the question and answer scenario;

Extract the answer data, and obtain target guidance data;

Generating a confrontation network model through a text obtained through pre-training, and generating target text data according to the target guidance data;

The target guide data is sentence beginning data of the target text data.

In order to achieve the above objective, this application also provides a text generation device, including:

The collection unit is used to collect the answer data generated by the business object in the question and answer scenario;

An obtaining unit, configured to extract the answer data and obtain target guidance data;

A generating unit, configured to generate a confrontation network model through pre-trained text and generate target text data according to the target guidance data;

The target guide data is sentence beginning data of the target text data.

To achieve the above objective, the present application also provides a computer device, the computer device including a memory, a processor, and computer readable instructions stored in the memory and running on the processor, and the processor executes the computer The method for generating text when the instruction is readable includes the following steps:

Extract the answer data, and obtain target guidance data;

The target guide data is sentence beginning data of the target text data.

In order to achieve the foregoing objective, the present application also provides a computer-readable storage medium on which computer-readable instructions are stored. When the computer-readable instructions are executed by a processor, the method for generating text includes the following steps:

Extract the answer data, and obtain target guidance data;

The target guide data is sentence beginning data of the target text data.

The beneficial effects of the above technical solutions:

In this technical solution, the text generation method, device, computer equipment and readable storage medium generate the target text data according to the target guidance data (for example: sentence beginning data) through the text generation confrontation network model obtained by pre-training, which solves the problem that the discrete output cannot be updated. The problem of using the text generation confrontation network model can be used to generate text sentences (for example: text questions) based on the data at the beginning of the sentence.

Description of the drawings

FIG. 1 is a method flowchart of an embodiment of the text generation method described in this application;

FIG. 2 is a flow diagram of an embodiment of obtaining a text generation confrontation network model;

FIG. 3 is a method flowchart of an embodiment of training the initial confrontation network model according to the sample guide set and the sample text set to obtain the text generation confrontation network model;

FIG. 4 is a block diagram of an embodiment of the text generation device described in this application;

FIG. 5 is a schematic diagram of the hardware architecture of an embodiment of the computer device described in this application.

Detailed ways

In order to make the purpose, technical solutions, and advantages of this application clearer and clearer, the following further describes the application in detail with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described here are only used to explain the application, and are not used to limit the application. Based on the embodiments in this application, all other embodiments obtained by those of ordinary skill in the art without creative work shall fall within the protection scope of this application.

The text generation method, device, computer equipment, and readable storage medium provided in this application are suitable for insurance, finance and other business fields, and provide an open text question that can be automatically generated for loan systems, insurance systems, and financial systems to facilitate testing candidates The text generation method of thinking ability. This application generates target text data based on the target guidance data (for example: sentence beginning data) through the pre-trained text generation confrontation network model, which solves the problem of non-renewable discrete output, and realizes that the text generation confrontation network model can be based on the sentence start data The purpose of generating text sentences (for example: text questions).

Example one

Please refer to Fig. 1, a text generation method of this embodiment includes the following steps:

S1. Collect the answer data generated by the business object in the question and answer scenario;

In this step, the business target can be a user who is consulting for business, or a buyer of an online trading platform, or an interviewer in the interview process. The answer data can be collected by collecting equipment (for example: audio receiving device, microphone or mobile terminal with recording function, etc.).

The text generation method in this embodiment is mainly used in conversation scenarios (at least two users), and the question text is generated based on the answer information of the target object for the target object to answer the question text. For example, the text generation method is applied to an interview. In the scenario, an open-ended text question is generated based on the keywords provided by the interviewer.

S2. Extract the answer data and obtain target guidance data;

In step S2, the answer data can be semantically analyzed to extract keywords in the answer data, and the keywords can be used as target guidance data; the answer data can be analyzed to extract nouns in the answer data, and the nouns can be used as target guidance data. data.

It should be noted that the target guidance data can be keywords or the first words of a sentence.

S3. Generate a confrontation network model through the text obtained by pre-training and generate target text data according to the target guidance data;

It should be noted that: the target guide data is the sentence beginning data of the target text data. For example: the target guidance data is: "Today"; the target text data is: "How is the weather today?". The target guiding data is that the first word of the sentence can be two words or three words, which is not limited here.

Referring to FIG. 2, before step S3 is executed, the step of obtaining the text generation confrontation network model may include:

S31. Obtain a sample guide set and a sample text set, the sample guide set including at least one sample guide data, the sample text set includes at least one sample text data, and the sample guide data is sentence beginning data of the sample text data ；

In this embodiment, the sample guide set is a sequence composed of sample guide data (sentence beginning data); the sample text set is a sequence of real text data composed of sample text data (complete sentences). The sample guide data is the sentence beginning data of the real text data.

S32. Training the initial confrontation network model according to the sample guide set and the sample text set, and obtain a text generation confrontation network model.

At present, in the image processing process of the generated confrontation network, the pixel value of each point of the generated image is continuous. Therefore, the calculation graph of the entire network, from the weight of the generator to its output, and then into the weight and output classification of the discriminator, are all It is differentiable (differentiable and differentiable), the error can be backpropagated normally, and the gradient and weight can be updated normally. However, in the process of text generation, the generator actually outputs a sequence. Each round outputs the probability distribution of the next word in the vocabulary based on the generated text sequence, and then selects the word with the highest probability, this "choice" The process is non-differentiable. The generator outputs discrete tokens. During the training process, the error is propagated back here. It is not possible to update the pixel values on each token as in the image generation task, thereby updating the generator. The weight value of. On the other hand, the discriminator can directly receive the input of a complete text sequence and output the true or false of the sentence, but it cannot judge the half of the unfinished sentences generated by the generator, which makes the discriminator unable to generate every word in the text sequence. Provide supervision on the training of the generator.

Therefore, in the training process of the generative confrontation network model of this embodiment, in order to solve the non-differentiable problem caused by the discrete output of the generator, in this embodiment, the text sequence generation process is regarded as a sequence decision process, and The policy gradient method in reinforcement learning uses the judgment result of the discriminator as a reward, part of the text generated by the generator as the state, and the generator as an agent to predict the next Words are actions, and generators are policies that need to be updated. In this way, the problem of nondifferentiable loss function of discrete output is solved. For the judgment method of the unfinished sequence, this embodiment adopts Monte Carlo search. Based on the generated sequence, the generator continues to generate until the sequence is completed, and the discriminator judges the sequence and simulates multiple times. The mean value of the final reward is used as the estimate of the reward of the current unfinished sequence.

It should be noted that: the initial confrontation network model includes a generator and a discriminator; as shown in FIG. 3, in step S32, the initial confrontation network model is trained according to the sample guide set and the sample text set, and The steps to get the text generation confrontation network model include:

As an example and not a limitation, the generator can use a long short-term memory network (LSTM) of the output sequence to generate a text sequence from a given initial state; the discriminator can use a two-class long short-term memory network for receiving and generating The output text and real text of the device can be used to determine whether the output text is true or false.

S321. Generate at least one sample text data according to at least one sample guidance data in the sample guidance set through the generator;

Further, the steps in step S321 may include:

Calculating by the generator and according to the sample guidance data, obtaining the first sample word with the highest probability in the vocabulary, and adding the first sample word to the end of the sample guidance data;

Through the generator and calculation based on the first sample word, the second sample word with the highest probability in the vocabulary is obtained, the second sample word is added to the end of the first sample word, and the loop is executed The above steps (and so on) until the sample text data of the preset length is obtained.

In this step, the generator G _θ and the discriminator D _{φ are} initialized; the sample guide data is the real text set S={X _1～T }, the sentence length of each real text in the real text set is T, and the length is less than T Fill the end with zeros; the sample guiding set is the word set {Y ₁ }.

The word set {Y ₁ } is input to the generator G _θ , and _{the input layer of the generator G θ} maps the input words to the corresponding label information (tokenized) of the corresponding words in the vocabulary for embedding and representation. In practical applications Send (y ₁ ,y ₂ ,...,y _t-1 ) as input to the generator G _θ , and the generator G _θ according to the input data, the softmax classifier outputs the probability of the next word in each word in the vocabulary, and The word with the highest probability is taken as y _t , and so on, and the loop processing is performed until the end of the sentence y _T , thereby obtaining a set of generated sample text sets {Y _1～T } of length T (length is less than zero padding).

Among them, (y ₁ , y ₂ ,..., y _t-1 ) represents an incomplete sentence consisting of t-1 words, y ₁ represents the first word in a sentence; y ₂ represents the sentence in a sentence The second word; y _t-1 means the t-1 word in a sentence; y _{T is the T} word in a sentence (end of sentence).

In the step, only the generator G _θ is used to pass in a word y ₁ , and the generator G _θ embeds it and transmits it to the LSTM, and outputs the generated token sequence and its corresponding words in the vocabulary to obtain the generated Text sequence (y ₁ ,y ₂ ,...,y _T ).

S322. Use Monte Carlo simulation to simulate the at least one sample text data and obtain multiple sample simulation text data;

Further, the steps of step S322 include:

Monte Carlo simulation is used to simulate the words in each sample text data one by one, and generate multiple sample simulated text data corresponding to the sample text data.

In this implementation, for _{each sequence in the sample text set {Y 1～T} }, take the (y ₁ ,y ₂ ,...,y _T ) sequence as an example, traverse each word y _{t in the} sequence for N times Monte Carlo simulation is different from the previous selection of the word with the highest probability as y _t , here each time the generator G _{θ is used} to sample according to the multinomial distribution of the output words, and repeat until the end of the sentence y _{T is} reached, thereby obtaining N different completes Sample simulation text collection {Y _1～T ¹ ,Y _1～T ² ,...,Y _1～T ^N }.

It should be noted that the simulation times of words located at different positions in the sentence in the sample text set can be the same or different.

S323. Recognize the multiple sample simulated text data according to the target text data in the sample text set by the discriminator, and update the parameter values of the generator according to the recognition result;

Further, step S323 may include:

Recognizing the multiple sample simulated text data by the discriminator and according to the target text data in the sample text set, and obtaining a state value function according to the recognition result;

An objective function is calculated according to the state value function, and parameter values of the generator are updated according to the objective function.

In the embodiment, the acquired sample simulation text set {Y _1～T ¹ ,Y _1～T ² ,...,Y _1～T ^N } is input into the discriminator D _φ for binary classification, and each sample simulation text Compare with the corresponding real text. If it is consistent, it means that the sample simulation text generated by the generator is real (mark 1); if it is inconsistent, it means that the sample simulation text generated by the generator is fake (mark 0). For complete sentences, the _{output result of the discriminator D φ is} directly used as the state value; for incomplete sentences, the discriminant results of N complete sentences obtained by Monte Carlo simulation are averaged. In summary, the state value function can be expressed as:

Among them, i represents the simulation times of Monte Carlo simulation.

According to the state value function, _{the parameter θ of the generator G θ} is updated. The objective function of the generator is to produce a more realistic sample to deceive the discriminator as much as possible, that is, to maximize the reward it obtains _{under the strategy G θ:}

Among them, G _θ (y _t |Y _1～t-1 ) represents the output of the strategy, which can be regarded as a probability in essence, outputting _{the probability value of y t} in the vocabulary; Y _1～t-1 are all occurrences of _{y t} Value. Parameter [theta] is a weight parameter generator G _θ is; the parameter generator G _θ in J (θ) update, in other words, from the policy gradient J (θ):

Among them, α _θ is the learning rate.

S324. Update the discriminator based on the updated generator and according to the loss function;

In this step, use the updated generator G _θ to generate a set of text sequences {Y _1～T }, and at the same time select the same number of text sequence sets {X from the real text set S={X _1～T} _1～T }, input into the discriminator D _φ to classify, and the loss function is a two-class logarithmic loss function:

The parameters of D _φ are updated on J(φ):

Among them, α _φ is the learning rate.

S325. Update the generator and the discriminator cyclically until the initial confrontation network model meets a preset convergence condition, and obtain the text generation confrontation network model composed of the updated generator.

In this step, in each round of training, the generator is repeatedly trained n _G times, and the discriminator is repeatedly trained n _D times until the model meets the preset convergence conditions. For example, the preset convergence condition is n _D > n _G to ensure that the discriminator can correctly guide the generator to update.

In step S3, the step of generating a confrontation network model obtained by pre-training the text and generating target text data according to the target guidance data includes:

The generator of the text generation confrontation network model is used to calculate the target guidance data to obtain the first sample word with the highest probability in the vocabulary, and the first sample word is added to the target guidance data. end;

Use the generator to calculate the first sample word, obtain the second sample word with the highest probability in the vocabulary, add the second sample word to the end of the first sample word, and execute the above in a loop Steps (and so on) until the target text data of the preset length is obtained. In this way, the target text data for questioning is generated according to the answer data, and the purpose of open question and answer based on the answer of the business object is realized, and it is convenient to test the temporary response ability of the business object to the open question.

In this embodiment, the text generation method is based on adversarial long-term and short-term memory networks and policy gradients, and uses the LSTM-based discriminator-generator structure to accurately generate text sequences and determine the authenticity of text; with the help of adversarial training, The discriminator can dynamically update its parameters, continuously improve the recognition ability, and provide appropriate guidance for the generator, which has more potential than evaluating the quality of the generated text purely based on other static benchmarks; with the help of the idea of reinforcement learning, the sequence generation process is transformed into a sequence The decision-making process solves the non-differentiable problem of the loss function caused by the discrete output, making it possible to generate the training of the confrontation network; use Monte Carlo search to simulate the strategy to obtain the complete sequence of each step and its scoring result in the discriminator , The average value is used as the reward value of the current time step, which solves the problem of the inability to directly obtain the reward of the unfinished sequence; in addition, only the generator part needs to be retained in the training phase, and other techniques such as Gumbel-softmax to deal with discretization are not differentiable In contrast, there is no need to train additional parameters, and the model occupies less memory.

Example two

As shown in Fig. 4, the present application also provides a text generation device 1, including: a collection unit 11, an acquisition unit 12, and a generation unit 13, wherein:

The collection unit 11 is used to collect the answer data generated by the business object in the question and answer scenario;

The business target can be consulting users for business consulting, buyers of online trading platforms, or interviewers in the interview process. The answer data can be collected by collecting equipment (for example: audio receiving device, microphone or mobile terminal with recording function, etc.).

The text generating device 1 in this embodiment is mainly used in conversation scenarios (at least two types of users), and generates a question text based on the answer information of the target object for the target object to answer the question text, for example: the text generating device 1 applies In the interview scenario, open-ended text questions are generated based on the keywords provided by the interviewer.

The obtaining unit 12 is configured to extract the answer data and obtain target guidance data;

The acquisition unit 12 can perform semantic analysis on the answer data to extract keywords in the answer data, and use the keywords as target guidance data; analyze the answer data to extract the nouns in the answer data, and use the noun as the target guidance data. data.

The generating unit 13 is configured to generate a confrontation network model through pre-trained text and generate target text data according to the target guidance data;

The target guide data is sentence beginning data of the target text data.

Specifically, the generating unit 13 uses the generator of the text generation confrontation network model to calculate according to the target guidance data, obtains the first sample word with the highest probability in the vocabulary, and adds the first sample word to all the words. The end of the target guide data;

The generator calculates according to the first sample word, obtains the second sample word with the highest probability in the vocabulary, adds the second sample word to the end of the first sample word, and so on until Obtain the target text data of a preset length.

In this embodiment, the text generation device 1 is based on the adversarial long-term and short-term memory network and policy gradients, and uses the LSTM-based discriminator-generator structure, which can accurately realize the task of generating text sequences and judging the authenticity of the text; with the help of adversarial training , The discriminator can dynamically update its parameters, continuously improve the recognition ability, and provide appropriate guidance for the generator, which has more potential than evaluating the quality of the generated text based purely on other static benchmarks; with the help of the idea of reinforcement learning, the sequence generation process is transformed into The sequence decision process solves the non-differentiable problem of the loss function caused by the discrete output, making it possible to generate the training of the adversarial network; using Monte Carlo search to simulate the strategy to obtain the complete sequence of each step and its score in the discriminator As a result, the average value is used as the reward value of the current time step, which solves the problem of not being able to directly obtain the reward of the unfinished sequence; in addition, only the generator part needs to be retained in the training phase, and other processing discretization such as Gumbel-softmax is not differentiable Compared with techniques, no additional parameters are required for training, and the model occupies less memory.

Example three

In order to achieve the above purpose, the present application also provides a computer device 2 which includes a plurality of computer devices 2. The components of the text generating device 1 of the second embodiment can be dispersed in different computer devices 2. The computer device 2 It can be a smartphone, tablet, laptop, desktop computer, rack server, blade server, tower server, or rack server (including independent servers, or server clusters composed of multiple servers) that executes the program. . The computer device 2 of this embodiment at least includes but is not limited to: a memory 21, a processor 23, a network interface 22, and a text generation device 1 (refer to FIG. 5) that can be communicatively connected to each other through a system bus. It should be pointed out that FIG. 5 only shows the computer device 2 with components, but it should be understood that it is not required to implement all the illustrated components, and more or fewer components may be implemented instead.

In this embodiment, the memory 21 includes at least one type of computer-readable storage medium, and the readable storage medium includes flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access Memory (RAM), Static Random Access Memory (SRAM), Read Only Memory (ROM), Electrically Erasable Programmable Read Only Memory (EEPROM), Programmable Read Only Memory (PROM), Magnetic Memory, Magnetic Disk, Optical Disk, etc. In some embodiments, the memory 21 may be an internal storage unit of the computer device 2, for example, the hard disk or memory of the computer device 2. In other embodiments, the memory 21 may also be an external storage device of the computer device 2, such as a plug-in hard disk, a smart media card (SMC), and a secure digital (Secure Digital, SMC) equipped on the computer device 2. SD) card, flash card (Flash Card), etc. Of course, the memory 21 may also include both the internal storage unit of the computer device 2 and its external storage device. In this embodiment, the memory 21 is generally used to store an operating system and various application software installed in the computer device 2, for example, the program code of the text generation method in the first embodiment. In addition, the memory 21 can also be used to temporarily store various types of data that have been output or will be output.

In some embodiments, the processor 23 may be a central processing unit (Central Processing Unit, CPU), a controller, a microcontroller, a microprocessor, or other data processing chips. The processor 23 is generally used to control the overall operation of the computer device 2, for example, to perform data interaction or communication-related control and processing with the computer device 2. In this embodiment, the processor 23 is used to run the program code or processing data stored in the memory 21, for example, to run the text generating device 1 and the like.

The network interface 22 may include a wireless network interface or a wired network interface, and the network interface 22 is generally used to establish a communication connection between the computer device 2 and other computer devices 2. For example, the network interface 22 is used to connect the computer device 2 with an external terminal through a network, and establish a data transmission channel and a communication connection between the computer device 2 and the external terminal. The network may be Intranet, Internet, Global System of Mobile Communication (GSM), Wideband Code Division Multiple Access (WCDMA), 4G network, 5G Network, Bluetooth (Bluetooth), Wi-Fi and other wireless or wired networks.

It should be pointed out that FIG. 5 only shows the computer device 2 with components 21-23, but it should be understood that it is not required to implement all the components shown, and more or fewer components may be implemented instead.

In this embodiment, the text generating device 1 stored in the memory 21 may also be divided into one or more program modules, and the one or more program modules are stored in the memory 21 and are composed of one or more program modules. It is executed by two processors (in this embodiment, the processor 23) to complete the application.

Embodiment four:

To achieve the above objective, the present application also provides a computer-readable storage medium, which includes multiple storage media, such as flash memory, hard disk, multimedia card, card-type memory (for example, SD or DX memory, etc.), random access memory (RAM ), static random access memory (SRAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), programmable read-only memory (PROM), magnetic memory, magnetic disks, optical disks, servers, App applications Shopping malls, etc., have computer-readable instructions stored thereon, and corresponding functions are realized when the programs are executed by the processor 23. The computer-readable storage medium of this embodiment is used to store the text generation device 1, and when executed by the processor 23, the text generation method of the first embodiment is implemented. The computer-readable storage medium may be non-volatile or volatile.

The serial numbers of the foregoing embodiments of the present application are for description only, and do not represent the superiority or inferiority of the embodiments.

Through the description of the above implementation manners, those skilled in the art can clearly understand that the above-mentioned embodiment method can be implemented by means of software plus the necessary general hardware platform, of course, it can also be implemented by hardware, but in many cases the former is better.的实施方式。

The above are only the preferred embodiments of the application, and do not limit the scope of the patent for this application. Any equivalent structure or equivalent process transformation made using the content of the description and drawings of the application, or directly or indirectly applied to other related technical fields , The same reason is included in the scope of patent protection of this application.

Claims

A text generation method, wherein, based on a question and answer scenario, the method includes the following steps:

Collect the answer data generated by the business object in the question and answer scenario;

Extract the answer data, and obtain target guidance data;

Generating a confrontation network model through a text obtained through pre-training, and generating target text data according to the target guidance data;

The target guide data is sentence beginning data of the target text data.
The text generation method according to claim 1, wherein before the step of generating a confrontation network model obtained by pre-training a text and generating target text data according to the target guidance data, the method comprises:

Acquiring a sample guidance set and a sample text set, the sample guidance set including at least one sample guidance data, the sample text set includes at least one sample text data, and the sample guidance data is sentence beginning data of the sample text data;

The initial confrontation network model is trained according to the sample guide set and the sample text set, and a text generation confrontation network model is obtained.
The text generation method according to claim 2, wherein the initial confrontation network model includes a generator and a discriminator, and the initial confrontation network model is trained according to the sample guide set and the sample text set to obtain The steps of text generation against the network model include:

Generating at least one sample text data by the generator and according to at least one sample guidance data in the sample guidance set;

Using Monte Carlo simulation to simulate the at least one sample text data and obtain a plurality of sample simulation text data;

Recognizing the plurality of sample simulated text data according to the target text data in the sample text set by the discriminator, and updating the parameter value of the generator according to the recognition result;

Update the discriminator based on the updated generator and according to the loss function;

The generator and the discriminator are cyclically updated until the initial confrontation network model meets a preset convergence condition, and the text generation confrontation network model composed of the updated generator is obtained.
The text generation method according to claim 3, wherein the step of generating at least one sample text data by the generator and according to at least one sample guidance data in the sample guidance set comprises:

Calculating by the generator and according to the sample guidance data, obtaining the first sample word with the highest probability in the vocabulary, and adding the first sample word to the end of the sample guidance data;

Calculating by the generator and according to the first sample word, obtaining the second sample word with the highest probability in the vocabulary, and adding the second sample word to the end of the first sample word;

Repeat the above steps until the sample text data of the preset length is obtained.
The text generation method according to claim 3, wherein the step of using Monte Carlo simulation to simulate the at least one sample text data and obtaining a plurality of sample simulation text data comprises:

Monte Carlo simulation is used to simulate the words in each sample text data one by one, and generate multiple sample simulated text data corresponding to the sample text data.
The text generation method according to claim 3, wherein the plurality of sample simulated text data is recognized by the discriminator and according to the target text data in the sample text set, and the generator is updated according to the recognition result The steps of parameter values include:

Recognizing the multiple sample simulated text data by the discriminator and according to the target text data in the sample text set, and obtaining a state value function according to the recognition result;

An objective function is calculated according to the state value function, and parameter values of the generator are updated according to the objective function.
The text generation method according to claim 3, wherein the step of generating a confrontation network model from the text obtained through pre-training and generating target text data according to the target guidance data comprises:

The generator of the text generation confrontation network model is used to calculate the target guidance data to obtain the first sample word with the highest probability in the vocabulary, and the first sample word is added to the target guidance data. end;

Using the generator to calculate the first sample word, obtain the second sample word with the highest probability in the vocabulary, and add the second sample word to the end of the first sample word;

Repeat the above steps until the target text data of the preset length is obtained.
A text generation device, wherein, based on a question and answer scenario, it includes:

The collection unit is used to collect the answer data generated by the business object in the question and answer scenario;

An obtaining unit, configured to extract the answer data and obtain target guidance data;

A generating unit, configured to generate a confrontation network model through pre-trained text and generate target text data according to the target guidance data;

The target guide data is sentence beginning data of the target text data.
A computer device comprising a memory, a processor, and computer readable instructions stored in the memory and running on the processor, wherein: the processor implements text generation when the computer readable instructions are executed The method includes the following steps:

Collect the answer data generated by the business object in the question and answer scenario;

Extract the answer data, and obtain target guidance data;

Generating a confrontation network model through a text obtained through pre-training, and generating target text data according to the target guidance data;

The target guide data is sentence beginning data of the target text data.
The computer device according to claim 9, wherein before the step of generating a confrontation network model obtained by pre-training the text and generating target text data according to the target guidance data, the method comprises:

Acquiring a sample guidance set and a sample text set, the sample guidance set including at least one sample guidance data, the sample text set includes at least one sample text data, and the sample guidance data is sentence beginning data of the sample text data;

The initial confrontation network model is trained according to the sample guide set and the sample text set, and a text generation confrontation network model is obtained.
The computer device according to claim 10, wherein the initial confrontation network model includes a generator and a discriminator, and the initial confrontation network model is trained according to the sample guide set and the sample text set, and the text is obtained The steps to generate a confrontation network model include:

Generating at least one sample text data by the generator and according to at least one sample guidance data in the sample guidance set;

Using Monte Carlo simulation to simulate the at least one sample text data and obtain a plurality of sample simulation text data;

Recognizing the plurality of sample simulated text data according to the target text data in the sample text set by the discriminator, and updating the parameter value of the generator according to the recognition result;

Update the discriminator based on the updated generator and according to the loss function;

The generator and the discriminator are cyclically updated until the initial confrontation network model meets a preset convergence condition, and the text generation confrontation network model composed of the updated generator is obtained.
11. The computer device according to claim 11, wherein the step of generating at least one sample text data by the generator and based on at least one sample guidance data in the sample guidance set comprises:

Calculating by the generator and according to the sample guidance data, obtaining the first sample word with the highest probability in the vocabulary, and adding the first sample word to the end of the sample guidance data;

Calculating by the generator and according to the first sample word, obtaining the second sample word with the highest probability in the vocabulary, and adding the second sample word to the end of the first sample word;

Repeat the above steps until the sample text data of the preset length is obtained.
The computer device according to claim 11, wherein the step of using Monte Carlo simulation to simulate the at least one sample text data and obtaining a plurality of sample simulation text data comprises:

Monte Carlo simulation is used to simulate the words in each sample text data one by one, and generate multiple sample simulated text data corresponding to the sample text data.
The computer device according to claim 11, wherein the plurality of sample simulated text data is recognized by the discriminator and according to the target text data in the sample text set, and the generator's data is updated according to the recognition result. The steps for parameter values include:

Recognizing the multiple sample simulated text data by the discriminator and according to the target text data in the sample text set, and obtaining a state value function according to the recognition result;

An objective function is calculated according to the state value function, and parameter values of the generator are updated according to the objective function.
A computer-readable storage medium having computer-readable instructions stored thereon, wherein: when the computer-readable instructions are executed by a processor, the method for generating text includes the following steps:

Collect the answer data generated by the business object in the question and answer scenario;

Extract the answer data, and obtain target guidance data;

Generating a confrontation network model through a text obtained through pre-training, and generating target text data according to the target guidance data;

The target guide data is sentence beginning data of the target text data.
15. The computer-readable storage medium according to claim 15, wherein before the step of generating a confrontation network model obtained by pre-training a text and generating target text data according to the target guidance data, the method comprises:

Acquiring a sample guidance set and a sample text set, the sample guidance set including at least one sample guidance data, the sample text set includes at least one sample text data, and the sample guidance data is sentence beginning data of the sample text data;

The initial confrontation network model is trained according to the sample guide set and the sample text set, and a text generation confrontation network model is obtained.
The computer-readable storage medium according to claim 16, wherein the initial confrontation network model includes a generator and a discriminator, and the initial confrontation network model is trained according to the sample guide set and the sample text set, And get the steps of text generation against the network model, including:

Generating at least one sample text data by the generator and according to at least one sample guidance data in the sample guidance set;

Using Monte Carlo simulation to simulate the at least one sample text data and obtain a plurality of sample simulation text data;

Recognizing the plurality of sample simulated text data according to the target text data in the sample text set by the discriminator, and updating the parameter value of the generator according to the recognition result;

Update the discriminator based on the updated generator and according to the loss function;

The generator and the discriminator are cyclically updated until the initial confrontation network model meets a preset convergence condition, and the text generation confrontation network model composed of the updated generator is obtained.
18. The computer-readable storage medium according to claim 17, wherein the step of generating at least one sample text data by the generator and based on at least one sample guidance data in the sample guidance set comprises:

Calculating by the generator and according to the sample guidance data, obtaining the first sample word with the highest probability in the vocabulary, and adding the first sample word to the end of the sample guidance data;

Calculating by the generator and according to the first sample word, obtaining the second sample word with the highest probability in the vocabulary, and adding the second sample word to the end of the first sample word;

Repeat the above steps until the sample text data of the preset length is obtained.
The computer-readable storage medium according to claim 17, wherein the step of using Monte Carlo simulation to simulate the at least one sample text data and obtaining a plurality of sample simulation text data comprises:

Monte Carlo simulation is used to simulate the words in each sample text data one by one, and generate multiple sample simulated text data corresponding to the sample text data.
18. The computer-readable storage medium according to claim 17, wherein the plurality of sample simulated text data is recognized by the discriminator and based on the target text data in the sample text set, and the plurality of sample simulated text data is updated according to the recognition result The steps to generate parameter values for the generator include:

Recognizing the multiple sample simulated text data by the discriminator and according to the target text data in the sample text set, and obtaining a state value function according to the recognition result;

An objective function is calculated according to the state value function, and parameter values of the generator are updated according to the objective function.