CN117744837A

CN117744837A - Model training and text detection method and device, storage medium and equipment

Info

Publication number: CN117744837A
Application number: CN202311873819.4A
Authority: CN
Inventors: 徐恪; 肖勇; 赵乙; 许卓尔; 孟昌华; 王维强
Original assignee: Alipay Hangzhou Information Technology Co Ltd
Current assignee: Alipay Hangzhou Information Technology Co Ltd
Priority date: 2023-12-29
Filing date: 2023-12-29
Publication date: 2024-03-22

Abstract

The specification discloses a model training, text detection method, device, storage medium and equipment, can be through arranging public text data sets on different internet platforms, simultaneously, through using a plurality of common dialogue large language models and rich prompt sets, construct the training data for training detection models, and because in the process of constructing the training data for training detection models, the training data is filtered for many times, the obtained training data is more similar to the scene of using the text generated by the large language models in actual scenes. In addition, in the process of training the detection model, the basis text output by the teacher large language model with a large parameter scale is used for learning the detection model, so that the accuracy of the trained detection model in recognizing the text generated by the large language model can be improved.

Description

Model training and text detection method and device, storage medium and equipment

Technical Field

The present disclosure relates to the field of internet application security technologies, and in particular, to a method, an apparatus, a storage medium, and a device for model training and text detection.

Background

The large language model is an artificial intelligent model based on deep learning technology, and after the large language model is trained by a large amount of training data, the large language model can be enabled to generate high-quality natural language text, for example: articles, stories, codes, poems, etc., and thus, the large language model is widely applied to the fields of question answering systems, machine translation, text understanding, etc.

However, since the similarity between the text generated by the large language model and the text written by the human being is high, it is generally difficult to recognize the text generated by the large language model, and the authenticity of the text generated by the large language model is difficult to be ensured, thereby bringing about such things as: the problems of hidden danger of privacy data of users, hidden danger of network safety, hidden danger of academic integrity, hidden danger reduction caused by user creation, and the like are solved.

Therefore, how to effectively identify the text generated by a large language model is a problem to be solved.

Disclosure of Invention

The specification provides a model training method, a text detection method, a device, a storage medium and equipment, which are used for partially solving the problem that texts generated by a large language model in the prior art cannot be effectively identified.

The technical scheme adopted in the specification is as follows:

the specification provides a model training method, comprising:

acquiring training data, the training data comprising: question data, label data, the question data comprising: the prompt is used for prompting the detection model to identify the sample text, and the tag data comprises: the method comprises the steps of actually detecting a result of a sample text and a predetermined basis text aiming at the actually detected result corresponding to the sample text;

inputting the problem data in the training data into a detection model to be trained, so that the detection model to be trained outputs a detection result aiming at the sample text according to the problem data to serve as a detection result to be verified, and obtains a basis text of the detection result to be verified output by the detection model to be trained to serve as a basis text to be verified;

and training the detection model to be trained by taking the deviation between the minimum to-be-verified detection result and the actual detection result of the sample text and the deviation between the minimum to-be-verified basis text and the preset basis text for obtaining the actual detection result corresponding to the sample text as optimization targets.

Optionally, before inputting the problem data in the training data into a detection model to be trained, the method further comprises:

inputting the prompt language into a preset embedded model, and generating a prompt vector aiming at the problem data through the embedded model;

inputting the problem data in the training data into a detection model to be trained, so that the detection model to be trained outputs a detection result aiming at the sample text according to the problem data to serve as a detection result to be verified, and obtains a basis text of the detection result to be verified, which is output by the detection model to be trained, and serves as the basis text to be verified, and specifically comprises the following steps:

inputting the problem data and the prompt vector into the detection model to be trained, so that the detection model to be trained outputs a detection result aiming at the sample text according to the problem data and the prompt vector to be used as a detection result to be verified, and obtains a basis text of the detection result to be verified, which is output by the detection model to be trained, to be used as the basis text to be verified.

Optionally, training the to-be-trained detection model with an optimization objective of minimizing a deviation between the to-be-verified detection result and the actual detection result of the sample text, and minimizing a deviation between the to-be-verified basis text and a predetermined basis text that obtains an actual detection result corresponding to the sample text, specifically including:

And training the detection model to be trained and the embedded model by taking the deviation between the minimum to-be-verified detection result and the actual detection result of the sample text and the deviation between the minimum to-be-verified basis text and the preset basis text for obtaining the actual detection result corresponding to the sample text as optimization targets.

Optionally, determining the basis text of the actual detection result corresponding to the sample text specifically includes:

inputting the sample text and the actual detection result of the sample text into a preset teacher model, so that the teacher model outputs at least one basis text for the actual detection result corresponding to the sample text according to the sample text and the actual detection result of the sample text.

Optionally, acquiring sample text specifically includes:

acquiring each first initial text, wherein each first initial text comprises: manually written short text and manually written long text;

determining a text title corresponding to each first initial text, and determining a prompt matched with the text title corresponding to the first initial text as a target prompt corresponding to the first initial text;

Inputting a text title corresponding to each first initial text and a target prompt corresponding to the first initial text into a preset large language model aiming at each first initial text, so that the large language model generates a second initial text corresponding to the first initial text according to the text title corresponding to the first initial text and the target prompt corresponding to the first initial text;

and obtaining a sample text according to each first initial text and each second initial text.

Optionally, determining a prompt matched with the text title corresponding to the first initial text as the target prompt corresponding to the first initial text specifically includes:

screening candidate prompt languages matched with the text titles corresponding to the first initial text from preset basic prompt languages;

according to a designated adjustment mode, the candidate prompt is adjusted to obtain a supplementary candidate prompt, and the designated adjustment mode comprises the following steps: at least one of synonym substitution and word order adjustment;

and determining a prompt matched with a text title corresponding to the first initial text as a target prompt corresponding to the first initial text according to the candidate prompt and the supplementary candidate prompt.

Optionally, obtaining a sample text according to each first initial text and each second initial text specifically includes:

inputting the second initial text and the first initial text corresponding to the second initial text into a preset semantic model aiming at each second initial text, so as to determine the semantic similarity between the second initial text and the first initial text corresponding to the second initial text through the semantic model, and taking the semantic similarity as the semantic similarity corresponding to the second initial text;

taking a second initial text with semantic similarity lower than a preset similarity threshold as a target second initial text;

and obtaining a sample text according to each first initial text and each target second initial text.

Optionally, the method further comprises:

obtaining each candidate detection model and a test set, wherein each candidate detection model is obtained by training the detection model to be trained according to the optimization target;

inputting each training data contained in the test set into each candidate detection model aiming at each candidate detection model to obtain a detection result output by the candidate detection model aiming at each training data contained in the test set;

Determining the accuracy corresponding to the candidate detection model according to the deviation between the detection result output by the candidate detection model aiming at each training data contained in the test set and the actual detection result corresponding to each training data contained in the test set;

and screening each candidate detection model according to the corresponding accuracy of each candidate detection model to obtain a target detection model.

The specification provides a text detection method, which specifically includes:

acquiring a text to be detected;

at least one round of detection is carried out on the text to be detected, and a detection result of the text to be detected is obtained; wherein,

for each round of detection, determining target data in the round of detection, inputting the target data into a pre-trained target detection model, so that the detection model outputs a detection result of the text to be detected in the round of detection according to the target data, and takes the detection result of the text to be detected in the round of detection, a detection result of each round of detection of the text to be detected before the round of detection and the text to be detected as target data of the next round of detection, wherein the target detection model is trained by the model training method;

Determining the step weight corresponding to the detection result in each round of detection according to the sequence of each round of detection;

and according to the step weight, fusing detection results in each round of detection to obtain a target detection result of the text to be detected, and executing tasks according to the target detection result.

The present specification provides a model training apparatus comprising:

the acquisition module is used for acquiring training data, wherein the training data comprises: question data, label data, the question data comprising: the prompt is used for prompting the detection model to identify the sample text, and the tag data comprises: the method comprises the steps of actually detecting a result of a sample text and a predetermined basis text aiming at the actually detected result corresponding to the sample text;

the detection module is used for inputting the problem data in the training data into a detection model to be trained, so that the detection model to be trained outputs a detection result aiming at the sample text according to the problem data to serve as a detection result to be verified, and obtains a basis text of the detection result to be verified output by the detection model to be trained to serve as a basis text to be verified;

The training module is used for training the detection model to be trained by taking the deviation between the minimum to-be-verified detection result and the actual detection result of the sample text and the deviation between the minimum to-be-verified basis text and the preset basis text for obtaining the actual detection result corresponding to the sample text as optimization targets.

The present specification provides a text detection apparatus including:

the text acquisition module is used for acquiring a text to be detected;

the detection module is used for carrying out at least one round of detection on the text to be detected to obtain a detection result of the text to be detected, wherein for each round of detection, target data in the round of detection are determined, the target data are input into a pre-trained target detection model, so that the detection model outputs the detection result of the text to be detected in the round of detection according to the target data, the detection result of the text to be detected in the round of detection and the text to be detected are used as target data of the next round of detection, and the target detection model is obtained through training of the model training method;

the determining module is used for determining the step weight corresponding to the detection result in each round of detection according to the sequence of each round of detection;

And the execution module is used for fusing the detection results in each round of detection according to the step weight to obtain the target detection result of the text to be detected, and executing tasks according to the target detection result.

The present specification provides a computer readable storage medium storing a computer program which when executed by a processor implements the model training, text detection method described above.

The present specification provides an electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the above model training, text detection methods when executing the program.

The above-mentioned at least one technical scheme that this specification adopted can reach following beneficial effect:

in the model training method provided in the present specification, training data is first obtained, where the training data includes: question data, label data, the question data includes: the prompt and the acquired sample text are used for prompting the detection model to identify the sample text, and the tag data comprise: inputting problem data in training data into a detection model to be trained according to the actual detection result of the sample text and a predetermined basis text of the actual detection result corresponding to the sample text, so that the detection model to be trained outputs the detection result for the sample text according to the problem data, and serves as a detection result to be verified, and the basis text of the detection result to be verified, which is output by the detection model to be trained, is obtained and serves as the basis text to be verified, so that deviation between the detection result to be verified and the actual detection result of the sample text is minimized, deviation between the basis text to be verified and the predetermined basis text of the actual detection result corresponding to the sample text is minimized, and training is carried out on the detection model to be trained.

According to the method, in the process of training the detection model, the basis text with the thinking guiding function is used for the detection model to learn, so that the recognition capability of the detection model for the text generated by the large language model can be improved, and the accuracy of recognizing the trained detection model for the text generated by the large language model can be improved.

Drawings

The accompanying drawings, which are included to provide a further understanding of the specification, illustrate and explain the exemplary embodiments of the present specification and their description, are not intended to limit the specification unduly. Attached at

In the figure:

FIG. 1 is a schematic flow chart of a model training method provided in the present specification;

fig. 2 is a schematic flow chart of a text detection method provided in the present specification;

FIG. 3 is a schematic diagram of a model training apparatus provided in the present specification;

fig. 4 is a schematic diagram of a text detection device provided in the present specification;

fig. 5 is a schematic diagram of an electronic device corresponding to fig. 1 provided in the present specification.

Detailed Description

For the purposes of making the objects, technical solutions and advantages of the present specification more apparent, the technical solutions of the present specification will be clearly and completely described below with reference to specific embodiments of the present specification and corresponding drawings. It will be apparent that the described embodiments are only some, but not all, of the embodiments of the present specification. All other embodiments, which can be made by one of ordinary skill in the art without undue burden from the present disclosure, are intended to be within the scope of the present disclosure.

The following describes in detail the technical solutions provided by the embodiments of the present specification with reference to the accompanying drawings.

Fig. 1 is a schematic flow chart of a model training method provided in the present specification, including the following steps:

s100: acquiring training data, the training data comprising: question data, label data, the question data comprising: the prompt is used for prompting the detection model to identify the sample text, and the tag data comprises: the method comprises the steps of actually detecting a sample text, and presetting a basis text of the actually detected result corresponding to the sample text.

In the specification, the service platform can obtain a detection result of whether the text to be recognized is the text generated by the large language model through the detection model by inputting the text to be recognized into the detection model. Before this, the service platform needs to train the detection model now, and the detection model can be used in the task of detecting the text to be identified, and the training method for the detection model is described in detail below.

The service platform needs to acquire training data for training the detection model before training the detection model, where the training data may include: question data, label data, where the question data includes: the prompt and the acquired sample text, wherein the tag data comprises: the above prompt is used for prompting the detection model to identify the sample text, for example: "i want you to take the role of news edit, i will provide you with a title, please first understand the title content, and then generate a piece of news", the above-mentioned evidence text of the actual detection result corresponding to the sample text can be understood as an explanation text of why the detection result of the sample text corresponds to the actual detection result of the sample text, for example: from the analysis of language style, this news appears to have no obvious subjective emotion or personal perspective, and grammar and vocabulary usage is also more natural, which may indicate that this news is machine-generated.

Specifically, the service platform may obtain the short text written manually and the long text written manually as the first initial text, determine, for each first initial text, a text title corresponding to the first initial text, and determine, as the target prompt corresponding to the first initial text, a prompt matched with the text title corresponding to the first initial text.

The method for obtaining the manually written short text by the service platform may be that the manually written short text is obtained from various short message sharing platforms, and the method for obtaining the manually written long text by the service platform may be that the manually written long text is obtained from various long text publishing platforms.

It should be noted that, since the short text generally includes a text title and a text body, the method for determining the text title of the short text by the service platform may be a rule-based character string matching method, and the text title is extracted from the short text, whereas the method for determining the text title of the long text by the service platform may be that the long text and a prompt extracted by the preset text title are input into a preset large dialogue language model, so as to determine the text title of the long text by the large dialogue language model, and the prompt extracted by the preset text title may be set according to actual requirements, for example: please first summarize the content in { news body }, generate summary text around 20 words, require that the answer only contain rewritten content, and do not need any supplementary explanation.

Further, the service platform may input, for each first initial text, a text title corresponding to the first initial text and a target prompt corresponding to the first initial text into a preset large language model, so that the large language model generates, according to the text title corresponding to the first initial text and the target prompt corresponding to the first initial text, a second initial text corresponding to the first initial text, where the large language model may be as follows: chat generation pre-training transformation model (Chat Generative Pre-Trained Transformer, chatGPT), chat generation language model (Chat Generative Language Model, chatGLM), and the like.

The method for obtaining the target prompt may be that each candidate prompt matched with the text title corresponding to the first initial text is selected from preset basic prompts, and the candidate prompt is used as the target prompt corresponding to the first initial text.

Further, in order to improve the quality of the second initial text output by the large language model, the service platform may further optimize each candidate prompt according to each prompt optimization criterion condition included in the preset prompt optimization criterion condition set, so as to obtain a target prompt corresponding to the first initial text, where the prompt optimization criterion condition may be: including clearly expressing task goals, splitting tasks into subtasks, etc.

The target prompt optimized by the method can be as follows: i want you to take the role of news editing, I will provide you with a title, please understand the title content first, then generate a piece of news, ask to contain six elements of news, time, place, person, event cause, event pass, event result, the number of words is no more than 150 words, the title is as follows { news title }, the output contains only the generated news content, does not need to distinguish the six elements, and does not need to add other explanatory words.

In an actual application scenario, since the quality difference of the second initial texts output by the large language model for different target prompt languages is large, in order to promote the richness of each second initial text generated by the large language model, the service platform may further adjust the candidate prompt language according to a specified adjustment mode to obtain a supplementary candidate prompt language, and determine, according to the candidate prompt language and the supplementary candidate prompt language, a prompt language matched with a text title corresponding to the first initial text as the target prompt language corresponding to the first initial text, where the specified adjustment mode includes: at least one of synonym replacement, word order adjustment, language style adjustment.

Further, since the quality difference of each second initial text output according to each text title and each target prompt through the large language model is large, in order to improve the quality of the sample text used for training the detection model, after obtaining each second initial data, the service platform may further determine, for each second initial data, whether the second initial data meets any one of the deletion criterion conditions included in the preset deletion criterion condition set, if so, delete the second initial text, and further obtain the sample text used for training the detection model according to each first initial text and each remaining second initial text.

The above-mentioned conditions of each deletion criterion may be such as: the answer content triggers a dialogue large language model fencing mechanism, the answer contains more than 20 repeated character strings, and the like.

Further, since the second initial text generated by the large language model may include a small text with obvious language characteristics generated by the machine, so that the second initial text can be easily identified by a reader, the service platform may further determine, according to a preset filter criterion set, whether at least part of the second initial text meets any one of the filter criterion conditions included in the filter criterion set, if so, perform text editing on the at least part of the text to obtain an edited second initial text, and may obtain a sample text for training the detection model according to each first initial text and each edited second initial text.

The above filtering criteria conditions may be as follows: format features (blank lines, fragments, etc.), description prefixes ("as a large language model", "good, i will accomplish the following tasks", etc.), alert related words ("news content", "place of occurrence", etc.).

In addition, in order to ensure that the second initial text generated through the large language model has semantic deviation with the second initial text, the service platform can input the second initial text and the first initial text corresponding to the second initial text into a preset semantic model for each second initial text, so as to determine the semantic similarity between the second initial text and the first initial text corresponding to the second initial text through the semantic model, and obtain a sample text according to each first initial text and each target second initial text, wherein the second initial text with the semantic similarity lower than a preset similarity threshold is used as the semantic similarity corresponding to the second initial text.

Further, after each sample text is obtained, the service platform can input the sample text and the actual detection result of the sample text into a preset teacher model for each sample text, so that the teacher model outputs at least one basis text for the actual detection result corresponding to the sample text according to the sample text and the actual detection result of the sample text.

The teacher model may be a large language model with a large number of parameters, for example: chat generates a Pre-training transformation model (Chat Generative Pre-Trained Transformer, chatGPT), generates a Pre-training transformation model 4 (generating Pre-Trained Transformer, gpt-4), and so on.

It should be noted that, since the teacher model takes the sample text and the actual detection result of the sample text as input, the number of the basis texts of the actual detection result corresponding to the obtained sample text may be multiple, so that when the detection model is trained based on multiple basis texts of the actual detection result corresponding to the sample text, the detection model can learn the basis texts of the actual detection result corresponding to multiple different sample texts, so that the accuracy of identifying the detection model for the text generated by the large language model can be improved.

In the present specification, the execution body for implementing the model training method may refer to a designated device such as a server provided on a service platform, or may refer to a terminal device such as a desktop computer or a notebook computer, and for convenience of description, the model training method provided in the present specification will be described below by taking the server as an example of the execution body.

S102: inputting the problem data in the training data into a detection model to be trained, so that the detection model to be trained outputs a detection result aiming at the sample text according to the problem data to serve as a detection result to be verified, and obtains a basis text of the detection result to be verified, which is output by the detection model to be trained, to serve as the basis text to be verified.

S104: and training the detection model to be trained by taking the deviation between the minimum to-be-verified detection result and the actual detection result of the sample text and the deviation between the minimum to-be-verified basis text and the preset basis text for obtaining the actual detection result corresponding to the sample text as optimization targets.

In this specification, the server may divide at least part of the training data in the training data into training sets, divide at least part of the training data from other training data except the training sets, use the training sets as test sets, and use other training data except the training sets and the test sets as verification sets, where a ratio of the number of each training data included in the test sets to the number of each training data included in the training sets may be set according to actual requirements, and preferably, a ratio of the number of each training data included in the test sets to the number of each training data included in the training sets may be 1:9.

Further, the server may input, for each training data included in the training set, problem data in the training data into the detection model to be trained, so that the detection model to be trained outputs a detection result for the sample text according to the problem data, as a detection result to be verified, and outputs a basis text for obtaining the detection result to be verified, as a basis text to be verified, so as to minimize a deviation between the detection result to be verified and an actual detection result of the sample text, and minimize a deviation between the basis text to be verified and a predetermined basis text for obtaining the actual detection result corresponding to the sample text, as an optimization target, and train the detection model to be trained.

Further, the server may include each training data in the training set, and use the test model trained by the training data as a candidate test model, further, the server may input each training data included in the test set into the candidate test model for each candidate test model to obtain a test result output by the candidate test model for each training data included in the test set, determine an accuracy corresponding to the candidate test model according to a deviation between the test result output by the candidate test model for each training data included in the test set and an actual test result corresponding to each training data included in the test set, and screen each candidate test model according to the accuracy corresponding to each candidate test model to obtain a target test model.

Of course, the server may divide the training set into a plurality of sub-training sets including a specified number of training data, and use the detection model trained by the sub-training set as a candidate detection model, further, the server may input, for each candidate detection model, each training data included in the test set into the candidate detection model to obtain a detection result output by the candidate detection model for each training data included in the test set, determine an accuracy corresponding to the candidate detection model according to a deviation between the detection result output by the candidate detection model for each training data included in the test set and an actual detection result corresponding to each training data included in the test set, and screen each candidate detection model according to the accuracy corresponding to each candidate detection model to obtain a target detection model, where the specified number may be set according to actual requirements, for example: every fifth training data may be set as a sub-training set.

It should be noted that, since the detection result output by the detection model includes the detection result for the sample text included in the training data and the corresponding basis text, the form of the detection result output by the detection model is various, so the server may determine the detection result of the detection model for the training data by using a method of matching the character strings, for example: if the detection result output by the detection model contains 'human writing', the detection model is considered to judge the sample text as the text written by human, and if the detection result output by the detection model contains 'machine generation', the detection model is considered to judge the sample text as the text generated by machine.

In addition, the need for fine tuning for a large number of specific parameters contained in the test model during the training process described above is reduced, while maintaining the generalization and language understanding capabilities of the test model described above. The server can also input a prompt into a preset embedded model, generate a prompt vector aiming at the problem data through the embedded model, and further input the problem data and the prompt vector into a detection model to be trained, so that the detection model to be trained outputs a detection result aiming at a sample text according to the problem data and the prompt vector to serve as a detection result to be verified, and the detection model to be trained is obtained to output a basis text of the detection result to be verified to serve as the basis text to be verified.

Further, the server may train the detection model to be trained and the embedded model with an optimization objective of minimizing a deviation between the detection result to be verified and the actual detection result of the sample text, and minimizing a deviation between the basis text to be verified and a predetermined basis text that obtains the actual detection result corresponding to the sample text.

It should be noted that the above-mentioned embedded model may be an additional embedded layer independent of the detection model, and the parameter amount of this embedded model is much smaller than the parameter amount of the detection model, after the server inputs the problem data and the prompt vector into the detection model to be trained, the detection model may splice the prompt vector with the feature vector extracted for the input problem data by itself, that is, use the prompt vector as the prefix of the feature vector extracted for the input problem data by itself, so that when training the detection model according to the deviation between the detection result to be verified and the actual detection result of the sample text, and the deviation between the text to be verified and the predetermined text to obtain the actual detection result corresponding to the sample text, only fine tuning may be performed for some parameters included in the detection model and parameters of the embedded model, without adjusting all parameters included in the detection model.

From the above, it can be seen that the server can sort the public text data sets on different internet platforms, and simultaneously construct the training data for training the detection model by using multiple common dialogs large language models and rich prompt sets, and because multiple times of filtering are performed on the second initial text in the process of constructing the training data for training the detection model, the obtained second initial text is closer to a scene of malicious use of the text generated by the large language model in the actual scene.

In addition, in the process of training the detection model, the basis text output by the teacher large language model with a large parameter scale is used for learning the detection model, so that the accuracy of the trained detection model in recognizing the text generated by the large language model can be improved.

For further explanation of the present specification, a method for performing text detection by using the detection model trained by the above method will be described in detail as shown in fig. 2.

Fig. 2 is a flow chart of a method for testing a database provided in the present specification, specifically including the following steps:

S200: and acquiring a text to be detected.

S202: and carrying out at least one round of detection on the text to be detected to obtain a detection result of the text to be detected.

S204: for each round of detection, determining target data in the round of detection, inputting the target data into a pre-trained target detection model, so that the detection model outputs a detection result of the text to be detected in the round of detection according to the target data, and takes the detection result of the text to be detected in the round of detection, a detection result of each round of detection of the text to be detected before the round of detection and the text to be detected as target data of the next round of detection, wherein the target detection model is trained by the model training method.

S206: and determining the step weight corresponding to the detection result in each round of detection according to the sequence of each round of detection.

S208: and according to the step weight, fusing detection results in each round of detection to obtain a target detection result of the text to be detected, and executing tasks according to the target detection result.

In the specification, the server can acquire the text to be detected, input the acquired text to be detected into a pre-trained detection model, further obtain a detection result of whether the text to be detected is a text generated by a large model through the detection model, and further identify the text to be detected according to the detection result of the text to be detected.

The text to be detected may be text input by the user, text obtained in batch from a specified database, or the like.

In an actual application scenario, in order to detect a text to be detected from a plurality of different angles to obtain a more accurate detection result, the server may further perform at least one round of detection on the text to be detected to obtain a detection result of the text to be detected, wherein for each round of detection, target data in the round of detection is determined, the target data is input into a target detection model trained in advance, so that the detection model outputs a detection result of the text to be detected in the round of detection according to the target data, and the detection result of the text to be detected in the round of detection, the detection result of each round of detection of the text to be detected before the round of detection, and the text to be detected serve as target data of the next round of detection until a preset termination condition is met, where the termination condition may be set according to actual requirements, for example: and when the detected number of rounds meets the preset maximum number of rounds, the termination condition is considered to be met, wherein the target detection model is obtained by training through the model training method.

As can be seen from the foregoing, in the above multi-round detection process for the text to be detected, the later detection may refer to the detection results of the previous multi-round detection, so the more accurate the detection result of the later detection is, and therefore, the server may determine the step weight corresponding to the detection result in each round of detection according to the sequence of each round of detection, and specifically may refer to the following formula:

w _i ＝α×(n-i)

in the above formula, w _i And (3) for the step weight corresponding to the ith detection result, alpha is a preset basic weight, and n is the maximum detection round number.

Further, the server may normalize the step weights corresponding to the detection results of each detection round, so as to obtain normalized step weights corresponding to the detection results of each detection round, and specifically may refer to the following formula:

and then the detection results in each round of detection can be fused according to the normalized step weight corresponding to the detection result of each round of detection to obtain the target detection result of the text to be detected, and task execution can be performed according to the target detection result, specifically, the following formula can be referred to:

in the formula, R is the detection result after fusion, b _i Is the detection result of the ith round (wherein, if the detection result of the ith round is written manually, b is _i The value of (b) may be 1, where b is the case if the detection result of the ith round is machine-written _i May have a value of 0).

Further, the server may obtain a target detection result according to the value of the fused detection result R, where if the value of the fused detection result R is greater than a preset detection threshold, it may determine that the target detection result of the text to be detected is a text written by man, and if not, it may determine that the target detection result of the text to be detected is a text written by a machine.

From the above, it can be seen that the step voting rule is applied in the detection process, and the accuracy of the detection result is ensured through multiple rounds of detection.

The above model training and text detection methods provided for one or more embodiments of the present disclosure further provide corresponding model training devices and text detection devices based on the same ideas, as shown in fig. 3 and fig. 4.

Fig. 3 is a schematic diagram of a model training device provided in the present specification, including:

an obtaining module 301, configured to obtain training data, where the training data includes: question data, label data, the question data comprising: the prompt is used for prompting the detection model to identify the sample text, and the tag data comprises: the method comprises the steps of actually detecting a result of a sample text and a predetermined basis text aiming at the actually detected result corresponding to the sample text;

The detection module 302 is configured to input the problem data in the training data into a detection model to be trained, so that the detection model to be trained outputs a detection result for the sample text according to the problem data, and the detection result is used as a detection result to be verified, and a basis text of the detection result to be verified output by the detection model to be trained is obtained and is used as a basis text to be verified;

the training module 303 is configured to train the to-be-trained detection model with a deviation between the to-be-verified detection result and the actual detection result of the sample text minimized, and a deviation between the to-be-verified basis text and a predetermined basis text that obtains the actual detection result corresponding to the sample text minimized as an optimization target.

Optionally, the detection module 302 is specifically configured to input the prompt into a preset embedding model, and generate, through the embedding model, a prompt vector for the problem data; inputting the problem data and the prompt vector into the detection model to be trained, so that the detection model to be trained outputs a detection result aiming at the sample text according to the problem data and the prompt vector to be used as a detection result to be verified, and obtains a basis text of the detection result to be verified, which is output by the detection model to be trained, to be used as the basis text to be verified.

Optionally, the training module 303 is specifically configured to train the to-be-trained detection model and the embedded model with a deviation between the to-be-verified detection result and the actual detection result of the sample text being minimized, and a deviation between the to-be-verified basis text and a predetermined basis text that obtains the actual detection result corresponding to the sample text being minimized as an optimization target.

Optionally, the obtaining module 301 is specifically configured to input the sample text and an actual detection result of the sample text into a preset teacher model, so that the teacher model outputs at least one reference text for the sample text corresponding to the actual detection result according to the sample text and the actual detection result of the sample text.

Optionally, the obtaining module 301 is specifically configured to obtain each first initial text, where each first initial text includes: manually written short text and manually written long text; determining a text title corresponding to each first initial text, and determining a prompt matched with the text title corresponding to the first initial text as a target prompt corresponding to the first initial text; inputting a text title corresponding to each first initial text and a target prompt corresponding to the first initial text into a preset large language model aiming at each first initial text, so that the large language model generates a second initial text corresponding to the first initial text according to the text title corresponding to the first initial text and the target prompt corresponding to the first initial text; and obtaining a sample text according to each first initial text and each second initial text.

Optionally, the obtaining module 301 is specifically configured to screen candidate prompt terms that match a text title corresponding to the first initial text from preset basic prompt terms; according to a designated adjustment mode, the candidate prompt is adjusted to obtain a supplementary candidate prompt, and the designated adjustment mode comprises the following steps: at least one of synonym substitution and word order adjustment; and determining a prompt matched with a text title corresponding to the first initial text as a target prompt corresponding to the first initial text according to the candidate prompt and the supplementary candidate prompt.

Optionally, the obtaining module 301 is specifically configured to, for each second initial text, input the second initial text and a first initial text corresponding to the second initial text into a preset semantic model, so as to determine, by using the semantic model, a semantic similarity between the second initial text and the first initial text corresponding to the second initial text, as the semantic similarity corresponding to the second initial text; taking a second initial text with semantic similarity lower than a preset similarity threshold as a target second initial text; and obtaining a sample text according to each first initial text and each target second initial text.

Optionally, the training module 303 is specifically configured to obtain each candidate detection model and a test set, where each candidate detection model is obtained by training the detection model to be trained according to the optimization objective; inputting each training data contained in the test set into each candidate detection model aiming at each candidate detection model to obtain a detection result output by the candidate detection model aiming at each training data contained in the test set; determining the accuracy corresponding to the candidate detection model according to the deviation between the detection result output by the candidate detection model aiming at each training data contained in the test set and the actual detection result corresponding to each training data contained in the test set; and screening each candidate detection model according to the corresponding accuracy of each candidate detection model to obtain a target detection model.

Fig. 4 is a schematic diagram of a text detection device provided in the present specification, including:

a text acquisition module 401, configured to acquire a text to be detected;

the detection module 402 is configured to perform at least one round of detection on the text to be detected to obtain a detection result of the text to be detected, wherein for each round of detection, target data in the round of detection is determined, the target data is input into a pre-trained target detection model, so that the detection model outputs the detection result of the text to be detected in the round of detection according to the target data, and the detection result of the text to be detected in the round of detection and the text to be detected are used as target data of the next round of detection, and the target detection model is obtained through training by the model training method;

A determining module 403, configured to determine a step weight corresponding to a detection result in each round of detection according to a sequence of each round of detection;

and the execution module 404 is configured to fuse the detection results in each round of detection according to the step weight, obtain a target detection result of the text to be detected, and execute a task according to the target detection result.

The present specification also provides a computer readable storage medium storing a computer program operable to perform a model training, text detection method as provided in fig. 1 above.

The present specification also provides a schematic structural diagram of an electronic device corresponding to fig. 1 shown in fig. 5. At the hardware level, as in fig. 5, the electronic device includes a processor, an internal bus, a network interface, a memory, and a non-volatile storage, although it may include hardware required for other services. The processor reads the corresponding computer program from the nonvolatile memory into the memory and then runs the computer program to realize the model training and text detection method of the figure 1. Of course, other implementations, such as logic devices or combinations of hardware and software, are not excluded from the present description, that is, the execution subject of the following processing flows is not limited to each logic unit, but may be hardware or logic devices.

In the 90 s of the 20 th century, improvements to one technology could clearly be distinguished as improvements in hardware (e.g., improvements to circuit structures such as diodes, transistors, switches, etc.) or software (improvements to the process flow). However, with the development of technology, many improvements of the current method flows can be regarded as direct improvements of hardware circuit structures. Designers almost always obtain corresponding hardware circuit structures by programming improved method flows into hardware circuits. Therefore, an improvement of a method flow cannot be said to be realized by a hardware entity module. For example, a programmable logic device (Programmable Logic Device, PLD) (e.g., field programmable gate array (Field Programmable Gate Array, FPGA)) is an integrated circuit whose logic function is determined by the programming of the device by a user. A designer programs to "integrate" a digital system onto a PLD without requiring the chip manufacturer to design and fabricate application-specific integrated circuit chips. Moreover, nowadays, instead of manually manufacturing integrated circuit chips, such programming is mostly implemented by using "logic compiler" software, which is similar to the software compiler used in program development and writing, and the original code before the compiling is also written in a specific programming language, which is called hardware description language (Hardware Description Language, HDL), but not just one of the hdds, but a plurality of kinds, such as ABEL (Advanced Boolean Expression Language), AHDL (Altera Hardware Description Language), confluence, CUPL (Cornell University Programming Language), HDCal, JHDL (Java Hardware Description Language), lava, lola, myHDL, PALASM, RHDL (Ruby Hardware Description Language), etc., VHDL (Very-High-Speed Integrated Circuit Hardware Description Language) and Verilog are currently most commonly used. It will also be apparent to those skilled in the art that a hardware circuit implementing the logic method flow can be readily obtained by merely slightly programming the method flow into an integrated circuit using several of the hardware description languages described above.

The controller may be implemented in any suitable manner, for example, the controller may take the form of, for example, a microprocessor or processor and a computer readable medium storing computer readable program code (e.g., software or firmware) executable by the (micro) processor, logic gates, switches, application specific integrated circuits (Application Specific Integrated Circuit, ASIC), programmable logic controllers, and embedded microcontrollers, examples of which include, but are not limited to, the following microcontrollers: ARC 625D, atmel AT91SAM, microchip PIC18F26K20, and Silicone Labs C8051F320, the memory controller may also be implemented as part of the control logic of the memory. Those skilled in the art will also appreciate that, in addition to implementing the controller in a pure computer readable program code, it is well possible to implement the same functionality by logically programming the method steps such that the controller is in the form of logic gates, switches, application specific integrated circuits, programmable logic controllers, embedded microcontrollers, etc. Such a controller may thus be regarded as a kind of hardware component, and means for performing various functions included therein may also be regarded as structures within the hardware component. Or even means for achieving the various functions may be regarded as either software modules implementing the methods or structures within hardware components.

The system, apparatus, module or unit set forth in the above embodiments may be implemented in particular by a computer chip or entity, or by a product having a certain function. One typical implementation is a computer. In particular, the computer may be, for example, a personal computer, a laptop computer, a cellular telephone, a camera phone, a smart phone, a personal digital assistant, a media player, a navigation device, an email device, a game console, a tablet computer, a wearable device, or a combination of any of these devices.

For convenience of description, the above devices are described as being functionally divided into various units, respectively. Of course, the functions of each element may be implemented in one or more software and/or hardware elements when implemented in the present specification.

It will be appreciated by those skilled in the art that embodiments of the present description may be provided as a method, system, or computer program product. Accordingly, the present specification may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present description can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.

The present description is described with reference to flowchart illustrations and/or block diagrams of methods, apparatus (systems) and computer program products according to embodiments of the specification. It will be understood that each flow and/or block of the flowchart illustrations and/or block diagrams, and combinations of flows and/or blocks in the flowchart illustrations and/or block diagrams, can be implemented by computer program instructions. These computer program instructions may be provided to a processor of a general purpose computer, special purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be stored in a computer-readable memory that can direct a computer or other programmable data processing apparatus to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instruction means which implement the function specified in the flowchart flow or flows and/or block diagram block or blocks.

These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer implemented process such that the instructions which execute on the computer or other programmable apparatus provide steps for implementing the functions specified in the flowchart flow or flows and/or block diagram block or blocks.

In one typical configuration, a computing device includes one or more processors (CPUs), input/output interfaces, network interfaces, and memory.

The memory may include volatile memory in a computer-readable medium, random Access Memory (RAM) and/or nonvolatile memory, such as Read Only Memory (ROM) or flash memory (flash RAM). Memory is an example of computer-readable media.

Computer readable media, including both non-transitory and non-transitory, removable and non-removable media, may implement information storage by any method or technology. The information may be computer readable instructions, data structures, modules of a program, or other data. Examples of storage media for a computer include, but are not limited to, phase change memory (PRAM), static Random Access Memory (SRAM), dynamic Random Access Memory (DRAM), other types of Random Access Memory (RAM), read Only Memory (ROM), electrically Erasable Programmable Read Only Memory (EEPROM), flash memory or other memory technology, compact disc read only memory (CD-ROM), digital Versatile Discs (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices, or any other non-transmission medium, which can be used to store information that can be accessed by a computing device. Computer-readable media, as defined herein, does not include transitory computer-readable media (transmission media), such as modulated data signals and carrier waves.

It should also be noted that the terms "comprises," "comprising," or any other variation thereof, are intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements does not include only those elements but may include other elements not expressly listed or inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one … …" does not exclude the presence of other like elements in a process, method, article or apparatus that comprises an element.

It will be appreciated by those skilled in the art that embodiments of the present description may be provided as a method, system, or computer program product. Accordingly, the present specification may take the form of an entirely hardware embodiment, an entirely software embodiment or an embodiment combining software and hardware aspects. Furthermore, the present description can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) having computer-usable program code embodied therein.

The description may be described in the general context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The specification may also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media including memory storage devices.

In this specification, each embodiment is described in a progressive manner, and identical and similar parts of each embodiment are all referred to each other, and each embodiment mainly describes differences from other embodiments. In particular, for system embodiments, since they are substantially similar to method embodiments, the description is relatively simple, as relevant to see a section of the description of method embodiments.

The foregoing is merely an example of the present specification and is not intended to limit the present specification. Various modifications and alterations to this specification will become apparent to those skilled in the art. Any modifications, equivalent substitutions, improvements, or the like, which are within the spirit and principles of the present description, are intended to be included within the scope of the claims of the present description.

Claims

1. A model training method, comprising:

2. The method of claim 1, prior to inputting the problem data in the training data into a detection model to be trained, the method further comprising:

3. The method according to claim 2, wherein training the test model to be trained with respect to minimizing a deviation between the test result to be verified and the actual test result of the sample text, and minimizing a deviation between the text-by-text to be verified and a predetermined text-by-text that yields the actual test result corresponding to the sample text, specifically includes:

4. The method of claim 1, wherein determining the basis text for the actual detection result corresponding to the sample text specifically comprises:

5. The method of claim 1, obtaining sample text, comprising:

6. The method according to claim 5, wherein determining a prompt matched with a text title corresponding to the first initial text as the target prompt corresponding to the first initial text comprises:

7. The method of claim 5, obtaining sample text according to each first initial text and each second initial text, specifically comprising:

8. The method of claim 1, the method further comprising:

9. A text detection method, comprising:

acquiring a text to be detected;

for each round of detection, determining target data in the round of detection, inputting the target data into a pre-trained target detection model, so that the detection model outputs detection results of the text to be detected in the round of detection according to the target data, and takes the detection results of the text to be detected in the round of detection, the detection results of each round of detection of the text to be detected before the round of detection and the text to be detected as target data of the next round of detection, wherein the target detection model is obtained by training according to the method of any one of claims 1-8;

10. A model training apparatus comprising:

11. A text detection device, comprising:

the text acquisition module is used for acquiring a text to be detected;

the detection module is used for carrying out at least one round of detection on the text to be detected to obtain a detection result of the text to be detected, wherein for each round of detection, target data in the round of detection are determined, the target data are input into a pre-trained target detection model, so that the detection model outputs the detection result of the text to be detected in the round of detection according to the target data, the detection result of the text to be detected in the round of detection and the text to be detected are used as target data of the next round of detection, and the target detection model is trained by the method according to any one of claims 1-8;

12. A computer readable storage medium storing a computer program which, when executed by a processor, implements the method of any one of the preceding claims 1 to 9.

13. An electronic device comprising a memory, a processor and a computer program stored on the memory and executable on the processor, the processor implementing the method of any of the preceding claims 1-9 when the program is executed.