CN115129866A - Training text generation method, model training device and electronic equipment - Google Patents

Abstract

The present application discloses a training text generation method, a model training method, a text recognition method, a device electronic device, and a computer-readable storage medium, wherein the training text is used for training a to-be-trained model to obtain a text recognition model, and the training text generates The method includes: acquiring guide text, the guide text is consistent with the semantic attributes of the target text, and the target text is the positive example text recognized by the text recognition model; inputting the guide text into a guide-based text generation model , obtain output text consistent with the semantic attributes of the guide text; determine training text according to the output text. In the present application, the output text is automatically generated by the guided text generation model, so that the training text is determined, and the training text can be obtained more quickly and efficiently.

Description

Training text generation method, model training method, device and electronic device

technical field

The present application relates to the field of computer technology, and in particular, to a training text generation method, a model training method, an apparatus, and an electronic device.

Background technique

The Internet is an important tool for people to live and work. With the increasing openness of the Internet, the Internet is filled with a large number of sensitive texts that are not suitable for display to users, such as users' online messages, users' conversation information on chat software, The information replied by the dialogue robot, etc. In order to create a green chat environment, these sensitive texts need to be identified and filtered out in advance.

In the related art, a text recognition model can be used to identify sensitive text. However, the text recognition model needs to be pre-trained with a large number of sensitive texts with different expressions. Since it is difficult to collect a large number of sensitive texts on the Internet, the efficiency of manually writing sensitive texts is very low, and the number of manual writing is limited. Therefore, how to quickly , Efficient acquisition of training text to train text recognition models is a problem that needs to be solved.

SUMMARY OF THE INVENTION

The present application provides a training text generation method, a model training method, a text recognition method, a device electronic device, and a computer-readable storage medium, which can acquire the training text more quickly and efficiently, so as to facilitate the training of a text recognition model. The specific plans are as follows:

In a first aspect, the present application provides a method for generating training text, where the training text is used to train a model to be trained to obtain a text recognition model, and the method includes:

Obtaining guide text, the guide text is consistent with the semantic attribute of the target text, and the target text is the positive example text identified by the text recognition model;

Inputting the guidance text into a guidance-based text generation model to obtain output text consistent with the semantic attributes of the guidance text;

Training text is determined from the output text.

Optionally, before inputting the guidance text into the guidance-based text generation model, the method further includes:

Get the question text;

Said inputting the guidance text into a guidance-based text generation model to obtain output text consistent with the semantic attributes of the guidance text, including:

The question text and the guidance text are input into a guidance-based dialogue generation model, and an output text for replying to the question text and consistent with the semantic attributes of the guidance text is obtained.

Optionally, the output text includes multiple pieces;

The determining of the training text according to the output text includes:

A training text is determined from a plurality of said output texts.

Optionally, the determining the training text from a plurality of the output texts includes:

The training text is determined by a first strategy, the first strategy includes: selecting a text containing at least one preset keyword from a plurality of the output texts as the training text, the preset keyword and the semantics of the target text properties are consistent;

Alternatively, the training text is determined through a second strategy, where the second strategy includes: selecting a first text from a plurality of the output texts or randomly selecting a text as the training text.

Optionally, selecting the first strategy to determine the probability of the training text is a first preset probability, and selecting the second strategy to determine the probability of the training text is a second preset probability, the first preset probability. The probability is greater than the second preset probability, and the sum of the first preset probability and the second preset probability is 1.

Optionally, the range of the first preset probability may be 0.7-0.9, and the range of the second preset probability may be 0.1-0.3.

Optionally, the guide text includes at least one guide word, and each of the guide words is consistent with the semantic attribute of the target text;

The preset keywords include: each of the guide words.

Optionally, the preset keywords further include: each first target word, the first target word is contained in any piece of the output text, is consistent with the semantic attributes of the target text, and is consistent with each Describe words with different leading words.

Optionally, the first strategy further includes: when none of the multiple pieces of the output text contains any of the preset keywords, selecting the first piece of the multiple pieces of the output text to determine the training text.

Optionally, the semantic attribute of the positive example text is semantically sensitive text, the semantic attribute of the target text is semantically sensitive text, and the text recognition model is used to recognize the text generated by the dialogue generation model.

In a second aspect, the embodiments of the present application also provide a method for training a text recognition model, including:

Acquiring training samples, the training samples include positive samples and negative samples, and the text corresponding to the positive samples includes: training text generated by any one of the training text generating methods described in the first aspect;

The model to be trained is trained using the training samples to obtain a text recognition model.

Optionally, the training method further includes:

Obtain a first text, where the first text is the text that the text recognition model has identified incorrectly, and the actual semantic attributes of the incorrectly identified text and the semantic attributes identified by the text recognition model for the incorrectly identified text different;

annotating the first text to obtain a first sample;

The text recognition model is optimally trained using the first sample.

Optionally, before using the first sample to optimize the training of the text recognition model, the training method further includes:

Obtain a second text, the second text contains a second target word, and the second text is opposite to the semantic attribute expressed by the first text, and the second target word is contained in the first text. words that are consistent with the semantic attributes expressed by the target text;

Annotating the second text to obtain a second sample, the second sample is opposite to the annotation information of the first sample;

The performing optimization training on the text recognition model using the first sample includes:

The text recognition model is optimally trained using the first sample and the second sample.

Optionally, the training samples include reply samples and question and answer splicing samples;

The text corresponding to the positive sample in the reply sample includes: the training text determined by the method of inputting the question text and the guide text into the guide-based dialogue generation model in the first aspect;

The text corresponding to the question and answer splicing sample is a spliced text, and the spliced text includes: a text formed by splicing the question text and the reply text corresponding to the question text.

In a third aspect, an embodiment of the present application also provides a text recognition method, which is characterized by comprising:

Get the text to be recognized;

The text to be recognized is input into a text recognition model, and a recognition result of the text to be recognized is obtained, wherein the text recognition model is obtained by training the training method described in any one of the first aspects.

Optionally, the text to be recognized is text generated by a dialogue generation model;

Alternatively, the text to be recognized is the text formed by splicing the user's question text and the text generated by the dialogue generation model, wherein the text recognition model is obtained by using the model training method described in the second aspect. The samples include the reply samples and the model training method when the question and answer splicing samples are trained.

In a fourth aspect, the present application also provides a training text generating device, the training text is used to train a model to be trained to obtain a text recognition model, and the device includes:

an information acquisition unit, used for acquiring guide text, the guide text is consistent with the semantic attribute of the target text, and the target text is the positive example text recognized by the text recognition model;

A text generation unit, for inputting the guidance text into a guidance-based text generation model to obtain output text consistent with the semantic attributes of the guidance text;

A text determination unit, configured to determine training text according to the output text.

Optionally, the device further includes:

a first text acquisition unit, used for acquiring question text;

The text generation unit is specifically configured to: input the question text and the guidance text into a guidance-based dialogue generation model, and obtain an output that is used to reply to the question text and is consistent with the semantic attributes of the guidance text text.

Optionally, the output text includes multiple pieces;

The text determining unit is specifically used for: determining training text from a plurality of the output texts.

Optionally, the text determination unit is specifically used for:

The training text is determined by a first strategy, the first strategy includes: selecting a text containing at least one preset keyword from a plurality of the output texts as the training text, the preset keyword and the semantics of the target text properties are consistent;

Alternatively, the training text is determined through a second strategy, where the second strategy includes: selecting a first text from a plurality of the output texts or randomly selecting a text as the training text.

Optionally, selecting the first strategy to determine the probability of the training text is a first preset probability, and selecting the second strategy to determine the probability of the training text is a second preset probability, the first preset probability. The probability is greater than the second preset probability, and the sum of the first preset probability and the second preset probability is 1.

Optionally, the range of the first preset probability may be 0.7-0.9, and the range of the second preset probability may be 0.1-0.3.

Optionally, the guide text includes at least one guide word, and each of the guide words is consistent with the semantic attribute of the target text;

The preset keywords include: each of the guide words.

Optionally, the preset keywords further include: each first target word, the first target word is contained in any piece of the output text, is consistent with the semantic attributes of the target text, and is consistent with each Describe words with different leading words.

Optionally, the first strategy further includes: when none of the multiple pieces of the output text contains any of the preset keywords, selecting the first piece of the multiple pieces of the output text to determine the training text.

Optionally, the semantic attribute of the positive example text is semantically sensitive text, the semantic attribute of the target text is semantically sensitive text, and the text recognition model is used to recognize the text generated by the dialogue generation model.

In a fifth aspect, the embodiment of the present application also provides a training device for a text recognition model, including:

A sample acquisition unit, configured to acquire training samples, the training samples include positive samples and negative samples, and the text corresponding to the positive samples includes: generated by the training text generating device according to any one of the fourth aspects training text;

A model training unit, configured to use the training samples to train the model to be trained to obtain a text recognition model.

Optionally, the training device further includes:

A second text acquisition unit, configured to acquire a first text, where the first text is the text that the text recognition model has identified incorrectly, and the actual semantic attribute of the incorrectly recognized text is related to the recognition error of the text by the text recognition model. The semantic attributes recognized by the text are different;

a sample labeling unit, configured to label the first text to obtain a first sample;

A model optimization unit, configured to perform optimization training on the text recognition model by using the first sample.

Optionally, the second text acquisition unit is further configured to:

Obtain a second text, the second text contains a second target word, and the second text is opposite to the semantic attribute expressed by the first text, and the second target word is contained in the first text. words that are consistent with the semantic attributes expressed by the target text;

The sample labeling unit is further configured to: label the second text to obtain a second sample, where the labeling information of the second sample is opposite to that of the first sample;

The model optimization unit is specifically configured to: use the first sample and the second sample to perform optimization training on the text recognition model.

Optionally, the training samples include reply samples and question and answer splicing samples;

The text corresponding to the positive sample in the reply sample includes: the training text determined by the method of inputting the question text and the guide text into the guide-based dialogue generation model in the first aspect;

The text corresponding to the question and answer splicing sample is a spliced text, and the spliced text includes: a text formed by splicing the question text and the reply text corresponding to the question text.

In a sixth aspect, an embodiment of the present application further provides a text recognition device, including:

a third text acquisition unit, used for acquiring the text to be recognized;

A text recognition unit, configured to input the text to be recognized into a text recognition model, and obtain a recognition result of the text to be recognized, wherein the text recognition model is performed by the training device according to any one of the fifth aspects obtained by training.

Optionally, the text to be recognized is text generated by a dialogue generation model;

Alternatively, the text to be recognized is the text formed by splicing the user's question text and the text generated by the dialogue generation model, wherein the text recognition model is obtained by using the training method described in the first aspect when the training sample includes a reply The samples and the question and answer splicing samples are obtained by training.

In a seventh aspect, an embodiment of the present application further provides an electronic device, including:

processor; and

The memory is used to store a data processing program, and after the electronic device is powered on and runs the program through the processor, the method according to any one of the first aspects is executed.

In an eighth aspect, an embodiment of the present application further provides an electronic device, including:

processor; and

The memory is used to store a data processing program, and after the electronic device is powered on and runs the program through the processor, the method according to any one of the second aspects is executed.

In a ninth aspect, an embodiment of the present application also provides an electronic device, including:

processor; and

The memory is used to store a data processing program, and after the electronic device is powered on and runs the program through the processor, the method according to any one of the third aspects is executed.

In a tenth aspect, an embodiment of the present application further provides a computer-readable storage medium storing a data processing program, where the program is executed by a processor to execute the method described in any one of the first aspect.

In an eleventh aspect, an embodiment of the present application further provides a computer-readable storage medium storing a data processing program, where the program is executed by a processor to execute the method described in any one of the second aspect.

In a twelfth aspect, an embodiment of the present application further provides a computer-readable storage medium, which stores a data processing program, and the program is executed by a processor to execute the method according to any one of the third aspects.

Compared with the prior art, the present application has the following advantages:

In the method for generating training text provided by the present application, after inputting the guiding text into the guiding-based text generation model, an output text consistent with the semantic attributes of the guiding text can be obtained. Since the guiding text is consistent with the semantic attributes of the target text, the target text is The positive text recognized by the text recognition model, therefore, the obtained output text is also consistent with the semantic attributes of the positive text recognized by the text recognition model. In this way, the training text determined according to the obtained output text can be used as positive text. Example sample text for training the model to be trained.

In the present application, the output text is automatically generated by the guided text generation model, so that the training text can be determined, and the training text can be obtained more quickly and efficiently. The diversity of the training text is also better, so that the recognition accuracy of the text recognition model obtained by training can be improved, so that the text recognition model can more accurately recognize the positive text.

Description of drawings

1 is a flowchart of a training text generation method provided by an embodiment of the present application;

2 is a flowchart of another example of a training text generation method provided by an embodiment of the present application;

3 is a flowchart of a text recognition model training method provided by an embodiment of the present application;

4 is a flowchart of another example of a text recognition model training method provided by an embodiment of the present application;

5 is a unit block diagram of a training text generating apparatus provided by an embodiment of the present application;

FIG. 6 is a schematic structural diagram of an electronic device for implementing a training text generation method provided by an embodiment of the present application.

Detailed ways

In the following description, numerous specific details are set forth in order to provide a thorough understanding of the present application. However, the present application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar promotions without violating the connotation of the present application. Therefore, the present application is not limited by the specific implementation disclosed below.

Intelligent chat technology can automatically reply to the questions raised by users, and its intelligent customer service and chat machines in the fields of e-commerce or public services

It is widely used in people, game chat and other fields.

For the user's question, the intelligent chatting device can retrieve the reply content corresponding to the user's question from the pre-stored question and answer database. Due to the limited number of questions stored in the question-and-answer database, this method cannot give corresponding responses to the questions that are not stored in the question-and-answer database, resulting in a more fluent question-and-answer with the user, and the content of the responses is relatively simple. Bad user experience.

With the development of deep learning, the intelligent chat scene has gradually begun to use the dialogue generation model to reply to the user's questions, which greatly improves the diversity of the reply content, and also supports multiple rounds of coherent chats, and the user experience is better.

However, the dialogue generation model is usually trained based on massive sample data, which inevitably contains some sensitive texts, such as abuse, violence and other sensitive texts. Therefore, the dialogue generation model will learn the expressions of these sensitive texts. As a result, the dialogue generation model may generate sensitive text to reply to the user. In order to create a green chat environment, the generated sensitive text needs to be identified and filtered out in advance.

In the related art, sensitive text can be filtered in a word-based or word-based manner. For example, texts containing sensitive words such as "do", "your mother" and "you were born" are directly filtered out. For example, if the intelligent dialogue device replies "I want to do it with you", it will be directly filtered out. Since most of the texts containing the word "do" are not sensitive texts, such as "cooking", "doing housework", "doing sports", etc., a whitelist table will also be set up and not filtered in the whitelist table text. For example, if "doing housework" is put into the whitelist, the reply of "I am doing housework at home today" from the smart conversation device will not be filtered out.

However, due to the huge number of sensitive words and the limited number of sensitive words listed, many sensitive texts will be missed, resulting in many sensitive output texts that cannot be filtered. Second, since the whitelist is also endless, many normal texts will be filtered out, resulting in a decrease in the quality of the conversation. In addition, this method can only filter out texts that display sensitive words, but cannot filter out texts that do not contain sensitive words but are semantically sensitive, such as "I want you".

In the related art, sensitive text can also be filtered in a way based on regular expressions. For example, when the person setting of the intelligent dialogue device is a child, the text replied by the intelligent dialogue device expressing that he wants to have a child or that he has a child is a sensitive text. In this case, the regular expression can be "has .* .* children" or "has .* .* children", where "." matches any character except a newline, and "*" matches the preceding 0 or more characters. When the smart dialogue device replies "Mother cat has given birth to 2 children" and "You have given birth to 1 child", it will be filtered out because it matches the regular expression "born.*children.*children".

However, due to the inability to exhaust all sensitive regular expressions, many sensitive texts cannot be recognized. In addition, many normal texts will be killed by mistake. For example, the above-mentioned "Mother cat gave birth to 2 children" will be killed by mistake. In addition, the regular expression method can only filter out the text that matches the regular expression, but cannot filter out the implicitly sensitive text that does not match the regular expression.

In order to improve the recognition accuracy of sensitive text, a recognition model can be used to recognize sensitive text, wherein the recognition model is a deep model. However, the recognition model needs to be pre-trained with a large number of sensitive texts with different expressions as samples. Since it is difficult to collect a large number of sensitive texts on the Internet, and the number and diversity of manually written sensitive texts are limited, how to efficiently Obtaining a large amount of training text (such as sensitive text) is a problem that needs to be solved.

In order to obtain a large amount of training text more quickly and efficiently, the present application provides a training text generation method, a text recognition model training method, a text recognition method, and apparatuses, electronic devices, and computer-readable storage corresponding to the methods The following provides embodiments to describe in detail the above-mentioned methods, apparatuses, electronic devices, and computer-readable storage media.

The first embodiment of the present application provides a method for generating training text, where the training text is used to train a model to be trained to obtain a text recognition model. In the embodiment of the present application, the execution body of the training text generation method is an electronic device, and the electronic device may be any electronic device with data processing capability, such as a desktop computer, a notebook computer, a tablet computer, a server, a mobile phone, and the like.

The above-mentioned models to be trained may include at least one of a bert model, a convolutional neural network, a logistic regression model, a K-nearest neighbor (K-Nearest Neighbor, KNN) model, a logistic regression model, and a two-class model, or may be any other depth Learning models.

The above text recognition model can determine whether the text to be recognized is positive text or negative text, and the above text recognition model can also be understood as a text classification model. The above text recognition model can be used to recognize Chinese texts, and can also recognize foreign language texts such as English, French, and German.

Positive example text refers to the text that the text recognition model needs to recognize. For example, if the text recognition model is used to recognize text containing place names, then the text containing place names is positive example text, and the text that does not contain place names is negative example text. The recognition model is used to identify texts with sensitive semantics. Then, the texts with sensitive semantics are positive example texts, and the texts with non-sensitive semantics are negative example texts.

As shown in FIG. 1 , the training text generation method provided by the embodiment of the present application includes the following steps S110-S130.

Step S110: Acquire guidance text.

The above-mentioned guide text is consistent with the semantic attributes of the target text, and the target text is the positive example text recognized by the text recognition model.

The above-mentioned guide text can be one or more paragraphs of text, one or more sentences of text, and one or more words. . In the embodiment of the present application, the guide text may be manually set and then input into the electronic device, and the electronic device may obtain the manually input guide text; or, the electronic device may automatically determine the guide text according to the target text. For example, the electronic device may automatically determine the guide text. Identify the semantic attribute of the target text, and determine the guide text consistent with the semantic attribute according to the semantic attribute of the target text.

The above-mentioned guiding text may be Chinese text or Chinese words, and may also be foreign language texts or foreign words such as English and German.

The semantic attribute of the target text can be understood as the type to which the semantics of the target text belong. Specifically, the semantic attribute of the target text may be at least one of semantic-sensitive text, confidential text, popular science text, academic text, and medical knowledge text, and the semantic attribute of the target text may also be other specific semantic attributes. Personnel can determine the semantic properties of the target text according to the semantic properties of the positive example text that the text recognition model needs to recognize, and this application does not limit the specific semantic properties of the target text.

For example, if the text recognition model is used to identify semantically sensitive text, the semantic attribute of the guide text can be sensitive text, and if the trained text recognition model is used to identify confidential text, the semantic attribute of the guide text can be classified as confidential text.

Optionally, the above-mentioned sensitive text may include violent text, pornographic text, language attack text or other unhealthy text.

For example, if the positive text identified by the text recognition model is text with sensitive attributes such as violence, pornography, and abuse, the guiding text may include one or more single-character words with sensitive semantic attributes such as violence, pornography, and abuse. or multiple words.

Each single-word or multi-word contained in the guide text may be called guide words, that is, the guide text contains one or more guide words, and each guide word is consistent with the semantic attributes of the target text. When the guide text contains one or more guide words, the semantics of the guide words are easier to obtain because the guide words are shorter. Therefore, when the guide text contains guide words, it is more convenient for the guide-based text generation model to generate the semantics of the guide words. Attribute consistent text.

The number of guiding words included in the guiding text may range from 5 to 15, for example, the number of guiding words included in the guiding text is 5, 8, 10, 12, 15, and so on. The number of guiding words should not be too much or too little. Too much will make the computational complexity of the guiding-based text generation model too high, which will affect the output efficiency of the output text, and even the output text cannot be obtained due to the operation error. The semantic properties of text and target text are quite different.

Guiding words can be used as prefix texts. Prefix texts refer to words that remain unchanged every time the text generation method is trained. The semantics of prefix texts are more stable and more guiding. That is to say, for training samples that generate the same semantic attribute, the prefix text remains unchanged every time the text is generated by the training text generation method provided in this application, so that the training text obtained each time is the same as the prefix text. have the same semantic properties.

Step S120: Input the guide text into the guide-based text generation model to obtain output text consistent with the semantic attributes of the guide text.

The above guidance-based text generation model is a pre-trained model. In the embodiments of the present application, deep models such as neural network models such as ELMO, OpenAIGPT, BERT, or OpenAIGPT-2 can be trained based on the guide samples and the text samples corresponding to the guide samples, thereby obtaining a guide-based text generation model. Among them, guide samples and text samples can be obtained and marked from novels, scripts, articles in magazines or journals and other materials, and those skilled in the art can train and obtain guide-based text generation models according to conventional model training methods. detail.

Step S130: Determine the training text according to the above output text.

In step S130, the output text can be directly determined as the training text, or the sentence expansion can be performed according to the output text to obtain the extended text consistent with the semantic attributes of the output text, and the output text and the extended text can be determined as the training samples, or, The training text is determined by other means based on the output text.

In the embodiment of the present application, since the semantic attributes of the guide text and the positive text recognized by the text recognition model are consistent, when the obtained output text is consistent with the semantic attributes of the guide text, the sample text determined according to the output text and the text The semantic properties of the positive texts identified by the recognition model are also consistent.

For example, when the semantic attribute of the guide text is abusive text, the semantic attribute of the output text is also abusive text, and the semantic attribute of the training text determined according to the output text is also abusive text, so the training text can be used as abusive text. The training model is trained according to the text corresponding to the positive sample, so that the text recognition model obtained by training can recognize the text with abusive attributes, and the text with abusive attributes is the positive example text.

In the present application, the output text is automatically generated by the guided text generation model, so that the training text can be determined, and the training text can be obtained more quickly and efficiently. The diversity of the training text is also better, so that the recognition accuracy of the text recognition model obtained by training can be improved, so that the text recognition model can more accurately recognize the positive text.

In an embodiment, as shown in FIG. 2 , before step S120, the following step S140 may also be included.

Step S140: Obtain the question text.

Step S120 can be implemented in the following step S121.

Step S121 : Input the question text and the guidance text into the guidance-based dialogue generation model, and obtain an output text for replying to the question text and consistent with the semantic attributes of the guidance text.

The above question texts are obtained through novel fragments, line scripts, social media chat records, etc. The question text acts as the text of the user's question. The question text can be inquiring text such as "What time is it?" and "Where is the company address?" The question text can also be chatting text such as "I wish you a good time", "We are good friends", and "The weather is nice". The question text can be one-sentence text or multiple-sentence text.

The question text may be text input by the user into the electronic device, or may be text selected by the electronic device from a stored text library.

In this embodiment, the text generation model in step S120 is the dialogue generation model in step S121.

The above-mentioned guidance-based dialogue generation model may train a deep model based on question samples, guidance samples, and response samples corresponding to the guidance samples and question samples, thereby obtaining a guidance-based text generation model. Among them, the question samples, guidance samples and reply samples can be obtained and marked from novels, line scripts, social media chat data and other materials, and those skilled in the art can train a text dialogue model based on guidance according to conventional model training methods. The application is no longer detailed.

In this embodiment, an output text is generated by a dialogue generation model based on guidance. Since the output text is a reply to the question text, the output text is more consistent with the content automatically replied by the intelligent chatting device. In this way, after the training text is determined according to the output text , the text recognition model trained based on the training text is more suitable for recognizing the chat information automatically generated by the intelligent chatting device, and also more suitable for recognizing the reply text generated by the intelligent chatting device through the dialogue generation model. In this way, when the intelligent chatting device generates unhealthy sensitive text such as abuse and violence through the dialogue generation model, the text recognition model can more accurately identify the sensitive text generated by the dialogue generation model.

Optionally, the output text derived from the guided dialogue generation model may include a piece of output text, in which case, the piece of output text may be determined as a training sample.

In an implementation manner, the above-mentioned output text may include multiple pieces of text, and step S130 may be implemented in the following step S131: determining training text from multiple pieces of output text.

The output text includes multiple pieces, that is, multiple pieces of output text are obtained through the guided-based text generation model.

Optionally, as shown in FIG. 2 , the training text may be determined from a plurality of output texts according to the following step S131a.

Step S131a: Select a text containing at least one preset keyword from a plurality of output texts and determine it as a training text.

In this embodiment of the present application, the method of determining the training text in step S131a may be determined as the first strategy.

The semantic attributes of the preset keywords are consistent with the semantic attributes of the target text. The above preset keywords may be words input by the user, and the preset keywords may include one or more keywords. For example, when the positive example text recognized by the text recognition model is semantically sensitive text, the preset keywords may include: moaning, beating, dating, beating, and other semantically sensitive words.

When there is more than one text containing preset keywords in multiple output texts, each output text containing preset keywords can be determined as training text, or the first text in each output text containing preset keywords can be determined as training text. One is determined as training text, or the output text that contains the most preset keywords is determined as training text.

In this embodiment, since the preset keywords are consistent with the semantic attributes of the positive example text recognized by the text generation model, it is easier for the output text containing the preset keywords to match the semantic attributes of the positive example text recognized by the text generation model Consistently, in this way, using the output text containing the preset keywords as the training text enables the trained text recognition model to recognize the positive text with higher accuracy.

Optionally, as shown in FIG. 2 , the training text may also be determined from a plurality of output texts according to the following step S131b or step S131c.

Step S131b: Select the first one from the plurality of output texts and determine it as the training text.

In this embodiment of the present application, the method of determining the training text in step S131b may be used as the second strategy.

Since the first one of the multiple output texts usually has a higher degree of matching with the guiding text, selecting the first output text as the training text to train the model to be trained enables the trained text recognition model to recognize the positive text higher accuracy.

Step S131c: Randomly select a text from a plurality of output texts and determine it as a training text.

In this embodiment, each output text may also be determined as training text, or each output text may be displayed so that the user can select from each output text, and then the text selected by the user is determined as training text . The present application does not specifically limit the manner of determining the training text from multiple pieces of output text.

In this embodiment, since the output text obtained by the semantic-based text model includes multiple pieces of output text, it is possible to flexibly determine the output text that is more consistent with the semantic attributes of the target text from the multiple pieces of output text as the training text, thereby making it possible to determine The semantic attributes of the training text and the target text are more consistent, so that the text recognition model determined by the training text can more accurately identify the positive text.

In one embodiment, as shown in FIG. 2 , the probability that the first strategy is selected to determine the training text is the first preset probability, the probability that the second strategy is selected to determine the training text is the second preset probability, and the first preset probability It is assumed that the probability is greater than the second preset probability, and the sum of the first preset probability and the second preset probability is 1.

That is, in this embodiment, the first strategy is selected with a first preset probability to determine the training text, and the second strategy is selected with a second preset probability to determine the training text.

The first preset probability is greater than the second preset probability, and the first preset probability and the second preset probability may be 0.8 and 0.2 respectively, or the first preset probability and the second preset probability may be 0.9 and 0.1 respectively. , or the first preset probability and the second preset probability are 0.6, 0.4, etc., respectively, the first preset probability is greater than the second preset probability, and the sum of the two is 1. This application does not limit two The specific value of the probability.

In a specific embodiment, the range of the first preset probability may be 0.7-0.9, and the range of the second preset probability may be 0.1-0.3. For example, the first preset probability and the second preset probability are respectively 0.7 and 0.3, the first preset probability and the second preset probability are respectively 0.8 and 0.2, and the first preset probability and the second preset probability are respectively 0.9 , 0.1. That is to say, the difference between the first preset probability and the second preset probability is relatively large. Generally speaking, the probability that the output text containing the preset keywords is consistent with the semantic attributes of the target text is much greater than the probability that the output text does not contain the preset keywords but is consistent with the semantic attributes of the target text. Therefore, the first preset probability ratio The second preset probability is relatively large, so that the obtained training text can be more consistent with the semantic attributes of the target text, and the diversity of the obtained training text can be better, so that the training text can be passed through the training text. The trained text recognition model can more accurately identify more diverse positive texts.

Due to the limited number of preset keywords, it is difficult for each preset keyword to contain all words that are consistent with the semantic attributes of the target text. Therefore, even if the output text does not contain any preset keywords, the output text may be consistent with the target text. have the same semantic properties. Since the probability that the first output text is consistent with the semantic attributes of the target text is relatively large, no matter whether the first output text contains preset keywords or not, the first output text may also be consistent with the semantic attributes of the target text. Consistent, in this embodiment, the first output text is selected as the training text with a smaller second preset probability, and training texts that do not contain preset keywords but are consistent with the semantic attributes of the target text can be added, so that the The diversity is better, so that the trained text recognition model can recognize more diverse positive texts.

When the output text contains a preset keyword, it means that the probability that the output text is consistent with the semantic attributes of the target text is very high. Therefore, the output text containing at least one preset keyword is determined with a larger first preset probability. In order to train the text, the determined training text can be made more consistent with the semantic attributes of the target text, so that the text recognition model obtained by training the training text can more accurately identify the positive example text.

The above-mentioned preset keywords may include each of the above-mentioned guide words. Since the guide words contained in the guide text are consistent with the semantics of the target text, each guide word is directly determined as a preset keyword, and the preset keyword is obtained quickly.

In a specific embodiment, the above-mentioned preset keywords may also include first target words, and the first target words are contained in any output text, are consistent with the semantic attributes of the target text, and are different from each guiding word. word.

In this embodiment, multiple pieces of output text may be displayed, so that the user can view each piece of output text. After viewing each piece of output text, the user can find words from each piece of output text that are consistent with the semantic attributes of the target text and different from each guide word, and input the found words into the electronic device, so that the electronic device can Get the words that the user found and typed.

Since the output text is generated based on the guide text, the output text has a high probability of being consistent with the semantic attributes of the target text. Therefore, the probability that the output text contains words that are consistent with the semantic attributes of the target text is also high. In this way, from the output text It is possible to filter out words that are consistent with the semantic properties of the target text and are different from each leading word.

Due to the limited number of guiding words, it is possible to filter out some texts containing other words that are semantically consistent with the target text to determine whether the output text is training text according to whether the output text contains guiding words. In this embodiment, the first target word is determined as the preset keyword, that is, the preset keyword is supplemented, and the preset keyword is more abundant and diverse. In this way, it is not easy to miss the text consistent with the target text in the output text. Make the training samples more diverse and rich.

In a specific embodiment, after step S131a, the following step S131c may also be included.

Step S131c: when each of the multiple output texts does not contain any preset keyword, determine the first one of the multiple output texts as a training sample.

In this embodiment of the present application, the manner in which the training text is determined in steps S131a and S131c may be jointly determined as the first strategy.

This embodiment can increase the number of acquired texts that are consistent with the semantic attributes of the target text.

In this embodiment of the present application, the user can perform steps S110 to S130 multiple times, and the guiding text each time is consistent with the semantic attributes of the positive example text recognized by the text generation model, but the guiding text each time is different. They are exactly the same, for example, the guiding texts may be partially or completely different each time the process is performed, so that by executing the training text generation method multiple times, more different training texts can be obtained. The text recognition model trained by training text can be used to identify sensitive texts such as abuse and yellowishness. It can be seen that various training texts can be generated by the method provided in this application.

As shown in FIG. 3 , a second embodiment of the present application provides a method for training a text recognition model, and the method includes the following steps S510 to S520.

Step S510: Obtain training samples.

The above-mentioned training samples include positive samples and negative samples, and the texts corresponding to the positive samples include: training texts generated by any one of the training text generating methods described in the first embodiment.

In this embodiment of the present application, the training text generated by any one of the training text generating methods described in the first embodiment may be marked as a positive sample, thereby obtaining a positive sample.

The text corresponding to the above negative samples can be obtained from novels, magazines, web pages and other materials. The text corresponding to the negative samples has opposite semantic properties to the text corresponding to the positive samples. For example, the text corresponding to the positive sample is semantically sensitive text, the text corresponding to the negative sample is semantically insensitive text, the text corresponding to the positive sample is the text containing the name of the person, then the text corresponding to the negative sample does not contain The text of the person's name.

When the text recognition model is used to identify whether the text generated by the dialogue generation model is sensitive text, the text corresponding to the positive sample may further include: a third text determined from the sensitive text filtered out of the historical chat information. For example, for an intelligent chat system, it may have been running for a period of time, and some texts have been filtered through regular expression-based filtering, word- or word-based filtering, etc., and these filtered texts The probability of being sensitive text is very large, so a large number of sensitive texts can be quickly obtained from these filtered sensitive texts.

Specifically, a sensitive third text can be manually selected from the above-mentioned filtered sensitive texts, and the electronic device determines the manually selected third text as the text corresponding to the positive sample, and marks the third text as the positive sample. .

The text corresponding to the positive sample may further include: a fourth text determined from unfiltered texts in the historical chat information. The unfiltered texts in the historical chat information may also contain sensitive texts. Therefore, a sensitive fourth text can be manually selected from the above unfiltered texts, and the electronic device determines the manually selected fourth text as a positive example. The text corresponding to the sample, and the fourth text is marked with the text positive sample.

Through the above method, a large number of positive samples can be quickly obtained.

When the text recognition model is used to identify whether the text generated by the dialogue generation model is sensitive text, the text corresponding to the negative sample may include: text determined from unfiltered texts in historical chat information. Since the non-sensitive texts in the historical chat information account for a larger proportion and the number is larger, the texts determined from the unfiltered texts are basically non-sensitive texts, so that a large number of non-sensitive texts can be easily obtained. , that is, the text corresponding to the negative sample. The electronic device may mark the text corresponding to the negative example as a negative example.

Step S520: Use the above training samples to train the model to be trained to obtain a text recognition model.

For the specific type of the model to be trained, reference may be made to the description of the first embodiment, which will not be repeated here.

In the embodiment of the present application, the text corresponding to the positive sample and the text corresponding to the negative sample can be input into the model to be trained for encoding, and the encoded text is classified by a binary classifier, and the classification result is the positive text (for example, sensitive text, text containing names) or negative texts (such as non-sensitive texts, texts not containing names), compare the classification results obtained with the label information of the texts corresponding to positive and negative samples to adjust the model to be trained of each parameter.

During the training process, the loss function can be a general binary classification cross-entropy loss function.

In this embodiment of the present application, the model to be trained may include a text encoding model and a binary classification model.

Since the training method of the text recognition model provided by the present application adopts the training text generation method described in the first embodiment to generate the text corresponding to the positive sample, it has beneficial effects corresponding to the first embodiment, which is not repeated here. Repeat.

In one embodiment, as shown in FIG. 4 , the training method may further include the following steps S530-S540.

Step S530: Acquire the first text, and annotate the first text to obtain a first sample.

The first text is an incorrectly recognized text by the text recognition model, and the actual semantic attribute of the incorrectly recognized text is different from the semantic attribute recognized by the text recognition model for the incorrectly recognized text. For example, the actual semantic attribute of the sentence "Cao Cao was a historical figure in the Three Kingdoms period" is insensitive semantics, but the semantic attribute recognized by the text recognition model for this sentence is sensitive semantics, then this sentence is recognized by the text recognition model. wrong text.

Step S540: Use the first sample to perform optimization training on the text recognition model.

Specifically, the first text is the text recognized by the trained text recognition model in the process of performing text recognition, and the present application can manually check the text recognized by the text recognition model to find the text recognized by mistake.

In this application, the above-mentioned first text can be manually marked, for example, marked as a positive sample or a negative sample.

In this embodiment, the text recognition model can be optimized and trained at preset time intervals, and the preset time interval can be one month, two months, or other time intervals. Alternatively, the text recognition model can also be optimized and trained in response to an optimization instruction triggered by the user.

In this embodiment, the text recognition model is optimized and trained by using the wrongly recognized text by the text recognition model, which can further improve the recognition accuracy of the text recognition model.

In a specific embodiment, as shown in FIG. 4 , before the above-mentioned step of using the first sample to optimize the training of the text recognition model, the above-mentioned training method may further include the following step S550.

Step S550: Acquire the second text, and mark the second text to obtain a second sample.

Wherein, the second text contains a second target word, and the second text is opposite to the semantic attribute expressed by the first text, and the second target word is contained in the first text and is consistent with the semantic attribute expressed by the target text. word. The label information of the second sample is opposite to that of the first sample.

The above step S540 can be implemented by the following step S541.

Step S541: Use the first sample and the second sample to perform optimization training on the text recognition model.

In this embodiment, the second sample can be manually determined, and the second sample can be input into the electronic device, so that the electronic device can acquire the second sample.

For example, for the scene where the text recognition model is used to recognize sensitive text, the text recognition model predicts the wrong first text as "Cao Cao is three famous people", the text recognition model predicts it as sensitive text, but the first text It is actually a non-sensitive text, so "Cao Cao is three famous people" is marked as a sensitive text, that is, marked as a positive sample, and the words contained in the first text that are consistent with the semantic attributes expressed by the target text are "Cao Cao". ”, that is, “Cao” is the second target word, then the second text that contains “Cao” and is opposite to the semantic attribute of the non-sensitive text “Cao Cao is three famous people” can be “How do you want to fuck” , the second text is sensitive text, which is opposite to the semantic properties of the first text, so "how do you want to do it" is marked as non-sensitive text, that is, marked as a negative example.

Since "Cao Cao is three famous people" contains the semantically sensitive word "Cao", the text recognition model recognizes this sentence as a sensitive text. In this example, "Cao Cao is three famous people" is marked as insensitive text. When the model is optimized for training, it is easy for the model to think that the text containing "Cao" is insensitive text. In this case, in order to avoid model overfitting To some words, you can add some second samples that are opposite to the labeling information of the first sample and contain the second target word "動", and optimize the training of the model through the first sample and the second sample. The phenomenon that the model is over-fitted to some words is well avoided, and the accuracy of model recognition is improved.

In one embodiment, the above-mentioned training samples may include reply samples and question-and-answer splicing samples.

The text corresponding to the positive sample in the reply sample includes: the text generated by the above-mentioned guidance-based dialogue generation model; the text corresponding to the question and answer splicing sample is spliced text, and the spliced text includes: splicing the question text and the reply text corresponding to the question text text formed later.

In this embodiment, the text corresponding to the positive sample in the reply sample may further include: reply text obtained from the intelligent chatting device, reply text determined from the dialogue of a novel or a script, and the like.

The text corresponding to the negative sample in the reply sample may include: text determined from resources such as novels, scripts, and online articles, and the text corresponding to the negative sample in the reply sample may also include: text generated by a dialogue generation model.

The user can manually annotate the positive samples and negative samples in the reply samples.

In this embodiment, the text corresponding to the reply sample is used to represent the reply to the question asked by the user.

The above spliced text can reflect the information obtained by splicing the user's question and the respondent's reply.

Dialogue information can be manually obtained from resources such as novels, scripts, and online articles, and then a question text and a reply text corresponding to the question text can be obtained from these dialogue information.

Alternatively, a reply text corresponding to the question text may also be generated by the dialogue generation model, and the question text and the text generated by the dialogue generation model may be determined as the reply text.

For example, if the question text is "do you miss me" and the reply text is "I miss you", the concatenated text can be "do you miss me, I miss you".

Since some users may deliberately chat about some sensitive topics, at this time, if the text replied by the chat device is identified separately, it may not be sensitive text, but the user's question and the reply from the smart reply device may be connected to look like sensitive text. If the user asks "do you want me", the chat device replies with "um, I want", the reply text "um, I want" is not sensitive text, but "do you want me, um, I want" is Sensitive text. In this embodiment, the question and answer splicing sample is used as a training sample, so that the text recognition model obtained by training can be used to recognize the question and answer splicing text, so that the text recognition model has a wider application range and can identify more types of target texts.

In the second embodiment, the parts different from the first embodiment are mainly explained and described, and the content of the same or similar parts as the first embodiment will not be repeated.

A third embodiment of the present application provides a text recognition method, comprising the following steps:

Get the text to be recognized;

The above text to be recognized is input into the text recognition model, and the recognition result of the text to be recognized is obtained.

Wherein, the above-mentioned text recognition model is obtained by training through any one of the training methods described in the second embodiment.

The above text to be recognized may be text generated by a dialogue generation model.

The above-mentioned text to be recognized may also be a text formed by splicing the user's question text and the text generated by the dialogue generation model. In this case, the training samples for training the text recognition model include the above-mentioned reply samples and the above-mentioned question-and-answer splicing samples.

The above text to be recognized may also be an online message, dialogue information on a chat tool, text in an online article, etc. The application does not limit the specific content of the text to be recognized.

The methods in the prior art include a sensitive word recognition method and a rule-based recognition method. Using the text recognition method provided by the present application, the obscure and sensitive text can be well recognized, and the phenomenon of misjudgment is not easy to occur, and the recognition of sensitive text is accurate. higher rate.

The text recognition model in the text recognition method of the third embodiment is obtained by training with the method provided by the second embodiment. Therefore, this embodiment has similar beneficial effects as the second embodiment, which will not be repeated here.

In the third embodiment, the different parts from the first embodiment and the second embodiment are mainly explained, and the content of the same or similar parts as the first embodiment and the second embodiment will not be repeated.

As shown in FIG. 5 , the fourth embodiment of the present application also provides a training text generation device, the training text is used to train a model to be trained to obtain a text recognition model, and the device includes:

an information acquisition unit 810, configured to acquire guide text, the guide text being consistent with the semantic attributes of the target text, and the target text is the positive example text recognized by the text recognition model;

A text generation unit 820, configured to input the guidance text into a guidance-based text generation model to obtain output text consistent with the semantic attributes of the guidance text;

A text determination unit 830, configured to determine training text according to the output text.

Optionally, the device further includes:

a first text obtaining unit, used for obtaining the question text;

The text generation unit is specifically configured to: input the question text and the guidance text into a guidance-based dialogue generation model, and obtain an output that is used to reply to the question text and is consistent with the semantic attributes of the guidance text text.

Optionally, the output text includes multiple pieces;

The text determining unit is specifically used for: determining training text from a plurality of the output texts.

Optionally, the text determination unit is specifically used for:

The training text is determined by a first strategy, the first strategy includes: selecting a text containing at least one preset keyword from a plurality of the output texts as the training text, the preset keyword and the semantics of the target text properties are consistent;

Alternatively, the training text is determined through a second strategy, where the second strategy includes: selecting a first text from a plurality of the output texts or randomly selecting a text as the training text.

Optionally, selecting the first strategy to determine the probability of the training text is a first preset probability, and selecting the second strategy to determine the probability of the training text is a second preset probability, the first preset probability. The probability is greater than the second preset probability, and the sum of the first preset probability and the second preset probability is 1.

Optionally, the range of the first preset probability may be 0.7-0.9, and the range of the second preset probability may be 0.1-0.3.

Optionally, the guide text includes at least one guide word, and each of the guide words is consistent with the semantic attribute of the target text;

The preset keywords include: each of the guide words.

Optionally, the preset keywords further include: each first target word, the first target word is contained in any piece of the output text, is consistent with the semantic attributes of the target text, and is consistent with each Describe words with different leading words.

Optionally, the first strategy further includes: when none of the multiple pieces of the output text contains any of the preset keywords, selecting the first piece of the multiple pieces of the output text to determine the training text.

Optionally, the semantic attribute of the positive example text is semantically sensitive text, the semantic attribute of the target text is semantically sensitive text, and the text recognition model is used to recognize the text generated by the dialogue generation model.

The fifth embodiment of the present application also provides a training device for a text recognition model, including:

A sample acquisition unit, configured to acquire training samples, the training samples include positive samples and negative samples, and the text corresponding to the positive samples includes: generated by the training text generating apparatus described in any one of the fourth embodiments; Generated training text;

A model training unit, configured to use the training samples to train the model to be trained to obtain a text recognition model.

Optionally, the training device further includes:

A second text acquisition unit, configured to acquire a first text, where the first text is the text that the text recognition model has identified incorrectly, and the actual semantic attribute of the incorrectly recognized text is related to the recognition error of the text by the text recognition model. The semantic attributes recognized by the text are different;

a sample labeling unit, configured to label the first text to obtain a first sample;

A model optimization unit, configured to perform optimization training on the text recognition model by using the first sample.

Optionally, the second text acquisition unit is further configured to:

Obtain a second text, the second text contains a second target word, and the second text is opposite to the semantic attribute expressed by the first text, and the second target word is contained in the first text. words that are consistent with the semantic attributes expressed by the target text;

The sample labeling unit is further configured to: label the second text to obtain a second sample, where the labeling information of the second sample is opposite to that of the first sample;

The model optimization unit is specifically configured to: use the first sample and the second sample to perform optimization training on the text recognition model.

Optionally, the training samples include reply samples and question and answer splicing samples;

The text corresponding to the positive sample in the reply sample includes: the training text determined by inputting the question text and the guide text into the guide-based dialogue generation model in the first embodiment;

The text corresponding to the question and answer splicing sample is a spliced text, and the spliced text includes: a text formed by splicing the question text and the reply text corresponding to the question text.

The sixth embodiment of the present application also provides a text recognition device, including:

a third text acquisition unit, used for acquiring the text to be recognized;

A text recognition unit, configured to input the text to be recognized into a text recognition model, and obtain a recognition result of the text to be recognized, wherein the text recognition model is trained by any one of the fifth embodiments device for training.

Optionally, the text to be recognized is text generated by a dialogue generation model;

Or, the text to be recognized is the text formed by splicing the user's question text and the text generated by the dialogue generation model, wherein the text recognition model is obtained by using the training samples in the first embodiment including reply samples and question and answer samples. It is obtained by training with the training method described when splicing samples.

Corresponding to the training text generation method provided in the first embodiment of the present application, the seventh embodiment of the present application further provides an electronic device for generating training text. As shown in FIG. 6 , the electronic device includes: a processor 901; and a memory 902 for storing a program of a training text generation method. After the device is powered on and runs the program of the data change response method through the processor, executes Follow the steps below:

Obtaining guide text, the guide text is consistent with the semantic attribute of the target text, and the target text is the positive example text identified by the text recognition model;

Inputting the guidance text into a guidance-based text generation model to obtain output text consistent with the semantic attributes of the guidance text;

Training text is determined from the output text.

Corresponding to the method for training a text recognition model provided by the second embodiment of the present application, the eighth embodiment of the present application further provides an electronic device for training a text recognition model. The electronic device comprises: a processor; and a memory for storing the program of the training method of the text recognition model, after the device is powered on and runs the program of the training method of the text recognition model through the processor, the following steps are performed:

Acquiring training samples, the training samples include positive samples and negative samples, and the text corresponding to the positive samples includes: training text generated by the training text generating method described in any one of the first embodiments;

The model to be trained is trained using the training samples to obtain a text recognition model.

Corresponding to the text recognition method provided in the third embodiment of the present application, the ninth embodiment of the present application further provides an electronic device for recognizing text. The electronic device comprises: a processor; and a memory for storing the program of the text recognition method, after the device is powered on and runs the program of the text recognition method through the processor, the following steps are performed:

Get the text to be recognized;

The to-be-recognized text is input into a text recognition model, and a recognition result of the to-be-recognized text is obtained, wherein the text recognition model is obtained by training using the training method described in any one of the second embodiments.

Corresponding to the training text generation method provided by the first embodiment of the present application, the tenth embodiment of the present application provides a computer-readable storage medium storing a program of the training text generation method, the program is run by a processor, and executes the following: Describe the steps:

Obtaining guide text, the guide text is consistent with the semantic attribute of the target text, and the target text is the positive example text identified by the text recognition model;

Inputting the guidance text into a guidance-based text generation model to obtain output text consistent with the semantic attributes of the guidance text;

Training text is determined from the output text.

It should be noted that, for the embodiments of the apparatus, electronic device, and computer-readable storage medium provided in the fourth embodiment to the tenth embodiment of the present application, for a detailed description, reference may be made to the related information of the first embodiment to the third embodiment of the present application. description, which will not be repeated here.

Although the present application is disclosed above with preferred embodiments, it is not intended to limit the present application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the present application The scope of protection shall be subject to the scope defined by the claims of this application.

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

Memory may include non-persistent memory in computer readable media, random access memory (RAM) and/or non-volatile memory in the form of, for example, read only memory (ROM) or flash memory (flash RAM). Memory is an example of a computer-readable medium.

1. Computer readable media includes both persistent and non-permanent, removable and non-removable media. Information storage can be implemented by any method or technology. Information may be computer readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), random access memory (RAM) of other attributes, 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 Disc (DVD) or other optical storage, Magnetic tape cartridges, magnetic tape magnetic disk storage or other magnetic storage media or any other non-transmission medium that can be used to store information that can be accessed by a computing device. Computer-readable media, as defined herein, excludes non-transitory computer-readable media, such as modulated data signals and carrier waves.

2. Those skilled in the art should understand that the embodiments of the present application may be provided as methods, systems or computer program products. Accordingly, the present application may take the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application may take the form of a computer program product embodied 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.

Although the present application is disclosed above with preferred embodiments, it is not intended to limit the present application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present application. Therefore, the present application The scope of protection shall be subject to the scope defined by the claims of this application.

Claims (18)

1. A method for generating a training text is characterized in that the training text is used for training a model to be trained to obtain a text recognition model, and the method comprises the following steps:

acquiring a guide text, wherein the semantic attributes of the guide text are consistent with those of a target text, and the target text is a regular text identified by the text identification model;

inputting the guide text into a guide-based text generation model to obtain an output text consistent with the semantic attribute of the guide text;

and determining a training text according to the output text.

2. The method of claim 1, wherein prior to the entering the guide text into a guide-based text generation model, the method further comprises:

obtaining a question text;

inputting the guide text into a guide-based text generation model to obtain an output text consistent with the semantic attribute of the guide text, wherein the method comprises the following steps:

and inputting the question text and the guide text into a guide-based dialog generation model to obtain an output text which is used for replying the question text and is consistent with the semantic attribute of the guide text.

3. The method of claim 2, wherein the output text comprises a plurality of pieces;

the determining a training text according to the output text includes:

a training text is determined from a plurality of the output texts.

4. The method of claim 3, wherein determining a training text from the plurality of output texts comprises:

determining a training text by a first strategy, the first strategy comprising: selecting a text containing at least one preset keyword from the output texts as a training text, wherein the preset keyword is consistent with the semantic attribute of the target text;

or determining the training text through a second strategy, wherein the second strategy comprises the following steps: and selecting a first text from the output texts or randomly selecting a text as a training text.

5. The method according to claim 4, wherein selecting the first strategy determines the probability of the training text to be a first preset probability, selecting the second strategy determines the probability of the training text to be a second preset probability, the first preset probability is greater than the second preset probability, and the sum of the first preset probability and the second preset probability is 1.

6. The method of claim 4, wherein the guide text comprises at least one guide word, each guide word being consistent with a semantic attribute of the target text;

the preset keywords include: each of the guidance words.

7. The method of claim 6, wherein the preset keywords further comprise: and each first target word is a word which is consistent with the semantic attribute of the target text and is different from each leading word.

8. The method of claim 4, wherein the first policy further comprises: and when the output texts do not contain any preset keyword, selecting a first output text of the output texts to determine a training text.

9. The method according to any one of claims 1 to 8, wherein the semantic attribute of the regular text is a semantically sensitive text, the semantic attribute of the target text is a semantically sensitive text, and the text recognition model is used for recognizing the text generated by the dialogue generation model.

10. A training method of a text recognition model is characterized by comprising the following steps:

obtaining training samples, wherein the training samples comprise positive example samples and negative example samples, and texts corresponding to the positive example samples comprise: a training text generated by the training text generation method of any one of claims 1 to 9;

and training the model to be trained by using the training sample to obtain a text recognition model.

11. The training method of claim 10, further comprising:

acquiring a first text, wherein the first text is a text which is recognized by the text recognition model and has an error, and the actual semantic attribute of the text which is recognized by the error is different from the semantic attribute recognized by the text recognition model for the text which is recognized by the error;

labeling the first text to obtain a first sample;

and performing optimization training on the text recognition model by using the first sample.

12. A training method as defined in claim 11, wherein prior to the optimally training the text recognition model using the first sample, the training method further comprises:

acquiring a second text, wherein the second text comprises a second target word, the semantic attribute of the second text is opposite to that of the first text, and the second target word is a word contained in the first text and consistent with the semantic attribute expressed by the target text;

labeling the second text to obtain a second sample, wherein the labeling information of the second sample is opposite to that of the first sample;

the optimally training the text recognition model using the first sample comprises:

optimally training the text recognition model using the first and second samples.

13. Training method according to any of claims 10 to 12, wherein the training samples comprise reply samples and question-answer splice samples;

the text corresponding to the normal sample in the reply sample comprises: a text generated by the training text generation method of any one of claims 2 to 8;

the text corresponding to the question-answer stitching sample is a stitching text, and the stitching text comprises: and splicing the question text and the reply text corresponding to the question text to form a text.

14. A text recognition method, comprising:

acquiring a text to be identified;

inputting the text to be recognized into a text recognition model to obtain a recognition result of the text to be recognized, wherein the text recognition model is obtained by training through the training method of any one of claims 10 to 13.

15. The text recognition method according to claim 14, wherein the text to be recognized is a text generated by a dialogue generating model;

or, the text to be recognized is a text formed by splicing a question text of a user and a text generated by a dialogue generating model, where the text recognition model is obtained by training through the training method of claim 13.

16. A training text generation apparatus, wherein the training text is used for training a model to be trained to obtain a text recognition model, the apparatus comprising:

the information acquisition unit is used for acquiring a guide text, the semantic attributes of the guide text are consistent with those of a target text, and the target text is a regular text recognized by the text recognition model;

the text generation unit is used for inputting the guide text into a guide-based text generation model to obtain an output text consistent with the semantic attribute of the guide text;

and the text determining unit is used for determining a training text according to the output text.

17. An electronic device, comprising:

a processor; and

a memory for storing a data processing program which, when powered on and executed by said processor, performs the method of any one of claims 1 to 15.

18. A computer-readable storage medium, in which a data processing program is stored, which program, when executed by a processor, performs the method according to any one of claims 1-15.

Images