CN115033682A

CN115033682A - Self-adaptive training method, device, equipment and medium of text generation model

Info

Publication number: CN115033682A
Application number: CN202210679590.XA
Authority: CN
Inventors: 苏雪琦; 王健宗; 程宁
Original assignee: Ping An Technology Shenzhen Co Ltd
Current assignee: Ping An Technology Shenzhen Co Ltd
Priority date: 2022-06-16
Filing date: 2022-06-16
Publication date: 2022-09-09

Abstract

The application is applicable to the technical field of artificial intelligence, and particularly relates to a self-adaptive training method, device, equipment and medium for a text generation model. The method comprises the steps of determining a content item text by positioning the position of a data item text in each text to be processed, covering the content item text in each text to be processed to form a corresponding first training text, wherein the covered content item text is a label of the first training text, pre-training a text generation model by using all the first training texts, inputting the data text in a verification table into the pre-trained text generation model, comparing the similarity of the output of the data text with the content text corresponding to the data text in the verification table, re-training the pre-trained text generation model according to a comparison result, automatically dividing a training set with labels in a positioning mode, automatically re-training the pre-trained text generation model in a verification table mode, and enabling the trained text generation model to be more suitable for a use scene and to have higher accuracy and adaptability.

Description

Self-adaptive training method, device, equipment and medium for text generation model

Technical Field

The application is applicable to the technical field of artificial intelligence, and particularly relates to a self-adaptive training method, device, equipment and medium for a text generation model.

Background

At present, daily financial bulletins are common functions in financial management applications, and are used for summarizing financial hotspots and trend trends of the day and providing financial management references for users. Generally, these briefs are written manually and released to the corresponding platform, but due to the characteristics of the financial stock/fund market, for example, the financial market contains massive data and news, and the financial market is changing all the time, how to extract key information from the massive data and create a highly summarized financial briefs is difficult to store, the financial briefs need to be edited and have very strong analysis capability, and the investigation of the market and the data analysis need to occupy a large amount of manpower and time, the timeliness of the financial briefs is difficult to guarantee for financial decision-making only depending on the manpower, and the operation and the editing have to face pain points of large data volume, difficult analysis, fast change, slow timeliness and the like when writing the briefs. At present, a manual writing mode is still adopted, financial briefs have high requirements on data analysis capability and timeliness, professional teams are often required to collect, analyze and write data into texts, the labor cost is high, the work efficiency is difficult to improve, in order to change the manual situation, a text generation mode using artificial intelligence is provided, but at present, a text generation model widely applied is mainly written according to a mode of nesting important information in a template, although the problems of high labor cost and low work efficiency can be relieved to a certain extent, the format of news generated in a template mode is fixed, for example, news manuscripts are generated according to a preset structure, the model fills in the template according to input materials, and because the content of the financial briefs per day has high requirements on data analysis, valuable information is difficult to generate only by means of information nesting, the text generation model obtained by training by adopting the conventional training method is not suitable for the financial news bulletin generation task, so that the accuracy rate of the trained text generation model is low and the applicability is poor. Therefore, how to reasonably train the text generation model to improve the accuracy and the applicability of the text generation model becomes an urgent problem to be solved.

Disclosure of Invention

In view of this, embodiments of the present application provide a method, an apparatus, a device, and a medium for adaptive training of a text generation model, so as to solve the problem that a text generation model is trained reasonably to improve accuracy and applicability of the text generation model.

In a first aspect, an embodiment of the present application provides an adaptive training method for a text generation model, where the adaptive training method includes:

acquiring N texts to be processed, positioning the position of a data item text in each text to be processed, and determining that a text corresponding to the position of the text to be processed behind the data item text is a content item text, wherein N is an integer greater than zero;

masking the content item text in each text to be processed to form corresponding first training texts, wherein the label of each first training text is the corresponding masked content item text;

training a text generation model by using all the first training texts, and updating parameters in the text generation model until a loss function is converged to obtain a pre-trained text generation model;

inputting the data text in the verification table into the pre-trained text generation model, and comparing the output of the pre-trained text generation model with the similarity of the content text corresponding to the data text in the verification table to obtain a comparison result;

and according to the comparison result, retraining the pre-trained text generation model to obtain a trained text generation model.

In one embodiment, locating the position of the data item text within each text to be processed comprises:

for any text to be processed, converting each word in the text to be processed into a word vector, and arranging all the word vectors into a text sequence according to the text sequence;

and carrying out similarity matching on all the word vectors in the text sequence and a target vector in sequence, and determining the word vector which is ranked most backwards in all the matched word vectors as the position of the data item text, wherein the target vector represents the data item text.

In one embodiment, masking the content item text in each text to be processed to form corresponding first training texts, wherein the labeling of each first training text as the corresponding masked content item text comprises:

for any text to be processed, extracting a content item text in the text to be processed;

and taking the corresponding content item text as the label of the residual text of the text to be processed after extraction, and forming a first training text corresponding to the text to be processed.

In an embodiment, training a text generation model by using all first training texts, and updating parameters in the text generation model until a loss function converges, to obtain a pre-trained text generation model includes:

labels of all first training texts are removed, the first training texts without labels are used for carrying out unsupervised training on the text generation model, parameters in the text generation model are updated until a first loss function is converged, and the unsupervised trained text generation model is obtained;

and retraining the unsupervised trained text generation model by using all the first training texts, and updating parameters in the unsupervised trained text generation model until a second loss function is converged to obtain a pre-trained text generation model.

In one embodiment, the unsupervised training of the text generation model using the unlabeled first training text and the updating of the parameters in the text generation model until the first loss function converges, and obtaining the unsupervised and trained text generation model includes:

coding and decoding a first unlabeled training text by using a text generation model, and randomly selecting K results from the decoding results as initial clustering centers, wherein K is an integer greater than zero;

calculating to obtain a first loss value according to the distance between other results and the initial clustering center and by combining a first loss function;

and updating the parameters in the text generation model by adopting a reverse gradient updating algorithm according to the first loss value to obtain a text generation model with updated parameters, and returning to the step of encoding and decoding the unmarked first training text until the first loss value obtained by calculation is minimum to obtain the unsupervised trained text generation model.

In an embodiment, inputting a data text in a validation table into the pre-trained text generation model, and performing similarity comparison between an output of the pre-trained text generation model and a content text corresponding to the data text in the validation table, where the comparison result includes:

inputting the data text in the verification table into the pre-trained text generation model, and outputting a prediction text;

performing vector coding on the content text corresponding to the data text in the verification table to obtain a target vector, and performing vector coding on the predicted text to obtain a predicted vector;

and calculating cosine similarity of the target vector and the prediction vector, and determining a corresponding comparison result according to the cosine similarity.

In an embodiment, determining the corresponding comparison result according to the cosine similarity includes:

when the cosine similarity meets a preset condition, determining that a comparison result is positive, and when the cosine similarity does not meet the preset condition, determining that the comparison result is negative;

according to the comparison result, retraining the pre-trained text generation model to obtain a trained text generation model, comprising:

counting the positive occupation ratio of the comparison result in the comparison result, and if the occupation ratio is smaller than a preset threshold value, taking the content text in the verification table as the label of the corresponding data text to form a second training text;

and retraining the pre-trained text generation model by using the second training text to obtain a trained text generation model.

In a second aspect, an embodiment of the present application provides an adaptive training apparatus for a text generation model, where the adaptive training apparatus includes:

the text preprocessing module is used for acquiring N texts to be processed, positioning the position of a data item text in each text to be processed, and determining that a text corresponding to the position in the text to be processed behind the data item text is a content item text, wherein N is an integer greater than zero;

the training text generation module is used for covering the content item text in each text to be processed to form corresponding first training texts, and the label of each first training text is the content item text which is correspondingly covered;

the model pre-training module is used for training the text generation model by using all the first training texts, updating parameters in the text generation model until a loss function is converged, and obtaining a pre-trained text generation model;

the similarity comparison module is used for inputting the data texts in the verification table into the pre-trained text generation model, and comparing the output of the pre-trained text generation model with the similarity of the content texts corresponding to the data texts in the verification table to obtain a comparison result;

and the retraining module is used for retraining the pre-trained text generation model according to the comparison result to obtain a trained text generation model.

In one embodiment, the text pre-processing module comprises:

the text sequence generating unit is used for converting each word in the text to be processed into a word vector aiming at any text to be processed and arranging all the word vectors into a text sequence according to the text sequence;

and the position determining unit is used for carrying out similarity matching on all the word vectors in the text sequence and a target vector in sequence, determining the word vector which is ranked most backwards in all the matched word vectors as the position of the data item text, and the target vector represents the data item text.

In one embodiment, the training text generation module comprises:

the text extraction unit is used for extracting the content item text in any text to be processed;

and the first text generation unit is used for taking the corresponding content item text as the label of the residual text after the extraction of the text to be processed to form a first training text corresponding to the text to be processed.

In one embodiment, the model pre-training module comprises:

the unsupervised training unit is used for eliminating labels of all the first training texts, performing unsupervised training on the text generation model by using the unlabeled first training texts, and updating parameters in the text generation model until the first loss function is converged to obtain the unsupervised trained text generation model;

and the supervision training unit is used for retraining the unsupervised trained text generation model by using all the first training texts, and updating parameters in the unsupervised trained text generation model until a second loss function is converged to obtain a pre-trained text generation model.

In one embodiment, the unsupervised training unit comprises:

the clustering center selecting subunit is used for encoding and decoding the unmarked first training text by using a text generation model, and randomly selecting K results from the decoding results as initial clustering centers, wherein K is an integer greater than zero;

the loss calculating subunit is used for calculating a first loss value according to the distance between the other results and the initial clustering center and by combining a first loss function;

and the updating subunit is used for updating the parameters in the text generation model by adopting a reverse gradient updating algorithm according to the first loss value to obtain the text generation model with updated parameters, and returning to the step of encoding and decoding the unmarked first training text until the first loss value obtained by calculation is minimum to obtain the unsupervised trained text generation model.

In one embodiment, the similarity comparison module comprises:

the predicted text unit is used for inputting the data text in the verification table into the pre-trained text generation model and outputting a predicted text;

the vector coding unit is used for carrying out vector coding on the content text corresponding to the data text in the verification table to obtain a target vector, and carrying out vector coding on the predicted text to obtain a predicted vector;

and the comparison result determining unit is used for calculating the cosine similarity between the target vector and the prediction vector and determining a corresponding comparison result according to the cosine similarity.

In an embodiment, the comparison result determination unit is specifically configured to

the retraining module comprises:

the second text generation unit is used for counting the proportion that the comparison result in the comparison result is positive, and if the proportion is smaller than a preset threshold value, the content text in the verification table is used as a label of the corresponding data text to form a second training text;

and the retraining unit is used for retraining the pre-trained text generation model by using the second training text to obtain a trained text generation model.

In a third aspect, an embodiment of the present application provides a computer device, which includes a processor, a memory, and a computer program stored in the memory and executable on the processor, and when the processor executes the computer program, the adaptive training method according to the first aspect is implemented.

In a fourth aspect, embodiments of the present application provide a computer-readable storage medium, which stores a computer program, and when the computer program is executed by a processor, the computer program implements the adaptive training method according to the first aspect.

Compared with the prior art, the embodiment of the application has the advantages that: the method comprises the steps of determining a text corresponding to a processed text and positioned behind a data item text as a content item text by positioning the position of the data item text in each processed text, covering the content item text in each processed text to form a corresponding first training text, marking each first training text as the corresponding covered content item text, training a text generation model by using all the first training texts, updating parameters in the text generation model until a loss function is converged to obtain a pre-trained text generation model, inputting the data text in a verification table into the pre-trained text generation model, comparing the output of the pre-trained text generation model with the content text corresponding to the data text in the verification table to obtain a comparison result, and retraining the pre-trained text generation model according to the comparison result, the trained text generation model is obtained, the text is automatically divided into a training set with labels in a positioning mode, the training set is used for pre-training the text generation model, meanwhile, the pre-trained text generation model is automatically retrained in a verification table mode, and the trained text generation model is more suitable for a use scene, so that the method has higher accuracy and adaptability.

Drawings

In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required to be used in the embodiments or the prior art description will be briefly described below, and it is obvious that the drawings in the following description are only some embodiments of the present application, and for those skilled in the art, other drawings may be obtained according to these drawings without inventive labor.

Fig. 1 is a schematic diagram of an application environment of an adaptive training method for a text-to-model according to an embodiment of the present application;

FIG. 2 is a schematic flowchart illustrating an adaptive training method for a text-generating model according to a second embodiment of the present application;

fig. 3 is a schematic flowchart of a method for adaptively training a text generation model according to a third embodiment of the present application;

fig. 4 is a schematic structural diagram of an adaptive training apparatus for generating a text model according to a fourth embodiment of the present application;

fig. 5 is a schematic structural diagram of a computer device according to a fifth embodiment of the present application.

Detailed Description

In the following description, for purposes of explanation and not limitation, specific details are set forth, such as particular system structures, techniques, etc. in order to provide a thorough understanding of the embodiments of the present application. It will be apparent, however, to one skilled in the art that the present application may be practiced in other embodiments that depart from these specific details. In other instances, detailed descriptions of well-known systems, devices, circuits, and methods are omitted so as not to obscure the description of the present application with unnecessary detail.

It will be understood that the terms "comprises" and/or "comprising," when used in this specification and the appended claims, specify the presence of stated features, integers, steps, operations, elements, and/or components, but do not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components, and/or groups thereof.

It should also be understood that the term "and/or" as used in this specification and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

As used in this specification and the appended claims, the term "if" may be interpreted contextually as "when", "upon" or "in response to" determining "or" in response to detecting ". Similarly, the phrase "if it is determined" or "if a [ described condition or event ] is detected" may be interpreted contextually to mean "upon determining" or "in response to determining" or "upon detecting [ described condition or event ]" or "in response to detecting [ described condition or event ]".

Furthermore, in the description of the present application and the appended claims, the terms "first," "second," "third," and the like are used for distinguishing between descriptions and not necessarily for describing or implying relative importance.

Reference throughout this specification to "one embodiment" or "some embodiments," or the like, means that a particular feature, structure, or characteristic described in connection with the embodiment is included in one or more embodiments of the present application. Thus, appearances of the phrases "in one embodiment," "in some embodiments," "in other embodiments," or the like, in various places throughout this specification are not necessarily all referring to the same embodiment, but rather "one or more but not all embodiments" unless specifically stated otherwise. The terms "comprising," "including," "having," and variations thereof mean "including, but not limited to," unless expressly specified otherwise.

The embodiment of the application can acquire and process related data based on an artificial intelligence technology. Among them, Artificial Intelligence (AI) is a theory, method, technique and application system that simulates, extends and expands human Intelligence using a digital computer or a machine controlled by a digital computer, senses the environment, acquires knowledge and uses the knowledge to obtain the best result.

The artificial intelligence infrastructure generally includes technologies such as sensors, dedicated artificial intelligence chips, cloud computing, distributed storage, big data processing technologies, operation/interaction systems, mechatronics, and the like. The artificial intelligence software technology mainly comprises a computer vision technology, a robot technology, a biological recognition technology, a voice processing technology, a natural language processing technology, machine learning/deep learning and the like.

It should be understood that, the sequence numbers of the steps in the following embodiments do not mean the execution sequence, and the execution sequence of each process should be determined by the function and the inherent logic of the process, and should not constitute any limitation to the implementation process of the embodiments of the present application.

In order to explain the technical means of the present application, the following description will be given by way of specific examples.

The adaptive training method for the text generation model provided by the embodiment of the present application can be applied to the application environment shown in fig. 1, in which a client communicates with a server. The client includes, but is not limited to, a palm top computer, a desktop computer, a notebook computer, an ultra-mobile personal computer (UMPC), a netbook, a cloud computing device, a Personal Digital Assistant (PDA), and other computing devices. The server can be implemented by an independent server or a server cluster composed of a plurality of servers.

Referring to fig. 2, which is a schematic flow chart of an adaptive training method for a text generating model according to a second embodiment of the present disclosure, the adaptive training method is applied to the server in fig. 1, a computer device corresponding to the server is connected to a corresponding database to obtain corresponding text data, verification table data, and the like in the database, and a corresponding model to be trained, that is, a text generating model, is stored in the server, where the text generating model may be applied to an application scenario such as a text abstract. As shown in fig. 2, the method for adaptively training a text generation model may include the following steps:

step S201, acquiring N texts to be processed, positioning the position of the data item text in each text to be processed, and determining that a text corresponding to the position after the data item text in the text to be processed is a content item text.

In the application, N is an integer greater than zero, the text to be processed is stored in the database, the text to be processed can be obtained from a corresponding website, an application program, a client and the like by means of crawling and the like, and the obtained text is set according to an actual use scene. For example, in the application scenario of financial text summarization, data is acquired from financial websites, various types of stock exchange software, and the like.

The text to be processed comprises text information and digital information, the text information is a content item text in the text to be processed and is represented in a text form, the digital information is a data item text in the text to be processed and is represented in a digital form, and the text to be processed comprises financial text information and financial data information corresponding to the application scene of the financial text abstract.

Positioning may refer to determining a position of the target information in the text, for example, the text is "XXXYXXX", where "Y" is the target information, and the position of the "Y" is the last three bits from the beginning of the text. The positioning can be performed by means of feature extraction, feature comparison and the like, so as to determine the position corresponding to the target information. For example, the position of the target feature may be determined by extracting the feature of each text word, comparing the feature with the feature corresponding to the target information, and determining the target feature.

Optionally, the positioning the position of the data item text in each text to be processed includes:

and carrying out similarity matching on all the word vectors in the text sequence and the target vector in sequence, determining the word vector which is ranked most back in all the matched word vectors as the position of the data item text, and representing the data item text by the target vector.

Specifically, for each word, a learning model such as one hot code, word2vec, skip-gram, and Bag Of Words (BOW), or a pre-training model (BERT), Bi-directional Short Term Memory (Bi-LSTM), etc. may be used to perform feature coding, so as to obtain a word vector for each word. And sequencing the word vectors according to the sequence of the words in the text to obtain a text sequence.

And sequentially comparing each word vector in the text sequence with the target vector according to the sequence to determine the word vector matched with the target similarity, wherein the word vector with the most rear sequence in all the matched word vectors is used as the position of the data item text.

In one embodiment, analyzing the alphanumeric features to obtain a target vector characterizing the numeric features, and accordingly, locating the position of the text of the data item within each text to be processed comprises:

performing word segmentation on any text to be processed to obtain word segmentation results;

and extracting a word vector of each word in the word segmentation result, and comparing the word vector with the target vector, so as to determine that the word corresponding to the word vector matched with the target vector is a data item text, wherein the last word of the word is the position of the data item text.

The word segmentation can be performed by adopting a maximum matching algorithm, a conditional random field model and the like, and the encoding of the word vector can adopt any feature encoding.

Step S202, covering the content item text in each text to be processed to form corresponding first training texts, wherein the label of each first training text is the corresponding covered content item text.

In this application, the content item text and the data item text may be determined in step S201, for the text to be processed, masking may refer to operations such as deleting and blank covering, so as to obtain the remaining text, and the masked text may be used as a label of the remaining text, so as to form a training text.

Optionally, masking the content item text in each text to be processed to form a corresponding first training text, where the labeling of each first training text as the correspondingly masked content item text includes:

extracting a content item text in the text to be processed aiming at any text to be processed;

The content item text in the text to be processed is extracted and is different from masking operations such as deletion, blank coverage and the like, the content item text can be obtained through extraction and is further used for the formation of a subsequent training text, and the operation of obtaining the content item text before other masking operations are adopted is not needed.

Step S203, training the text generation model by using all the first training texts, and updating parameters in the text generation model until the loss function is converged to obtain a pre-trained text generation model.

In the application, because the first training text is automatically generated in a self-adaptive mode, in order to ensure the accuracy of the text after the model is generated, the first training text is used for pre-training the first training text, and then the pre-trained text is retrained, so that the training text is generated in a self-adaptive mode, the training text construction efficiency is improved, and meanwhile, the accuracy of the text generation model training can still be ensured.

The training of the text generation model can adopt a plurality of different training modes, such as unsupervised training and supervised training, the unsupervised training can be to train an unlabeled training set by adopting a corresponding algorithm so as to find the potential structure of the group of data, the unsupervised training can be roughly divided into two categories of clustering and dimensionality reduction, the supervised training can be to know the relationship between input and output results according to the existing data set and then train to obtain an optimal model according to the known relationship, in the supervised learning, the training data should have both features and labels, then through the training, the machine can find the relationship between the features and the labels by itself, and then the labels can be judged when facing the unlabeled data.

Unsupervised training can comprise k-means Clustering, Balanced Iterative Reduction and Clustering (BIRCH), Gaussian mixture models and the like, has better training set adaptability, and can comprise k-nearest neighbor algorithm, decision tree algorithm, naive Bayes algorithm, logistic regression algorithm and the like, and has better adaptability.

Optionally, the training of the text generation model by using all the first training texts and the updating of the parameters in the text generation model are performed until the loss function converges, and obtaining the pre-trained text generation model includes:

labels of all first training texts are removed, the first training texts without labels are used for carrying out unsupervised training on the text generation model, parameters in the text generation model are updated until the first loss function is converged, and the unsupervised trained text generation model is obtained;

and retraining the unsupervised trained text generation model by using all the first training texts, and updating parameters in the unsupervised trained text generation model until the second loss function is converged to obtain the pre-trained text generation model.

The method has the advantages that the text generation model is trained in a mode of combining unsupervised training and supervised training, adaptability of the model can be guaranteed, and fitting performance of the model can be improved. Firstly, labels of a first training text are removed to obtain an unmarked first training text, the unmarked first training text is used for carrying out unsupervised training, and secondly, the first training text is used for retraining a text generation model which is trained unsupervised. Because the algorithms adopted in the two training processes are different, the adopted loss functions are different, and the condition for converging the loss functions needs to be achieved in the two training processes, for example, the condition for converging is that the loss value is minimum.

Optionally, the method for performing unsupervised training on the text generation model by using the unmarked first training text, and updating parameters in the text generation model until the first loss function converges to obtain the unsupervised and trained text generation model includes:

coding and decoding a first unlabelled training text by using a text generation model, and randomly selecting K results from the decoding results as initial clustering centers, wherein K is an integer greater than zero;

and updating the parameters in the text generation model by adopting a reverse gradient updating algorithm according to the first loss value to obtain a text generation model after the parameters are updated, and returning to the step of encoding and decoding the unmarked first training text until the first loss value obtained by calculation is minimum to obtain the unsupervised and trained text generation model. When all the first training texts are used for retraining the unsupervised trained text generation model, parameters in encoding and decoding can be fixed, so that the effect of rapid training is achieved.

The unsupervised training is carried out on the text generation model by adopting a k-means clustering algorithm, the model parameters are updated by adopting a reverse gradient updating algorithm, the model parameters comprise parameters during encoding and decoding, and the unsupervised trained text generation model has a higher encoding and decoding function. For the parameters of other layers in the text generation model, the training can be performed based on unsupervised training, and of course, the most important is to train the parameters of other layers through subsequent supervised training.

And step S204, inputting the data text in the verification table into the pre-trained text generation model, and comparing the output of the pre-trained text generation model with the similarity of the content text corresponding to the data text in the verification table to obtain a comparison result.

In the application, a verification text set for verifying a text generation model is preset in a verification table, and a data text and a content text in the verification table correspond to one verification text. The verification table can be set differently according to different application scenes, and the verification text in the verification table is the text stripped of the actual scene, so that the verification table can be set indiscriminately for different scenes.

Inputting the data text into a pre-trained text generation model, namely outputting a corresponding predicted text, comparing the similarity of the predicted text with a corresponding content text, wherein the comparison result is whether the predicted text is similar to the content text, if the predicted text is not similar to the content text, the pre-trained text generation model is inaccurate and needs to be improved, and if the predicted text is similar to the content text, the pre-trained text generation model has certain generation capacity.

And S205, retraining the pre-trained text generation model according to the comparison result to obtain the trained text generation model.

In the application, if the prediction text is not similar to the content text, the pre-trained text generation model needs to be retrained, the retraining can adopt a new training text, the new training text can be composed of a data text and a content text in a verification table, the number of the texts in the verification table is smaller than that of the texts in the training text, the effect of light training or fine adjustment training is achieved, and the trained text generation model is obtained.

In the embodiment of the application, the position of the data item text in each text to be processed is positioned, the text corresponding to the position behind the data item text in the text to be processed is determined to be the content item text, the content item text in each text to be processed is covered to form a corresponding first training text, each first training text is marked as the corresponding covered content item text, all the first training texts are used for training the text generation model, parameters in the text generation model are updated until a loss function is converged to obtain a pre-trained text generation model, the data text in the verification table is input into the pre-trained text generation model, the output of the pre-trained text generation model is compared with the similarity of the content text corresponding to the data text in the verification table to obtain a comparison result, and the pre-trained text generation model is retrained according to the comparison result, the trained text generation model is obtained, the text is automatically divided into a training set with labels in a positioning mode, the training set is used for pre-training the text generation model, meanwhile, the pre-trained text generation model is automatically retrained in a verification table mode, and the trained text generation model is more suitable for a use scene, so that the method has higher accuracy and adaptability.

Referring to fig. 3, which is a schematic flow chart of an adaptive training method for a text-to-model provided in the third embodiment of the present application, as shown in fig. 3, the adaptive training method may include the following steps:

step S301, obtaining N texts to be processed, positioning the position of the data item text in each text to be processed, and determining the text corresponding to the position behind the data item text in the texts to be processed as the content item text.

Step S302, the content item text in each text to be processed is covered to form a corresponding first training text, and the label of each first training text is the content item text which is correspondingly covered.

Step S303, training the text generation model by using all the first training texts, and updating parameters in the text generation model until the loss function is converged to obtain a pre-trained text generation model.

And S304, inputting the data text in the verification table into a pre-trained text generation model, and outputting a prediction text.

The contents of the steps S301 to S304 are the same as the contents of the steps S201 to S204 in the second embodiment, and the descriptions of the steps S201 to S204 may be specifically referred to, and are not repeated herein.

Step S305, carrying out vector coding on the content text corresponding to the data text in the verification table to obtain a target vector, and carrying out vector coding on the prediction text to obtain a prediction vector.

In the application, the vector coding of the content text can adopt sentence vector coding, and can also adopt character vector or word vector coding, and then the coding is fused to obtain the vector of the content text. Similarly, the same encoding method as that for the content text is also used for the predicted text to obtain the corresponding vector.

Step S306, calculating cosine similarity of the target vector and the prediction vector, and determining a corresponding comparison result according to the cosine similarity.

In the application, the cosine similarity is adopted to represent the similarity between two vectors, so that the similarity between the prediction vector and the target vector is determined, and the comparison result is determined. Comparing the cosine similarity value with the similarity threshold may determine whether the target vector is similar to the predicted vector, e.g., when the cosine similarity value is greater than the similarity threshold, determining that the target vector is similar to the predicted vector.

And step S307, retraining the pre-trained text generation model according to the comparison result to obtain the trained text generation model.

The content of the step S307 is the same as the content of the step S205 in the second embodiment, and reference may be specifically made to the description of the step S205, which is not repeated herein.

Optionally, determining, according to the cosine similarity, a corresponding comparison result includes:

when the cosine similarity meets a preset condition, determining that the comparison result is positive, and when the cosine similarity does not meet the preset condition, determining that the comparison result is negative;

according to the comparison result, retraining the pre-trained text generation model to obtain the trained text generation model comprises the following steps:

counting the comparison result in the comparison result as a positive ratio, and if the ratio is smaller than a preset threshold value, taking the content text in the verification table as the label of the corresponding data text to form a second training text;

and retraining the pre-trained text generation model by using the second training text to obtain the trained text generation model.

And quantifying the comparison result, counting the number of positive comparison results and the number of negative comparison results, determining the positive proportion, and indicating that the pre-trained text generation model needs to be finely adjusted when the positive proportion is smaller than a preset threshold value or the negative proportion is larger than a certain value. At the moment, a second training text is generated according to the verification table, and the pre-trained text generation model is retrained.

The embodiment of the application determines that a text corresponding to the position behind the data item text in the text to be processed is a content item text by positioning the position of the data item text in each text to be processed, covers the content item text in each text to be processed to form a corresponding first training text, marks each first training text as the corresponding covered content item text, trains a text generation model by using all the first training texts, updates parameters in the text generation model until a loss function is converged to obtain a pre-trained text generation model, inputs the data text in a verification table into the pre-trained text generation model, outputs a prediction text, performs vector coding on the content text corresponding to the data text in the verification table to obtain a target vector, performs vector coding on the prediction text to obtain a prediction vector, and calculates the cosine similarity between the target vector and the prediction vector, and determining a corresponding comparison result according to the cosine similarity, and retraining the pre-trained text generation model according to the comparison result to obtain the trained text generation model, so that the accuracy of the trained text generation model is improved.

Fig. 4 shows a block diagram of a structure of an adaptive training device for a text generation model according to a fourth embodiment of the present application, where the adaptive training device is applied to the server in fig. 1, a computer device corresponding to the server is connected to a corresponding database to obtain corresponding text data, validation table data, and the like in the database, and a corresponding model to be trained, that is, a text generation model, is stored in the server, and the text generation model can be applied to an application scenario such as a text abstract. For ease of illustration, only portions relevant to the embodiments of the present application are shown.

Referring to fig. 4, the adaptive training apparatus includes:

the text preprocessing module 41 is configured to obtain N texts to be processed, locate a position of a data item text in each text to be processed, and determine that a text corresponding to the position of the data item text in the text to be processed is a content item text, where N is an integer greater than zero;

a training text generation module 42, configured to mask the content item text in each to-be-processed text to form a corresponding first training text, where a label of each first training text is the correspondingly masked content item text;

the model pre-training module 43 is configured to train the text generation model using all the first training texts, and update parameters in the text generation model until the loss function converges to obtain a pre-trained text generation model;

the similarity comparison module 44 is configured to input the data text in the verification table into the pre-trained text generation model, and compare the output of the pre-trained text generation model with the similarity of the content text corresponding to the data text in the verification table to obtain a comparison result;

and the retraining module 45 is configured to retrain the pre-trained text generation model according to the comparison result to obtain a trained text generation model.

Optionally, the text preprocessing module 41 includes:

and the position determining unit is used for carrying out similarity matching on all the word vectors in the text sequence and the target vector in sequence, determining the word vector which is ranked most backwards in all the matched word vectors as the position of the data item text, and representing the data item text by the target vector.

Optionally, the training text generating module 42 includes:

the text extraction unit is used for extracting content item texts in the texts to be processed aiming at any texts to be processed;

Optionally, the model pre-training module 43 includes:

and the supervision training unit is used for retraining the unsupervised trained text generation model by using all the first training texts, updating parameters in the unsupervised trained text generation model until the second loss function is converged, and obtaining the pre-trained text generation model.

Optionally, the unsupervised training unit includes:

the loss calculating subunit is used for calculating a first loss value according to the distance between the other results and the initial clustering center and by combining the first loss function;

and the updating subunit is used for updating the parameters in the text generation model by adopting a reverse gradient updating algorithm according to the first loss value to obtain the text generation model after the parameters are updated, and returning to the step of executing the steps of encoding and decoding the unmarked first training text until the first loss value obtained by calculation is minimum to obtain the unsupervised trained text generation model.

Optionally, the similarity comparing module 44 includes:

the predicted text unit is used for inputting the data text in the verification table into a pre-trained text generation model and outputting a predicted text;

and the comparison result determining unit is used for calculating the cosine similarity of the target vector and the predicted vector and determining a corresponding comparison result according to the cosine similarity.

Optionally, the result determining unit is specifically configured to:

the retraining module 45 comprises:

the second text generation unit is used for counting the proportion that the comparison result in the comparison result is positive, and if the proportion is smaller than a preset threshold value, the content text in the verification table is used as the label of the corresponding data text to form a second training text;

and the retraining unit is used for retraining the pre-trained text generation model by using the second training text to obtain the trained text generation model.

It should be noted that, because the above-mentioned information interaction between the modules, the execution process, and other contents are based on the same concept, specific functions, and technical effects brought by the method embodiment of the present application may be specifically referred to a part of the method embodiment, and are not described herein again.

Fig. 5 is a schematic structural diagram of a computer device according to a fifth embodiment of the present application. As shown in fig. 5, the computer apparatus of this embodiment includes: at least one processor (only one shown in fig. 5), a memory, and a computer program stored in the memory and executable on the at least one processor, the processor when executing the computer program implementing the steps in any of the various text generation model adaptive training method embodiments described above.

The computer device may include, but is not limited to, a processor, a memory. It will be appreciated by those skilled in the art that fig. 5 is merely an example of a computer device and is not intended to be limiting, and that a computer device may include more or fewer components than those shown, or some components may be combined, or different components may be included, such as a network interface, a display screen, and input devices, etc.

The Processor may be a CPU, or other general purpose Processor, a Digital Signal Processor (DSP), an Application Specific Integrated Circuit (ASIC), a Field-Programmable Gate Array (FPGA) or other Programmable logic device, discrete Gate or transistor logic, discrete hardware components, or the like. A general purpose processor may be a microprocessor or the processor may be any conventional processor or the like.

The memory includes readable storage media, internal memory, etc., wherein the internal memory may be the internal memory of the computer device, and the internal memory provides an environment for the operating system and the execution of the computer-readable instructions in the readable storage media. The readable storage medium may be a hard disk of the computer device, and in other embodiments may also be an external storage device of the computer device, such as a plug-in hard disk, a Smart Media Card (SMC), a Secure Digital (SD) Card, a Flash memory Card (Flash Card), and the like, provided on the computer device. Further, the memory may also include both internal and external storage units of the computer device. The memory is used for storing an operating system, application programs, a BootLoader (BootLoader), data, and other programs, such as program codes of a computer program, and the like. The memory may also be used to temporarily store data that has been output or is to be output.

It will be apparent to those skilled in the art that, for convenience and brevity of description, only the above-mentioned division of the functional units and modules is illustrated, and in practical applications, the above-mentioned function distribution may be performed by different functional units and modules according to needs, that is, the internal structure of the apparatus is divided into different functional units or modules, so as to perform all or part of the functions described above. Each functional unit and module in the embodiments may be integrated in one processing unit, or each unit may exist alone physically, or two or more units are integrated in one unit, and the integrated unit may be implemented in a form of hardware, or in a form of software functional unit. In addition, specific names of the functional units and modules are only for convenience of distinguishing from each other, and are not used for limiting the protection scope of the present application. The specific working processes of the units and modules in the above-mentioned apparatus may refer to the corresponding processes in the foregoing method embodiments, and are not described herein again. The integrated unit, if implemented in the form of a software functional unit and sold or used as a stand-alone product, may be stored in a computer readable storage medium. Based on such understanding, all or part of the processes in the method of the embodiments described above may be implemented by instructing relevant hardware by a computer program, and the computer program may be stored in a computer readable storage medium, and when executed by a processor, the computer program may implement the steps of the embodiments of the method described above. Wherein the computer program comprises computer program code, which may be in the form of source code, object code, an executable file or some intermediate form, etc. The computer readable medium may include at least: any entity or device capable of carrying computer program code, recording medium, computer Memory, Read-Only Memory (ROM), Random Access Memory (RAM), electrical carrier signals, telecommunications signals, and software distribution media. Such as a usb-disk, a removable hard disk, a magnetic or optical disk, etc. In certain jurisdictions, computer-readable media may not be an electrical carrier signal or a telecommunications signal in accordance with legislative and patent practice.

When the computer program product runs on a computer device, the computer device is enabled to implement the steps in the method embodiments.

In the above embodiments, the descriptions of the respective embodiments have respective emphasis, and reference may be made to the related descriptions of other embodiments for parts that are not described or illustrated in a certain embodiment.

Those of ordinary skill in the art will appreciate that the various illustrative elements and algorithm steps described in connection with the embodiments disclosed herein may be implemented as electronic hardware or combinations of computer software and electronic hardware. Whether such functionality is implemented as hardware or software depends upon the particular application and design constraints imposed on the implementation. Skilled artisans may implement the described functionality in varying ways for each particular application, but such implementation decisions should not be interpreted as causing a departure from the scope of the present application.

In the embodiments provided in the present application, it should be understood that the disclosed apparatus/computer device and method may be implemented in other ways. For example, the above-described apparatus/computer device embodiments are merely illustrative, and for example, a module or a unit may be divided into only one logical function, and may be implemented in other ways, for example, multiple units or components may be combined or integrated into another system, or some features may be omitted, or not implemented. In addition, the shown or discussed mutual coupling or direct coupling or communication connection may be an indirect coupling or communication connection through some interfaces, devices or units, and may be in an electrical, mechanical or other form.

Units described as separate parts may or may not be physically separate, and parts displayed as units may or may not be physical units, may be located in one place, or may be distributed on a plurality of network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the solution of the embodiment.

The above embodiments are only used to illustrate the technical solutions of the present application, and not to limit the same; although the present application has been described in detail with reference to the foregoing embodiments, it should be understood by those of ordinary skill in the art that: the technical solutions described in the foregoing embodiments may still be modified, or some technical features may be equivalently replaced; such modifications and substitutions do not substantially depart from the spirit and scope of the embodiments of the present application and are intended to be included within the scope of the present application.

Claims

1. An adaptive training method for a text generation model, the adaptive training method comprising:

covering the content item text in each text to be processed to form corresponding first training texts, wherein the label of each first training text is the corresponding covered content item text;

2. The adaptive training method of claim 1, wherein locating the position of the data item text within each text to be processed comprises:

and carrying out similarity matching on all the word vectors in the text sequence and a target vector in sequence, and determining the word vector with the most back order in all the matched word vectors as the position of the data item text, wherein the target vector represents the data item text.

3. The adaptive training method of claim 1, wherein masking the content item text in each text to be processed to form corresponding first training texts, wherein labeling each first training text as corresponding masked content item text comprises:

extracting a content item text in any text to be processed;

and taking the corresponding content item text as a label of the residual text of the text to be processed after extraction, and forming a first training text corresponding to the text to be processed.

4. The adaptive training method according to claim 1, wherein training a text generation model using all first training texts and updating parameters in the text generation model until a loss function converges to obtain a pre-trained text generation model comprises:

5. The adaptive training method of claim 4, wherein the unsupervised training of the text generation model using the unlabeled first training text and the updating of the parameters in the text generation model until the first loss function converges, and obtaining the unsupervised and trained text generation model comprises:

6. The adaptive training method according to any one of claims 1 to 5, wherein inputting the data text in the validation table into the pre-trained text generation model, and comparing the output of the pre-trained text generation model with the similarity of the content text corresponding to the data text in the validation table, and obtaining the comparison result comprises:

carrying out vector coding on the content text corresponding to the data text in the verification table to obtain a target vector, and carrying out vector coding on the prediction text to obtain a prediction vector;

7. The adaptive training method of claim 6, wherein determining a corresponding comparison result according to the cosine similarity comprises:

counting the proportion of the comparison result which is positive in the comparison result, and if the proportion is smaller than a preset threshold value, taking the content text in the verification table as the label of the corresponding data text to form a second training text;

8. An adaptive training device for a text generation model, the adaptive training device comprising:

the similarity comparison module is used for inputting the data texts in the verification table into the pre-trained text generation model and comparing the output of the pre-trained text generation model with the similarity of the content texts corresponding to the data texts in the verification table to obtain a comparison result;

9. A computer device, characterized in that the computer device comprises a processor, a memory and a computer program stored in the memory and executable on the processor, the processor implementing the adaptive training method according to any of claims 1 to 7 when executing the computer program.

10. A computer-readable storage medium, in which a computer program is stored which, when being executed by a processor, carries out the adaptive training method according to any one of claims 1 to 7.