WO2023115914A1

WO2023115914A1 - Method and device for generating document having consistent writing style, and storage medium

Info

Publication number: WO2023115914A1
Application number: PCT/CN2022/105318
Authority: WO
Inventors: 罗清彩; 孙善宝; 蒋梦梦; 张晖; 解萌; 于晓艳; 张鑫
Original assignee: 山东浪潮科学研究院有限公司
Priority date: 2021-12-20
Filing date: 2022-07-13
Publication date: 2023-06-29
Also published as: CN114239600A

Abstract

Disclosed are a method and device for generating a document having a consistent writing style, and a storage medium. The method comprises: obtaining a main author document by means of a collaborative writing platform, and inputting the main author document into a document encoder to generate a context vector; obtaining a secondary author document by means of the collaborative writing platform, and inputting the secondary author document into the document encoder to generate a document statement vector set; inputting the context vector and the document statement vector set into a collaborative writing model to generate cooperative documents having the same style as the main author; and generating a plurality of cooperative documents, evaluating according to an author writing style discriminator and the author content rationality discriminator, and selecting a cooperative document having the highest similarity as a final document.

Description

A method, device, and storage medium for generating consistent writing style documents

technical field

The present application relates to the field of natural language processing, and in particular to a method, device, and storage medium for generating consistent writing style documents.

Background technique

In recent years, reinforcement learning technology has received widespread attention, especially in combination with deep learning, which has brought great progress to the field of artificial intelligence. Reinforcement learning is different from traditional supervised learning, mainly in the reinforcement signal. In reinforcement learning, the reinforcement signal provided by the environment is an evaluation of the quality of the generated action.

With the development of the Internet, more and more attention has been paid to digital content production, and collaborative writing has gradually become an important means of content production. Collaborative writing or cooperative writing refers to a writing plan that is completed by multiple people instead of individuals alone. It is mostly in the form of crowdsourcing and distribution to achieve efficient content production and collaboration.

Due to the different writing styles of the authors participating in the collaborative writing, in terms of format or content, the final document works often have inconsistent content and style, which affects the reading quality of readers.

Based on this, there is a need for a solution that can make the document content and style of collaborative writing consistent, so as to better improve the user's reading experience.

Contents of the invention

This application provides a method, device, and storage medium for generating consistent writing style documents, which solves the technical problem of inconsistency in writing style and writing content when multiple people write collaboratively.

A method of generating a consistent writing style document comprising:

Obtaining the main author's document through the collaborative writing platform, inputting the main author's document into the document encoder, and generating a context vector;

Obtain the sub-author document through the collaborative writing platform, input the sub-author document into the document encoder, and generate a set of document sentence vectors;

Inputting the context vector and the set of document sentence vectors into a collaborative writing model to generate a collaborative document with the same style as the main author;

Multiple collaborative documents are generated, evaluated according to the author's writing style discriminator and the author's content plausibility discriminator, and the collaborative document with the highest similarity is selected as the final document.

In an embodiment of the present application, before generating the collaborative document through the collaborative writing model, the method further includes: pre-training the bidirectional coding representation transformer BERT, including: collecting the author's text in the collaborative writing platform The content forms a document library; select the author's document in the document library, and train the BERT general model through the author's document to obtain a personalized BERT author model; according to the BERT general model and the BERT author model, trained to obtain an author content generator; according to the BERT author model and a linear classifier, trained to obtain an author writing style discriminator; according to the BERT author model and the linear classifier, trained to obtain an author content rationality discriminator.

In an embodiment of the present application, before generating the collaborative document through the collaborative writing model, the method further includes: training the collaborative writing model, including: downloading the document on the collaborative writing platform, and The author document is input into the document encoder to generate the context vector of the main author document; the author document is input into the document encoder to generate the statement vector set of the author document; the context vector and the statement vector set are passed through The author content generator constructs a sentence sequence; inputs the sentence sequence into the collaborative writing model, trains the collaborative writing model, and generates a collaborative document; interacts the generated collaborative document with a feedback environment , obtain a feedback result; transmit the feedback result to the collaborative writing model, the collaborative writing model updates network parameters according to the feedback result, and trains to obtain a further optimized collaborative writing model.

In one embodiment of the present application, the generated collaborative document is interacted with the feedback environment to obtain the feedback result, which specifically includes: forming a feedback environment based on the author's writing style discriminator, the author's content rationality discriminator, and reader feedback, determining A reward function: interacting the generated cooperation document with a feedback environment, and calculating a feedback result according to the reward function.

In an embodiment of the present application, the author's writing style discriminator and the author's content reasonable discriminator perform evaluation, and select the collaborative document with the highest similarity as the final document, which specifically includes: discriminating according to the author's writing style Judging the similarity between each cooperative document and the author's writing style by a device; judging the similarity between each cooperative document and the author's writing content according to the author's content reasonable discriminator; Weighting processing is performed to obtain the final similarity value of each cooperative document; the similarity value of each document is compared, and the cooperative document corresponding to the highest similarity value is determined as the final document.

In one embodiment of the present application, the training of the collaborative writing model specifically includes: training the collaborative writing model through the A3C algorithm; using multiple worker threads to adopt the same network structure as the global model public neural network , generating a cooperation document; inputting the cooperation document into the feedback environment to obtain a feedback result; forming a final document generation strategy according to the feedback result.

In one embodiment of the present application, before the collection of the author's text content in the collaborative writing platform, the method further includes: the collaborative writing platform receives the author's registration, checks and marks the author's identity; Receive documents uploaded by authors on the collaborative writing platform; when a document uploaded by a newly registered author appears, automatically obtain the document of the new author for training.

An apparatus for generating consistent writing style documents, comprising:

at least one processor; and,

a memory communicatively coupled to the at least one processor; wherein,

The memory stores instructions executable by the at least one processor, the instructions being executable by the at least one processor to enable the at least one processor to:

In an embodiment of the present application, the at least one processor is further configured to: perform pre-training on the bidirectional encoding representation transformer BERT, including: collecting the text content of the author in the collaborative writing platform to form a document library; Select the author's document in the document library, and train the BERT general model through the author's document to obtain a personalized BERT author model; according to the BERT general model and the BERT author model, train the author content generator ; According to the BERT author model and the linear classifier, the author's writing style discriminator is obtained through training; according to the BERT author model and the linear classifier, the author's content rationality discriminator is obtained through training.

A non-volatile storage medium storing computer-executable instructions, wherein the computer-executable instructions are set to:

This application provides a method, device, and storage medium for generating consistent writing style documents, which at least include the following beneficial effects: a document library is formed by collecting the text content of a large number of collaborative writing participants, and a BERT model is used to construct the author's personalized Model, and use reinforcement learning to realize model training through collaborative writing platform and interactive feedback environment, forming a consistent style collaborative writing model, which can generate documents with uniform document format, consistent content style, and more accurate and reasonable documents. Compared with the traditional method of unifying the content style, the cooperation documents generated by the model formed by deep learning and reinforcement learning can better discover the internal semantic connection of the article and more accurately simulate the author's writing style; the BERT model is used for prediction The training can form a targeted language model based on the author's actual writing documents. On the one hand, it improves the training efficiency and makes reasonable use of existing resources. On the other hand, it can better meet the individual needs of the field; using intensive learning training , making effective use of collaborative writing platform resources, using deep learning discriminators such as author writing content discriminators and author writing style discriminators to judge the rationality and effectiveness of generated documents, and at the same time using the platform's actual reader feedback to train the model, improving While improving the model training effect, it can form a consistent style document generation model that is more in line with the real user reading experience.

Description of drawings

The drawings described here are used to provide a further understanding of the application and constitute a part of the application. The schematic embodiments and descriptions of the application are used to explain the application and do not constitute an improper limitation to the application. In the attached picture:

FIG. 1 is a schematic diagram of the steps of a method for generating a consistent writing style document provided by an embodiment of the present application;

Fig. 2 is a schematic diagram of model training provided by the embodiment of the present application;

Fig. 3 is a schematic composition diagram of a device for generating consistent writing style documents provided by an embodiment of the present application.

Detailed ways

In order to make the purpose, technical solution and advantages of the present application clearer, the following will give a clear and complete description of the present application in conjunction with specific embodiments of the present application. Apparently, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in this application, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the scope of protection of this application.

In recent years, reinforcement learning technology has received widespread attention, especially in combination with deep learning, which has brought great progress to the field of artificial intelligence. Reinforcement learning is different from traditional supervised learning, mainly in the reinforcement signal. The reinforcement signal provided by the environment in reinforcement learning is an evaluation of the quality of the generated action (usually a scalar signal), rather than telling the reinforcement learning system RLS (reinforcement learning system) how to produce the correct action. Reinforcement learning continuously learns to make optimal actions in different environments through the task of interaction between the agent and the environment, and uses these perception generation strategies to create higher machine intelligence. Reinforcement learning has been applied in the fields of robot control, autonomous driving, and recommender systems, and has surpassed human performance in many areas.

Bidirectional Encoder Representation from Transformers (Bidirectional Encoder Representation from Transformers, BERT), that is, the encoder Encoder of the bidirectional transformation Transformer, compared with the traditional natural language processing mode, BERT is a revolutionary natural language processing mode, in natural language processing It has important applications in the field and has also inspired many existing computer logic frameworks and training methods. In particular, its ability to abstract continuous long sequence features has become one of the most important language processing models at present.

In one embodiment of the present application, the text content of a large number of collaborative writing participants is collected to form a document library, the BERT model is used to train documents of different authors to form a personalized BERT author model, and the BERT author model is used to construct a specific The author's author content generator model, author writing style discriminator model and author content reasonable discriminator model use reinforcement learning to train the collaborative writing model through the collaborative writing platform and the interactive feedback environment to form a consistent style collaborative writing model and generate Documents with uniform format, consistent content and style, and more accurate and reasonable documents. A detailed description will be given below.

FIG. 1 is a schematic diagram of the steps of a method for generating a consistent writing style document provided in an embodiment of the present application, which may include the following steps:

S101: Obtain the main author document through the collaborative writing platform, input the main author document into the document encoder, and generate a context vector.

S102: Obtain the sub-author document through the collaborative writing platform, input the sub-author document into a document encoder, and generate a set of document sentence vectors.

S103: Input the context vector and the document sentence vector set into the collaborative writing model to generate a collaborative document with the same style as the main author.

S104: Generate a plurality of collaborative documents, evaluate according to the author's writing style discriminator and the author's content plausibility discriminator, and select the collaborative document with the highest similarity as the final document.

In one embodiment of the present application, before the cooperative document is generated through the collaborative writing model, it is necessary to pre-train the bidirectional encoding representation transformer BERT.

In one embodiment of the present application, before collecting the author's text content on the collaborative writing platform, the collaborative writing platform receives the author's registration, verifies and marks the author's identity; receives the document uploaded by the author on the collaborative writing platform ; When a document uploaded by a newly registered author appears, automatically obtain the document of the new author for training.

Collect the author's text content on the collaborative writing platform to form a document library, and then use the documents in the document library to train the various models used to form an available model; first, pre-train the bidirectional encoding representation transformer BERT to form the author's personality Optimized BERT model.

As shown in Figure 2, select the author's document in the document library, input the selected author's document into the BERT general model, and perform personalized training on the BERT general model to obtain a personalized BERT author model.

According to the BERT general model and the BERT author model, the author content generator is trained. The core of the author content generator ContentGen is the Transformer model. The content generator formed based on the BERT language model training is used to generate sentences that conform to the author's style.

According to the BERT author model and linear classifier, the author's writing style discriminator is trained. The core of the author's writing style discriminator StyleClzfier is composed of a BERT language model and a linear classifier to determine whether the input document comes from the author's writing style.

According to the BERT author model and linear classifier, the author content rationality discriminator is trained. The core of the author content rationality discriminator ContClzfier is composed of a BERT language model and a linear classifier to judge whether the semantics of the input document is reasonable.

In an embodiment of the present application, before the collaborative document is generated through the collaborative writing model, the collaborative writing model needs to be trained. The core of the collaborative writing model π is a neural network model based on the author content generator and format generator. By inputting multiple documents participating in collaborative writing (main author document T, secondary author documents X1~Xn), the main author is selected for training and generates Consistency Style Collaboration Document C, using the A3C training method to interact with the interactive feedback environment, and finally form a generation strategy. The interactive feedback environment is composed of the author's writing style discriminator StyleClzfier, the author's content rationality discriminator ContClzfier and the reader feedback ReaderClzfier. The reader feedback is the feedback of specific readers' direct evaluation of the content.

Specifically, download the document on the collaborative writing platform, input the main author document into the document encoder, and generate the context vector of the main author document. The collaborative writing platform is deployed in the cloud data center to provide services such as author registration review, reader management, online collaborative editing, and proofreading. The cloud data center where it is located provides computing, storage, network and other cloud infrastructure services to realize document collection and provide BERT Basic models such as language models, while providing the computing power, storage and environment required for deep learning and reinforcement learning training tasks.

Input the document encoder from the author document to generate a sentence vector set from the author document; pass the context vector and sentence vector set through the author content generator to construct a sentence sequence; input the sentence sequence into the collaborative writing model, and train the collaborative writing through the A3C algorithm The model uses multiple worker threads to adopt the same network structure as the global model public neural network to generate collaborative documents; input the collaborative documents into the feedback environment to obtain feedback results; form the final document generation strategy based on the feedback results.

The feedback environment is formed according to the author's writing style discriminator, the author's content rationality discriminator and the reader's feedback, and the reward function is determined; the generated cooperation document is interacted with the feedback environment, and the feedback result is calculated according to the reward function.

The feedback results are transmitted to the collaborative writing model, and the collaborative writing model updates the network parameters according to the feedback results, and the further optimized collaborative writing model is trained.

In one embodiment of the present application, the author's writing style discriminator and the author's content plausibility discriminator are evaluated, and the collaborative document with the highest similarity is selected as the final document.

Specifically, according to the author's writing style discriminator, the similarity between each cooperative document and the author's writing style is judged; according to the author's content reasonable discriminator, the similarity between each cooperative document and the author's writing content is judged; Values are weighted to obtain the final similarity value of each cooperative document; the similarity values of each document are compared, and the cooperative document corresponding to the highest similarity value is determined as the final document.

For example, writing styles include romanticism, unrestrained style, postmodern style, documentary style, ideological genre, youth literature, network literature, romance style, critical style, pure literature style, etc. The writing style of the main author's document is a critical style, the writing content is a criticism around social phenomena, and the writing format is a total score format, and the content of the secondary author's document is a total score format or other formats. Then, after the cooperative document is generated according to the collaborative writing model, judge the similarity between the style of the cooperative document and the critical style, the similarity between the writing content and the social phenomenon, the similarity between the writing format and the total score format, and weight these similarities , to obtain the similarity value of each document, compare the similarity values of each document, and determine the cooperation document corresponding to the highest similarity value as the final document.

The above is a method for generating a consistent writing style document provided by the embodiment of the present application. Based on the same inventive idea, the embodiment of the present application also provides a corresponding device for generating a consistent writing style document, as shown in FIG. 3 .

This embodiment provides a device for generating consistent writing style documents, including:

at least one processor; and,

memory communicatively coupled to at least one processor; wherein,

The memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to:

Obtain the main author's document through the collaborative writing platform, input the main author's document into the document encoder, and generate a context vector;

Obtain the author's document through the collaborative writing platform, input the author's document into the document encoder, and generate a set of document sentence vectors;

Input the context vector and document statement vector set into the collaborative writing model to generate a collaborative document with the same style as the main author;

Multiple collaborative documents are generated, evaluated according to the author's writing style discriminator and author's content plausibility discriminator, and the collaborative document with the highest similarity is selected as the final document.

In one embodiment of the present application, at least one processor is also used to: perform pre-training on the bidirectional encoding representation transformer BERT, including: collect the text content of the author in the collaborative writing platform to form a document library; select from the document library The author's document, the BERT general model is trained through the author's document, and the personalized BERT author model is obtained; according to the BERT general model and the BERT author model, the author content generator is trained; according to the BERT author model and the linear classifier, the training is obtained The author's writing style discriminator; according to the BERT author model and linear classifier, the author's content rationality discriminator is trained.

Based on the same idea, some embodiments of the present application also provide media corresponding to the above methods and devices.

Some embodiments of the present application provide a storage medium for generating consistent writing style documents, which store computer-executable instructions, and the computer-executable instructions are set to:

Each embodiment in the present application is described in a progressive manner, the same and similar parts of each embodiment can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the method and medium embodiments, since they are basically similar to the method embodiments, the description is relatively simple, and for relevant parts, please refer to the descriptions of the method embodiments.

The methods and media provided in the embodiments of the present application correspond to the methods one by one, therefore, the methods and media also have beneficial technical effects similar to their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, therefore, The beneficial technical effects of the method and medium will not be repeated here.

It should also be noted that the term "comprises," "comprises," or any other variation thereof is intended to cover a non-exclusive inclusion such that a process, method, commodity or method that includes a set of elements includes not only those elements, but also includes not expressly included. other elements listed, or also include elements inherent to the process method commodity or method. Without further limitations, an element qualified by the phrase "comprising a ..." does not preclude the presence of additional identical elements in the process method article or method comprising the element.

The above are only examples of the present application, and are not intended to limit the present application. For those skilled in the art, various modifications and changes may occur in this application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included within the scope of the claims of the present application.

Claims

A method of generating a consistent writing style document, comprising:

Obtaining the main author's document through the collaborative writing platform, inputting the main author's document into the document encoder, and generating a context vector;

Obtain the sub-author document through the collaborative writing platform, input the sub-author document into the document encoder, and generate a set of document sentence vectors;

Inputting the context vector and the set of document sentence vectors into a collaborative writing model to generate a collaborative document with the same style as the main author;

Multiple collaborative documents are generated, evaluated according to the author's writing style discriminator and the author's content plausibility discriminator, and the collaborative document with the highest similarity is selected as the final document.
The method according to claim 1, wherein, before generating a collaborative document through the collaborative writing model, the method further comprises:

Pre-train the bidirectional encoding representation transformer BERT, including:

Collect the author's text content in the collaborative writing platform to form a document library;

Select the author's document in the document library, and train the BERT general model through the author's document to obtain a personalized BERT author model;

According to the BERT general model and the BERT author model, the author content generator is obtained through training;

According to the BERT author model and linear classifier, train the author's writing style discriminator;

According to the BERT author model and the linear classifier, an author content plausibility discriminator is obtained through training.
The method according to claim 2, wherein, before generating a collaborative document through the collaborative writing model, the method further comprises:

The collaborative writing model is trained, including:

Downloading documents on the collaborative writing platform, inputting the main author's document into a document encoder to generate a context vector of the main author's document;

inputting the author document into the document encoder to generate a set of sentence vectors from the author document;

passing the context vector and the sentence vector set through the author content generator to construct a sentence sequence;

inputting the statement sequence into the collaborative writing model, training the collaborative writing model, and generating a collaborative document;

Interacting the generated cooperation document with the feedback environment to obtain a feedback result;

The feedback result is transmitted to the collaborative writing model, and the collaborative writing model updates network parameters according to the feedback result, and is trained to obtain a further optimized collaborative writing model.
The method according to claim 3, wherein the generated cooperation document is interacted with a feedback environment to obtain a feedback result, which specifically includes:

According to the author's writing style discriminator, author's content reasonable discriminator and reader feedback to form a feedback environment, determine the reward function;

The generated cooperation document is interacted with a feedback environment, and a feedback result is obtained through calculation according to the reward function.
The method according to claim 1, wherein the evaluation is performed according to the author's writing style discriminator and the author's content rational discriminator, and the collaborative document with the highest similarity is selected as the final document, specifically comprising:

Judging the similarity between each collaborative document and the author's writing style according to the author's writing style discriminator;

Judging the similarity between each collaborative document and the author's writing content according to the author's content reasonable discriminator;

performing weighting processing on the author's writing style and the author's content calculation value of each cooperative document to obtain the final similarity value of each cooperative document;

The similarity values of the various documents are compared, and the cooperation document corresponding to the highest similarity value is determined as the final document.
The method according to claim 3, wherein the training of the collaborative writing model specifically includes:

Training the collaborative writing model through the A3C algorithm;

Use multiple worker threads to adopt the same network structure as the global model public neural network to generate collaborative documents;

inputting the cooperation document into the feedback environment to obtain a feedback result;

A final document generation strategy is formed according to the feedback results.
The method according to claim 2, wherein, before the collection of the author's text content in the collaborative writing platform, the method also includes:

The collaborative writing platform receives the author's registration, reviews and marks the author's identity;

Receive documents uploaded by authors on the collaborative writing platform;

When a document uploaded by a newly registered author appears, the document of the new author is automatically obtained for training.
An apparatus for generating consistent writing style documents, comprising:

at least one processor; and,

a memory communicatively coupled to the at least one processor; wherein,

The memory stores instructions executable by the at least one processor, the instructions being executable by the at least one processor to enable the at least one processor to:

Obtaining the main author's document through the collaborative writing platform, inputting the main author's document into the document encoder, and generating a context vector;

Obtain the sub-author document through the collaborative writing platform, input the sub-author document into the document encoder, and generate a set of document sentence vectors;

Inputting the context vector and the set of document sentence vectors into a collaborative writing model to generate a collaborative document with the same style as the main author;

Multiple collaborative documents are generated, evaluated according to the author's writing style discriminator and the author's content plausibility discriminator, and the collaborative document with the highest similarity is selected as the final document.
The device according to claim 8, wherein the at least one processor is further configured to:

Pre-train the bidirectional encoding representation transformer BERT, including:

Collect the author's text content in the collaborative writing platform to form a document library;

Select the author's document in the document library, and train the BERT general model through the author's document to obtain a personalized BERT author model;

According to the BERT general model and the BERT author model, the author content generator is obtained through training;

According to the BERT author model and linear classifier, train the author's writing style discriminator;

According to the BERT author model and the linear classifier, an author content plausibility discriminator is obtained through training.
A non-volatile storage medium storing computer-executable instructions, wherein the computer-executable instructions are set to:

Obtaining the main author's document through the collaborative writing platform, inputting the main author's document into the document encoder, and generating a context vector;

Obtain the sub-author document through the collaborative writing platform, input the sub-author document into the document encoder, and generate a set of document sentence vectors;

Inputting the context vector and the set of document sentence vectors into a collaborative writing model to generate a collaborative document with the same style as the main author;

Multiple collaborative documents are generated, evaluated according to the author's writing style discriminator and the author's content plausibility discriminator, and the collaborative document with the highest similarity is selected as the final document.