WO2023241226A1

WO2023241226A1 - Statement generation method and device and storage medium

Info

Publication number: WO2023241226A1
Application number: PCT/CN2023/090386
Authority: WO
Inventors: 蒋炜; 段新宇; 王喆锋; 怀宝兴
Original assignee: 华为云计算技术有限公司
Priority date: 2022-06-17
Filing date: 2023-04-24
Publication date: 2023-12-21
Also published as: CN117291181A

Abstract

A statement generation method and device and a storage medium, relating to the technical field of artificial intelligence. In the method, after a plurality of first candidate similar statements of a reference statement are generated by means of a first statement generation model, at least one second candidate similar statement can be obtained on the basis of a modification operation of a user on at least one first candidate similar statement, and then the first statement generation model is updated by means of the at least one second candidate similar statement. That is, according to the method, the statement generation model can be optimized in real time by means of the candidate similar statement modified by the user; on this basis, more high-quality similar statements can be generated by means of the optimized statement generation model and the reference statement, thereby effectively expanding a corpus.

Description

Statement generation method, device and storage medium

This application claims priority to the Chinese patent application with application number 202210693429.8 and the invention title "Sentence Generation Method, Device and Storage Medium" submitted on June 17, 2022, the entire content of which is incorporated into this application by reference.

Technical field

This application relates to the field of artificial intelligence (AI), and in particular to a sentence generation method, device and storage medium.

Background technique

Currently, in the field of natural language processing, the training and use of AI models often rely on corpora containing a large number of similar sentences. For example, in an automatic question and answer system based on a retrieval question and answer algorithm, a large number of similar sentences in the corpus can be used to train the question matching model. Subsequently, after receiving the target question input by the user, the question in the corpus can also be matched with the target question through the trained question matching model, and then the answer corresponding to the matching question of the target question can be The answer to the target question. Among them, the more similar questions to the target question in the corpus, the easier it is to obtain matching questions with higher accuracy during the matching process.

In summary, it can be seen that the training and use of AI models in the field of natural language processing will be affected by the number and quality of similar sentences contained in the corpus. Based on this, there is an urgent need to provide a method that can automatically generate high-quality similar sentences to expand the corpus and provide data support for the training and use of AI models.

Contents of the invention

Embodiments of the present application provide a sentence generation method, device and storage medium, which can automatically generate high-quality similar sentences to achieve effective expansion of the corpus. The technical solutions are as follows:

In a first aspect, a sentence generation method is provided. The method includes: based on a reference sentence, generating a plurality of first candidate similar sentences through a first sentence generation model; based on a user's modification operation for at least one first candidate similar sentence, Generating at least one second candidate similar statement, the plurality of first candidate similar statements including the at least one first candidate similar statement; using the at least one second candidate similar statement to update the first statement generation model, Obtain the second sentence generation model.

In this embodiment of the present application, after generating a plurality of first candidate similar sentences of the reference sentence through the first sentence generation model, at least one second candidate similar sentence can be obtained based on the user's modification operation on at least one first candidate similar sentence. , and then use the at least one second candidate similar sentence to update the first sentence generation model. That is to say, the embodiment of the present application can optimize the sentence generation model in real time through the candidate similar sentences modified by the user. On this basis, more high-quality similar sentences can be generated based on the optimized sentence generation model and reference sentences. In this way, not only the corpus is effectively expanded, but also the accuracy of the sentence generation model can be continuously improved.

Optionally, the above sentence generation model is an artificial intelligence model. For example, the sentence generation model can be a neural network model or a decision tree model. For example, the sentence generation model can be UniLM, a unified pre-trained language model for natural language understanding and generation.

Optionally, after generating a plurality of first candidate similar sentences based on the reference sentence, the method further includes: displaying the plurality of first candidate similar sentences to the user. In this way, the user can browse the plurality of first candidate similar sentences and modify, confirm or delete each first candidate similar sentence.

Optionally, after updating the first sentence generation model to obtain the second sentence generation model, the method further includes: generating a plurality of third candidate similar sentences based on the reference sentence and the second sentence generation model; The plurality of third candidate similar statements are displayed to the user.

After the first sentence generation model is optimized based on the first candidate similar sentences modified by the user, the plurality of third candidate similar sentences generated by the optimized sentence generation model will be more accurate. On this basis, multiple third candidate similar sentences are continued to be displayed to the user, so that the user can continue to modify the displayed candidate similar sentences, thereby continuing to optimize the sentence generation model. In this way, on the one hand, the generation results of the sentence generation model can be made more and more accurate, and at the same time, more similar sentences can be obtained to expand the corpus.

Optionally, the modification operation includes one or more of the following: adding character operations, deleting character operations, exchanging character operations, replacing character operations, and rewriting statement operations. Among them, the operations of adding characters, deleting characters, exchanging characters, and replacing characters can be used to modify the words in the statement when there are lexical problems in the statement, and the rewriting statement operation can be used when the statement has semantic problems. Rewrite the statement below.

Optionally, the implementation process of updating the first sentence generation model using the at least one second candidate similar sentence to obtain the second sentence generation model may include: obtaining the at least one first candidate similar sentence Difference information between each first candidate similar sentence and the corresponding modified second candidate similar sentence; based on the difference information, update the parameters of the first sentence generation model to obtain the second sentence generation Model.

A second aspect provides a sentence generation device, which is used to implement the sentence generation method described in the first aspect. Wherein, the statement generating device may include at least one module. Exemplarily, the at least one module may include a generation module, a modification module and an update module.

Among them, the generation module is used to generate a plurality of first candidate similar sentences based on the reference sentence through the first sentence generation model; the modification module is used to generate at least one second candidate similar sentence based on the user's modification operation of at least one first candidate similar sentence. Candidate similar sentences, the plurality of first candidate similar sentences include the at least one first candidate similar sentence; an update module for updating the first sentence generation model using the at least one second candidate similar sentence, Obtain the second sentence generation model.

Optionally, the sentence generation model is an artificial intelligence model.

Optionally, the device further includes: a display module. Wherein, the display module is used to display the plurality of first candidate similar sentences to the user; the generation module is also used to generate a plurality of first candidate similar sentences based on the reference sentence and the second sentence generation model. Three candidate similar sentences; the display module is also used to display the plurality of third candidate similar sentences to the user.

Optionally, the modification operation includes one or more of adding character operation, deleting character operation, exchanging character operation, replacing character operation, and rewriting statement operation.

Optionally, the update module is mainly configured to: obtain difference information between each first candidate similar statement in the at least one first candidate similar statement and the corresponding modified second candidate similar statement; based on the The difference information is obtained, and parameters of the first sentence generation model are updated to obtain the second sentence generation model.

In a third aspect, a computer device is provided. The structure of the computer device includes a processor and a memory. The memory is used to store at least one program instruction that supports the computer device to execute the statement generation method provided in the first aspect. code, and storage of data involved in implementing the statement generation method provided in the first aspect. The processor is configured to execute program instructions or code stored in the memory.

In a fourth aspect, a computer-readable storage medium is provided. Instructions are stored in the computer-readable storage medium. When the instructions are run on a computer device, the computer device executes the statement generation method described in the first aspect. .

A fifth aspect provides a computer program product containing instructions that, when run on a computer, causes the computer to execute the statement generation method described in the first aspect.

The technical effects obtained by the above-mentioned second aspect, third aspect, fourth aspect and fifth aspect are similar to those obtained by the corresponding technical means in the first aspect, and will not be described again here.

The technical solution provided by the embodiments of this application has at least the following beneficial effects:

In this embodiment of the present application, after generating a plurality of first candidate similar sentences of the reference sentence through the first sentence generation model, at least one second candidate similar sentence can be obtained based on the user's modification operation on at least one first candidate similar sentence. , and then use the at least one second candidate similar sentence to update the first sentence generation model. That is to say, the embodiment of the present application can optimize the sentence generation model in real time through the candidate similar sentences modified by the user. On this basis, more high-quality similar sentences can be generated based on the optimized sentence generation model and reference sentences, thereby achieving An effective expansion of the corpus.

Description of the drawings

Figure 1 is a schematic diagram of the implementation environment of a statement generation method provided by an embodiment of the present application;

Figure 2 is a schematic structural diagram of a computer device provided by an embodiment of the present application;

Figure 3 is a flow chart of a statement generation method provided by an embodiment of the present application;

Figure 4 is a schematic diagram of a display page and a modification page of a first candidate similar statement provided by an embodiment of the present application;

Figure 5 is a schematic flow chart of interaction between a user and a terminal device provided by an embodiment of the present application;

Figure 6 is a schematic diagram of a display page of a second candidate similar sentence provided by an embodiment of the present application;

Figure 7 is an exemplary flow chart of a statement generation process provided by an embodiment of the present application;

Figure 8 is a schematic structural diagram of a sentence generation device provided by an embodiment of the present application.

Detailed ways

In order to make the objectives, technical solutions, and advantages of the embodiments of the present application clearer, the embodiments of the present application will be described in further detail below with reference to the accompanying drawings.

Before explaining the embodiments of the present application in detail, the application scenarios of the embodiments of the present application are first introduced.

Currently, in the field of natural language processing, the processing of text information often relies on corpora containing a large number of similar sentences. The sentence generation method provided by the embodiment of the present application can be used to generate similar sentences of the reference sentence, thereby effectively expanding the corpus, and thereby providing data support for the processing of text information.

For example, automatic question and answer technology is a research hotspot in the field of natural language. This technology is widely used in various Internet products, such as question and answer robots, chat robots, intelligent customer service robots, etc. Among them, retrieval question answering is an important branch of automatic question answering. In a retrieval question answering system, a large number of similar sentences in the corpus can be used to match questions The model is trained. Subsequently, after receiving the target question input by the user, the question in the corpus can also be matched with the target question through the trained question matching model, and then the answer corresponding to the matching question of the target question can be The answer to the target question. Among them, the more similar questions to the target question in the corpus, the easier it is to obtain matching questions with higher accuracy during the matching process. Based on this, in this scenario, the sentence generation method provided by the embodiment of the present application can be used to automatically generate a large number of high-quality similar sentences based on the reference sentences to expand the corpus, thereby improving the accuracy of automatic question and answer.

For another example, when building a document retrieval system, after the user uploads a reference document, the document retrieval system can extract the key content in the document, such as extracting the title, table of contents, etc. of the document, and build a corpus based on the extracted key content. On this basis, the document retrieval system can also use the key content extracted from the document as a reference sentence, and then use the sentence generation method provided by the embodiment of the present application to generate similar sentences to the reference sentence, thereby expanding the corpus. In this way, when the document retrieval system subsequently receives a retrieval statement for document retrieval, it can use the retrieval model to search for statements matching the retrieval statement among the extracted key content and generated similar statements, thereby improving the accuracy of retrieval. In addition, these similar sentences in the corpus can also be used to train the retrieval model to improve the accuracy of the retrieval model.

The above are two exemplary application scenarios of the sentence generation method provided by the embodiments of the present application. Of course, the sentence generation method can also be applied in other scenarios, such as synonymous sentence generation, sentence rewriting, etc. The embodiments of the present application are No restrictions.

Next, the implementation environment of the statement generation method provided by the embodiment of this application is introduced.

The statement generation method provided by the embodiments of this application can be implemented by a computer device in any environment. For example, the sentence generation method can be implemented through a terminal device, and the terminal device can be a device such as a desktop computer, a notebook computer, a smart phone, a tablet computer, etc.

Alternatively, the statement generation method can also be implemented by multiple computer devices. For example, it can be implemented by the client device 101 and the server 102 shown in FIG. 1 .

Among them, the client device 101 is used to provide an interactive interface, and interact with the user through the interactive interface to obtain the reference sentence. Afterwards, the reference sentence is sent to the server 102.

A statement generation model is deployed on the server 102. Based on this, after receiving the reference sentence sent by the client device 101, the server 102 generates a plurality of candidate similar sentences through the sentence generation model. Afterwards, the plurality of candidate similar sentences are sent to the client device 101.

The client device 101 is also configured to present the plurality of candidate similar sentences to the user. Based on the user's modification operation on the displayed at least one candidate similar sentence, at least one modified candidate similar sentence is generated, and the at least one modified candidate similar sentence is sent to the server 102 . In addition, the client device 101 may also send the candidate similar sentence to the server 102 as a similar sentence of the reference sentence based on the user's confirmation operation on any displayed candidate similar sentence.

The server 102 is also configured to add similar sentences to the reference sentence to the database when receiving similar sentences to the reference sentence confirmed by the user. When at least one modified candidate similar sentence is received, the sentence generation model is updated using the at least one modified candidate similar sentence, thereby obtaining an updated sentence generation model. Afterwards, the server 102 can continue to use the updated sentence generation model to generate multiple candidate similar sentences based on the reference sentences, and again send the multiple candidate similar sentences to the client device 101 for display, to repeat the above steps until termination is reached. conditions. The termination condition may be that the similar statements added to the reference statements in the database by the server 102 reach the reference threshold, or the termination condition may be that the number of update rounds of the statement generation model by the server 102 reaches the reference number of rounds, etc.

It should be noted that the above-mentioned server 102 can be deployed in a cloud environment. For example, the server 102 may be a server in a cloud data center. Alternatively, the server 102 may also be an edge computing device in an edge environment. For example, the server 102 may be a computing device in a data center of an organization. Alternatively, the server 102 may also be a computing device in other scenarios or types, which is not limited in the embodiments of this application.

In addition, the above-mentioned client device 101 can be a device such as a laptop computer, a smart phone, a tablet computer, etc.

Figure 2 is a schematic structural diagram of a computer device provided by an embodiment of the present application. Regardless of whether the sentence generation method provided by the embodiment of the present application is implemented by one computer device or multiple computer devices, the computer device can be the computer device 200 shown in FIG. 2 . Exemplarily, the computer device 200 includes: a processor 201, a communication bus 202, a memory 203 and at least one communication interface 204.

Among them, the processor 201 can be a general central processing unit (CPU), an application-specific integrated circuit (ASIC), a graphics processing unit (GPU) or any combination thereof. The processor 201 may include one or more chips, and the processor 201 may include an AI accelerator, such as a neural network processor (neural processing unit, NPU).

Communication bus 202 may include a path that carries information between various components of computer device 200 (eg, processor 201, memory 203, communication interface 204).

The memory 203 may be a read-only memory (ROM) or other type of static storage device that can store static information and instructions, a random access memory (random access memory (RAM)) or other type of static storage device that can store information and instructions. Type of dynamic storage device, it can also be electrically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disk storage, optical disc Storage (including compact discs, laser discs, optical discs, digital versatile discs, Blu-ray discs, etc.), magnetic disk storage media or other magnetic storage devices, or can be used to carry or store desired program code in the form of instructions or data structures and can be used by Any other media accessed by a computer, but not limited to this. The memory 203 may exist independently and be connected to the processor 201 through the communication bus 202. The memory 203 may also be integrated with the processor 201. The memory 203 can store computer instructions. When the computer instructions stored in the memory 203 are executed by the processor 201, the statement generation method provided by the embodiment of the present application can be implemented. In addition, the memory 203 may also store data required by the processor during execution of the above method as well as intermediate data and/or result data generated.

The communication interface 204 uses any device such as a transceiver to communicate with other devices or communication networks, such as Ethernet, wireless access network (RAN), wireless local area networks (WLAN), etc.

As an embodiment, the processor 201 may include one or more CPUs.

As an embodiment, the computer device 200 may include multiple processors. Each of these processors may be a single-CPU processor or a multi-CPU processor. A processor here may refer to one or more devices, circuits, and/or processing cores for processing data (eg, computer program instructions).

As an embodiment, the computer device 200 may also include an output device 205 and an input device 206. Output device 205 communicates with processor 201 and can display information in a variety of ways. For example, the output device 205 may be a liquid crystal display (LCD), a light emitting diode (LED) display device, a cathode ray tube (CRT) display device, a projector, etc. . Input device 206 communicates with processor 201 and may receive user input in a variety of ways. For example, the input device 206 may be a mouse, a keyboard, a touch screen device, a sensing device, or the like.

Next, the statement generation method provided by the embodiment of this application is introduced.

Figure 3 is a flow chart of a statement generation method provided by an embodiment of the present application. This method can be applied to one or more computer devices. The following description takes the method applied to one terminal device as an example. Referring to Figure 3, the method includes the following steps:

Step 301: Based on the reference sentence, generate a plurality of first candidate similar sentences through the first sentence generation model.

In this embodiment of the present application, an interactive interface is displayed on the terminal device. The interactive interface may include a text input box. The user may enter one or more reference sentences in the text input box. Correspondingly, the terminal device may obtain user input. reference sentence. Alternatively, the interactive interface may include a document upload option. After detecting the user's selection of the document upload option, the user receives the document uploaded by the user and extracts one or more reference sentences from the document. Alternatively, the interactive interface includes a voice option. After detecting the user's selection of the voice option, the terminal device turns on the audio collection device to collect the user's voice information, and then processes the voice information through a voice recognition model, thereby Get one or more reference statements.

Optionally, in this embodiment of the present application, the terminal device can also obtain the reference sentence through other methods. For example, get reference statements from a database that stores a large number of statements. The database can be deployed on the terminal device or on other devices, for example, in a background server.

For any reference sentence obtained, the terminal device can perform text domain processing operations such as word segmentation and stop word removal on the reference sentence. Afterwards, the terminal device can use the processed reference sentence as an input of the first sentence generation model, process the reference sentence through the first sentence generation model, and thereby output a plurality of first candidate similar sentences. The number of output first candidate similar sentences may be set by the user, or may be a default number. For example, the number of output first candidate similar sentences is 10.

In this embodiment of the present application, the first sentence generation model may be an AI model. For example, the first sentence generation model may be a neural network model or a decision tree model. When the first sentence generation model is a neural network model, the neural network model can be a sequence-to-sequence ( seq2seq) model. In addition, the first sentence generation model may be a sentence generation model pre-trained using a large amount of unlabeled data and open source similar problem sets. Alternatively, the first sentence generation model may also be a sentence generation model that has been updated using candidate similar sentences modified by the user.

Step 302: Generate at least one second candidate similar sentence based on the user's modification operation on at least one first candidate similar sentence.

After generating a plurality of first candidate similar sentences through the first sentence generation model, the terminal device may display the plurality of first candidate similar sentences. The user may perform a modification operation on at least one first candidate similar statement among the displayed plurality of first candidate similar statements. Correspondingly, the terminal device generates at least one second candidate similar sentence based on the user's modification operation on the at least one first candidate similar sentence. The modification operation includes one or more of adding character operation, deleting character operation, exchanging character operation, replacing character operation, and rewriting statement operation.

For example, the terminal device may display the plurality of first candidate similar sentences in the interactive interface, and display the operation options corresponding to each first candidate similar sentence. The operation options may include modification options, deletion options and confirmation options. The user can perform a selection operation on the modification option corresponding to a first candidate similar sentence. In response to the selection operation, the terminal device can detect the user's modification operation on the first candidate similar sentence, and then generate a second modification based on the modification operation. wait Choose similar sentences.

In one possible implementation, in response to the user selecting a modification option corresponding to a first candidate similar sentence, the terminal device displays the modification sub-option corresponding to the first candidate similar sentence. The modification of sub-options may include adding sub-options, deleting sub-options, exchanging sub-options, replacing sub-options and rewriting sub-options. The user can perform one or more of adding character operations, deleting character operations, swapping character operations, replacing character operations, and rewriting statement operations based on the displayed modification sub-options. Correspondingly, the terminal device can generate a modified sentence, that is, a second candidate similar sentence, based on the modification operation performed by the user on the first candidate similar sentence.

For example, the user can perform a selection operation on the add sub-option corresponding to the first candidate similar sentence. In response to the selection operation, the terminal device can receive a character inserted by the user at a certain position in the first candidate similar sentence. , the first candidate similar statement is updated based on the inserted character.

Alternatively, the user can perform a selection operation on the deletion sub-option corresponding to the first candidate similar sentence. In response to the selection operation, the terminal device can detect the characters selected and deleted by the user, and then based on the characters deleted by the user, the first candidate is similar. statement to update.

Alternatively, the user can perform a selection operation on the exchange sub-option corresponding to the first candidate similar sentence. In response to the selection operation, the terminal device can detect that the user has selected and dragged one or more characters in the first candidate similar sentence. Operation, based on the user's drag operation, the selected character or characters are moved to the position specified by the user.

Alternatively, the user can perform a selection operation on the replacement sub-option corresponding to the first candidate similar sentence. In response to the selection operation, the terminal device can detect the character selected by the user in the first candidate similar sentence and display a text input box. The characters input by the user in the text input box are obtained, and the selected characters in the first candidate similar sentence are replaced with the characters input by the user, thereby obtaining the updated first candidate similar sentence.

Alternatively, the user can perform a selection operation on the rewriting sub-option corresponding to the first candidate similar sentence. In response to the selection operation, the terminal device can display a text input box and obtain the rewriting sentence input by the user in the text input box. Treat the rewritten statement as the updated statement.

It should be noted that through the above method, the user can sequentially perform at least one modification operation on the first candidate similar sentence through at least one modification sub-option. Each time the terminal device receives a modification operation on the first candidate similar sentence by the user, , based on this modification operation, update the most recently updated first candidate similar statement. On this basis, the interactive interface of the terminal device may also include an end modification option, so that after the user completes the modification, the user can select the end modification option. When the terminal device detects the selection operation of the end modification option, the first candidate similar statement after the last update is used as the second candidate similar statement.

For example, the first sentence generation model generates n first candidate similar sentences based on the reference sentence, which are respectively sentence 1-1, sentence 1-2, ..., and sentence 1-n. Referring to Figure 4, the terminal device displays a display page in which the n first candidate similar sentences are displayed, and corresponding modification options, confirmation options, and deletion options are displayed after each sentence. Users can browse each statement. If the user is partially satisfied with a certain statement, for example, if the user believes that statements 1-3 can be modified to become similar statements that meet the requirements, the user can select the modification options displayed after statements 1-3. In response to the selection operation of the modification option, the terminal device may display the modification interface of statements 1-3. The modification interface displays statements 1-3, multiple modification sub-options, and an end modification option. The multiple modification sub-options include adding sub-options, deleting sub-options, exchanging sub-options, replacing sub-options, and rewriting sub-options. Users can refer to the method introduced previously to perform modification operations on statements 1-3 based on the multiple modification sub-options. After completing the modification, the user can perform a selection operation on the end modification option. In response to the selection operation on the end modification option, the terminal device can display the modification results in the display page. Sentences 1-3'. In addition, the terminal device can also display a continue option and an end option in the display page. If the user has completed modifying the displayed statement in the display page, the user can select the continue option. In response to the selection operation of the continue option, the terminal device may perform step 303. If the user no longer needs to generate similar sentences for the reference sentence, the user can also instruct the terminal device to end updating the sentence generation model and end generating candidate similar sentences for the reference sentence by clicking the end option.

In another possible implementation, in response to the user selecting a modification option corresponding to a first candidate similar statement, the terminal device can change the first candidate similar statement from an uneditable state to an editable state. When the first candidate similar sentence is in an editable state, the user can insert characters into the first candidate similar sentence, delete, replace, exchange, etc. characters in the first candidate similar sentence. The terminal device can update the first candidate similar sentence according to the modification operation performed by the user, thereby obtaining the modified sentence, that is, the second candidate similar sentence.

Optionally, the user can also perform a selection operation on the confirmation option of a first candidate similar sentence. In response to the selection operation on the confirmation option, the terminal device can store the first candidate similar sentence as a similar sentence of the reference sentence. to the database. The database may be a corpus storing a large number of statements.

Optionally, the user can also perform a selection operation on the deletion option of a certain first candidate similar sentence. In response to the selection operation on the deletion option, the terminal device can directly delete the first candidate similar sentence.

FIG. 5 is a schematic flowchart of interaction between a user and a terminal device according to an embodiment of the present application. As shown in Figure 5, for a certain candidate similar statement displayed on the terminal device, such as statement n, the user determines whether the statement n is satisfactory. If satisfied, click the confirmation option, and then the terminal device adds the statement n to the database. If you are not satisfied, click the delete option, and then the terminal device deletes the statement n. If satisfied with the n part of the statement, the user clicks the modification option. Afterwards, the user determines the problem with statement n. If there is a semantic problem, the user can click the reselect sub-option and enter the rewritten data to obtain a modified similar statement. If there is a lexical problem, the user can click one or more of add sub-option, delete sub-option, modify sub-option and exchange sub-option to modify the statement n. After that, if the modification is completed, click on the End Modification option. Afterwards, the terminal device generates a modified similar statement based on the user's modification operation.

Through the method introduced above, the user can modify at least one first candidate similar sentence displayed on the terminal device, thereby obtaining at least one second candidate similar sentence.

Optionally, in some possible situations, after the terminal device displays a plurality of first candidate similar sentences, if the user is satisfied with the plurality of first candidate similar sentences, the user can select the plurality of first candidate similar sentences. The confirmation operation is all performed, so that the terminal device can add the plurality of first candidate similar sentences to the database as similar sentences of the reference sentence. In this case, the end device will not update the statement generation model. Alternatively, if the user is satisfied with some of the displayed first candidate similar sentences and is dissatisfied with some of the sentences, the user can perform a confirmation operation on the satisfied first candidate similar sentences and not on the dissatisfied first candidate similar sentences. No modification operations are performed. In this way, the terminal device can add the first candidate similar statement confirmed by the user to the database as a similar statement of the reference statement based on the user operation, and discard the first candidate similar statement for which the user has not performed a modification operation. In this case, the terminal device also does not update the sentence generation model.

Step 303: Update the first sentence generation model using at least one second candidate similar sentence to obtain a second sentence generation model.

After obtaining at least one second candidate similar sentence, the terminal device can obtain the difference information between each first candidate similar sentence in the at least one first candidate similar sentence and the corresponding modified second candidate similar sentence; based on the Difference Based on the different information, the parameters of the first sentence generation model are updated to obtain the second sentence generation model.

It can be known from the foregoing introduction that at least one second candidate similar sentence is obtained by modifying at least one first candidate similar sentence. The at least one first candidate similar sentence is the actual output value of the first sentence generation model, and the at least one modified second candidate similar sentence can be used as the expected output value of the first sentence generation model. Based on this, the terminal device may calculate an error value between each first candidate similar sentence in the at least one first candidate similar sentence and the corresponding modified second candidate similar sentence, that is, calculate each actual output value The error value between the error value and the corresponding expected output value is used as the difference information, and then based on the difference information, the parameters of the first sentence generation model are updated to obtain the second sentence generation model.

It should be noted that in the embodiment of the present application, after obtaining the error value between each first candidate similar sentence and the corresponding second candidate similar sentence, taking the first sentence generation model as a neural network model as an example, the The terminal device can use the cross-entropy loss function based on autoregression to determine the total error value based on the calculated at least one error value. If the total error value is greater than the error threshold, the terminal device can determine the total error value from the neural network model based on the total error value. The output layer of the neural network model begins to perform backpropagation to update the parameters of each layer of the neural network model layer by layer, thereby obtaining the second sentence generation model.

Of course, if the first sentence generation model is another type of AI model, the terminal device can also update the parameters of the first sentence generation model through other update methods to obtain the second sentence generation model. For example, Newton's method can be used to update the parameters of the first sentence generation model. The first sentence generation model is updated, which will not be described again in the embodiment of this application.

After obtaining the second sentence generation model, the terminal device can use the aforementioned reference sentence as the input of the second sentence generation model, process the reference sentence through the second sentence generation model, and thereby output a plurality of third similar candidates. statement. Afterwards, the terminal device can display the plurality of third candidate similar sentences to the user in the method introduced in step 302, and decide whether to use the displayed candidate similar sentences as a reference sentence based on the user's operation on the displayed candidate similar sentences. Add similar statements to the database, delete or modify candidate similar statements. If the terminal device obtains at least one modified candidate similar sentence based on the user's modification operation of at least one third candidate similar sentence, the second sentence generation model can continue to be updated, and so on, until the terminal device modifies the sentence The number of update rounds of the generated model reaches the specified number of rounds, or until the similar sentences of the reference sentence stored by the terminal device in the database reach the specified number, the terminal device can stop updating the sentence generation model and stop generating by sentences. The model generates candidate similar sentences for the reference sentence.

For example, after the second sentence generation model is updated based on at least one second candidate similar sentence, the second sentence generation model generates n third candidate similar sentences, namely sentence 2-1, sentence 2-2, and sentence 2. -3,..., statement 2-n. Referring to Figure 6, the terminal device can display the n third candidate similar sentences, and display corresponding operation options after each sentence for the user to perform the corresponding operation.

Optionally, in this embodiment of the present application, after the first sentence generation model is updated using the at least one second candidate similar sentence, since the at least one second candidate similar sentence is a similar sentence that meets the requirements modified by the user, statement, therefore, the terminal device can also store the at least one second candidate similar statement into the database.

Based on the sentence generation method provided in the above embodiments, the embodiment of this application provides an exemplary flow chart of the sentence generation process. Referring to Figure 7, the sentence generation model first generates similar sentences 1 to n based on the reference sentence, and then displays the similar sentences 1 to n on the interactive interface. Users can provide feedback on the displayed statements through the operation options provided in the interactive interface. Among them, statements that the user confirms to be satisfactory can be directly stored in the database. Statements that users are not satisfied with will be deleted directly. For the statement modified by the user, the modified statement can be used to update the parameters of the statement generation model, and the modified statement can be stored in the database.

In this embodiment of the present application, after generating a plurality of first candidate similar sentences of the reference sentence through the first sentence generation model, at least one second candidate similar sentence can be obtained based on the user's modification operation on at least one first candidate similar sentence. , and then use the at least one second candidate similar sentence to update the first sentence generation model. That is to say, the embodiments of the present application can optimize the sentence generation model in real time through user-modified candidate similar sentences. On this basis, new higher-quality similar sentences can be generated again based on the optimized sentence generation model and reference sentences. In this way, not only the corpus is effectively expanded, but also the accuracy of the sentence generation model can be continuously improved.

In addition, in this embodiment of the present application, the sentence generation model deployed on the terminal device can be pre-trained through a large amount of unlabeled data and open source labeled similar question sets. In this way, the generalization ability and cold start ability of the model can be improved.

It should be noted that the implementation process of the method is introduced above by taking the terminal device executing the statement generation method as an example. When the sentence generation method is executed by multiple computer devices, for example, by the client device and the server shown in Figure 1, then the above embodiment shows candidate similar sentences, interaction with the user, and acquisition based on the user's modification operation. The step of obtaining the modified candidate similar sentences can be implemented by the client device, or the step of obtaining the modified candidate similar sentences based on the user's modification operation can also be implemented by the server. In addition, the steps of generating candidate similar sentences based on the reference sentences and updating the parameters of the sentence generation model based on the modified candidate similar sentences can be implemented by the server.

Next, the sentence generation device provided by the embodiment of the present application will be introduced.

Referring to Figure 8, this embodiment of the present application provides a statement generation device 800. The device 800 includes a generation module 801, a modification module 802 and an update module 803.

Generating module 801, used to perform step 301 in the above embodiment;

Modify module 802, used to perform step 302 in the above embodiment;

The update module 803 is used to perform step 303 in the above embodiment.

The generation module 801, the modification module 802 and the update module 803 can be deployed on a computer device and executed by a processor on the computer device. Alternatively, the generation module 801, modification module 802 and update module 803 can be distributed on multiple computer devices and executed by processors of multiple computer devices to jointly implement the above statement generation method.

Optionally, the sentence generation model is an artificial intelligence model.

Optionally, referring to Figure 8, the device 800 also includes:

Display module 804, used to display multiple first candidate similar sentences to the user;

The generation module 801 is also used to generate a plurality of third candidate similar sentences based on the reference sentence and the second sentence generation model;

The display module 804 is also used to display multiple third candidate similar sentences to the user.

Optionally, the modification operation includes one or more of the following:

Add character operations, delete character operations, exchange character operations, replace character operations, and rewrite statement operations.

Optionally, the update module 803 is mainly used for:

Obtaining difference information between each first candidate similar statement in the at least one first candidate similar statement and the corresponding modified second candidate similar statement;

Based on the difference information, the parameters of the first sentence generation model are updated to obtain the second sentence generation model.

To sum up, in the embodiment of the present application, multiple first candidates of the reference sentence are generated through the first sentence generation model. After generating similar sentences, at least one second candidate similar sentence can be obtained based on the user's modification operation on at least one first candidate similar sentence, and then the at least one second candidate similar sentence can be used to update the first sentence generation model. That is to say, the embodiments of the present application can optimize the sentence generation model in real time through user-modified candidate similar sentences. On this basis, new higher-quality similar sentences can be generated again based on the optimized sentence generation model and reference sentences. In this way, not only the corpus is effectively expanded, but also the accuracy of the sentence generation model can be continuously improved.

It should be noted that when the sentence generation device provided in the above embodiment generates similar sentences to the reference sentence, the division of the above functional modules is only used as an example. In practical applications, the above functions can be allocated to different functions as needed. Module completion means dividing the internal structure of the device into different functional modules to complete all or part of the functions described above. In addition, the sentence generation device provided by the above embodiments and the sentence generation method embodiments belong to the same concept. Please refer to the method embodiments for the specific implementation process, which will not be described again here.

In the above embodiments, it may be implemented in whole or in part by software, hardware, firmware, or any combination thereof. When implemented using software, it may be implemented in whole or in part in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer instructions are loaded and executed on the computer, the processes or functions described in the embodiments of the present application are generated in whole or in part. The computer may be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions may be stored in or transmitted from one computer-readable storage medium to another, e.g., the computer instructions may be transferred from a website, computer, server, or data center Transmission to another website, computer, server or data center through wired (such as coaxial cable, optical fiber, digital subscriber line (DSL)) or wireless (such as infrared, wireless, microwave, etc.) means. The computer-readable storage medium may be any available medium that can be accessed by a computer or a data storage device such as a server, data center, etc. that contains one or more available media integrated. The available media may be magnetic media (such as floppy disks, hard disks, tapes), optical media (such as digital versatile discs (DVD)), or semiconductor media (such as solid state disks (SSD) )wait.

Those of ordinary skill in the art can understand that all or part of the steps to implement the above embodiments can be completed by hardware, or can be completed by instructing relevant hardware through a program. The program can be stored in a computer-readable storage medium. The above-mentioned The storage media mentioned can be read-only memory, magnetic disks or optical disks, etc.

The above description is not intended to limit the embodiments of the present application. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the embodiments of the present application shall be included in the protection scope of the embodiments of the present application.

Claims

A sentence generation method, characterized in that the method includes:

Based on the reference sentence, generate multiple first candidate similar sentences through the first sentence generation model;

Generate at least one second candidate similar statement based on the user's modification operation on at least one first candidate similar statement, the plurality of first candidate similar statements including the at least one first candidate similar statement;

The first sentence generation model is updated using the at least one second candidate similar sentence to obtain a second sentence generation model.
The method according to claim 1, characterized in that the sentence generation model is an artificial intelligence model.
The method according to claim 1 or 2, characterized in that after generating a plurality of first candidate similar sentences, the method further includes:

displaying the plurality of first candidate similar sentences to the user;

The method also includes:

Generate a plurality of third candidate similar sentences based on the reference sentence and the second sentence generation model;

The plurality of third candidate similar statements are displayed to the user.
The method according to any one of claims 1 to 3, characterized in that the modification operation includes one or more of the following:

Add character operations, delete character operations, exchange character operations, replace character operations, and rewrite statement operations.
The method according to any one of claims 1 to 4, characterized in that, using the at least one second candidate similar sentence to update the first sentence generation model to obtain a second sentence generation model includes:

Obtaining difference information between each first candidate similar statement in the at least one first candidate similar statement and the corresponding modified second candidate similar statement;

Based on the difference information, parameters of the first sentence generation model are updated to obtain the second sentence generation model.
A sentence generation device, characterized in that the device includes:

A generation module, configured to generate multiple first candidate similar sentences through the first sentence generation model based on the reference sentence;

A modification module configured to generate at least one second candidate similar statement based on the user's modification operation on at least one first candidate similar statement, the plurality of first candidate similar statements including the at least one first candidate similar statement;

An update module, configured to update the first sentence generation model using the at least one second candidate similar sentence to obtain a second sentence generation model.
The device according to claim 6, wherein the sentence generation model is an artificial intelligence model.
The device according to claim 6 or 7, characterized in that the device further includes:

a display module, configured to display the plurality of first candidate similar sentences to the user;

The generation module is also configured to generate a plurality of third candidate similar sentences based on the reference sentence and the second sentence generation model;

The display module is also configured to display the plurality of third candidate similar sentences to the user.
The device according to any one of claims 6 to 8, characterized in that the modification operation includes one or more of the following:

Add character operations, delete character operations, exchange character operations, replace character operations, and rewrite statement operations.
The device according to any one of claims 6 to 9, characterized in that the update module is mainly used for:

Obtaining difference information between each first candidate similar statement in the at least one first candidate similar statement and the corresponding modified second candidate similar statement;

Based on the difference information, parameters of the first sentence generation model are updated to obtain the second sentence generation model.
A computer device, characterized in that the computer device includes a processor and a memory, at least one program instruction or code is stored in the memory, and the at least one program instruction or code is loaded and executed by the processor to The computer device is caused to implement the statement generation method described in any one of claims 1-5.
A computer-readable storage medium, characterized in that instructions are stored in the computer-readable storage medium, and when the instructions are run on a computer device, the computer device executes the method described in any one of claims 1-5. Statement generation method.