TWI822258B - Method and electiacal device for automatical generating text data and for revising text data - Google Patents

Abstract

A method for automatically generating text data by an electronic device is provided. The method includes receiving a plurality of text parameters, generating the text data including a plurality of sentences based on the text parameters, determining whether the text data includes an adjusting sentence set to be adjusted, and adjusting the adjusting sentence when the adjusting sentence exists, wherein the determining further includes obtaining a plurality of sentence phrases and their word classes from a specific sentence based on a first language model to determine whether the specific sentence is set as the adjusting sentence; and obtaining a plurality of neighboring phrases and their word classes from a plurality of neighboring sentences based on the first language model to determine a sentence relationship based on a second language model for determining whether one of the neighboring sentences is set as the adjusted sentence.

Description

Automatic text generation method and text modification method and electronic device thereof

The present invention relates to the field of automatic text processing, and in particular, to a method for automatic text generation and an electronic device thereof.

Although various text generation technologies are improving with each passing day, existing text generation technologies are still unable to use a large amount of text for model training due to too little training data for special fields with less training data, resulting in insufficient smoothness of generated sentences. Or the context is inconsistent and requires a lot of extra time to correct. For example: if you want to generate a climate-related explanation document, the content and quantity of climate-related documents around the user may be small, resulting in insufficient training text information required by the text generation technology, and the generated text will be The sentence is not fluent or the context is incoherent.

The present invention is completed in view of the above problems, and provides a method for automatically generating text through at least one electronic device and an electronic device thereof.

In order to solve the above problem, the automatic text generation method includes: receiving multiple text parameters; generating text information based on the multiple text parameters, wherein the text information includes multiple sentences sub; confirming whether there is a sentence to be adjusted in the plurality of sentences, wherein the confirmation of the sentence to be adjusted includes: through a first language model, obtaining a plurality of single-sentence vocabulary and their respective words from a specific sentence in the plurality of sentences. Part-of-speech participles, and based on the plurality of part-partial speech parts of the plurality of single-sentence words, confirm whether the specific sentence is set as the sentence to be adjusted; and through the first language model, from the plurality of adjacent sentences in the plurality of sentences The sentence obtains multiple adjacent words and respective participles of speech, and based on the multiple participles of the multiple adjacent words, a sentence correlation between the multiple adjacent sentences is confirmed through a second language model, based on the sentence Relevance confirms whether one of the plurality of adjacent sentences is set as the sentence to be adjusted; and when the text information is confirmed to have the sentence to be adjusted, the sentence to be adjusted is adjusted to complete the text information .

In one embodiment, the method further includes: obtaining a grammatical relationship based on the distribution of the plurality of word participles of the plurality of single-sentence words in the specific sentence; and based on the grammatical relationship, confirming whether the specific sentence is Set as the sentence to be adjusted.

In one embodiment, the method further includes: receiving a plurality of grammatical structures; confirming the sentence structure completeness of the specific sentence based on the multiple grammatical structures and the grammatical relationship; and confirming the sentence structure completeness based on the sentence structure. Whether a specific sentence is the sentence to be adjusted.

In one embodiment, the method further includes: filtering out a plurality of target words from the plurality of adjacent words according to the plurality of word participles of the plurality of adjacent words; and performing a correlation on the plurality of target words. Analyze and generate the sentence relevance of the multiple adjacent sentences.

In one embodiment, the method further includes: confirming a plurality of lexical correlations between the plurality of target words according to the second language model; and confirming a plurality of lexical correlations between the plurality of target words according to the plurality of lexical correlations. Whether there is at least one set of related words between

In one embodiment, the method further includes: when there is at least one group of related words between the plurality of target words, confirming the sentence correlation between the plurality of adjacent sentences based on the at least one group of related words. ; and when there is no at least one set of related words between the plurality of target words. When, confirm the sentence relevance between the plurality of adjacent sentences, and confirm that one of the plurality of adjacent sentences is the sentence to be adjusted.

In one embodiment, when the number of the at least two adjacent sentences is more than 3, one of the sentences is an intermediate sentence among the plurality of sentences.

In one embodiment, the method further includes: when the text information is confirmed to contain the sentence to be adjusted, determining whether to regenerate a new sentence to replace the sentence to be adjusted or directly delete the sentence to be adjusted based on the text parameters. sentence, wherein the text information and the regenerated new sentence are generated based on a text model and the text parameters.

In one embodiment, the electronic device includes a storage and at least one processor, at least one instruction is stored in the storage, and when the at least one instruction is executed by the at least one processor, the text is automatically generated. method.

The present invention is completed in view of the above problems, and provides a text modification method performed by at least one electronic device and its electronic device.

In order to solve the above problem, the text modification method includes: receiving a text information and a text parameter, wherein the text information includes multiple sentences; confirming whether there is a sentence to be adjusted in the multiple sentences, wherein the confirmation of the sentence to be adjusted is The method includes: using a first language model to obtain a plurality of single-sentence words and their respective part-of-speech parts from a specific sentence among the plurality of sentences, and based on the plurality of part-word parts of speech of the multiple single-sentence words, confirm whether the specific one is The sentence is set as the sentence to be adjusted; and through the first language model, a plurality of adjacent words and their respective parts of speech are obtained from a plurality of adjacent sentences in the plurality of sentences, and based on the plurality of adjacent words Part-of-speech segmentation, confirming a sentence correlation between the plurality of adjacent sentences through a second language model, and confirming whether to set one of the plurality of adjacent sentences as the sentence to be adjusted based on the sentence correlation; And when it is confirmed that the text information contains the sentence to be adjusted, it is determined based on the text parameters whether a new sentence needs to be generated to replace the sentence to be adjusted.

In one embodiment, the method further includes: obtaining a grammatical relationship based on the distribution of the plurality of word participles of the plurality of single-sentence words in the specific sentence; and based on the grammatical relationship, confirming whether the specific sentence is Set as the sentence to be adjusted.

In one embodiment, the method further includes: receiving a plurality of grammatical structures; confirming the sentence structure completeness of the specific sentence based on the multiple grammatical structures and the grammatical relationship; and confirming the sentence structure completeness based on the sentence structure. Whether a specific sentence is the sentence to be adjusted.

In one embodiment, the method further includes: filtering out a plurality of target words from the plurality of adjacent words according to the plurality of word participles of the plurality of adjacent words; and performing a correlation on the plurality of target words. Analyze and generate the sentence relevance of the multiple adjacent sentences.

In one embodiment, the method further includes: confirming a plurality of lexical correlations between the plurality of target words according to the second language model; and confirming a plurality of lexical correlations between the plurality of target words according to the plurality of lexical correlations. Whether there is at least one set of related words between

In one embodiment, the method further includes: when there is at least one group of related words between the plurality of target words, confirming the sentence correlation between the plurality of adjacent sentences based on the at least one group of related words. ; and when there is no at least one group of related words between the multiple target words, confirm the sentence relevance between the multiple adjacent sentences, and confirm that one of the multiple adjacent sentences is The sentence needs to be adjusted.

In one embodiment, when the number of the at least two adjacent sentences is more than 3, one of the sentences is an intermediate sentence among the plurality of sentences.

In one embodiment, the method further includes: when the text information is confirmed to contain the sentence to be adjusted, determining whether to regenerate a new sentence to replace the sentence to be adjusted or directly delete the sentence to be adjusted based on the text parameters. sentence, wherein the text information and the regenerated new sentence are generated based on a text model and the text parameters.

In one embodiment, the electronic device includes a storage and at least one processor, at least one instruction is stored in the storage, and when the at least one instruction is executed by the at least one processor, the text modification method is implemented. .

Through the above method, the automatic text generation method or the text modification method can determine the irregular and unfluent sentences in the text information through the first language model and the second language model, and can confirm that there are inconsistencies between the sentences in the text information. Coherence, a problem that occurs when a model that reinforces text generation lacks sufficient training text. Therefore, the technology of the present invention can analyze and correct sentences that are not smooth enough or have incoherent context, so that the output article can still maintain smooth content while complying with the required theme, thereby reducing the subsequent correction of article content. required time.

On the other hand, through the above method, the automatic text generation method and the text modification method can also select different training texts according to different usage objects and usage scenarios under the first language model and the second language model, thereby making the generated text It can better match the actual required use objects and usage situations. In addition, it is even possible to modify a ready-made text information, so that the text information originally corresponding to one user object can be converted into text information corresponding to another different user object.

10, 20: Text system

11, 21, 31: Electronic devices

111, 211, 311: receiving device

112, 212, 312: Transmission device

12, 22: Network equipment

121, 221: receiving unit

122, 222: Transmission unit

123, 313: Storage

124, 314: Processor

1000: Text processing program

1100:Text generation module

1200: Statement processing module

1210: Intra-sentence processing module

1220: Inter-sentence processing module

231:First processing server

232: Secondary processing server

23n:nth processing server

400, 500, 600, 700: Method

S410-S440, S510-S530, S610-S630, S710-S740: steps

In order to more clearly explain the embodiments of the present application or the technical solutions in the prior art, the drawings needed to be used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings in the following description are only for the purpose of describing the embodiments or the prior art. For some embodiments of the application, those of ordinary skill in the art can also obtain other drawings based on these drawings without exerting creative efforts.

1A is a block diagram of a text system according to the present invention that performs an automatic text generation method and/or a text modification method through at least one electronic device.

FIG. 1B is a block diagram of a network device that performs an automatic text generation method and/or a text modification method according to the present invention.

FIG. 2 is a block diagram of another text system according to the present invention that performs an automatic text generation method and/or a text modification method through at least one electronic device.

FIG. 3 is a block diagram of an electronic device for performing an automatic text generation method and/or a text modification method according to the present invention.

FIG. 4 is a flow chart of a method for automatically generating text through at least one electronic device according to the present invention.

FIG. 5 is a flow chart of a text modification method provided by at least one electronic device according to the present invention.

FIG. 6 is a flow chart of an intra-sentence analysis method provided by the present invention through at least one electronic device.

FIG. 7 is a flow chart of an inter-sentence analysis method provided by the present invention through at least one electronic device.

The following description contains specific information related to exemplary embodiments of the present invention. The drawings and detailed description accompanying this disclosure are merely exemplary embodiments. However, the present invention is not limited to these exemplary embodiments. Other variations and embodiments of the invention will occur to those skilled in the art. Unless otherwise stated, the same or corresponding elements in the drawings may be designated by the same or corresponding reference numerals. Furthermore, the drawings and illustrations in this disclosure are generally not to scale and are not intended to correspond to actual relative sizes.

For purposes of consistency and ease of understanding, identical features are identified by reference numbers in the illustrative drawings (although in some examples they are not). However, features in different embodiments may differ in other respects and therefore should not be narrowly limited to those shown in the drawings.

Terms such as "at least one embodiment", "an embodiment", "multiple embodiments", "different embodiments", "some embodiments", "this embodiment", etc. may indicate the implementation of the invention so described. Modes may include specific features, structures, or characteristics, but not every possible embodiment of the invention must include a particular feature, structure, or characteristic. Furthermore, repeated use of the phrases "in one embodiment" and "in this embodiment" do not necessarily refer to the same embodiment, although they may be. Furthermore, phrases such as "embodiments" used in connection with "the present invention" do not mean that all embodiments of the invention must include a particular feature, structure or characteristic, and should be understood to mean "at least some embodiments of the invention" ” includes the specific features, structures or characteristics described. The term "coupled" is defined as a connection, whether directly or indirectly through intervening components, and is not necessarily limited to physical connections. When the term "includes" is used, it means "including but not limited to," which expressly indicates the open inclusion or relationship of stated combinations, groups, series, and equivalents.

Additionally, for purposes of explanation and not limitation, specific details such as functional entities, technologies, protocols, standards, etc. are set forth to provide an understanding of the described technologies. In other examples, detailed descriptions of well-known methods, techniques, systems, architectures, etc. are omitted to avoid obscuring the narrative with unnecessary detail.

The terms "first", "second" and "third" in the description of the present invention and the above-mentioned drawings are used to distinguish different objects, rather than describing a specific sequence. Furthermore, the term "includes" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or device that consists of a sequence of steps or modules is not limited to the steps listed. steps or modules, but optionally also includes steps or modules not listed, or optionally also includes other steps or modules that are inherent to the process, method, product, or apparatus.

Those skilled in the art will immediately recognize that any computing function or algorithm described in this invention may be implemented by hardware, software, or a combination of software and hardware. The modules corresponding to the described functions may be software, hardware, firmware, or any combination thereof. Software implementations may include computer-executable instructions stored on computer-readable media such as memory or other types of storage. For example, one or more microprocessors or general purpose computers with communications processing capabilities may be programmed with corresponding executable instructions to design and execute the described network functions or algorithms. A processor, microprocessor or general-purpose computer may be formed by an Application Specific Integrated Circuit (ASIC), a programmable logic array, and/or use one or more Digital Signal Processors (DSP). Although several of the exemplary embodiments described in this specification are directed to software installed and executed on computer hardware, alternative exemplary embodiments in which the embodiments are implemented in firmware or hardware or a combination of hardware and software are also contemplated herein. within the scope of the invention.

The storage device can be a computer-readable medium, including but not limited to RAM (Random Access Memory), ROM (Read Only Memory), and EPROM (Erasable Programmable Memory). Read-Only Memory), electrically erasable programmable read-only memory EEPROM (Electrically Erasable Programmable Read-Only Memory), flash memory, compact disc read-only memory CD ROM (Compact Disc Read-Only Memory), magnetic cartridge , magnetic tape, disk memory, or any other equivalent medium capable of storing computer-readable instructions.

The coupling between the devices of the present invention can adopt customized protocols or follow existing standards or de facto standards, including but not limited to Ethernet, IEEE 802.11 or IEEE 802.15 series, wireless USB or telecommunications standards (including but not limited to GSM (Global) System for Mobile Communications, Global System for Mobile Communications), CDMA2000 (Code Division Multiple Access, code division multiple access technology), TD-SCDMA (Time Division-Synchronization Code Division Multiple Access, time division synchronization code division multiple access technology), WiMAX (World Interoperability for Microwave Access, global microwave access interoperability), 3GPP-LTE (Long Term Evolution, long-term evolution technology) or TD-LTE (Time Division Long Term Evolution, time division long-term evolution technology)). In addition, each apparatus of the present invention may also each include any device configured to transmit and/or store data to and receive data from a computer-readable medium. Furthermore, each device of the present invention may include a computer system interface that enables data to be stored on or received from the storage device. For example, each device of the present invention may include a chip set supporting peripheral component interconnect (Peripheral Component Interconnect, PCI) and high-speed peripheral component interconnect (Peripheral Component Interconnect Express, PCIe) bus protocols, dedicated bus protocols, and universal serial bus (Universal serial bus). Serial Bus (USB) protocol, I2C, or any other logical and physical structure that can be used to interconnect peer devices.

The present invention will be described in further detail below with reference to the embodiments of the drawings.

Please refer to FIG. 1A . FIG. 1A is a block diagram of a text system 10 that performs an automatic text generation method and/or a text modification method through at least one electronic device according to the present invention. In one embodiment, the text system 10 includes an electronic device 11 and a network device 12 . The electronic device 11 may further include a receiving device 111 and a transmitting device 112 . In one embodiment, when executing the automatic text generation method or the text modification method, the receiving device 111 may be used to receive information input from a user. In one embodiment, when executing the automatic text generation method or the text modification method, the receiving device 111 and the transmitting device 112 can be coupled with the network device 12 to thereby transmit information from the transmitting device 112 to the network device 12 and receiving data from the network device 12 through the receiving device 111 information. In one embodiment, after the text system 10 completes the automatic text generation method or the text modification method, the electronic device 11 may stop coupling with the network device 12 .

Please refer to FIG. 1B . FIG. 1B is a block diagram of a network device 12 that performs an automatic text generation method and/or a text modification method according to the present invention. The network device 12 may include, but is not limited to, a single network server, a server group composed of multiple network servers, or a cloud composed of a large number of hosts or network servers based on cloud computing. The network where the network device 12 is located includes but is not limited to the Internet, wide area network, metropolitan area network, local area network, virtual private network (Virtual Private Network, VPN), etc. The network device 12 includes a receiving unit 121, a transmitting unit 122, a storage 123 and a processor 124.

The processor 124 and the storage 123 are coupled to each other. The storage 123 stores a plurality of instructions for the processor 124 to execute the automatic text generation method or the text modification method according to the plurality of instructions stored in the storage 123 . In order to complete the automatic text generation method or the text modification method of the present invention, the storage 123 stores a text processing program 1000. In the embodiment, the text processing program 1000 further includes a text generation module 1100 and a sentence processing module 1200. The text generation module 1100 can be used to generate text information including multiple sentences, and generate new sentences to be inserted into the text information. The sentence processing module 1200 can be used to determine whether there are sentences that need to be adjusted among the multiple sentences of the text information, and set the sentences to be adjusted. The sentence processing module 1200 further includes an intra-sentence processing module 1210 and an inter-sentence processing module 1220. The intra-sentence processing module 1210 can analyze each single sentence in the text information to confirm whether each single sentence is set as the sentence to be adjusted. The inter-sentence processing module 1220 can analyze multiple adjacent sentences in the text information to confirm whether there is a sentence between the adjacent sentences that needs to be set as the sentence to be adjusted.

The receiving unit 121 and the transmitting unit 122 may use customized protocols or follow existing standards or actual standards, including but not limited to Ethernet, IEEE 802.11 or IEEE 802.15 series, wireless USB or telecommunications standards (including but not limited to GSM, CDMA2000, TD-SCDMA, WiMAX, 3GPP-LTE or TD-LTE) to transmit data with the receiving device 111 and the transmitting device 112 of the electronic device 11 . In one implementation, the transmitted data may include text information and a plurality of text parameters set for the text information.

Please refer to FIG. 2. FIG. 2 is a block diagram of another text system 20 of the present invention that performs an automatic text generation method and/or a text modification method through at least one electronic device. In one embodiment, the text system 20 includes an electronic device 21, a network device 22, and at least one processing server. In one embodiment, the number of the at least one processing server may be n, and the value n may be a positive integer greater than zero, such as 1, 2, 3, or the like. In one embodiment, the at least one processing server may include a first processing server 231, a second processing server 232, ... and an nth processing server 23n.

The electronic device 21 may further include a receiving device 211 and a transmitting device 212 . In one embodiment, when executing the automatic text generation method or the text modification method, the receiving device 211 may be used to receive information input from a user. In one embodiment, when executing the automatic text generation method or the text modification method, the receiving device 211 and the transmitting device 212 can be coupled with the network device 22, thereby transmitting information from the transmitting device 212 to the network device 22. and receiving information from the network device 22 through the receiving device 211. In one embodiment, after the text system 20 completes the automatic text generation method or the text modification method, the electronic device 21 may stop coupling with the network device 22 .

The network device 22 may further include a receiving unit 221 and a transmitting unit 222. In one embodiment, when executing the automatic text generation method or the text modification method, the receiving unit 221 may be used to receive user input information from the electronic device 21 . In one embodiment, when executing the automatic text generation method or the text modification method, the receiving unit 221 and the transmitting unit 222 may be coupled to the at least one processing server, thereby transmitting information from the transmitting unit 222 to The at least one processing server receives information from the at least one processing server through the receiving unit 221 . In one embodiment, the transmitting unit 222 can also transmit the processing results of the automatic text generation method or the text modification method to the electronic device 21 . In one embodiment, after the text system 20 completes the text automatic generation method or the text modification method, the network device 22 may stop coupling with the electronic device 21 and the at least one processing server.

The at least one processing server may each include a corresponding processing component (not shown), a storage component (not shown), a receiving component (not shown), and a transmitting component (not shown). In one embodiment, when executing the text automatic generation method or the text modification method, each receiving element and transmitting element of the at least one processing server can be coupled to the network device 22, and thereby receive the network device 22. 22 and return the processed results to the network device 22. In addition, after the respective receiving components of the at least one processing server receive the information from the network device 22, each processing component will use the received information to generate the corresponding processing program according to the processing program stored in the corresponding storage component. Process the results. In one embodiment, after the at least one processing server completes its respective processing and returns the processing results to the network device 22, the at least one processing server may stop coupling with the network device 22.

Please refer to FIG. 1B and FIG. 2 together. The text processing program 1000 can be split into multiple subroutines, and the multiple subroutines are dispersed in the network device 22 and the at least one processing server. In one embodiment, the network device 22 also has a program distribution module (not shown). The program allocation module is used to allocate which receiving object the transmission unit 222 should transmit the required transmission information to according to the distribution status of the multiple subprograms.

In one embodiment, the subroutine of the network device 22 only includes the program distribution module, and the at least one processing server may each include a text generation module 1100, an intra-sentence processing module 1210, and an inter-sentence processing module. At least one of 1220. For example, when the value n is equal to 3, the subroutine of the first processing server 231 includes the text generation module 1100, and the second processing server 231 The subroutine of 232 includes an intra-sentence processing module 1210, and the subroutine of the nth processing server 23n includes an inter-sentence processing module 1220. In another embodiment, when the value n is equal to 2, the subroutine of the first processing server 231 includes the text generation module 1100, and the subroutine of the nth processing server 23n includes the intra-sentence processing module 1210 and the sentence processing module 1210. Intermediate processing module 1220. In other words, the subroutine of the nth processing server 23n includes the statement processing module 1200. In yet another embodiment, when the value n is equal to 2, the subroutine of the first processing server 231 includes the text generation module 1100 and the intra-sentence processing module 1210, and the subroutine of the nth processing server 23n includes The inter-sentence processing module 1220, or the subroutine of the n-th processing server 23n includes both the text generation module 1100 and the inter-sentence processing module 1220, whereby the intra-sentence processing module 1210 and the inter-sentence processing module 1220 are configured in different After the processing server completes the judgment, it can directly adjust the text information through its corresponding text generation module 1100. In addition, in one embodiment, when the value n is equal to 1, the subroutine of the first processing server 231 includes a text generation module 1100, an intra-sentence processing module 1210, and an inter-sentence processing module 1220. In other words, the subroutine of the first processing server 231 includes the text generation module 1100 and the sentence processing module 1200.

In one embodiment, the subroutine of the network device 22 may also include other modules in addition to the program distribution module. In other words, the text generation module 1100, the intra-sentence processing module 1210, and the inter-sentence processing module 1220 can each be distributed in the network device 22 and the at least one processing server. For example, when the value n is equal to 2, the subroutine of the network device 22 includes the text generation module 1100, and the subroutine of the first processing server 231 includes the intra-sentence processing module 1210, or the first processing server The subroutine of 231 includes both the text generation module 1100 and the intra-sentence processing module 1210, the subroutine of the nth processing server 23n includes the inter-sentence processing module 1220, or the subroutine of the nth processing server 23n simultaneously It includes a text generation module 1100 and an inter-sentence processing module 1220. In another embodiment, when the value n is equal to 1, the subroutine of the network device 22 includes the text generation module 1100, and the subroutine of the first processing server 231 includes a statement processing module. 1200, or the subroutine of the first processing server 231 includes both the text generation module 1100 and the statement processing module 1200. In yet another embodiment, when the value n is equal to 1, the subroutine of the network device 22 includes the statement processing module 1200, and the subroutine of the first processing server 231 includes the text generation module 1100. In all the above embodiments, it only means that each module in the text processing program 1000 can be detached and distributed to the network device 22 and the at least one processing server. However, the actual dispersion method is not limited to the above-exemplified dispersion method, and other dispersion methods can also be used to distribute the load of each module on different processing devices.

In one implementation, each module in the text processing program 1000 may also be processed by the electronic device 21 itself. In one embodiment, the electronic device 21 has a storage device (not shown) to store at least one of the subroutines. For example, the electronic device 21 may have a text processing program 1000 to generate the text information by itself and transmit it to the network device 22 through the transmission device 212 . In another embodiment, the electronic device 21 may have at least one module among the sentence processing modules 1200 to perform sentence processing on the text information provided from the network device 22 or other sources.

Please refer to FIG. 3. FIG. 3 is a block diagram of an electronic device 31 that performs an automatic text generation method and/or a text modification method according to the present invention. In one embodiment, the electronic device 31 includes a mobile phone, a tablet computer, a desktop computer, a notebook computer, or other electronic equipment, which is not limited here. The electronic device 31 includes a receiving device 311 , a transmitting device 312 , a storage 313 and a processor 314 .

The processor 314 and the storage 313 are coupled to each other. The storage 313 stores a plurality of instructions for the processor 314 to execute the automatic text generation method or the text modification method according to the plurality of instructions stored in the storage 313 . Please refer to FIG. 1B and FIG. 3 together. In order to complete the automatic text generation method or the text modification method of the present invention, the storage 313 can separately store the text processing program 1000. In the embodiment, the text processing program 1000 shown in FIG. 3 is the same as the text processing program 1000 shown in FIG. 1B. This handler is the same as 1000. That is to say, the electronic device 31 can complete the automatic text generation method or the text modification method by itself through the text processing program 1000 in the memory.

The receiving device 311 and the transmitting device 312 may utilize customized protocols or follow existing standards or actual standards, including but not limited to Ethernet, IEEE 802.11 or IEEE 802.15 series, wireless USB or telecommunications standards (including but not limited to GSM, CDMA2000, TD -SCDMA, WiMAX, 3GPP-LTE or TD-LTE) to transmit data with other electronic devices. In one embodiment, the transmitted data may include multiple language models required by the automatic text generation method or the text modification method, or multiple model training sets required for training the multiple language models. In one embodiment, the receiving device 311 can also receive input information from the user.

FIG. 4 is a flow chart of a method 400 for automatically generating text through at least one electronic device according to the present invention. Because there are multiple ways to perform the automatic text generation method 400, the automatic text generation method 400 shown in FIG. 4 is only an example. The automatic text generation method 400 can be executed using the configuration shown in FIGS. 1A, 1B, 2, and 3 or other configurations. While describing the automatic text generation method 400, please refer to FIGS. 1A, 1B, and FIG. 2 and the various components in Figure 3. Each step shown in Figure 4 may represent one or more processes, methods or subroutines executed, and the order of each step may be adjusted arbitrarily, without making the essence of the automatic text generation method 400 deviate from the automatic text generation method. 400 range of technical solutions.

In step S410, multiple text parameters are received.

In one embodiment, the user inputs text parameters into the electronic device so that the text generation module 1100 can generate text information based on the text parameters. In one embodiment, when the text generation module 1100 is located in the network device 12, the electronic device 11 can receive the text parameters through the receiving device 111 and provide them to the network device 12 through the transmitting device 112. In another embodiment, when the text generation module 1100 is located in the first processing server 231, the electronic device 21 can receive the text parameter through the receiving device 211 and provide it to the network device through the transmitting device 212. 22 of the receiving unit 221 and provided to the first processing server 231 through the transmitting unit 222. In another embodiment, when the text generation module 1100 is located on the electronic device 31 itself, the electronic device 31 can receive the text parameters through the receiving device 311 and then perform subsequent steps.

In step S420, a text information is generated according to the plurality of text parameters, wherein the text information includes a plurality of sentences.

In one embodiment, after the text generation module 1100 receives the text parameter, the text generation module 1100 generates the text information based on the text parameter and through a text model. In one implementation, the text model can be established through a first neural network architecture, and the first neural network architecture can be a Transformer-based Model. For example, the first neural network architecture may be a generative pre-trained transformer (GPT) model, such as GPT-1, GPT-2, and GPT-3. In one implementation, the text parameter may be a keyword, a key sentence, or a required number of words in the text, which are used to limit the content or number of words of the generated text information.

In one implementation, the text model can be trained by providing a first training data set to the first neural network architecture in advance. In one implementation, the first training data set can be prepared by selecting the required topics in advance according to the user's needs. In one implementation, if the selected topic has abundant information, a more complete text model can be trained, so that the text model can generate a more complete and fluent article. However, if the selected topic contains less information, it will be difficult for the trained text model to produce a complete and smooth article.

In one embodiment, even if the selected topic information is small and the trained text model cannot generate a complete and smooth article, the automatic text generation method 400 can still rely on the judgment and adjustment of subsequent steps to generate a complete and smooth article. Shun article. In one implementation, if the selected topic has rich data, the trained text model can also generate a complete and fluent article. At this time, the automatic text generation method 400 can still rely on the judgment and adjustment in subsequent steps to further adjust the generated text information into another form to adapt to another text requirement.

In step S430, it is confirmed whether there is a sentence to be adjusted among the plurality of sentences.

In one embodiment, after the sentence processing module 1200 receives the text information, the sentence processing module 1200 determines whether there are any errors in the sentences that need to be adjusted based on the first language model and/or the second language model. sentence. In one implementation, the first language model can be established through a second neural network architecture, and the second language model can be established through a third neural network architecture. In one embodiment, the second neural network architecture may be a Recurrent Neural Network (RNN), and the third neural network architecture may be a Recurrent Neural Network (NN).

In one implementation, the sentence processing module 1200 can analyze each sentence in the text information through the first language model. In one implementation, the sentence processing module 1200 can analyze each sentence one by one through the first language model. If one of the sentences is confirmed to be the sentence to be adjusted, the sentence to be adjusted can be deleted directly, or a sentence to be adjusted can be regenerated through the text generation module 1100 to replace the sentence to be adjusted. In one implementation, the sentence processing module 1200 can still use the first language model to determine again whether the regenerated sentence needs to be set as the sentence to be adjusted.

In one embodiment, the sentence processing module 1200 can analyze multiple adjacent sentences in the text information through the second language model. In one implementation, the plurality of adjacent sentences may be two adjacent sentences. In another embodiment, the plurality of adjacent sentences may be three or more consecutive sentences. When the sentence processing module 1200 confirms that the sentence meanings of the multiple adjacent sentences are discontinuous, the sentence processing module 1200 can set one of the sentences as the sentence to be adjusted, and then directly delete the sentence to be adjusted, or through the text The generation module 1100 regenerates a sentence to replace the sentence to be adjusted. In one embodiment, when the plurality of adjacent sentences are two adjacent sentences At this time, the sentence processing module 1200 can preset that the previous sentence or the following sentence should be selected as the sentence to be adjusted. In one embodiment, when the plurality of adjacent sentences are three consecutive sentences, the sentence processing module 1200 can preset to select the previous sentence, the middle sentence or the following sentence as the sentence to be adjusted. In one embodiment, the sentence processing module 1200 can still use the first language model and/or the second language model to determine again whether the regenerated sentence needs to be set as the sentence to be adjusted.

In one embodiment, the visual text requirement of the sentence processing module 1200 only uses the first language model to determine whether a single sentence needs to be modified. In another embodiment, the visual text requirement of the sentence processing module 1200 only uses the second language model to determine whether correction is needed between multiple adjacent sentences. In yet another embodiment, the sentence processing module 1200 uses the first language model and the second language model sequentially according to text requirements to determine whether the text information needs to be modified.

In one embodiment, if the sentence processing module 1200 uses the first language model and the second language model to determine whether the text information needs to be modified, the sentence processing module 1200 can first use the first language model to determine one by one. Whether each sentence itself needs to be modified, after the multiple sentences of the text information have completed the judgment of the first language model, the sentence processing module 1200 can then use the second language model to judge one by one whether there are differences between adjacent sentences. Needs correction.

In another embodiment, if the sentence processing module 1200 uses the first language model and the second language model to determine whether the text information needs to be modified, the sentence processing module 1200 can first use the first language model to determine Whether the first two sentences themselves need to be modified, then the sentence processing module 1200 can then use the second language model to determine whether the first two sentences need to be modified. Then, the sentence processing module 1200 can first use the first language model to determine whether the next sentence after the previous two sentences needs to be modified, and then use the second language model to determine the latter of the previous two sentences. Is there any need to modify between the sentence and the next sentence? just. In the above embodiment, intra-sentence analysis and inter-sentence analysis can be used alternately to make the text information complete and smooth.

In step S440, when it is confirmed that the sentence to be adjusted exists in the text information, the sentence to be adjusted is adjusted to complete the text information.

In one embodiment, when the text information is confirmed to contain the sentence to be adjusted, the text generation module 1100 can directly delete the sentence to be adjusted, or generate a new sentence to replace the sentence to be adjusted, to complete the process. Text information.

In one embodiment, if the text generation module 1100 wants to generate a new sentence to replace the sentence to be adjusted, the text generation module 1100 can directly delete the sentence to be adjusted, and then the text generation module 1100 further based on the text The parameter determines whether the new sentence needs to be generated. In one embodiment, if the text parameter includes the required word count of the text, the text generation module 1100 may determine whether the difference between the word count of the text information after deleting the sentence to be adjusted and the required word count of the text exceeds A preset word count threshold. In one embodiment, if the difference exceeds the preset word count threshold, the text generation module 1100 can use at least one sentence before and/or at least one sentence after the sentence to be adjusted as a key sentence, and use the text model to generate The new sentence replaces the original sentence to be adjusted. In another embodiment, the key sentence may include the first two sentences or the last two sentences of the sentence to be adjusted. In another embodiment, the key sentence may include the first two sentences and the last sentence of the sentence to be adjusted. If the difference does not exceed the preset word count threshold, the text generation module 1100 may not generate a new sentence.

In one embodiment, after the text information is adjusted to form a text file, the text file can be sent back to the electronic device. In one embodiment, if the text information is completed on the network device 12, the network device 12 returns it to the electronic device 11. If the text information is completed on the nth processing server 23n, the nth processing server 23n will first send it back to the network device 22, thereby transmitting back to the electronic device 21. If the text information is completed on the electronic device 31, the electronic device 31 directly obtains the text file.

In one embodiment, when the new sentence is generated, the new sentence still needs to be analyzed by the sentence processing module 1200 again. Therefore, after analysis by the sentence processing module 1200, it is determined that the entire content of the text file no longer has the content to be adjusted. When the sentence is executed, the text information is adjusted and the text file is formed.

In the above manner, the automatic text generation method 400 can determine the irregular and unfluent sentences in the generated text information through the first language model and/or the second language model, and can confirm the sentences of the text information. Problems that occur when the model of text generation lacks sufficient training text to reinforce the incoherence between text.

FIG. 5 is a flow chart of a text modification method 500 provided by the present invention through at least one electronic device. Because there are multiple ways to perform the text modification method 500, the text modification method 500 shown in Figure 5 is only an example. The text modification method 500 may be executed using the configuration shown in FIGS. 1A, 1B, 2 and 3 or other configurations. While describing the text modification method 500, please refer to FIGS. 1A, 1B, 2 and 3. Various components in Figure 3. Each step shown in FIG. 5 may represent one or more processes, methods or subroutines that are executed, and the order of each step may be adjusted arbitrarily without departing from the essence of the text modification method 500 . Scope of technical solutions.

In step S510, a text parameter and text information including a plurality of sentences are received.

In one embodiment, when the user wants to modify the text information originally held, the user can input the text parameters to the electronic device, so that the text generation module 1100 can adjust the text information according to the text parameters. In one embodiment, when the text generation module 1100 is located in the network device 12, the electronic device 11 can receive the text parameter through the receiving device 111, and through The transmitting device 112 provides the text parameters and the text information to the network device 12 . In another embodiment, when the text generation module 1100 is located in the first processing server 231, the electronic device 21 can receive the text parameters through the receiving device 211, and provide the text parameters and the text through the transmitting device 212. The information is provided to the receiving unit 221 of the network device 22, and the text parameter and the text information are provided to the first processing server 231 through the transmitting unit 222. In another embodiment, when the text generation module 1100 is located on the electronic device 31 itself, the electronic device 31 can receive the text parameters through the receiving device 311 and then perform subsequent steps.

In step S520, it is confirmed whether there is a sentence to be adjusted among the plurality of sentences.

In one implementation, step S520 may be completely identical to step S430 of the automatic text generation method 400. In one embodiment, after the sentence processing module 1200 receives the text information, the sentence processing module 1200 determines whether there are any errors in the sentences that need to be adjusted based on the first language model and/or the second language model. sentence. In one implementation, the first language model can be established through a second neural network architecture, and the second language model can be established through a third neural network architecture. In one embodiment, the second neural network architecture may be a Recurrent Neural Network (RNN), and the third neural network architecture may be a Recurrent Neural Network (NN).

In one embodiment, the sentence processing module 1200 can analyze each sentence in the text information through the first language model. In one implementation, the sentence processing module 1200 can analyze each sentence one by one through the first language model. If one of the sentences is confirmed to be the sentence to be adjusted, the sentence to be adjusted can be deleted directly, or a sentence to be adjusted can be regenerated through the text generation module 1100 to replace the sentence to be adjusted. In one implementation, the sentence processing module 1200 can still use the first language model to determine again whether the regenerated sentence needs to be set as the sentence to be adjusted.

In one embodiment, the sentence processing module 1200 can analyze multiple adjacent sentences in the text information through the second language model. When the sentence processing module 1200 confirms that the sentence meanings of the multiple adjacent sentences are discontinuous, the sentence processing module 1200 can set one of the sentences as the sentence to be adjusted, and then directly delete the sentence to be adjusted, or through the text The generation module 1100 regenerates a sentence to replace the sentence to be adjusted. In one embodiment, the sentence processing module 1200 can still use the first language model and/or the second language model to determine again whether the regenerated sentence needs to be set as the sentence to be adjusted.

In step S530, when it is confirmed that the sentence to be adjusted exists in the text information, it is determined based on the text parameters whether a new sentence needs to be generated to replace the sentence to be adjusted.

In one embodiment, when the text information is confirmed to contain the sentence to be adjusted, the text generation module 1100 can determine whether to directly delete the sentence to be adjusted based on the text parameters, or to generate a new sentence to replace the sentence. The sentence needs to be adjusted to complete the text information.

In one embodiment, if the text parameter includes a text required word count, the text generation module 1100 may determine whether the difference between the word count of the text information after deleting the sentence to be adjusted and the text required word count exceeds A preset word count threshold. In one embodiment, if the difference exceeds the preset word count threshold, the text generation module 1100 can use at least one sentence before and/or at least one sentence after the sentence to be adjusted as a key sentence, and use the text model to generate The new sentence replaces the original sentence to be adjusted. In another embodiment, the key sentence may include the first two sentences or the last two sentences of the sentence to be adjusted. In another embodiment, the key sentence may include the first two sentences and the last sentence of the sentence to be adjusted. If the difference does not exceed the preset word count threshold, the text generation module 1100 may not generate a new sentence.

In one embodiment, after the text information is adjusted to form a text file, the text file can be sent back to the electronic device. In one embodiment, if the text information is completed on the network device 12, the network device 12 returns it to the electronic device 11. If the text information is at the nth position If the processing server 23n completes, the nth processing server 23n will first send back to the network device 22, thereby sending back to the electronic device 21. If the text information is completed in the electronic device 31, the electronic device 31 directly obtains the text file.

In one embodiment, when the new sentence is generated, the new sentence still needs to be analyzed by the sentence processing module 1200 again. Therefore, after analysis by the sentence processing module 1200, it is determined that the entire content of the text file no longer has the content to be adjusted. When the sentence is executed, the text information is adjusted and the text file is formed.

FIG. 6 is a flow chart of an intra-sentence analysis method 600 performed by at least one electronic device according to the present invention. Because there are multiple ways to perform the intra-sentence analysis method 600, the intra-sentence analysis method 600 shown in Figure 6 is only an example. The intra-sentence analysis method 600 can be executed using the configuration shown in FIGS. 1A, 1B, 2, and 3 or other configurations. While describing the intra-sentence analysis method 600, please refer to FIGS. 1A, 1B, and FIG. 2 and the various components in Figure 3. Each step shown in Figure 6 may represent one or more processes, methods, or subroutines that are executed, and the order of each step may be adjusted arbitrarily without departing from the essence of the intra-sentence analysis method 600. 600 range of technical solutions. In one embodiment, the intra-sentence analysis method 600 can be used in step S430 of the automatic text generation method 400. In another implementation, the intra-sentence analysis method 600 may be used in step S520 of the text modification method 500.

In step S610, text information including a plurality of sentences is received.

In one embodiment, whether it is the automatic text generation method 400 or the text modification method 500, the intra-sentence processing module 1210 in the sentence processing module 1200 must first obtain the text information. In one embodiment, when implementing the automatic text generation method 400, the text information is generated through the text generation module 1100 and based on a text parameter and a text model. In one embodiment, when implementing the text modification method 500, the text information may be a text generation module The text information generated by 1100 can also be text information taken from the Internet, or text information input by the user.

In one embodiment, if the intra-sentence processing module 1210 is located in the network device 12, the text information can be generated by the text generation module 1100 in the network device 12, or directly provided through the electronic device 11. In another embodiment, if the intra-sentence processing module 1210 is located in one of the at least one processing server (for example, the second processing server 232), the text information can be generated by the electronic system where the text generation module 1100 is located. Devices (such as the network device 22 and one of the at least one processing server) are generated, or provided directly by the electronic device 21 through the network device 22 . In another embodiment, if the intra-sentence processing module 1210 is located in the electronic device 31 , the text information can be generated by the text generation module 1100 or directly obtained from the electronic device 31 .

In step S620, through a first language model, a plurality of single-sentence words and their respective part-of-speech parts are obtained from a specific sentence among the plurality of sentences.

In one embodiment, the intra-sentence processing module 1210 analyzes the multiple sentences one by one to confirm whether there are sentences to be adjusted in the multiple sentences. Therefore, the intra-sentence processing module 1210 may first analyze the first sentence among the plurality of sentences, and then analyze the second sentence until all sentences are analyzed and the text information is adjusted.

In one embodiment, the intra-sentence processing module 1210 can perform word segmentation processing on the specific sentence through the first language model, thereby obtaining the plurality of single-sentence words in the specific sentence. In addition, the intra-sentence processing module 1210 can obtain the part-of-speech of each of the plurality of single-sentence words at the same time through the first language model.

In one implementation, the first language model can be established through a neural network architecture. In one implementation, the neural network architecture may be a Recurrent Neural Network (RNN). In one embodiment, the neural network architecture can integrate Electra Small model, biaffine (BiAffine) model and conditional random field CRF (Condition Random Field) model to obtain the multiple single-sentence words and their respective word segmentation parts of speech.

In one implementation, the first language model can be trained by providing a language training data set to the neural network architecture in advance. In one implementation, the language training data set can select the required topics for training in advance based on the text information expected to be received later, and a large amount of training texts with no limit on topics are still required for the neural network architecture to train. , so that the first language model can correctly perform word segmentation and give correct part-of-speech judgments to the plurality of single-sentence words generated by the word segmentation results.

In one embodiment, if the text information is generated through a text model of the automatic text generation method 400, a text training set used by the text model can be directly used as a part of the language training data set to facilitate the first The language model identifies the words used in the topic of the text information. In one embodiment, the additional large amount of training texts included in the language training data set are not limited to the content of the topic. With the assistance of a large amount of training text, the first language model can correctly understand the meaning and part-of-speech of each word in the sentence structure without being restricted by the subject of the text information when identifying words and determining part-of-speech.

In one embodiment, the plurality of participle parts of speech of the plurality of single-sentence words may include subjects (agent subjects or subjects), verbs, and objects (object subjects or objects). In one embodiment, the plurality of participle parts of speech may further include parts of speech with grammatical relationships such as predicate, object, predicate, attributive, adverbial, complement, and head.

In step S630, it is confirmed whether the specific sentence is set as the sentence to be adjusted based on the plurality of word participles of the plurality of single-sentence words.

In one embodiment, the intra-sentence processing module 1210 obtains a grammatical relationship of the specific sentence based on the distribution of the plurality of word participles of the plurality of single-sentence words in the specific sentence. For example: the order of the plurality of participles of a specific sentence is subject, verb When words, objects (indirect objects) and objects (direct objects) are used, the grammatical relationship of a specific sentence can be represented by subject, verb, object and object. In one embodiment, the grammatical relationship of the specific sentence can be used to confirm whether the specific sentence is set as the sentence to be adjusted.

In one embodiment, the intra-sentence processing module 1210 may receive multiple grammatical structures. In one embodiment, the plurality of grammatical structures are preset allowed grammatical relationships. Therefore, the intra-sentence processing module 1210 can confirm the completeness of the sentence structure of the specific sentence based on a comparison result between the grammatical relationship of the specific sentence and the plurality of grammatical structures. In one embodiment, the intra-sentence processing module 1210 can determine whether the specific sentence is the sentence to be adjusted based on the completeness of the sentence structure.

In one embodiment, when the multiple grammatical structures are different from the grammatical relationship of the specific sentence, the intra-sentence processing module 1210 may determine that the completeness of the sentence structure of the specific sentence is insufficient, and set the specific sentence. One sentence is the sentence to be adjusted. In one embodiment, when one of the plurality of grammatical structures is the same as the grammatical relationship of the specific sentence, the intra-sentence processing module 1210 may determine that the specific sentence has sufficient completeness of the sentence structure, and It is determined that the specific sentence is not the sentence to be adjusted.

In one embodiment, when the grammatical relationship between the multiple grammatical structures and the specific sentence is different, the intra-sentence processing module 1210 can select one of the multiple grammatical structures that is consistent with the specific sentence. A specific grammatical structure with the closest grammatical relationship, and compares the grammatical relationship of the specific sentence with the specific grammatical structure. If the difference between the two is only a function word, the intra-sentence processing module 1210 can determine that the specific grammatical structure A sentence has sufficient structural integrity of the sentence, and it is determined that the specific sentence is not the sentence to be adjusted. On the contrary, if the difference between the two is more than one function word, the intra-sentence processing module 1210 may determine that the sentence structure of the specific sentence is insufficient and set the specific sentence as the sentence to be adjusted. In the embodiment, the function word may be an adjective, an adverb, a preposition, an auxiliary, an interjection, etc. that serve as attributives or adverbials.

In one implementation, the preset settings of the plurality of grammatical structures may also be determined based on the requirements of the text information. For example: if the text information is required to be used in children's books, then the multiple grammatical structures selected can be composed of the simplest multiple grammatical structures. In addition, if the text in a specific field requires supplements, the idiomatic grammatical structures in the specific field can also be supplemented (especially the grammatical structures that are not commonly used, but are idiomatic in the specific field) to prevent the text information from being included in the text. Sentences were marked inappropriately.

In one embodiment, when the specific sentence is determined not to be the sentence to be adjusted, the intra-sentence processing module 1210 can continue to confirm the next sentence until the multiple sentences of the text information are all confirmed. In one implementation, when the specific sentence is determined to be the sentence to be adjusted, the sentence to be adjusted can be directly deleted, or a new sentence can be generated to replace the sentence to be adjusted to complete the text information.

In one embodiment, when the specific sentence is determined to be the sentence to be adjusted, the intra-sentence processing module 1210 may provide the sentence to be adjusted to a text generation module 1100. The text generation module 1100 may determine whether the difference between the word count of the text information after deleting the sentence to be adjusted and the word count of a text requirement exceeds a preset word count threshold. In one embodiment, if the difference exceeds the preset word count threshold, the text generation module 1100 can use at least one sentence before and/or at least one sentence after the sentence to be adjusted as a key sentence, and use the text model to generate The new sentence replaces the original sentence to be adjusted. If the difference does not exceed the preset word count threshold, the text generation module 1100 may not generate a new sentence.

In one embodiment, when the new sentence is generated, the new sentence still needs to be analyzed by the intra-sentence processing module 1210 again. Therefore, the intra-sentence processing module 1210 analyzes and determines that the new sentence does not need to be set as the pending sentence. When adjusting a sentence, the intra-sentence processing module 1210 may continue to analyze subsequent sentences in the text information.

In one embodiment, the text parameters may include the required word count of the text, the preset word count threshold, and/or a structure indication indicating which grammatical structures are preset, etc. The text parameters may be provided to the in-sentence processing module 1210 through the electronic device.

FIG. 7 is a flow chart of an inter-sentence analysis method 700 provided by the present invention through at least one electronic device. Because there are multiple ways to perform the inter-sentence analysis method 700, the inter-sentence analysis method 700 shown in FIG. 7 is only an example. The inter-sentence analysis method 700 can be executed using the configuration shown in FIGS. 1A, 1B, 2, and 3 or other configurations. While describing the inter-sentence analysis method 700, please refer to FIGS. 1A, 1B, and FIG. 2 and the various components in Figure 3. Each step shown in Figure 7 may represent one or more processes, methods or subroutines executed, and the order of each step may be adjusted arbitrarily without departing from the essence of the inter-sentence analysis method 700. 700 range of technical solutions. In one implementation, the inter-sentence analysis method 700 can be used in step S430 of the automatic text generation method 400. In another implementation, the inter-sentence analysis method 700 may be used in step S520 of the text modification method 500. In another embodiment, the inter-sentence analysis method 700 can be used sequentially or alternately with the intra-sentence analysis method 600 to form a sentence analysis method.

In step S710, text information including a plurality of sentences is received.

In one embodiment, whether it is the automatic text generation method 400 or the text modification method 500, the inter-sentence processing module 1220 in the sentence processing module 1200 must first obtain the text information. In one embodiment, when implementing the automatic text generation method 400, the text information is generated through the text generation module 1100 and based on a text parameter and a text model. In one embodiment, when implementing the text modification method 500, the text information may be generated by the text generation module 1100, may be text information obtained from the Internet, or may be text information input by the user.

In one embodiment, if the inter-sentence processing module 1220 is located in the network device 12, the text information can be generated by the text generation module 1100 in the network device 12, or directly provided through the electronic device 11. In another embodiment, if the inter-sentence processing module 1220 is located in one of the at least one processing server (for example: the second processing server 232), then the text information can be generated by the electronic system where the text generation module 1100 is located. Devices (such as the network device 22 and one of the at least one processing server) are generated, or provided directly by the electronic device 21 through the network device 22 . In another embodiment, if the inter-sentence processing module 1220 is located in the electronic device 31 , the text information can be generated by the text generation module 1100 or directly obtained from the electronic device 31 .

In step S720, a first language model is used to obtain a plurality of adjacent words and respective part-of-speech parts from a plurality of adjacent sentences in the plurality of sentences.

In one implementation, the inter-sentence processing module 1220 will gradually analyze the multiple sentences to confirm whether there are sentences to be adjusted in the multiple sentences. Therefore, the inter-sentence processing module 1220 may first analyze the first few adjacent sentences in the plurality of sentences, and then remove the first sentence from the first few adjacent sentences and add the subsequent sentences. The analysis is performed again until all sentences have been analyzed by the inter-sentence processing module 1220, and the adjustment of the text information is completed.

In one embodiment, the inter-sentence processing module 1220 can perform word segmentation processing on the adjacent sentences through the first language model, thereby obtaining the adjacent words in the adjacent sentences. In addition, the inter-sentence processing module 1220 can simultaneously obtain the part-of-speech of each of the plurality of adjacent words through the first language model. In one embodiment, when the intra-sentence analysis method 600 and the inter-sentence analysis method 700 are executed together, the plurality of single-sentence vocabulary and corresponding part-of-speech words obtained by the intra-sentence processing module 1210 can be directly stored in a temporary register. (not shown), so that the inter-sentence processing module 1220 directly corresponds to the multiple adjacent sentences to determine the multiple adjacent words and their part-of-speech participles.

In one implementation, the first language model can be established through a neural network architecture. In one embodiment, the neural network architecture may be a recurrent neural network RNN. In one implementation, the neural network architecture can integrate the Electra Small model, the BiAffine model and the CRF model to obtain the plurality of adjacent words and their respective word segmentation parts of speech.

In one embodiment, the first language model used in the intra-sentence analysis method 600 may be the same as the first language model used in the inter-sentence analysis method 700 . In other words, the language training data set used in the intra-sentence analysis method 600 may be the same as the first language training data set used by the first language model in the inter-sentence analysis method 700 . In one implementation, the first language training data set can be trained in advance by selecting the required topics based on the text information expected to be received later. A large amount of training texts with no limit on topics are still required for the neural network architecture. Training is performed to facilitate the first language model to correctly perform word segmentation and provide correct part-of-speech judgments for the plurality of adjacent words generated as a result of the word segmentation.

In one embodiment, if the text information is generated through a text model of the automatic text generation method 400, a text training set used by the text model can be directly used as a part of the first language training data set. In one embodiment, the additional large amount of training text included in the first language training data set is not limited to the content of the topic. With the assistance of a large amount of training text, the first language model can correctly understand the meaning and part-of-speech of each word in the sentence structure without being restricted by the subject of the text information when identifying words and determining part-of-speech.

In one embodiment, the plurality of participle parts of speech of the plurality of adjacent words may include subjects (agent subjects or subjects), verbs, and objects (object subjects or objects). In one embodiment, the plurality of participle parts of speech may further include parts of speech with grammatical relationships such as predicate, object, predicate, attributive, adverbial, complement, and head.

In step S730, a sentence correlation between the plurality of adjacent sentences is confirmed through a second language model based on the plurality of word participles of the plurality of adjacent words.

In one embodiment, the inter-sentence processing module 1220 selects a plurality of target words from the plurality of adjacent words according to the plurality of word participles of the plurality of adjacent words, and performs a correlation on the plurality of target words. Analyze and generate the sentence relevance of the multiple adjacent sentences.

In one embodiment, the inter-sentence processing module 1220 may pre-store a plurality of desired parts of speech, and compare the plurality of word participles of the plurality of adjacent words with the plurality of desired parts of speech. When the plurality of word participles of the plurality of adjacent words have the same situation as the plurality of desired parts of speech, the plurality of adjacent words belonging to the plurality of desired parts of speech are regarded as the plurality of target words. For example: assuming that the plurality of participle parts of speech of the plurality of adjacent words include subjects, predicates, objects, attributives and adverbials, and the plurality of desired parts of speech are subjects, verb predicates and objects, then the inter-sentence processing module 1220 From the plurality of adjacent words, words serving as subjects, verb predicates and objects are regarded as the plurality of target words. In one embodiment, the plurality of desired parts of speech can be subjects and objects.

In one embodiment, the inter-sentence processing module 1220 confirms a plurality of word correlations between the plurality of target words based on the second language model, and confirms a plurality of word correlations between the plurality of target words based on the plurality of word correlations. Whether there is at least one set of related words between In one implementation, the second language model can also be established through another neural network architecture. In one implementation, the other neural network architecture may be a recurrent neural network NN (Neural Network). In one embodiment, the other neural network architecture can use a Skip-Gram model to obtain whether there is a word-to-word association between the plurality of adjacent words.

In one embodiment, the first language training data set used in the first language model may be the same as the second language training data set used in the second language model. In other words, the second language training data set used by the second language model can not only select the required topics for training in advance based on the text information expected to be received subsequently, but also needs to provide a large amount of training texts with no limit on topics. The neural network architecture is trained to facilitate the second language model to correctly A correlation judgment is made, and a correct judgment is made on the plurality of vocabulary correlations between the plurality of target words.

In one implementation, depending on the usage objects and usage occasions of the text information, the first language training data set used by the first language model may be slightly different from the second language training data set used by the second language model. . If the object and occasion of use of the text information belong to formal occasions in a special professional and technical field, the second language training data set needs more training texts in the special professional and technical field than the first language training data set to facilitate Correctly determine the multiple word correlations between the multiple target words (especially when the source of the text data is not the text generation module 1100, the source of the words will be broader and deeper). However, if the use context of the text information is for children to be used in formal occasions (for example, children's exhibition venues), the second language training data set will require more special expertise in other fields than the first language training data set. Children's books are used as supplements to avoid overly difficult and difficult-to-associate connections between technical words, which may affect children's reading fluency.

In one embodiment, the inter-sentence processing module 1220 can determine the relevance of the plurality of words for the plurality of target words belonging to different sentences according to the plurality of desired parts of speech, thereby generating words between different sentences. multiple associated values. For example, the inter-sentence processing module 1220 can perform the plurality of word correlation judgments on the words in the plurality of target words that belong to the same predicate and belong to different sentences. In one embodiment, the inter-sentence processing module 1220 can perform the multiple word correlation judgments on the words that belong to the same subject in the multiple target words and belong to different sentences, and determine the relevance of the words in the multiple target words that belong to the same object. And the vocabulary belonging to different sentences is judged on the relevance of the multiple words. In one embodiment, the inter-sentence processing module 1220 can only perform the word correlation determination on the words that belong to the subject or the object and belong to different sentences among the target words. In one embodiment, each of the plurality of associated values uses at least two of the plurality of target words for analysis.

In one embodiment, the inter-sentence processing module 1220 can determine whether there is at least one group of related words between the plurality of target words based on the plurality of related values. In one implementation, the correlation value determined by the inter-sentence processing module 1220 may be within a range. When a specific associated value among the multiple associated values is greater than a word association threshold, the multiple specific target words forming the specific associated value are regarded as a group of associated words. However, when none of the plurality of correlation values is greater than the word correlation threshold, the plurality of target words are regarded as not being related words. In one implementation, the size of the word association threshold can also be determined according to the usage object. For example: for use by children, the degree of understanding of the words can be considered to increase the requirements for the relevance of the vocabulary, so as to avoid not understanding the relevance of the preceding and following sentences because the deep-seated connection between the words is too difficult. As for formal occasions in special professional and technical fields, the appropriate association threshold of the word can be selected after weighing the actual needs. In one implementation, the range interval may be 0-10, and the word association threshold may be 6.

In one embodiment, when there is at least one group of related words between the plurality of target words, the inter-sentence processing module 1220 can confirm the sentence correlation between the plurality of adjacent sentences based on the at least one group of related words. . Since a set of associated words is determined based on the associated value, and the calculation of the associated value requires capturing the multiple target words in different sentences, the multiple target words in the set of associated words should belong to different sentences. To help determine the sentence relevance between the multiple adjacent sentences.

In one embodiment, the inter-sentence processing module 1220 can directly calculate the number of the at least one group of related words in the plurality of adjacent sentences, and compare the number with a related number threshold and/or a repetition number threshold. If the number of the at least one group of related words in the plurality of adjacent sentences is not less than the association quantity threshold, it can be determined that the plurality of adjacent sentences have a sentence correlation degree. If the number of the at least one group of related words in the plurality of adjacent sentences is less than the association quantity threshold, it can be determined that the plurality of adjacent sentences do not have the sentence correlation degree. In one embodiment, the association number threshold The value can be 2. If the number of at least one group of related words in the multiple adjacent sentences is equal to or greater than a repetition quantity threshold, it can be determined that the repetition between the multiple adjacent sentences may be too high, and the number of the multiple adjacent sentences is determined. There is a high degree of repetition of sentences in between. If the number of the at least one group of related words in the multiple adjacent sentences is less than the repetition number threshold, it can be determined that there will not be excessive repetition between the multiple adjacent sentences, and it is determined that the multiple adjacent sentences There is no high degree of repetition of the sentence between them. In one implementation, the repetition number threshold may be 4. In one embodiment, the inter-sentence processing module 1220 can directly sum up the correlation degree of the at least one group of related words to determine whether there is the sentence correlation degree between the plurality of adjacent sentences and/or the sentence is highly repetitive. Spend. In the embodiment, the summing result is directly affected by the number of the at least one group of associated words. Therefore, the associated quantity threshold and the repetition quantity threshold for the quantity result may have a corresponding relationship with an associated summing threshold and a repetition summing threshold for the summing result.

In one embodiment, the inter-sentence processing module 1220 can determine the sentence correlation between the plurality of adjacent sentences based on the sentence correlation and/or the sentence high repetition degree.

In one embodiment, when there is no at least one set of related words between the plurality of target words, the inter-sentence processing module 1220 can directly confirm that there is no sentence correlation between the plurality of adjacent sentences.

In step S740, it is confirmed whether one of the plurality of adjacent sentences is set as the sentence to be adjusted based on the sentence correlation.

In one embodiment, when the inter-sentence processing module 1220 confirms that there is no sentence correlation between the plurality of adjacent sentences, the inter-sentence processing module 1220 can set one of the plurality of adjacent sentences to Sentence to be adjusted.

In one embodiment, the inter-sentence processing module 1220 can determine the sentence correlation between the plurality of adjacent sentences based on the sentence correlation and/or the sentence high repetition degree. Therefore, sentence The interprocessing module 1220 may determine whether to set one of the plurality of adjacent sentences as the sentence to be adjusted based on at least one of the sentence relevance and the sentence high repetition.

In one embodiment, when the inter-sentence processing module 1220 confirms that the multiple adjacent sentences have the sentence relevance and do not have the high degree of repetition of the sentences, the inter-sentence processing module 1220 can confirm that the multiple adjacent sentences The sentence does not contain the sentence to be adjusted and needs to be modified. In one embodiment, when the inter-sentence processing module 1220 confirms that there is no sentence correlation between the plurality of adjacent sentences, the inter-sentence processing module 1220 may set one of the plurality of adjacent sentences to This is the sentence to be adjusted. In one embodiment, when the inter-sentence processing module 1220 confirms that there is a high degree of repetition of the sentence between the plurality of adjacent sentences, the inter-sentence processing module 1220 can set one of the plurality of adjacent sentences. This is the sentence to be adjusted.

In one embodiment, when the number of adjacent sentences is 2, one of the sentences is a subsequent sentence among the plurality of sentences. In another embodiment, when the number of adjacent sentences is 2, one of the sentences is a previous sentence in the plurality of sentences. In one embodiment, when the number of adjacent sentences is more than three, one of the sentences is a preceding sentence, a middle sentence or a following sentence among the plurality of sentences.

In one embodiment, when the plurality of adjacent sentences is confirmed not to have the sentence to be adjusted, the inter-sentence processing module 1220 may remove the frontmost sentence from the plurality of adjacent sentences and add the plurality of adjacent sentences. The next sentence after the adjacent sentences is used to perform another set of analysis between the adjacent sentences until the multiple sentences of the text information are all confirmed. In one embodiment, when one of the plurality of adjacent sentences is determined to be the sentence to be adjusted, the sentence to be adjusted can be directly deleted, or a new sentence can be generated to replace the sentence to be adjusted. Complete this text message.

In one embodiment, when the text information is confirmed to contain the sentence to be adjusted, it is determined based on the text parameters whether to regenerate a new sentence to replace the sentence to be adjusted or to directly delete the sentence to be adjusted, wherein the text information and the regenerated new sentences are all based on the text model and the text parameters generated. In one embodiment, when one of the plurality of adjacent sentences is determined to be the sentence to be adjusted, the text generation module 1100 can determine the number of words of the text information after deleting the sentence to be adjusted and the number of words in the text. Whether the difference in word count required for a text in the parameter exceeds a preset word count threshold. In one embodiment, if the difference exceeds the preset word count threshold, the text generation module 1100 can use at least one sentence before and/or at least one sentence after the sentence to be adjusted as a key sentence, and use the text model to generate The new sentence replaces the original sentence to be adjusted. If the difference does not exceed the preset word count threshold, the text generation module 1100 may not generate a new sentence.

In one embodiment, when the new sentence is generated, the new sentence still needs to be analyzed by the inter-sentence processing module 1220 again, so the analysis by the inter-sentence processing module 1220 determines that the new sentence does not need to be set as the pending sentence. When adjusting sentences, the inter-sentence processing module 1220 can continue to analyze subsequent sentences in the text information. In one embodiment, when the new sentence is generated, the new sentence can also be submitted to the intra-sentence processing module 1210 to perform the intra-sentence analysis method 600, so the intra-sentence processing module 1210 and the inter-sentence processing module 1220 analyze When it is considered that the new sentence does not need to be set as the sentence to be adjusted, the intra-sentence processing module 1210 and the inter-sentence processing module 1220 can continue to analyze subsequent sentences in the text information.

In one embodiment, the text parameter may include the required number of words in the text, the preset word number threshold, a structure indication that may indicate which grammatical structures are preset, and a structure indication that may indicate which plurality of desired parts of speech are stored in advance. Part-of-speech indication, an association indication that can indicate the lexical association threshold, a quantity indication that can indicate the association quantity threshold and the repetition quantity threshold, and/or a summation indication that can indicate the association sum threshold and the repetition sum threshold. wait. The text parameters can be provided to the inter-sentence processing module 1220 through the electronic device.

In one implementation, the various neural networks used in the present invention can use a Gaussian Error Linear Unit (GELU) as the activation function.

As mentioned above, the above embodiments are only used to illustrate the technical solution of the present invention, but not to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the foregoing. The technical solutions described in each embodiment may be modified, or some of the technical features may be equivalently replaced; however, these modifications or substitutions do not cause the essence of the corresponding technical solutions to depart from the scope of the technical solutions of each embodiment of the present invention.

To sum up, this invention meets the requirements for an invention patent and a patent application should be filed in accordance with the law. However, the above are only preferred embodiments of the present invention. Any equivalent modifications or changes made by those familiar with the art of this application in accordance with the creative spirit of this application should be included in the scope of the following patent applications.

400:Method

S410-S440: Steps

Claims (14)

A text automatic generation method performed by at least one electronic device. The text automatic generation method includes: receiving a plurality of text parameters; generating a text information based on the plurality of text parameters, wherein the text information includes a plurality of sentences; confirming the plurality of text parameters. Whether there is a sentence to be adjusted in a sentence, wherein the confirmation of the sentence to be adjusted includes: using a first language model to obtain multiple single-sentence vocabulary and respective part-of-speech parts from a specific sentence in the multiple sentences, and based on The plurality of word participles of the plurality of single-sentence words are used to confirm whether the specific sentence is set as the sentence to be adjusted; through the first language model, a plurality of adjacent words are obtained from the plurality of adjacent sentences in the plurality of sentences. and respective part-of-speech participles; based on the multiple part-parts of speech of the multiple adjacent words, select a plurality of target words from the multiple adjacent words; through a second language model, perform a correlation on the multiple target words Analyze and confirm multiple word correlations between the multiple target words; based on the multiple word correlations, confirm whether there is at least one group of related words between the multiple target words to confirm the relationship between the multiple adjacent sentences. There is a sentence correlation between; and based on the sentence correlation, it is confirmed whether one of the multiple adjacent sentences is set as the sentence to be adjusted; and when the text information is confirmed to have the sentence to be adjusted, Then adjust the sentence to be adjusted to complete the text information. The automatic text generation method described in request item 1 further includes: Obtain a grammatical relationship based on the distribution of the plurality of word participles of the plurality of single-sentence words in the specific sentence; and based on the grammatical relationship, confirm whether the specific sentence is set as the sentence to be adjusted. The method for automatically generating text as described in claim 2 further includes: receiving multiple grammatical structures; confirming the completeness of a specific sentence based on the multiple grammatical structures and the grammatical relationship; and based on the completeness of the sentence structure. , confirm whether the specific sentence is the sentence to be adjusted. The automatic text generation method as described in claim 1, further comprising: when there is at least one group of related words between the plurality of target words, confirming the relationship between the plurality of adjacent sentences based on the at least one group of related words. Sentence correlation; and when there is no at least one group of related words between the plurality of target words, confirm the sentence correlation between the plurality of adjacent sentences, and confirm the one among the plurality of adjacent sentences. One sentence is the sentence to be adjusted. The automatic text generation method as described in claim 1, wherein: when the number of the at least two adjacent sentences is more than 3, the one sentence is an intermediate sentence among the plurality of sentences. The automatic text generation method as described in request item 1 further includes: when the text information is confirmed to contain the sentence to be adjusted, determining whether to regenerate a new sentence based on the text parameters to replace the sentence to be adjusted or to delete it directly In the sentence to be adjusted, the text information and the regenerated new sentence are generated based on a text model and the text parameters. An electronic device, characterized in that the electronic device includes a storage and at least one processor, at least one instruction is stored in the storage, and when the at least one instruction is executed by the at least one processor, it implements claim 1 To the automatic text generation method described in any one of 6. A text modification method performed by at least one electronic device. The text modification method includes: receiving a text information and a text parameter, wherein the text information includes a plurality of sentences; confirming whether there is a sentence to be adjusted in the plurality of sentences, The confirmation of the sentence to be adjusted includes: using a first language model to obtain a plurality of single-sentence vocabulary and their respective part-of-speech parts from a specific sentence among the plurality of sentences, and based on the plurality of word segmentations of the plurality of single-sentence vocabulary Part of speech, confirm whether the specific sentence is set as the sentence to be adjusted; through the first language model, obtain multiple adjacent words and respective part of speech from multiple adjacent sentences in the multiple sentences; based on the multiple The multiple word participles of adjacent words are used to select multiple target words from the multiple adjacent words; through a second language model, a correlation analysis is performed on the multiple target words to confirm the relationship between the multiple target words. Multiple lexical correlations; based on the multiple lexical correlations, confirm whether there is at least one group of related words between the multiple target words, to confirm a sentence correlation between the multiple adjacent sentences; and based on the Sentence correlation confirms whether one of the multiple adjacent sentences is set as the sentence to be adjusted; and when the text information is confirmed to have the sentence to be adjusted, it is determined based on the text parameters whether a new sentence needs to be generated. sentence to replace the sentence to be adjusted. The text modification method as described in claim 8 further includes: obtaining a grammatical relationship based on the distribution of the plurality of word participles of the plurality of single-sentence words in the specific sentence; and based on the grammatical relationship, confirming whether the specific A sentence is set as the sentence to be adjusted. The text modification method as described in claim 9 further includes: receiving multiple grammatical structures; confirming the completeness of the sentence structure of the specific sentence based on the multiple grammatical structures and the grammatical relationship; and based on the completeness of the sentence structure, Confirm whether the specific sentence is the sentence to be adjusted. The text modification method according to claim 8, further comprising: when there is at least one group of related words between the plurality of target words, confirming the sentence between the plurality of adjacent sentences based on the at least one group of related words. Correlation; and when there is no at least one group of related words between the plurality of target words, confirm the sentence correlation between the plurality of adjacent sentences, and confirm the one of the plurality of adjacent sentences. The sentence is the sentence to be adjusted. The text modification method as described in claim 8, wherein: when the number of the at least two adjacent sentences is more than 3, the one sentence is an intermediate sentence among the plurality of sentences. The text modification method as described in claim 8, further comprising: when the new sentence is generated, replacing the sentence to be adjusted with the new sentence, wherein the new sentence is generated based on a text model; and when the new sentence is not generated , directly delete the sentence to be adjusted. An electronic device, characterized in that the electronic device includes a storage and at least one processor, at least one instruction is stored in the storage, and when the at least one instruction is executed by the at least one processor, it implements claim 8 to the text modification method described in any one of 13.

Images