CN112036155A - Text generation method, text generation device and computer readable storage medium - Google Patents

Abstract

The present disclosure relates to a text generation method, a text generation device, and a computer-readable storage medium. The method includes: acquiring target text; determining a target position in the target text where rhetorical text is to be inserted; generating the rhetorical text according to the semantics of the text within a preset range of the target position; inserting the rhetorical text into the target location. In this way, the rhetorical text can be automatically generated without the need for the user to input the ontology and the metaphor, which makes the generated rhetorical text more diverse and improves the intelligence of the generated rhetorical text. In addition, when generating rhetorical text, the semantics of the text within the preset range of the target position are referred to, so that the semantics of the generated rhetorical text is more matched with the semantics of the text within the preset range of the target position, and the modification of the target text is improved. accuracy, thereby improving the information content of the modified text.

Description

Text generation method, text generation device, and computer-readable storage medium

technical field

The present disclosure relates to the field of computer technologies, and in particular, to a text generation method, a text generation device, and a computer-readable storage medium.

Background technique

With the rapid development of science and technology, a large number of advanced technologies continue to emerge, and natural language processing technology is one of the important technologies. In the wide application of natural language processing technology, automatic text generation has become a hot research topic. Text generation has made great progress under the development of artificial intelligence and neural networks, such as automatic summary generation, automatic title generation, AI poetry writing, human-computer dialogue, text style conversion, story generation, etc. There is little research or product coverage on the automated use of human language writing skills.

In the related art, there are many related studies in the academic circle around the task of story generation. Among them, the story generation uses the encoder-decoder framework model. Given a piece of text as the starting input model, the model will continue to generate a suitable story expansion text or ending text. Among them, the model combines the corresponding emotion, protagonist setting, and instruction map as additional input at the moment of generating the unfolding text or ending text of the story to enhance the results of auxiliary story generation.

The above model for story generation will more or less generate some sentences containing modified sentences, but its shortcomings are obvious: it is end-to-end generation, and it is completely uncontrollable about what kind of rhetorical writing technique is used. In addition, for the existing end-to-end method to generate metaphors, the input must be a pair of metaphor ontology and metaphor, so that the model can generate the corresponding metaphor text. In other words, the technique is limited to the generation of figurative texts given an ontology and a metaphor. Moreover, when using this technology to generate figurative text, it does not refer to contextual semantics, and it cannot guarantee that the generated text is faithful to the original text.

SUMMARY OF THE INVENTION

To overcome the problems existing in the related art, the present disclosure provides a text generation method, a text generation device and a computer-readable storage medium.

According to a first aspect of the embodiments of the present disclosure, there is provided a text generation method, including:

get the target text;

determining the target position of the rhetorical text to be inserted in the target text;

generating the rhetorical text according to the semantics of the text within the preset range of the target position;

Inserting the rhetorical text into the target location.

Optionally, the determining the target position of the rhetorical text to be inserted in the target text includes:

Perform word segmentation processing on the target text to obtain a word segmentation result of the target text, wherein the word segmentation result includes each word segmentation and the location of the word segmentation;

The word segmentation result is input into the modification probability prediction model, and the modification probability corresponding to each position output by the modification probability prediction model is obtained, wherein the modification probability corresponding to the position is used to indicate that the rhetoric to be generated is inserted at the position the probability of the text;

The position corresponding to the maximum modification probability is determined as the target position.

Optionally, generating the rhetorical text according to the semantics of the text within the preset range of the target position includes:

determining the semantic vector corresponding to the target position according to the semantics of the text within the preset range of the target position;

The rhetorical text is generated from the semantic vector.

Optionally, the generating the rhetorical text according to the semantic vector includes:

The semantic vector is input into a text generation model to obtain the rhetorical text output by the text generation model.

Optionally, the text generation model includes a rhetorical type generation sub-model and a rhetorical text generation sub-model;

The inputting the semantic vector into the text generation model to obtain the rhetorical text output by the text generation model includes:

inputting the semantic vector into the rhetoric type generation sub-model, to obtain the rhetorical type vector of the rhetorical text to be generated output by the rhetoric type generation sub-model;

Inputting the semantic vector and the rhetorical type vector into the rhetorical text generation sub-model to obtain the rhetorical text output by the rhetorical text generation sub-model.

Optionally, the text generation model is obtained by:

obtaining the original text after removing the rhetorical text, and the semantic vector corresponding to the position of the removed rhetorical text in the original text;

The semantic vector corresponding to the position is used as the model input parameter, and the removed rhetorical text is used as the model output parameter, and the neural network is trained to obtain the text generation model.

Optionally, the method is applied to a server, and the server runs a modified probability prediction model and a text generation model;

The obtaining the target text includes: obtaining, by the server, the target text uploaded by the client;

The determining of the target position of the rhetorical text to be inserted in the target text includes: the server performs a word segmentation process on the target text, obtains a word segmentation result of the target text, and invokes the modification probability prediction model to perform a word segmentation on the target text. The word segmentation result is processed to obtain the modification probability corresponding to the position of each word segmentation, and the position corresponding to the maximum modification probability is determined as the target position;

The generating the rhetorical text according to the semantics of the text within the preset range of the target position includes: the server determining the target position according to the semantics of the text within the preset range of the target position corresponding semantic vector, and call the text generation model to process the semantic vector to obtain the rhetorical text;

The method further includes: after the server inserts the rhetorical text into the target position, sending a new text including the rhetorical text to the client.

According to a second aspect of the embodiments of the present disclosure, there is provided a text generation apparatus, including:

a first obtaining module, configured to obtain the target text;

a first determining module, configured to determine a target position in the target text where the rhetorical text is to be inserted;

a generating module configured to generate the rhetorical text according to the semantics of the text within the preset range of the target position;

An insertion module configured to insert the rhetorical text into the target location.

Optionally, the first determining module includes:

A word segmentation processing submodule, configured to perform word segmentation processing on the target text to obtain a word segmentation result of the target text, wherein the word segmentation result includes each word segmentation and the location of the word segmentation;

The first input sub-module is configured to input the word segmentation result into the modification probability prediction model, and obtain the modification probability corresponding to each position output by the modification probability prediction model, wherein the modification probability corresponding to the position is used to characterize the modification probability in the modification probability prediction model. the probability of inserting the rhetorical text to be generated at the position;

The first determination submodule is configured to determine the position corresponding to the maximum modification probability as the target position.

Optionally, the generation module includes:

a second determination submodule, configured to determine a semantic vector corresponding to the target position according to the semantics of the text within the preset range of the target position;

A generating submodule is configured to generate the rhetorical text according to the semantic vector.

Optionally, the generating submodule includes:

The second input sub-module is configured to input the semantic vector into a text generation model to obtain the rhetorical text output by the text generation model.

Optionally, the text generation model includes a rhetorical type generation sub-model and a rhetorical text generation sub-model;

The second input submodule includes:

a third input sub-module, configured to input the semantic vector into the rhetoric type generation sub-model, to obtain a rhetorical type vector of the rhetorical text to be generated output by the rhetoric type generation sub-model;

The fourth input sub-module is configured to input the semantic vector and the rhetorical type vector into the rhetorical text generation sub-model to obtain the rhetorical text output by the rhetorical text generation sub-model.

Optionally, the device further includes:

a first obtaining module, configured to obtain the original text after removing the rhetorical text, and the semantic vector corresponding to the position of the removed rhetorical text in the original text;

The training module is configured to use the semantic vector corresponding to the position as a model input parameter, and use the removed rhetorical text as a model output parameter to train a neural network to obtain the text generation model.

Optionally, the apparatus is applied to a server, and the server runs a modified probability prediction model and a text generation model;

The first obtaining module is configured to: the server obtains the target text uploaded by the client;

The first determining module is configured to: the server performs word segmentation processing on the target text, obtains a word segmentation result of the target text, and invokes the modification probability prediction model to process the word segmentation result, and obtains each word segmentation The modification probability corresponding to the position where it is located, and the position corresponding to the maximum modification probability is determined as the target position;

The generation module is configured to: the server determines a semantic vector corresponding to the target position according to the semantics of the text within the preset range of the target position, and invokes the text generation model to perform a function on the semantic vector. processing to obtain the rhetorical text;

The device also includes:

The sending module is configured to send the new text including the rhetorical text to the client after the server inserts the rhetorical text into the target position.

According to a third aspect of the embodiments of the present disclosure, there is provided a text device, comprising:

processor;

memory for storing processor-executable instructions;

wherein the processor is configured to:

get the target text;

determining the target position of the rhetorical text to be inserted in the target text;

generating the rhetorical text according to the semantics of the text within the preset range of the target position;

Inserting the rhetorical text into the target location.

According to a fourth aspect of the embodiments of the present disclosure, there is provided a computer-readable storage medium on which computer program instructions are stored, and when the program instructions are executed by a processor, implement the steps of the text generation method provided in the first aspect of the present disclosure.

The technical solutions provided by the embodiments of the present disclosure may include the following beneficial effects:

By adopting the above technical solution, first, the target position in the target text where the rhetorical text is to be inserted is determined, then the rhetorical text is generated according to the semantics of the text within the preset range of the target position, and finally, the generated rhetorical text is inserted into the target Location. In this way, the rhetorical text can be automatically generated without the need for the user to input the ontology and the metaphor, which makes the generated rhetorical text more diverse and improves the intelligence of the generated rhetorical text. In addition, when generating rhetorical text, the semantics of the text within the preset range of the target position are referred to, so that the semantics of the generated rhetorical text is more matched with the semantics of the text within the preset range of the target position, and the modification of the target text is improved. accuracy, thereby improving the information content of the modified text.

It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure.

Description of drawings

The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments consistent with the disclosure and together with the description serve to explain the principles of the disclosure.

FIG. 1 is a schematic diagram of a model for generating figurative text in the related art.

Fig. 2 is a flow chart of a text generation method according to an exemplary embodiment.

FIG. 3 is a flowchart of step S12 shown in the embodiment shown in FIG. 2 .

FIG. 4 is a flowchart of step S13 shown in the embodiment shown in FIG. 2 .

Fig. 5 is a flow chart of a method for obtaining rhetorical text according to an exemplary embodiment.

Fig. 6 is an interaction diagram between a client and a server in a text generation method according to an exemplary embodiment.

Fig. 7 is a block diagram of a text generating apparatus according to an exemplary embodiment.

Fig. 8 is a block diagram of a text generating apparatus according to an exemplary embodiment.

Fig. 9 is a block diagram of a text generating apparatus according to an exemplary embodiment.

Detailed ways

Exemplary embodiments will be described in detail herein, examples of which are illustrated in the accompanying drawings. Where the following description refers to the drawings, the same numerals in different drawings refer to the same or similar elements unless otherwise indicated. The implementations described in the illustrative examples below are not intended to represent all implementations consistent with this disclosure. Rather, they are merely examples of apparatus and methods consistent with some aspects of the present disclosure as recited in the appended claims.

In the generation of unconditional (without contextual semantics) figurative text, the forward and backward generation strategy is adopted. Illustratively, models for generating figurative text may include backward models and forward models. First, the backward model is used to generate sentences forward until the beginning of the sentence; then the forward model is used to generate sentences backward until the end of the sentence.

FIG. 1 is a schematic diagram of a model for generating figurative text in the related art. As shown in Figure 1, the input is a pair of ontology and metaphor, where the ontology is "enjoyed" and the metaphor is "devoured". First, the backward model is generated forward to the end of the sentence. In Figure 1, in the backward model, the probability of predicting the sentence head as "fire" in the α term is 0.4, the probability of "she" is 0.3, the probability of "flood" is 0.1, and the probability of predicting the sentence in the beta term is 0.4. The probability of the first being "fire" is 0.1, the probability of being "she" is 0.4, and the probability of being "flood" is 0.01. Combining the results of the α term and the β term, the final sentence start is "she". At this point the sentences are "shedevoured" and "sheenjoyed".

After that, the forward model is used to generate sentences backwards until the end of the sentence. As shown in Figure 1, the sentence "shedevoured" and the sentence "she enjoyed" are entered in the forward model respectively, and the forward model generates the text "his" and "novels" in turn, and ends after generating "novels". The final generated sentence is "she devoured his novel", that is, using "devoured" as a metaphor for "enjoyed".

In the related technical way of generating figurative text, the user needs to input ontology and figurative body, and rhetorical text cannot be automatically generated without given ontology and figurative body, resulting in a low degree of intelligence in text generation. Moreover, referring to context semantics when generating figurative text may result in that the generated figurative text cannot match the context well.

In view of this, the present disclosure provides a text generation method, a text generation apparatus, and a computer-readable storage medium.

Before introducing the text generation method, text generation device, and computer-readable storage medium provided by the present disclosure in detail, the application scenarios involved in the present disclosure are first introduced. In the first embodiment, the application scenario may be a scenario in which the client generates rhetorical text for the target text. The client can be, for example, a mobile terminal such as a smart phone, a tablet computer, a smart watch, a smart bracelet, a PDA (English: Personal Digital Assistant, Chinese: Personal Digital Assistant), or a fixed terminal such as a desktop computer. In the second embodiment, the application scenario is a scenario in which the user inputs target text through the client, the server receives the target text from the client, and generates rhetorical text for the target text. The server may be implemented by an independent server or a server cluster composed of multiple physical servers.

Fig. 2 is a flowchart of a text generation method according to an exemplary embodiment. As shown in Fig. 1 , the text generation method may include the following steps.

In step S11, the target text is acquired.

In the present disclosure, the target text may be text that needs to be modified with rhetorical text. For example, the target text may be relatively simple and straightforward narrative text written by a novelist. Also, the target text may be at least one of sentences, words, paragraphs, and the like. The present disclosure does not limit the length of the target text.

In a possible embodiment, the target text may be a sentence in a literary work to be generated. Among them, literary works refer to works of art that use language as a tool to visually reflect objective reality or express the spiritual world of writers. Literary works may include at least one of poems, essays, novels, plays, fables, fairy tales, and the like.

It is worth noting that, in the present disclosure, a literary work to be generated refers to a literary work that includes the target text but has not been completely generated. The text generation method in each embodiment of the present application is used to automatically modify and polish the target text, so as to realize the creation of the literary work. Instead of knowing the target text of an existing literary work, query the rhetorical text in the existing literary work.

In another possible embodiment, the target text may be a sentence in a dialogue exchange.

It is worth noting that, in the application scenario shown in the above-mentioned first embodiment, the client has the function of editing text, and the user can directly edit a relatively simple and straightforward narrative text in the client. In this way, the client can Get the target text. Or, the user inputs the target text in other text-editable devices, and then sends the edited target text to the client through the device, so that the client receives the target text and performs subsequent generation rhetoric based on the target text Text and steps to insert rhetorical text.

In the application scenario shown in the second embodiment above, the client is connected to the server in communication, the user edits a relatively simple and straightforward narrative text in the client, and then the client sends the text to the server. In this way, the server The target text can be received, and the subsequent steps of generating rhetorical text and inserting rhetorical text are performed based on the target text.

In step S12, the target position of the rhetorical text to be inserted in the target text is determined.

In the present disclosure, the target position refers to the position where the generated rhetorical text is to be inserted, that is, the text located before or after the target position needs to be modified using rhetorical devices. The target position where the rhetorical text is to be inserted may be any position in the target text. For example, the target position may be the position at the beginning of the sentence, the position in the middle of the sentence, or the position at the end of the sentence in the target text. This disclosure does not specifically limit this.

In step S13, rhetorical text is generated according to the semantics of the text within the preset range of the target position.

The rhetorical text generated in the present disclosure is to modify the words and sentences in the target text, so that the modified words and sentences can accurately and vividly express the objective reality or the spiritual world of the writer. Therefore, in the present disclosure, it is necessary to The rhetorical text is generated so that the semantics of the rhetorical text matches the semantics of the text within the preset range of the target location. The preset range may be from the previous position of the target position to the next position of the target position, or from the previous position of the target position to the target position, or from the target position to the next position of the target position. It should be noted that there may also be multiple front or rear positions, for example, if the target position is 5, the preset range may be [2, 8] and so on.

In step 14, the rhetorical text is inserted at the target location.

After the rhetorical text is generated, the rhetorical text is inserted into the target position, so that the rhetorical text can express the meaning of the target text more appropriately.

By adopting the above technical solution, first, the target position in the target text where the rhetorical text is to be inserted is determined, then the rhetorical text is generated according to the semantics of the text within the preset range of the target position, and finally, the generated rhetorical text is inserted into the target Location. In this way, the rhetorical text can be automatically generated without the need for the user to input the ontology and the metaphor, which makes the generated rhetorical text more diverse and improves the intelligence of the generated rhetorical text. In addition, when generating rhetorical text, the semantics of the text within the preset range of the target position are referred to, so that the semantics of the generated rhetorical text is more matched with the semantics of the text within the preset range of the target position, and the modification of the target text is improved. accuracy, thereby improving the information content of the modified text.

In order to facilitate those skilled in the art to better understand the text generation method provided by the present disclosure, a complete example will be described below.

First, the specific implementation of determining the target position of the rhetorical text to be inserted in the target text will be described.

In one embodiment, the user determines the target position in the target text and inputs the target position, so that the device executing the text generation method determines the target position.

In another embodiment, the target location can be predicted by means of machine learning. For example, FIG. 3 is a flowchart of step S12 shown in the embodiment shown in FIG. 2 . As shown in FIG. 2 , step S12 may include the following steps.

In step S121, a word segmentation process is performed on the target text to obtain a word segmentation result of the target text, wherein the word segmentation result includes each word segment and the location where the word segment is located.

The word segmentation technology in the related art can be used to segment the target, and the present disclosure does not limit the specific manner of the word segmentation processing.

In a possible manner, the target text is Chinese text, and after word segmentation processing is performed on the Chinese text, each character in the Chinese text is a word segment. For example, the target text is "I ate dinner today, and it feels delicious." The participle after word segmentation is "I/today/day/eat/dine/dinner/meal/, /feeling/feeling/very/good/eat /." The position of each participle is from 1 to 14, that is, the position of the participle "I" in the target text is 1, the position of "Jin" in the target text is 2, ... "." in the target text The position is 14.

In step 122, the word segmentation result is input into the modification probability prediction model, and the modification probability corresponding to each position output by the modification probability prediction model is obtained, wherein the modification probability corresponding to the position is used to represent the probability of inserting the rhetorical text to be generated at the position .

The modification probability prediction model is a model obtained by pre-training, and can be obtained by the following training methods: first, for each sample target text, obtain the sample word segmentation result of the sample target text, and the modification probability label corresponding to the position of each word segmentation. After that, the sample word segmentation result is used as the model input parameter, and the modification probability corresponding to each position is used as the model output parameter, and the neural network is trained to obtain the modification probability prediction model.

Considering the powerful performance of BERT (Bidirectional Encoder Representation from Transformers), and pre-training in large-scale corpus through unsupervised training, it can better adapt to downstream tasks that lack labeled data, such as automatic modification of text, etc. , therefore, in a preferred manner, BERT can be used as the modification probability prediction model. The specific training method may refer to the training method of the existing neural network model, which is not limited here.

The modification probability label corresponding to the position of the word segmentation can be expressed in the form of a vector, and in the vector, only the position marked by the user that needs to be inserted into the rhetorical text corresponds to the modification probability of 1, and the modification probability corresponding to other positions is 0.

After the modification probability prediction model is obtained by training, the modification probability corresponding to each position output by the modification probability prediction model can be obtained by simply inputting the word segmentation result of the target text into the modification probability prediction model.

It is worth noting that the word segmentation result can also be encoded to encode a vector sequence of length L, where the length L is equal to the number of positions. For example, if the word segmentation result includes positions 1 to 14, then the length of the encoded vector sequence is L is 14. In addition, the Softmax function can be used to obtain a probability vector of length L in the modified probability prediction model.

In step 123, the position corresponding to the maximum modification probability among the positions is determined as the target position.

As mentioned above, the modification probability corresponding to the position is used to represent the probability of inserting the rhetorical text to be generated at the position. In the disclosure, after obtaining the modification probability corresponding to each position, the size of each modification probability is compared, and the position corresponding to the largest modification probability among the positions is determined as the target position.

Continuing to use the above example, the target text is "I had dinner today and it was delicious." Input the word segmentation result of the target text into the modification probability prediction model pre-trained according to the above method, and the modification probability prediction model can be Output the probability vector representing the modification probability corresponding to each position, and the probability vector is [0.1, 0.01, 0.02, 0.02, 0.05, 0.2, 0.01, 0.04, 0.05, 0.1, 0, 0, 0, 0.4], then the target position That is position 14, ie, the position where rhetorical text needs to be inserted. It is worth noting that inserting rhetorical text at this position can be inserting rhetorical text at the previous position of the position, or inserting rhetorical text at the latter position of the character. Users can set it according to actual needs. There is no specific limitation in the disclosure. In practical applications, rhetorical text is mostly used to modify the previous sentence. Therefore, in the present disclosure, the rhetorical text is inserted into the position after the target position, so as to modify the words and sentences at the target position or the position before the target position. retouch.

For example, if the subsequently generated rhetorical text is "like the delicacy of mountains and seas", after inserting the rhetorical text into the target position, the obtained text is "I ate dinner today and it feels delicious. It is like delicacy of mountains and seas".

Next, step S13 in FIG. 2 will be described. FIG. 4 is a flowchart of step S13 shown in the embodiment shown in FIG. 2 . As shown in FIG. 4 , step S13 may include the following steps.

In step 131, the semantic vector corresponding to the target position is determined according to the semantics of the text within the preset range of the target position;

In step 132, rhetorical text is generated from the semantic vector.

In order to ensure that the generated rhetorical text can be faithful to the target text to the greatest extent, in the present disclosure, a semantic vector corresponding to the target position can be determined according to the semantics of the text within the target position range, and the rhetorical text can be generated based on the semantic vector.

In the present disclosure, the semantic vector corresponding to the target position is used to represent the semantic vector of the text within the preset range of the target position. It can be understood that the semantic vector is a low-dimensional expression of the text within the preset range of the target position, and covers the feature information of the text within the preset range of the target position. The semantic vector is a sentence vector, that is, a semantic vector may include multiple word vectors.

For example, each word segment in the text within the preset range of the target position can be encoded to obtain the semantics corresponding to each word segment, and then the semantics corresponding to multiple word segmentations within the preset range of the target location are synthesized, The semantic vector corresponding to the target position is determined.

In a possible way, rhetorical text can be generated by means of machine learning. For example, the semantic vector is input into the text generation model, and the rhetorical text output by the text generation model is obtained.

The text generation model can be obtained in the following ways: first, the original text after removing the rhetorical text and the semantic vector corresponding to the position of the removed rhetorical text in the original text are obtained. Illustratively, the semantic vector may be determined with reference to the manner described in FIG. 4 . After that, the semantic vector corresponding to the position is used as the model input parameter, and the removed rhetorical text is used as the model output parameter, and the neural network is trained to obtain the text generation model. The specific training method can refer to the existing neural network model training method, which is not limited here.

At this point, the text generation model after training can be obtained, and the rhetorical text can be determined by using the text generation model.

In practical application, rhetorical type refers to the type of rhetorical method, which can be of many kinds, for example, metaphor, personification, repetition, parallelism, and truth, etc. In order to further improve the accuracy of the generated rhetorical text, in the present disclosure, the rhetorical type may also be determined first, and then the rhetorical text is generated based on the rhetorical type and the semantic vector.

In one manner, the rhetoric type may be input by the user, and after acquiring the rhetoric type input by the user, the device executing the method for generating text generates rhetorical text according to the rhetoric type and the semantic vector.

In another manner, the device performing the method of generating text may automatically generate a rhetorical type, and then generate a rhetorical text according to the rhetoric type and the semantic vector.

In this manner, the text generation model includes a rhetorical type generation sub-model and a rhetorical text generation sub-model. As shown in FIG. 5 , inputting the semantic vector into the text generation model, and obtaining the rhetorical text output by the text generation model may further include steps S51 and S52.

In step S51, the semantic vector is input into the rhetoric type generation sub-model, and the rhetorical type vector of the rhetorical text to be generated output by the rhetoric type generation sub-model is obtained.

In this disclosure, a rhetorical type vector is a vectorization of rhetorical type adoption. The rhetorical type vector may include a vector representing a specific rhetorical type, and the rhetorical type vector may also include a vector representing no rhetorical type. That is, the rhetorical type vector can represent the meaning of no rhetorical type, and is not limited to representing only a specific rhetorical type, because the target text may not need to be modified with the rhetorical text.

In one embodiment, the preset rhetorical types include: metaphor, personification, repetition, parallelism, and realism. Combine the specific rhetoric type to illustrate the substantial form of the rhetorical type vector. For example, the rhetorical type vector representing metaphor can be [1,0,0,0], and the rhetorical type vector representing personification can be [0,1, 0,0].

The above text generation model may include two sub-models, namely, a rhetorical type generation sub-model for generating rhetorical type vectors for representing rhetorical types and a rhetorical text generation sub-model for generating rhetorical texts.

Among them, the rhetorical type generation sub-model can be trained by the following training methods: first, obtain the sample semantic vector corresponding to the position of the removed rhetorical text in the original text, and the sample rhetorical type vector of the removed rhetorical text; The sample semantic vector is used as the model input parameter, the sample rhetorical type vector is used as the model output parameter, and the neural network model is trained to obtain the rhetorical type generation sub-model. The specific training method can refer to the existing neural network model training method, which is not limited here.

In step S52, the semantic vector and the rhetorical type vector are input into the rhetorical text generation sub-model to obtain the rhetorical text output by the rhetorical text generation sub-model.

For example, the semantic vector and the rhetorical type vector can be combined into one vector, which is input to the rhetorical text generation sub-model, and then the rhetorical text generation sub-model decodes the combined vector to generate the rhetorical text. Among them, the rhetorical text generation sub-model may include a transformer decoder. It should be noted that the present disclosure does not specifically limit the combination manner.

In a possible way, it can be directly combined into a vector by splicing, and the sequence of splicing is not limited. For example, the rhetorical type vector can be concatenated after the semantic vector, or the rhetorical type vector can be concatenated before the semantic vector. In another possible way, the vector elements can be recombined to form a vector. For example, if the rhetorical type vector is [0,1,0,0], and the semantic vector is [1,1,1,0,0,1,0], then, by combining the vector elements, the combined The vector of is [1,1,1,0,0,1,0,1,0,0,0]. It can be understood that the reasonable fusion of the semantic vector and the rhetorical type vector can make the rhetorical device more appropriately reflected and improve the accuracy of the generated rhetorical text.

The rhetorical text generation sub-model can be trained by the following training methods: First, obtain the sample semantic vector, the sample rhetoric type vector, and the removed rhetorical text; The text is used as the model output parameter, and the neural network model is trained to obtain a rhetorical text generation sub-model. The specific training method can refer to the existing neural network model training method, which is not limited here.

It is worth noting that the above modification probability prediction model and text generation model can be the same model, and in this model, the target position can be predicted, and rhetorical text can be further generated. Just input the word segmentation result of the target text into the model, the model can output the target position and the rhetorical text, and then the electronic device executing the text generation method can insert the rhetorical text into the target position. In this case, the model is obtained by training the neural network with the word segmentation result of the sample target text as model input parameters, and the target position and rhetorical text of the sample target text as model output parameters.

Fig. 6 is an interaction diagram between a client and a server in a text generation method according to an exemplary embodiment. Among them, the server runs a modified probability prediction model and a text generation model. As shown in Figure 6, the method includes the following steps.

In step S61, the client uploads the target text. For example, the user can directly edit and generate the target text in the client, and then the client uploads the target text.

In step S62, the server obtains the target file uploaded by the client.

In step S63, the server performs word segmentation processing on the target text, obtains the word segmentation result of the target text, and calls the modification probability prediction model to process the word segmentation result, obtains the modification probability corresponding to the position of each word segmentation, and assigns the position corresponding to the maximum modification probability Determined as the target location. The specific implementation manner of determining the target position may refer to the manner described in FIG. 3 , which will not be repeated here.

In step S64, the server determines the semantic vector corresponding to the target position according to the semantics of the text within the preset range of the target position, and invokes the text generation model to process the semantic vector to obtain rhetorical text. The specific implementation manner of obtaining the rhetorical text may refer to the manner described in FIG. 4 , which will not be repeated here.

In step S65, the server inserts the rhetorical text into the target position.

In step S66, the server sends the new text including the rhetorical text to the client.

Based on the same inventive concept, the present disclosure also provides a text generation device. Fig. 7 is a block diagram of a text generating apparatus according to an exemplary embodiment. As shown in Figure 7, the text generating apparatus may include:

a first obtaining module 701, configured to obtain target text;

a first determining module 702, configured to determine a target position in the target text where the rhetorical text is to be inserted;

A generating module 703, configured to generate the rhetorical text according to the semantics of the text within the preset range of the target position;

An inserting module 704 is configured to insert the rhetorical text into the target location.

Optionally, the first determining module 702 includes:

A word segmentation processing submodule, configured to perform word segmentation processing on the target text to obtain a word segmentation result of the target text, wherein the word segmentation result includes each word segmentation and the location of the word segmentation;

The first input sub-module is configured to input the word segmentation result into the modification probability prediction model, and obtain the modification probability corresponding to each position output by the modification probability prediction model, wherein the modification probability corresponding to the position is used to characterize the modification probability in the modification probability prediction model. the probability of inserting the rhetorical text to be generated at the position;

The first determination submodule is configured to determine the position corresponding to the maximum modification probability as the target position.

Optionally, the generating module 703 includes:

a second determination submodule, configured to determine a semantic vector corresponding to the target position according to the semantics of the text within the preset range of the target position;

A generating submodule is configured to generate the rhetorical text according to the semantic vector.

Optionally, the generating submodule includes:

The second input sub-module is configured to input the semantic vector into a text generation model to obtain the rhetorical text output by the text generation model.

Optionally, the text generation model includes a rhetorical type generation sub-model and a rhetorical text generation sub-model;

The second input submodule includes:

a third input sub-module, configured to input the semantic vector into the rhetoric type generation sub-model, to obtain a rhetorical type vector of the rhetorical text to be generated output by the rhetoric type generation sub-model;

The fourth input sub-module is configured to input the semantic vector and the rhetorical type vector into the rhetorical text generation sub-model to obtain the rhetorical text output by the rhetorical text generation sub-model.

Optionally, the device further includes:

a first obtaining module, configured to obtain the original text after removing the rhetorical text, and the semantic vector corresponding to the position of the removed rhetorical text in the original text;

The training module is configured to use the semantic vector corresponding to the position as a model input parameter, and use the removed rhetorical text as a model output parameter to train a neural network to obtain the text generation model.

Optionally, the apparatus is applied to a server, and the server runs a modified probability prediction model and a text generation model;

The first obtaining module 701 is configured to: the server obtains the target text uploaded by the client;

The first determination module 702 is configured to: the server performs word segmentation processing on the target text, obtains the word segmentation result of the target text, and invokes the modification probability prediction model to process the word segmentation result, and obtains each word segmentation result. The modification probability corresponding to the position where the word segmentation is located, and the position corresponding to the maximum modification probability is determined as the target position;

The generation module 703 is configured to: the server determines the semantic vector corresponding to the target position according to the semantics of the text within the preset range of the target position, and invokes the text generation model to perform a translation on the semantic vector. processing to obtain the rhetorical text;

The device also includes:

The sending module is configured to send the new text including the rhetorical text to the client after the server inserts the rhetorical text into the target position.

Regarding the apparatus in the above-mentioned embodiment, the specific manner in which each module performs operations has been described in detail in the embodiment of the method, and will not be described in detail here.

The present disclosure also provides a computer-readable storage medium on which computer program instructions are stored, and when the program instructions are executed by a processor, implement the steps of the text generation method provided by the present disclosure.

Fig. 8 is a block diagram of an apparatus for generating text according to an exemplary embodiment. For example, apparatus 800 may be a mobile phone, computer, digital broadcast terminal, messaging device, game console, tablet device, medical device, fitness device, personal digital assistant, and the like.

8, the apparatus 800 may include one or more of the following components: a processing component 802, a memory 804, a power component 806, a multimedia component 808, an audio component 810, an input/output (I/O) interface 812, a sensor component 814, and communication component 816.

The processing component 802 generally controls the overall operation of the device 800, such as operations associated with display, phone calls, data communications, camera operations, and recording operations. The processing component 802 may include one or more processors 820 to execute instructions to perform all or part of the steps of the text generation method described above. Additionally, processing component 802 may include one or more modules that facilitate interaction between processing component 802 and other components. For example, processing component 802 may include a multimedia module to facilitate interaction between multimedia component 808 and processing component 802.

Memory 804 is configured to store various types of data to support operations at device 800 . Examples of such data include instructions for any application or method operating on device 800, contact data, phonebook data, messages, pictures, videos, and the like. Memory 804 may be implemented by any type of volatile or nonvolatile storage device or combination thereof, such as static random access memory (SRAM), electrically erasable programmable read only memory (EEPROM), erasable Programmable Read Only Memory (EPROM), Programmable Read Only Memory (PROM), Read Only Memory (ROM), Magnetic Memory, Flash Memory, Magnetic or Optical Disk.

Power component 806 provides power to various components of device 800 . Power components 806 may include a power management system, one or more power sources, and other components associated with generating, managing, and distributing power to device 800 .

Multimedia component 808 includes a screen that provides an output interface between the device 800 and the user. In some embodiments, the screen may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen may be implemented as a touch screen to receive input signals from a user. The touch panel includes one or more touch sensors to sense touch, swipe, and gestures on the touch panel. The touch sensor may not only sense the boundaries of a touch or swipe action, but also detect the duration and pressure associated with the touch or swipe action. In some embodiments, the multimedia component 808 includes a front-facing camera and/or a rear-facing camera. When the apparatus 800 is in an operation mode, such as a shooting mode or a video mode, the front camera and/or the rear camera may receive external multimedia data. Each of the front and rear cameras can be a fixed optical lens system or have focal length and optical zoom capability.

Audio component 810 is configured to output and/or input audio signals. For example, audio component 810 includes a microphone (MIC) that is configured to receive external audio signals when device 800 is in operating modes, such as call mode, recording mode, and voice recognition mode. The received audio signal may be further stored in memory 804 or transmitted via communication component 816 . In some embodiments, audio component 810 also includes a speaker for outputting audio signals.

The I/O interface 812 provides an interface between the processing component 802 and a peripheral interface module, which may be a keyboard, a click wheel, a button, or the like. These buttons may include, but are not limited to: home button, volume buttons, start button, and lock button.

Sensor assembly 814 includes one or more sensors for providing status assessment of various aspects of device 800 . For example, the sensor assembly 814 can detect the open/closed state of the device 800, the relative positioning of components, such as the display and keypad of the device 800, and the sensor assembly 814 can also detect a change in the position of the device 800 or a component of the device 800 , the presence or absence of user contact with the device 800 , the orientation or acceleration/deceleration of the device 800 and the temperature change of the device 800 . Sensor assembly 814 may include a proximity sensor configured to detect the presence of nearby objects in the absence of any physical contact. Sensor assembly 814 may also include a light sensor, such as a CMOS or CCD image sensor, for use in imaging applications. In some embodiments, the sensor assembly 814 may also include an acceleration sensor, a gyroscope sensor, a magnetic sensor, a pressure sensor, or a temperature sensor.

Communication component 816 is configured to facilitate wired or wireless communication between apparatus 800 and other devices. Device 800 may access wireless networks based on communication standards, such as WiFi, 2G or 3G, or a combination thereof. In one exemplary embodiment, the communication component 816 receives broadcast signals or broadcast related information from an external broadcast management system via a broadcast channel. In an exemplary embodiment, the communication component 816 also includes a near field communication (NFC) module to facilitate short-range communication. For example, the NFC module may be implemented based on radio frequency identification (RFID) technology, infrared data association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology and other technologies.

In an exemplary embodiment, apparatus 800 may be implemented by one or more application specific integrated circuits (ASICs), digital signal processors (DSPs), digital signal processing devices (DSPDs), programmable logic devices (PLDs), field programmable A gate array (FPGA), controller, microcontroller, microprocessor or other electronic component implementation for implementing the text generation method described above.

In an exemplary embodiment, a non-transitory computer-readable storage medium including instructions, such as a memory 804 including instructions, is also provided, and the instructions are executable by the processor 820 of the apparatus 800 to complete the text generation method described above. For example, the non-transitory computer-readable storage medium may be ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, optical data storage device, and the like.

In another exemplary embodiment, there is also provided a computer program product comprising a computer program executable by a programmable apparatus, the computer program having, when executed by the programmable apparatus, for performing the above The code section of the text generation method.

FIG. 9 is a block diagram of a text generating apparatus 1900 according to an exemplary embodiment. For example, the apparatus 1900 may be provided as a server. 9, apparatus 1900 includes processing component 1922, which further includes one or more processors, and a memory resource represented by memory 1932 for storing instructions executable by processing component 1922, such as application programs. An application program stored in memory 1932 may include one or more modules, each corresponding to a set of instructions. Additionally, the processing component 1922 is configured to execute instructions to perform the text generation method described above.

The device 1900 may also include a power supply assembly 1926 configured to perform power management of the device 1900, a wired or wireless network interface 1950 configured to connect the device 1900 to a network, and an input output (I/O) interface 1958. Device 1900 may operate based on an operating system stored in memory 1932, such as Windows Server ^™ , MacOS X ^™ , Unix ^™ , Linux ^™ , FreeBSD ^™ or the like.

Other embodiments of the present disclosure will readily occur to those skilled in the art upon consideration of the specification and practice of the present disclosure. This application is intended to cover any variations, uses, or adaptations of the present disclosure that follow the general principles of the present disclosure and include common knowledge or techniques in the technical field not disclosed by the present disclosure . The specification and examples are to be regarded as exemplary only, with the true scope and spirit of the disclosure being indicated by the following claims.

It is to be understood that the present disclosure is not limited to the precise structures described above and illustrated in the accompanying drawings, and that various modifications and changes may be made without departing from the scope thereof. The scope of the present disclosure is limited only by the appended claims.

Claims (10)

1. A text generation method, comprising:

acquiring a target text;

determining a target position of a retrieval text to be inserted in the target text;

generating the revision text according to the semantics of the text in the preset range of the target position;

and inserting the revise text into the target position.

2. The method of claim 1, wherein the determining a target position of the target text where the paraphrase text is to be inserted comprises:

performing word segmentation processing on the target text to obtain word segmentation results of the target text, wherein the word segmentation results comprise all words and positions of the words;

inputting the word segmentation result into a modification probability prediction model to obtain modification probabilities corresponding to positions output by the modification probability prediction model, wherein the modification probabilities corresponding to the positions are used for representing the probability of inserting the modification text to be generated at the positions;

and determining the position corresponding to the maximum modification probability as the target position.

3. The method of claim 1, wherein the generating the modified text according to semantics of the text within the preset range of the target position comprises:

determining a semantic vector corresponding to the target position according to the semantics of the text within the preset range of the target position;

and generating the revision text according to the semantic vector.

4. The method of claim 3, wherein said generating the pruned text from the semantic vector comprises:

and inputting the semantic vector into a text generation model to obtain the revision text output by the text generation model.

5. The method of claim 4, wherein the text generation model comprises a revisiting type generation submodel and a revising text generation submodel;

the inputting the semantic vector into a text generation model to obtain the revision text output by the text generation model includes:

inputting the semantic vector into the revision type generation submodel to obtain a revision type vector of the revision text to be generated, which is output by the revision type generation submodel;

and inputting the semantic vector and the retrieval type vector to the retrieval text generation submodel to obtain the retrieval text output by the retrieval text generation submodel.

6. The method of claim 4, wherein the text generation model is obtained by:

acquiring an original text from which a modifying text is removed and a semantic vector corresponding to the position of the removed modifying text in the original text;

and taking the semantic vector corresponding to the position as a model input parameter, taking the removed modifying text as a model output parameter, and training a neural network to obtain the text generation model.

7. The method of claim 1, applied to a server running a decorated probability prediction model and a text generation model;

the acquiring of the target text comprises: the server acquires a target text uploaded by a client;

the determining the target position of the retrieval text to be inserted in the target text comprises: the server carries out word segmentation processing on the target text to obtain word segmentation results of the target text, calls the modification probability prediction model to process the word segmentation results to obtain modification probabilities corresponding to positions of the word segments, and determines the position corresponding to the maximum modification probability as the target position;

the generating the revision text according to the semantics of the text in the preset range of the target position includes: the server determines a semantic vector corresponding to the target position according to the semantics of the text in the preset range of the target position, and calls the text generation model to process the semantic vector to obtain the revision text;

the method further comprises the following steps: and after the server inserts the retrieval text into the target position, sending a new text comprising the retrieval text to the client.

8. A text generation apparatus, comprising:

a first obtaining module configured to obtain a target text;

the first determining module is configured to determine a target position of a retrieval text to be inserted in the target text;

the generating module is configured to generate the revision text according to the semantics of the text in the preset range of the target position;

an insertion module configured to insert the thesaurus text into the target location.

9. A text apparatus, comprising:

a processor;

a memory for storing processor-executable instructions;

wherein the processor is configured to:

acquiring a target text;

determining a target position of a retrieval text to be inserted in the target text;

generating the revision text according to the semantics of the text in the preset range of the target position;

and inserting the revise text into the target position.

10. A computer-readable storage medium, on which computer program instructions are stored, which program instructions, when executed by a processor, carry out the steps of the method according to any one of claims 1 to 7.

Images