JP2020187729A - Text processing method, apparatus, device, and storage medium - Google Patents

Abstract

To provide a text processing method configured to reduce duplicated sentences in an output text, a text processing apparatus, a text processing device, and a computer-readable storage medium.SOLUTION: A text processing method includes: acquiring an input text; analyzing the input text to obtain an analysis result corresponding to the input text; dividing the input text into a plurality of sections by means of clustering; and generating an output text on the basis of the sections and the analysis result.SELECTED DRAWING: Figure 1

Description

The present application relates to the field of text processing, in particular to text processing methods, text processing devices, text processing devices, and computer-readable storage media.

Deep neural networks are large-scale multi-parameter optimization tools. Deep neural networks utilize large amounts of training data and can learn hidden features that are difficult to summarize from the data, so face detection, image semantic segmentation, text summarization, object detection, motion tracking, natural language translation, and much more. Can complete complex tasks.

Text abstract extraction refers to generating a text abstract by high-level summarization and abstraction of text content that has a clear meaning. Traditional text summarization extraction methods rely heavily on the specific content of the text (meaning of expression, sentence structure, rhetorical technique, narrative style, etc.). Therefore, performance will differ when different text summarization methods are applied to different texts (such as texts of different lengths). In addition, in the text summary extraction process, different expressions of the same information may appear multiple times in a sentence, which leads to duplication of the summary sentences extracted in the summary extraction process.

In view of the above problems, the present disclosure provides a text processing method, a text processing device, a text processing device, and a computer-readable storage medium.

According to one aspect of the present disclosure, a text processing method based on a neural network is provided. The method is to obtain the input text, analyze the input text, obtain the analysis result corresponding to the input text, and divide the input text into a plurality of sections by using clustering. , The generation of output text based on the plurality of sections and the analysis results.

According to one aspect of the present disclosure, dividing the input text into a plurality of sections by utilizing clustering includes an initialization step for initializing a plurality of central sentences corresponding to the plurality of sections and the input text. By calculating the similarity between the constituent sentence of the above and the plurality of central sentences, the constituent sentences of the input text are assigned to the sections corresponding to the plurality of central sentences based on the similarity, and the plurality of sections are assigned. An update step for updating the constituent statements included in the above, and a determination step for determining the constituent statement having the maximum total similarity as a new central sentence by calculating the similarity between the constituent statements in the plurality of sections. And a repetitive step of repeating the above update step and determination step until the new central sentence does not change.

According to one aspect of the present disclosure, analyzing the input text and obtaining the analysis result corresponding to the input text analyzes all the constituent sentences of the input text and of each sentence of all the constituent sentences. It includes acquiring the weight as the analysis result.

According to one aspect of the present disclosure, generating output text based on the plurality of sections and the analysis results is performed in each of the plurality of sections based on the weight of each sentence of all the constituent sentences. This includes selecting the constituent sentence having the maximum sentence weight in the section as the output result corresponding to the section, and combining the output results of the plurality of sections to generate the output text.

According to one aspect of the present disclosure, the neural network includes one text processing layer, and the number of a plurality of sections obtained by dividing the input text by clustering is a predetermined target sentence of the output text of the text processing layer. Determined by the number.

According to one aspect of the present disclosure, the neural network includes a cascade of N (N ≧ 2) text processing layers. The number of sections obtained by the nth text processing layer in the cascaded N text processing layers dividing the input text by clustering is a predetermined target sentence of the output text of the nth text processing layer. Determined by the number.

According to another aspect of the present disclosure, a text processing device based on a neural network is provided. The text processing device uses an acquisition unit for acquiring input text, an analysis unit for analyzing the input text and acquiring an analysis result corresponding to the input text, and clustering to divide the input text into a plurality of sections. It includes a division unit for dividing, the plurality of sections, and a generation unit for generating output text based on the analysis result.

According to another aspect of the present disclosure, the division portion initializes a plurality of central sentences corresponding to the plurality of sections and calculates the similarity between the constituent sentence of the input text and the plurality of central sentences. As a result, the constituent sentences in the input text are assigned to the sections corresponding to the plurality of central sentences based on the similarity, the constituent sentences included in the plurality of sections are updated, and the constituent sentences in the plurality of sections are each configured. By calculating the similarity between sentences, the constituent sentence having the maximum total similarity is determined as a new central sentence, and the above process is repeated until the new central sentence does not change.

According to another aspect of the present disclosure, the analysis unit analyzes all the constituent sentences of the input text and obtains the weight of each sentence of all the constituent sentences as an analysis result.

According to another aspect of the present disclosure, the output unit is based on the sentence weights of all the constituent sentences, and in each of the plurality of sections, the constituent sentence in which the sentence weight is maximized in the section. Is selected as the output result corresponding to the section, and the output results of multiple sections are combined to generate the output text.

According to another aspect of the present disclosure, the neural network includes one text processing layer, and the number of sections obtained by dividing the input text by clustering is a predetermined number of output texts of the text processing layer. Determined by the target number of sentences.

According to another aspect of the present disclosure, the neural network includes a cascaded N (N ≧ 2) text processing layers. The number of sections obtained by the nth text processing layer in the cascaded N text processing layers dividing the input text by clustering is a predetermined target sentence of the output text of the nth text processing layer. Determined by the number.

According to yet another aspect of the present disclosure, a text processing device based on a neural network is provided. The text processing device includes a memory configured to store computer-readable instructions and a processor configured to execute the computer-readable instructions stored in the memory. When the computer-readable instruction is executed by the processor, the input text is acquired, the input text is analyzed, and the analysis result corresponding to the input text is acquired, and the clustering is utilized. Dividing the input text into a plurality of sections and generating the output text based on the plurality of sections and the analysis result are executed.

According to yet another aspect of the present disclosure, a computer-readable storage medium is provided. Computer-readable instructions are stored in the computer-readable storage medium. When the computer-readable instruction is executed by the computer, the computer performs a text processing method. The method is to obtain the input text, analyze the input text, obtain the analysis result corresponding to the input text, and divide the input text into a plurality of sections by using clustering. , The generation of output text based on the plurality of sections and the analysis results.

In the above aspect of the present disclosure, it is possible to obtain the effect of reducing duplicate sentences in the acquired output text by dividing the input text into a plurality of sections by clustering and acquiring the analysis results corresponding to the plurality of sections. Yes, so the output text is more concise and clear.

By describing the embodiments of the present disclosure in more detail with reference to the drawings, the above and other objects, features and advantages of the present invention will become clearer. The drawings are intended to provide a further understanding of the embodiments of the invention and are incorporated herein by reference and are used in an interpretation of the present disclosure along with the embodiments of the present disclosure and are not restrictions on the present disclosure. In drawings, the same reference number always refers to the same part or step.

It is a flowchart of the text processing method by one Embodiment of this disclosure. It is a schematic diagram which acquires the weight of the constituent sentence of the input text based on the sentence vector and the vocabulary vector by one Embodiment of this disclosure. It is a schematic diagram which applies clustering to a text processing method based on a neural network according to one Embodiment of this disclosure. It is a flowchart of the method of dividing the input text by using clustering according to the embodiment of this disclosure. It is a schematic diagram which divides an input text by using clustering according to embodiment of this disclosure. FIG. 5 is an exemplary schematic diagram for obtaining analysis results corresponding to each of a plurality of sections according to an embodiment of the present disclosure. It is a schematic diagram of the text processing apparatus according to the embodiment of this disclosure. It is a schematic diagram of the text processing device according to the embodiment of this disclosure. It is a schematic diagram of a computer-readable storage medium according to the embodiment of the present disclosure.

The technical solutions according to the embodiments of the present disclosure will be clearly and completely described below together with the drawings in the embodiments of the present disclosure. It is clear that the embodiments described are only a part of the embodiments of the present disclosure and not all embodiments. All other embodiments obtained by one of ordinary skill in the art based on the embodiments of the present disclosure without creative work are all within the scope of the present disclosure.

First, a text processing method 100 for realizing one embodiment of the present disclosure will be described with reference to FIG. In the present disclosure, by dividing the input text into a plurality of sections using clustering, the constituent sentences of the same type are divided together, and then the analysis results extracted from the plurality of sections are combined as the output text. To. This can serve the purpose of removing duplicate sentences in the input text, thus making the output text more concise and clear.

As shown in FIG. 1, in step S101, the input text is acquired.

The input text is the original text that is expected to be processed, and the original text produces a summary of the expected number of characters or the expected number of sentences.

In step S102, the input text is analyzed, and the analysis result corresponding to the input text is acquired.

Traditionally, there are many different approaches. For example, there are various methods for text summary extraction, such as position method, phrase method, title method, and keyword method. In step S102, for example, different weights are assigned to each text processing method, all the constituent sentences of the input text are analyzed for each text processing method, and the weights of all the constituent sentences of the input text are acquired as the analysis result. May be good.

Alternatively, the weight of each constituent sentence may be obtained by using the sentence vector of the constituent sentence of the input text. Instead, use all the constituent sentences of the input text to get the corresponding vocabulary vector, analyze the input text using different processing methods, combine the resulting constituent sentence vector with the vocabulary vector, and then The feature amount may be extracted by using a method such as VGG-16 to obtain the weight of each constituent sentence. By analyzing based on the sentence vector and the vocabulary vector, it is possible to subdivide from the depth of sentence granularity to the depth of vocabulary granularity, and thus the recognizability of a special sentence can be improved. Further, by using VGG-16, the relationship between the feature values can be better grasped, and more detailed features can be extracted.

FIG. 2 is a schematic diagram for acquiring the weights of the constituent sentences of the input text based on the sentence vector and the vocabulary vector. As shown in FIG. 2, first, all the constituent sentences 70 ({S ₁ , S ₂ , S ₃ , ..., S ₇ }) of the input text are input to the neural network. Next, in the neural network, all the constituent sentences 70 of the input text are analyzed by using different processing methods, and the sentence vector 71 is generated for each constituent sentence. Next, the vocabulary vector 72 of all the constituent sentences of the input text is acquired. Then, based on the acquired sentence vector 71 and the vocabulary vector 72, the feature amount 74 ({V ₁ , V ₂ , V ₃ , ..., V ₇ }) is extracted using VGG-16, and the feature amount is said. Reference numeral 74 corresponds to the weight of each constituent sentence 70 of the input text.

It should be understood that the above method for obtaining the weight of the constituent sentence is not limited to this, and another appropriate method may be adopted for obtaining the weight of the constituent sentence.

Next, in step S103, the input text is divided into a plurality of sections by using clustering.

Here, the number of sections obtained by dividing the input text by clustering is determined by a predetermined target number of sentences of the output text. For example, when the predetermined target number of sentences of the desired output text is M, the number of a plurality of sections obtained by dividing the input text by clustering is also M.

FIG. 3 shows a schematic diagram in which clustering is applied to a text processing method based on a neural network. As shown in FIG. 3, assuming that the neural network contains N (N ≧ 2) cascaded text processing layers, the nth text processing layer of the cascaded N text processing layers is clustered. The number of sections obtained by dividing the input text is determined by a predetermined target number of sentences of the output text of the nth text processing layer. For example, the number of a plurality of sections obtained by dividing the input text 51 by the text processing layer 1 (represented by 52 in the figure) by using the clustering 50 is the output text 1 of the text processing layer 1 (in the figure). It may be determined by a predetermined target number of sentences (represented by 53), or the text processing layer 2 (represented by 54 in the figure) utilizes clustering 56 to input text (that is, output text 1 in the figure). The number of the plurality of sections obtained by dividing the above may be determined by a predetermined target number of sentences of the output text 2 (represented by 55 in the figure) of the text processing layer 2.

The neural network may contain only one text processing layer, so that the number of sections obtained by clustering the input text is determined by a predetermined target number of sentences in the output text of the text processing layer. Please understand that.

After obtaining the number of sections obtained by dividing the input text by clustering, the operation of dividing the input text by clustering is performed.

FIG. 4 is a flowchart of the method 200 for dividing the input text by using clustering according to the embodiment of the present disclosure.

As shown in FIG. 4, in step S201, a plurality of central sentences corresponding to the plurality of sections are initialized.

When the number of a plurality of sections is M, it is necessary to select one central sentence for each section, so it is necessary to select M central sentences. For example, the first M sentences of the input text can be selected as the central sentence at the time of initialization, or M sentences can be randomly selected as the central sentence at the time of initialization from the input text. It should be understood that the method of initializing the central sentence is not limited to this, and other appropriate methods may be adopted to initialize the central sentence.

In step S202, by calculating the similarity between the constituent sentences of the input text and the plurality of central sentences, the constituent sentences of the input text are assigned to the sections corresponding to the plurality of central sentences based on the similarity, and the plurality of sections are assigned. Update the constituent statements contained in.

For example, the similarity (or distance) can be calculated by the Rouge-1 method. For example, the similarity f _ij of the i-th sentence and the j-th sentence can be expressed as follows.

However, S _i represents the number of p vocabularies in the i-th sentence (for example, p = 2 represents two vocabularies, p = 3 represents three vocabularies), and R _ij represents the i-th sentence. The number of repeating p vocabularies in the jth sentence is shown. The larger f _{ij, the} higher the similarity between the i-th sentence and the j-th sentence. It should be understood that the method of calculating similarity is not limited to this, and other suitable methods can be adopted to calculate similarity.

Next, by calculating the similarity between the constituent sentence in the input text and the plurality of central sentences, the constituent sentence of the input text having a high degree of similarity is assigned to the section corresponding to the central sentence.

For example, the constituent sentence of the input text is T = {t ₁ , t ₂ , t ₃ , ..., t _z }, and the number of a plurality of sections obtained by dividing the input text by clustering is M = 2 (that is, that is). Assuming that the first two sentences t ₁ and t ₂ are selected as the central sentences of the two sections obtained by the first section and the second section), then the rest of the input text is constructed. Similarity between the sentence {t ₃ , t ₄ , ..., t _z } and the central sentence t ₁ , {f ₃₁ , f ₄₁ , ..., f _{z 1} }, and {t ₃ , t ₄ , ..., t _z } similarity _{{f 32, f 42, ...} , f z2} between the center sentence _{t 2} and by calculating the _high center sentence similarity configuration statement _{t 3} by comparing _{f 31} and _{f 32} (For example, if f ₃₁ > f ₃₂ , the constituent sentence t ₃ may be assigned to the first section, and if f ₃₁ <f ₃₂ , the constituent sentence t ₃ may be assigned to the second section). , F ₄₁ and f ₄₂ may be compared to assign the constituent sentence t ₄ to the section corresponding to the central sentence having a high degree of similarity.

After assigning all the constructs, update the constructs contained in multiple sections.

Next, in step S203, the constituent sentence having the maximum total similarity is determined as the new central sentence by calculating the similarity between the constituent sentences in each of the plurality of sections.

For example, assuming that the constructs contained in the first section are {t ₃ , t ₅ , t ₆ }, the similarity {f ₃₅ , f ₃₆ } between each construct {t ₃ , t ₅ , t ₆ } , F ₅₆ }, the total similarity (eg, the total similarity of t ₃ is (f ₃₅ + f ₃₆ )) and the total similarity of t ₅ is (f ₅₃ + f ₅₆ ). , T _{6 with} the largest sum of similarity (f ₆₅ + f ₆₃ )) is selected as the new central sentence in this section.

Next, in step S204, the steps 202 and 203 are repeated until the new central sentence does not change or the number of repetitions reaches a predetermined threshold value, that is, until each of the plurality of sections becomes stable. In this way, the constituent statements of each section are determined, that is, clustering is completed. It should be understood that the clustering method is not limited to the calculation of similarity, and other appropriate methods may be adopted for clustering.

In addition, each constituent sentence in a plurality of sections may be marked. For example, each constituent sentence included in the first section may be marked as "0", and each constituent sentence included in the second section may be marked as "1". Constituents marked with the same mark belong to the same section.

FIG. 5 is a schematic diagram for dividing the input text by using clustering according to the embodiment of the present disclosure. As shown in FIG. 5, the constituent sentence 21 of the input text is T = {t ₁ , t ₂ , t ₃ , ..., T ₇ }, and the number of a plurality of sections obtained by dividing the input text by clustering is It is assumed that M = 3 (1st section 22, 2nd section 23 and 3rd section 24), and the central sentences selected for the initialization of the 3 sections are t ₁ , t ₂ and t ₃ , respectively. To do. Next, the similarity is calculated by calculating the similarity between the remaining constituent sentences of the input text {t ₄ , t ₅ , t ₆ , t ₇ } and each central sentence {t ₁ , t ₂ , t ₃ }. Based on the degree, the constituent sentences {t ₄ , t ₅ , t ₆ , t ₇ } of the input text can be assigned to the sections corresponding to the central sentences {t ₁ , t ₂ , t ₃ }, respectively. For example, as shown in FIG. 5, t ₄ is assigned to the first section 25, t ₆ and t ₇ are assigned to the second section 26, and t ₅ is assigned to the third section 27. Next, in each section, the similarity between each constituent sentence is calculated, and the constituent sentence having the maximum total similarity is determined as a new central sentence. For example, in section 26, the similarity between t ₂ , t ₆ , and t ₇ is calculated, and t ₆ with the maximum total similarity is determined as the new central sentence (represented by a circle in the figure). To do. Repeat the above process until the central sentence does not change.

After the input text is divided into multiple sections by clustering, that is, after the highly similar constituents in the input text are assigned to one section, the next processing is performed on the assigned multiple sections. Get the output text.

Returning to FIG. 1, in step S104, output text is generated based on the plurality of sections and the analysis result.

For example, using the analysis result corresponding to the input text acquired in step S102, in each of the plurality of sections acquired in step S103, the constituent sentence having the maximum sentence weight in the section corresponds to the section. Select as the output result, obtain the analysis result corresponding to each section, and then combine the output results of multiple sections to generate the output text.

FIG. 6 shows an exemplary schematic diagram for obtaining analysis results corresponding to each of the plurality of sections.

As shown in FIG. 6, first, the text 60 is input to the neural network. After that, the input text 60 may be divided into a plurality of sections by using the clustering 65, and the number of the plurality of sections is determined by a predetermined target number of sentences of the output text 66 of the text processing layer 61. At the same time, all the constituent sentences of the input text are analyzed using different processing methods 64, a sentence vector of each constituent sentence is generated, and the vocabulary vector 62 of each constituent sentence is acquired by all the constituent sentences of the input text 60. Next, based on the acquired sentence vector and vocabulary vector, a feature amount is extracted using VGG-16, and the feature amount corresponds to the weight of each constituent sentence of the input text 60 and is divided by clustering. It also corresponds to the weight of each constituent sentence in the plurality of sections obtained by. Next, in each of the plurality of sections, the constituent sentence having the maximum sentence weight is selected as the analysis result corresponding to the section. Finally, the output results of the plurality of sections are combined to generate the output text 66.

By analyzing based on the sentence vector and the vocabulary vector based on the present disclosure, it is possible to subdivide from the depth of sentence granularity to the depth of vocabulary granularity, thereby improving the recognition of special sentences. Can be done. Further, by using VGG-16, the relationship between the feature values can be better grasped, and more detailed features can be extracted. Furthermore, by using clustering to divide the input text into multiple sections, the purpose is to divide the same type of constituent sentences into the same section, and extract summaries from multiple sections to remove duplicate sentences. Can be achieved.

Hereinafter, the text processing apparatus 1000 according to the embodiment of the present disclosure will be described with reference to FIG. 7. FIG. 7 is a schematic view of the text processing device 1000 according to the embodiment of the present disclosure. Since the function of the text processing apparatus of this embodiment is the same as the details of the method described above with reference to FIG. 1, detailed description of the same contents will be omitted here for convenience.

As shown in FIG. 7, the text processing apparatus 1000 according to the embodiment of the present disclosure includes an acquisition unit 1001, an analysis unit 1002, a division unit 1003, and a generation unit 1004. Although the text processing apparatus 1000 is shown in FIG. 7 to include only four components, this is merely an example, and the text processing apparatus 1000 also includes one or more other components. You may. Since these parts have nothing to do with the idea of the present invention, they are omitted here.

As shown in FIG. 7, the acquisition unit 1001 acquires the input text.

The input text is the original text that is expected to be processed, and the original text produces a summary of the expected vocabulary or expected number of sentences.

The analysis unit 1002 analyzes the input text and acquires the analysis result corresponding to the input text.

Traditionally, there are many different processing methods. For example, for text summary extraction, there are various methods such as position method, phrase method, title method, and keyword method. For example, the analysis unit 1002 assigns different weights to each text processing method, analyzes all the constituent sentences of the input text by each text processing method, and acquires the weights of all the constituent sentences of the input text as the analysis result. May be good.

Alternatively, the analysis unit 1002 may acquire the weight of each constituent sentence by using the sentence vector of the constituent sentence of the input text. Instead, the analysis unit 1002 uses all the constituent sentences of the input text to obtain the corresponding vocabulary vector, and analyzes the input text using different processing methods to obtain the constituent sentence vector as the vocabulary vector. Then, the feature amount may be extracted by using a method such as VGG-16 to obtain the weight of each constituent sentence. By performing analysis based on the sentence vector and the vocabulary vector, it is possible to subdivide from the depth of sentence granularity to the depth of vocabulary granularity, and thus the recognition system of a special sentence can be improved. Further, by using VGG-16, the relationship between the feature values can be better grasped, and thus more detailed features can be extracted.

FIG. 2 is a schematic diagram for acquiring the weights of the constituent sentences of the input text based on the sentence vector and the vocabulary vector. As shown in FIG. 2, first, all the constituent sentences 70 ({S ₁ , S ₂ , S ₃ , ..., S ₇ }) of the input text are input to the neural network. Then, in the neural network, the analysis unit 1002 analyzes all the constituent sentences 70 of the input text by using different processing methods, and generates the sentence vector 71 of each constituent sentence. Next, the analysis unit 1002 acquires the vocabulary vector 72 of all the constituent sentences of the input text. Next, the analysis unit 1002 uses the VGG-16 based on the acquired sentence vector 71 and the vocabulary vector 72, and the feature amount 74 ({V ₁ , V ₂ , ,) corresponding to each constituent sentence 70 of the input text. V _{3, ...,} extracts the _V 7}).

It should be understood that the above method for obtaining the weight of the constituent sentence is not limited to this, and another appropriate method may be adopted for obtaining the weight of the constituent sentence.

Next, the division unit 1003 divides the input text into a plurality of sections using clustering.

Here, the number of a plurality of sections obtained by dividing the input text by the dividing unit 1003 by clustering is determined by a predetermined target number of sentences of the output text. For example, when the predetermined target number of sentences of the desired output text is M, the number of a plurality of sections obtained by dividing the input text by clustering is also M.

FIG. 3 shows a schematic diagram in which clustering is applied to a text processing method based on a neural network. As shown in FIG. 3, assuming that the neural network contains N (N ≧ 2) cascaded text processing layers, the nth text processing layer of the cascaded N text processing layers clusters. The number of sections obtained by dividing the input text by using it is determined by a predetermined target number of sentences of the output text of the nth text processing layer. For example, the number of a plurality of sections obtained by dividing the input text 51 by the text processing layer 1 (represented by 52 in the figure) using the clustering 50 is the output text 1 of the text processing layer 1 (53 in the figure). It may be determined by a predetermined target number of sentences (represented by), or the text processing layer 2 (represented by 54 in the figure) divides the input text (that is, the output text 1 in the figure) by using clustering 56. The number of the plurality of sections thus obtained may be determined by a predetermined target number of sentences of the output text 2 (represented by 55 in the figure) of the text processing layer 2.

The neural network may contain only one text processing layer, in which case the number of sections obtained by clustering the input text depends on a predetermined target number of output texts in the text processing layer. Please be aware that it will be decided.

After acquiring the number of a plurality of sections obtained by dividing the input text by clustering, the division unit 1003 performs an operation of dividing the input text by using clustering.

FIG. 4 is a flowchart of the method 200 for dividing the input text by using clustering according to the embodiment of the present disclosure.

As shown in FIG. 4, in step S201, the division unit 1003 initializes a plurality of central sentences corresponding to the plurality of sections.

When the number of a plurality of sections is M, it is necessary to select one central sentence for each section, so it is necessary to select M central sentences. For example, the first M sentences of the input text can be selected as the central sentence at the time of initialization, or M sentences can be randomly selected as the central sentence at the time of initialization from the input text. It should be recognized that the method of initializing the central sentence is not limited to this, and other appropriate methods can be adopted to initialize the central sentence.

In step S202, the division unit 1003 divides the composition sentence of the input text into sections corresponding to the plurality of central sentences based on the similarity by calculating the similarity between the composition sentence of the input text and the plurality of central sentences. And update the constructs contained in multiple sections.

For example, the similarity (or distance) may be calculated by the Rouge-1 method. For example, the similarity f _ij of the i-th sentence and j th minute can be expressed as follows.

However, S _i represents the number of p vocabularies in the i-th sentence (for example, p = 2 represents two vocabularies, p = 3 represents three vocabularies), and R _ij represents the i-th sentence. The number of repeating p vocabularies in the jth sentence is shown. The larger f _{ij, the} higher the similarity between the i-th sentence and the j-th sentence. It should be understood that the method for calculating the similarity is not limited to this, and other appropriate methods may be adopted for calculating the similarity.

Next, by calculating the similarity between the constituent sentence in the input text and the plurality of central sentences, the constituent sentence of the input text having a high degree of similarity is assigned to the section corresponding to the central sentence.

For example, the constituent sentence of the input text is T = {t ₁ , t ₂ , t ₃ , ..., t _z }, and the number of a plurality of sections obtained by dividing the input text by clustering is M = 2 (that is, that is). Assuming that the first two sentences t ₁ and t ₂ are selected as the central sentences of the two sections obtained by the first section and the second section), then the rest of the input text is constructed. Similarity between the sentence {t ₃ , t ₄ , ..., t _z } and the central sentence t ₁ , {f ₃₁ , f ₄₁ , ..., f _{z 1} }, and {t ₃ , t ₄ , ..., t _z } similarity _{{f 32, f 42, ...} , f z2} between the center sentence _{t 2} and by calculating the _high center sentence similarity configuration statement _{t 3} by comparing _{f 31} and _{f 32} (For example, if f ₃₁ > f ₃₂ , the constituent sentence t ₃ may be assigned to the first section, and if f ₃₁ <f ₃₂ , the constituent sentence t ₃ may be assigned to the second section). , F ₄₁ and f ₄₂ may be compared to assign the constituent sentence t ₄ to the section corresponding to the central sentence having a high degree of similarity.

After assigning all the constructs, update the constructs contained in multiple sections.

Next, in step S203, in each of the plurality of sections, the division unit 1003 calculates the similarity between the constituent sentences to determine the constituent sentence having the maximum total similarity as the new central sentence.

For example, assuming that the constructs contained in the first section are {t ₃ , t ₅ , t ₆ }, the similarity {t ₃₅ , t ₃₆ between each construct {t ₃ , t ₅ , t ₆ } , T ₅₆ }, the total similarity (eg, the total similarity of t ₃ is (f ₃₅ + f ₃₆ )) and the total similarity of t ₅ is (f ₅₃ + f ₅₆ ). , T _{6 with} the largest sum of similarity (f ₆₅ + f ₆₃ )) is selected as the new central sentence in this section.

Next, in step S204, the above process is repeated until the new central sentence does not change or the number of repetitions reaches a predetermined threshold, that is, until each of the plurality of sections becomes stable. In this way, the constituent statements of each section are determined, that is, clustering is completed. It should be understood that the clustering method is not limited to the calculation of similarity, and other appropriate methods may be adopted for clustering.

In addition, each constituent sentence in a plurality of sections may be marked. For example, each constituent sentence included in the first section may be marked as "0", and each constituent sentence included in the second section may be marked as "1". Constituents marked with the same mark belong to the same section.

FIG. 5 is a schematic diagram for dividing the input text by using clustering according to the embodiment of the present disclosure. As shown in FIG. 5, the constituent sentence 21 of the input text is T = {t ₁ , t ₂ , t ₃ , ..., T ₇ }, and the number of a plurality of sections obtained by dividing the input text by clustering is It is assumed that M = 3 (1st section 22, 2nd section 23 and 3rd section 24), and the central sentences selected for the initialization of the 3 sections are t ₁ , t ₂ and t ₃ , respectively. To do. Next, the similarity is calculated by calculating the similarity between the remaining constituent sentences of the input text {t ₄ , t ₅ , t ₆ , t ₇ } and each central sentence {t ₁ , t ₂ , t ₃ }. Based on the degree, the constituent sentences {t ₄ , t ₅ , t ₆ , t ₇ } of the input text can be assigned to the corresponding sections of the central sentence {t ₁ , t ₂ , t ₃ }, respectively. For example, as shown in FIG. 5, t ₄ is assigned to the first section 25, t ₆ and t ₇ are assigned to the second section 26, and t ₅ is assigned to the third section 27. Next, in each section, the similarity between each constituent sentence is calculated, and the constituent sentence having the maximum total similarity is determined as a new central sentence. For example, the determination in the second section 26, t _{2, t 6,} and calculates the similarity between t _7, as total new center sentence t ₆ is the maximum degree of similarity (represented by circles in the drawing) To do. Repeat the above process until the central sentence does not change.

After the input text has been divided into multiple sections by clustering, that is, after the highly similar constituents in the input text have been assigned to the same section, then the assigned sections are processed. Get the output text.

Next, the generation unit 1004 generates an output text based on the plurality of sections and the analysis result.

For example, the generation unit 1004 uses the analysis result corresponding to the acquired input text, and in each of the acquired plurality of sections, the constituent sentence having the maximum sentence weight of the section is output as the output result corresponding to the section. After selecting and acquiring the analysis results corresponding to each section, the output results of multiple sections are combined to generate the output text.

FIG. 6 shows an exemplary schematic diagram for obtaining analysis results corresponding to each of the plurality of sections.

As shown in FIG. 6, first, the text 60 is input to the neural network. After that, the division unit 1003 may divide the input text 60 into a plurality of sections by using the clustering 65, and the number of the plurality of sections depends on the predetermined target number of sentences of the output text 66 by the text processing layer 61. It is determined. At the same time, the analysis unit 1002 analyzes all the constituent sentences of the input text using different processing methods 64, generates a sentence vector of each constituent sentence, and vocabulary of each constituent sentence by all the constituent sentences of the input text 60. The vector 62 is acquired, and then, based on the acquired sentence vector and vocabulary vector, the feature amount is extracted using VGG-16, and the feature amount corresponds to the weight of each constituent sentence of the input text 60. Also, it corresponds to the weight of each constituent sentence in a plurality of sections obtained by dividing by clustering. Next, in each of the plurality of sections, the constituent sentence having the maximum sentence weight is selected as the analysis result corresponding to the section. Finally, the generation unit 1004 combines the output results of the plurality of sections to generate the output text 66.

Hereinafter, the text processing apparatus 1100 according to the embodiment of the present disclosure will be described with reference to FIG. FIG. 8 is a schematic diagram of a text processing device according to the embodiment of the present disclosure. Since the function of the text processing apparatus of this embodiment is the same as the details of the method described above with reference to FIG. 1, detailed description of the same contents will be omitted here for convenience.

As shown in FIG. 8, the text processing device 1100 includes a memory 1101 and a processor 1102. Although the text processing device 1100 is shown in FIG. 8 to include only two devices, this is merely an example, and the text processing device 1100 also includes one or more other devices. May be good. Since these devices have nothing to do with the idea of the present invention, they are omitted here.

The text processing device 1100 based on the neural network of the present disclosure is configured to execute a memory 1101 configured to store computer-readable instructions and the computer-readable instructions stored in the memory. It includes a processor 1102. When the processor 1102 executes the computer-readable instruction, it obtains an input text, analyzes the input text, obtains an analysis result corresponding to the input text, and utilizes clustering. Dividing the input text into a plurality of sections and generating an output text based on the plurality of sections and the analysis result are executed.

Here, dividing the input text into a plurality of sections by using clustering includes an initialization step for initializing a plurality of central sentences corresponding to the plurality of sections, a constituent sentence of the input text, and the plurality of sections. By calculating the similarity with the central sentence of, the constituent sentences of the input text are assigned to the sections corresponding to the plurality of central sentences based on the similarity, and the constituent sentences included in the plurality of sections are updated. By calculating the similarity between each constituent sentence in the update step and the plurality of sections, the new central sentence remains unchanged from the determination step in which the constituent sentence having the maximum total similarity is determined as the new central sentence. It includes repeating the above update step and determination step until.

Here, to analyze the input text and obtain the analysis result corresponding to the input text, the analysis result analyzes all the constituent sentences of the input text and determines the weight of each sentence of all the constituent sentences. Including getting as.

Here, generating output text based on a plurality of sections and analysis results means that in each of the plurality of sections, the sentence weight is the maximum in the section, based on the sentence weights of all the constituent sentences. This includes selecting the constituent sentence to be the output result corresponding to the section, and combining the output results of a plurality of sections to generate the output text.

Here, the neural network may include only one text processing layer, and the number of a plurality of sections obtained by dividing the input text by clustering is determined by a predetermined target number of sentences of the output text of the text processing layer. To.

Here, the neural network may include N (N ≧ 2) cascaded text processing layers, in which case the nth text processing layer in the cascaded N text processing layers is clustered. The number of the plurality of sections obtained by dividing the input text is determined by a predetermined target number of sentences of the output text of the nth text processing layer.

FIG. 9 is a schematic diagram of a computer-readable storage medium according to the embodiment of the present disclosure.

As shown in FIG. 9, the present disclosure further comprises a computer-readable storage medium 1200 in which a computer-readable instruction 1201 is stored. When the computer-readable instruction is executed by the computer, the computer performs a text processing method. This method involves obtaining the input text, analyzing the input text, obtaining the analysis result corresponding to the input text, and using clustering to divide the input text into a plurality of sections. , The generation of output text based on the plurality of sections and the analysis results.

<Hardware configuration>
The block diagram used in the description of the above embodiment shows a block of functional units. These functional blocks (components) are realized by any combination of hardware and / or software. Further, the means for realizing each functional block is not particularly limited. That is, each functional block may be realized by one physically and / or logically coupled device, or directly and / or indirectly by two or more physically and / or logically separated devices. (For example, wired and / or wirelessly) may be connected and realized by these plurality of devices.

In the following description, the word "device" can be read as a circuit, device, unit, or the like. For example, although only one processor is shown, there may be multiple processors. Further, the processing may be executed by one processor, or the processing may be executed simultaneously, sequentially, or by other methods on one or more processors. The processor may be mounted on one or more chips.

The processor, for example, runs an operating system to control the entire computer. The processor may be composed of a central processing unit (CPU: Central Processing Unit) including an interface with peripheral devices, a control device, an arithmetic unit, a register, and the like.

Further, the processor reads a program (program code), software module, data, etc. from the storage and / or communication device into the memory, and executes various processes according to these. As the program, a program that causes a computer to execute at least a part of the operations described in the above embodiment is used. For example, the control unit may be realized by a control program stored in a memory and running on a processor, and may be realized for other functional blocks as well. The memory is a computer-readable storage medium, such as ROM (Read Only Memory), EPROM (Erasable ProgRAMmable ROM), EPROM (Electrically EPROM), RAM (Random Access Memory), or at least one of other suitable storage media. It may be composed of one. The memory may be referred to as a register, a cache, a main memory (main storage device), or the like. The memory can store a program (program code), a software module, or the like that can be executed to implement the wireless communication method according to the embodiment of the present disclosure.

The storage is a computer-readable storage medium, such as a flexible disk, a floppy (registered trademark) disk, a magneto-optical disk (for example, a compact disk (CD-ROM (Compact Disk ROM))), a digital versatile disk, or a Blu. -Ray® disks), removable disks, hard disk drives, smart cards, flash memory devices (eg cards, sticks, key drivers), magnetic stripes, databases, servers, and other suitable It may be composed of at least one storage medium. The storage may be referred to as auxiliary storage.

The input device is an input device (for example, a keyboard, a mouse, a microphone, a switch, a button, a sensor, etc.) that receives an input from the outside. The output device is an output device (for example, a display, a speaker, an LED (Light Emitting Diode) lamp, etc.) that outputs to the outside. The input device and the output device may have an integrated configuration (for example, a touch panel).

Further, the information, parameters, etc. described in the present specification may be represented by an absolute value, a relative value from a predetermined value, or another corresponding information. .. For example, the radio resource may be indicated by a predetermined index. Further, mathematical formulas and the like using these parameters may differ from those expressly disclosed herein.

The names used for parameters and the like in the present specification are not limited in any respect.

The input / output information, signals, etc. may be stored in a specific location (for example, a memory) or may be managed by a management table. Input / output information, signals, etc. can be overwritten, updated, or added. The output information, signals, etc. may be deleted. The input information, signals, etc. may be transmitted to other devices.

The notification of information is not limited to the embodiments / embodiments described herein, and may be made by other methods. For example, information notification includes physical layer signaling (for example, downlink control information (DCI), uplink control information (UCI)), upper layer signaling (eg, RRC (Radio Resource Control) signaling, etc.). It is implemented by broadcast information (Master Information Block (MIB), System Information Block (SIB), Media Access Control (MAC: Medium Access Control) signaling), other signals, or a combination thereof. You may.

In addition, the notification of predetermined information (for example, the notification of "being X") is not limited to the explicit notification, but implicitly (for example, by not notifying the predetermined information or another). It may be done (by notification of information).

The determination may be made by a value represented by 1 bit (0 or 1), or by a boolean value represented by true (true) or false (false). , May be done by numerical comparison (eg, comparison with a given value).

Software is an instruction, instruction set, code, code segment, program code, program, subprogram, software module, whether called software, firmware, middleware, microcode, hardware description language, or another name. , Applications, software applications, software packages, routines, subroutines, objects, executable files, execution threads, procedures, functions, etc. should be broadly interpreted to mean.

Further, software, instructions, information and the like may be transmitted and received via a transmission medium. For example, the software uses wired technology (coaxial cable, fiber optic cable, twist pair, digital subscriber line (DSL: Digital Subscriber Line), etc.) and / or wireless technology (infrared, microwave, etc.) to create a website, server. , Or when transmitted from other remote sources, these wired and / or wireless technologies are included within the definition of transmission medium.

The terms "system" and "network" as used herein are used interchangeably.

Each aspect / embodiment described in the present specification may be used alone, in combination, or may be switched and used according to the execution. Further, the order of the processing procedures, sequences, flowcharts, etc. of each aspect / embodiment described in the present specification may be changed as long as there is no contradiction. For example, the methods described herein present elements of various steps in an exemplary order, and are not limited to the particular order presented.

Each aspect / embodiment described herein includes LTE (Long Term Evolution), LTE-A (LTE-Advanced), LTE-B (LTE-Beyond), SUPER 3G, IMT-Advanced, 4G (4th generation mobile). communication system), 5G (5th generation mobile communication system), FRA (Future Radio Access), New-RAT (Radio Access Technology), NR (New LTE), NR (New LTE), LTE GSM® (Global System for Mobile communications), CDMA2000, UMB (Ultra Mobile Broadband), IEEE 802.11 (Wi-Fi®), LTE 802.16 (WiMAX) .20, UWB (Ultra-WideBand), Bluetooth®, and other systems that utilize suitable wireless communication methods and / or extended next-generation systems based on them.

The phrase "based on" as used herein does not mean "based on" unless otherwise stated. In other words, the statement "based on" means both "based only" and "at least based on".

Any reference to elements using designations such as "first", "second" as used herein does not generally limit the quantity or order of those elements. These designations can be used herein as a convenient way to distinguish between two or more elements. Thus, references to the first and second elements do not mean that only two elements can be employed or that the first element must somehow precede the second element.

The term "determining" as used herein may include a wide variety of actions. For example, a "decision" is a calculation, computing, processing, deriving, investigating, searching up (eg, a table, database or another data). It may be regarded as "judgment (decision)" of search in structure, ascertaining, and the like. In addition, "judgment (decision)" includes receiving (for example, receiving information), transmitting (transmitting) (for example, transmitting information), input (input), output (output), and access (for example). It may be regarded as "decision" (for example, accessing data in memory) (accessing) or the like. In addition, "judgment (decision)" is considered to be "judgment (decision)" such as solving, selecting, selecting, erection (setting), and comparing (comparing). May be good. That is, "judgment (decision)" may be regarded as "judgment (decision)" of some action.

As used herein, the terms "connected", "coupled", or any variation thereof, may be any direct or indirect connection between two or more elements or. It means a bond and can include the presence of one or more intermediate elements between two elements that are "connected" or "bonded" to each other. The connection or connection between the elements may be physical, logical, or a combination thereof. For example, "connection" may be read as "access." As used herein, the two elements are by using one or more wires, cables and / or printed electrical connections, and, as some non-limiting and non-comprehensive examples, radio frequencies. It can be considered to be "connected" or "coupled" to each other by using electromagnetic energy or the like having wavelengths in the region, microwave region and / or light (both visible and invisible) regions.

As used herein or in the claims, "inclusion," "comprising," and variations thereof, these terms are as comprehensive as the term "comprising." Is intended to be. Furthermore, the term "or" as used herein or in the claims is intended not to be an exclusive OR.

Although the present disclosure has been described in detail above, it will be apparent to those skilled in the art that the present disclosure is not limited to the embodiments described herein. The present disclosure can be implemented as a calibration and modification mode without departing from the spirit and scope of the present disclosure, which is determined by the description of the claims. Therefore, the description herein is for purposes of illustration only and has no limiting implications for the present disclosure.

Claims (14)

A text processing method based on a neural network
To get the input text and
To analyze the input text and obtain the analysis result corresponding to the input text,
Using clustering to divide the input text into multiple sections
A text processing method comprising generating output text based on the plurality of sections and the analysis results. Dividing the input text into a plurality of sections using the clustering
An initialization step that initializes a plurality of central sentences corresponding to the plurality of sections,
By calculating the similarity between the constituent sentence of the input text and the plurality of central sentences, the constituent sentences of the input text are assigned to the sections corresponding to the plurality of central sentences based on the similarity, and the plurality of said Update steps to update the constructs contained in the section,
In the plurality of sections, a determination step of determining the composition sentence having the maximum total similarity as a new central sentence by calculating the similarity between the composition sentences, and
The text processing method according to claim 1, further comprising a repeating step of repeating the update step and the determination step until the new central sentence remains unchanged. Analyzing the input text and obtaining the analysis result corresponding to the input text
The text processing method according to claim 1, which comprises analyzing all the constituent sentences of the input text and acquiring the weights of the sentences of each of the constituent sentences as the analysis result. Generating output text based on the plurality of sections and the results of the analysis
Based on the sentence weights of all the constituent sentences, in each of the plurality of sections, the constituent sentence having the maximum sentence weight in the section is selected as the output result corresponding to the section.
The text processing method according to claim 3, comprising combining the output results of a plurality of sections to generate output text. The neural network includes one text processing layer, and the number of a plurality of sections obtained by dividing the input text by clustering is determined by a predetermined target number of sentences of the output text of the text processing layer. The text processing method according to any one of 1 to 4. The neural network includes cascaded N (N ≧ 2) text processing layers.
The number of sections obtained by dividing the input text by the nth text processing layer in the cascaded N text processing layers is a predetermined target sentence of the output text of the nth text processing layer. The text processing method according to any one of claims 1 to 4, which is determined by a number. A text processing device based on a neural network
The acquisition part that acquires the input text, and
An analysis unit that analyzes the input text and acquires the analysis result corresponding to the input text.
A division part that divides the input text into a plurality of sections using clustering,
A text processing apparatus including the plurality of sections and a generation unit that generates output text based on the analysis result. The divided portion
An initialization step that initializes a plurality of central sentences corresponding to the plurality of sections,
By calculating the similarity between the constituent sentence of the input text and the plurality of central sentences, the constituent sentences of the input text are assigned to the sections corresponding to the plurality of central sentences based on the similarity, and the plurality of said Update steps to update the constructs contained in the section,
In each of the plurality of sections, a determination step of determining the composition sentence having the maximum total similarity as a new central sentence by calculating the similarity between the composition sentences, and
The text processing apparatus according to claim 7, wherein the update step and the repeating step of repeating the determination step are performed until the new central sentence does not change. The text processing apparatus according to claim 7, wherein the analysis unit analyzes all the constituent sentences of the input text and acquires the weight of each sentence of all the constituent sentences as the analysis result. The generator
Based on the sentence weights of all the constituent sentences, in each of the plurality of sections, the constituent sentence having the maximum sentence weight of the section is selected as the output result corresponding to the section.
The text processing apparatus according to claim 9, wherein the output results of a plurality of sections are combined to generate output text. The neural network includes one text processing layer, and the number of a plurality of sections obtained by dividing the input text by clustering is determined by a predetermined target number of sentences of the output text of the text processing layer. The text processing apparatus according to any one of 7 to 10. The neural network includes cascaded N (N ≧ 2) text processing layers.
The number of sections obtained by dividing the input text by the nth text processing layer in the cascaded N text processing layers is a predetermined target sentence of the output text of the nth text processing layer. The text processing apparatus according to any one of claims 7 to 10, which is determined by a number. A text processing device based on a neural network
Memory configured to store computer-readable instructions,
A processor configured to execute the computer-readable instructions stored in the memory.
With
Here, when the processor executes the computer-readable instruction,
To get the input text and
To analyze the input text and obtain the analysis result corresponding to the input text,
Using clustering to divide the input text into multiple sections
A text processing device that generates and performs output text based on the plurality of sections and the analysis results. A computer-readable storage medium that stores computer-readable instructions.
When the computer-readable instruction is executed by the computer, the computer
To get the input text and
To analyze the input text and obtain the analysis result corresponding to the input text,
Using clustering to divide the input text into multiple sections
A computer-readable storage medium that performs text processing methods, including generating output text based on the plurality of sections and the results of the analysis.

Images