Data for learning generation method, data for learning generation program, data for learning generation device, and composition output device

Abstract

To provide: a data for learning generation method for generating data for learning including a suitable output composition; a data for learning generation program; a data for learning generation device; and a composition output device.SOLUTION: A generation method for learning data includes: calculating degree of similarity in a plurality of sentences included in an output sentence among data sets built as data for learning used for a machine learning model outputting a sentence; deleting a specific sentence selected based on the degree of similarity from the output composition.SELECTED DRAWING: Figure 4

Description

[Technical field]

[0001] The present disclosure relates to a learning data generation method, a learning data generation program, a learning data generation device, and a text output device. [Background technology]

[0002] In recent years, for natural language processing tasks that output sentences, sentence output models have been proposed that extract important parts of input information and reconstruct the sentences. For example, when article information is input into a model, a summary of the article is output from the model. The output summary is then used as the headline, outline, or brief description of the article. Furthermore, for natural language processing tasks, sentence output models have been proposed that output multimodal (e.g., text, image, video, audio, etc.) input information as text information. The output sentences are used, for example, for image or video captions. In such tasks, it is required to output sentences that contain the necessary information without excess or deficiency.

[0003] For example, Patent Literature 1 discloses a learning device that improves the accuracy of machine learning. This learning device generates input and output pairs as positive examples and generates pairs with changed combinations of input and output as negative examples, and uses the positive and negative examples to learn a filter model. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] JP 2021-140558 A Summary of the Invention [Problem to be solved by the invention]

[0005] However, the learning device of Patent Document 1 is designed to devise a combination of learning data sets, and does not optimize the output sentences of the learning data themselves. Therefore, there is a risk that the sentences output from the model may contain excessive sentences with overlapping contents (for example, similar meanings, redundant expressions, etc.).

[0006] An object of the present disclosure is to provide a learning data generation method, a learning data generation program, a learning data generation device, and a sentence output device for generating learning data including appropriate output sentences. [Means for solving the problem]

[0007] The training data generation method according to the present disclosure includes calculating a similarity between a plurality of sentences included in an output sentence from a dataset constructed as training data to be used in a machine learning model that outputs sentences, and deleting from the output sentence a specific sentence selected based on the similarity.

[0008] The learning data generation program of the present disclosure causes a computer to calculate the similarity of multiple sentences included in an output sentence from a dataset constructed as learning data used in a machine learning model that outputs sentences, and delete from the output sentence a specific sentence selected based on the similarity.

[0009] The learning data generation device of the present disclosure includes a calculation unit that calculates the similarity of multiple sentences included in an output sentence from a dataset constructed as learning data used in a machine learning model that outputs sentences, and a correction unit that deletes a specific sentence selected based on the similarity from the output sentence.

[0010] The sentence output device of the present disclosure includes an acquisition unit that acquires input information, and an output unit that inputs the input information to a sentence output model trained using learning data generated by the above-mentioned learning data generation method, and outputs an output sentence from the sentence output model. Effect of the Invention

[0011] According to the present disclosure, it is possible to generate learning data that includes appropriate output sentences. [Brief description of the drawings]

[0012] [Figure 1] FIG. 1 is a diagram showing an overview of the first embodiment. [Diagram 2] FIG. 2 is a block diagram showing a functional configuration of the learning data generating device. [Diagram 3] FIG. 3 is a block diagram showing a hardware configuration of the learning data generating device. [Figure 4] FIG. 4 is a flowchart showing the learning data generation process. [Diagram 5] FIG. 5 is a diagram showing how similarities between a plurality of sentences are calculated. [Figure 6] FIG. 6 is a diagram showing the calculation results of the similarity. [Figure 7] FIG. 7 shows the corrected output sentence. [Figure 8] FIG. 8 is a diagram showing an overview of the third embodiment. [Figure 9] FIG. 9 is a block diagram showing a functional configuration of the model generating device. [Figure 10] FIG. 10 is a flowchart showing the process of generating a text output model. [Figure 11] FIG. 11 is a block diagram showing the functional configuration of the text output device. [Figure 12] FIG. 12 is a flowchart showing the text output process. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] Hereinafter, embodiments of the present disclosure will be described with reference to the drawings.

[0014] (Embodiment 1) An overview of the present disclosure will be described. For example, as shown in FIG. 1, when a dataset is constructed as learning data used for learning a machine learning model that outputs sentences, a learning data generation device 1 calculates the similarity of multiple sentences included in output sentences of the dataset. For example, the learning data may be composed of a dataset of input information and output sentences that summarize the input information. Then, the learning data generation device 1 selects a specific sentence from the multiple sentences based on the calculated similarity, and deletes the specific sentence from the output sentence.

[0015] Generally, when training data is constructed for use in a training model, sentences with extremely low similarity, such as sentences that are completely unrelated, are removed. On the other hand, training data may contain sentences with extremely high similarity, such as sentences with the same meaning or redundant expressions. Therefore, specific sentences with high similarity are deleted from the training data to correct the training data. Then, by performing machine learning using the corrected training data, a sentence output model that outputs appropriate sentences without excess or deficiency can be generated.

[0016] Next, a detailed description will be given of the configuration of the learning data generation device 1. As shown in FIG.

[0017] The calculation unit 2 calculates the similarity of a plurality of sentences included in an output sentence from a data set constructed as learning data. Here, the learning data is used for learning a machine learning model that outputs sentences, and is composed of a data set of input information and output sentences. The input information is information that can be made into a sentence, and may be composed of, for example, character information, image information, video information, or audio information. The input information may be multimodal information or single-modal information. The output sentence is composed of character information including a plurality of sentences created based on the input information, and may be composed of, for example, a summary sentence that summarizes the input information. Here, a sentence may indicate one including a plurality of coherent sentences. For example, a sentence may be composed of a series of sentences that are related to each other. Also, one sentence may indicate a portion separated by a period.

[0018] The calculation unit 2 may also calculate the similarity between a plurality of sentences based on the word matching rate between the plurality of sentences. For example, correct may be calculated based on the following formula 1, where the numerator of formula 1 indicates the number of matching words between two reference sentences (e.g., the first sentence and the second sentence), and the denominator of formula 1 indicates the number of words included in one of the two reference sentences (e.g., the first sentence or the second sentence).

[0019]

number

[0020] The correction unit 3 deletes from the output text a specific sentence selected based on the similarity calculated by the calculation unit 2. In this way, the learning data is corrected so as not to include excessive sentences with overlapping contents (e.g., similar meanings, redundant expressions, etc.).

[0021] Next, a detailed description will be given of the hardware configuration of the training data generation device 1. For example, as shown in Fig. 3, the training data generation device 1 has a storage device 4, a processor 5, a user interface (UI) device 6, and a communication device 7, which are interconnected via a bus B.

[0022] The programs or instructions for implementing the various functions and processes of the learning data generation device 1 may be downloaded from an external device via a network, etc. Also, the programs or instructions for implementing the various functions and processes of the learning data generation device 1 may be provided from a removable storage medium such as a CD-ROM (Compact Disk-Read Only Memory) or a flash memory.

[0023] The storage device 4 may be realized by one or more non-transitory storage media, such as, for example, random access memory, flash memory, or a hard disk drive, and may store installed programs or instructions as well as files or data used in the execution of the programs or instructions.

[0024] The processor 5 may be realized by one or more central processing units (CPUs), graphics processing units (GPUs), processing circuitry, etc., each of which may be composed of one or more processor cores. The processor 5 executes various functions and processes of the learning data generation device 1 in accordance with data such as programs, instructions, programs, or parameters (e.g., parameters required for executing instructions) stored in the storage device 4.

[0025] The user interface device 6 may be composed of input devices such as a keyboard, mouse, camera, or microphone, output devices such as a display, speaker, headset, or printer, and input / output devices such as a smartphone, tablet, or touch panel.

[0026] It should be noted that the above-described hardware configuration is merely an example, and the learning data generation device 1 according to the present disclosure may be realized by any other appropriate hardware configuration.

[0027] Next, the process of generating learning data by the learning data generating device 1 will be described with reference to the flowchart shown in FIG.

[0028] First, in step S1, the calculation unit 2 shown in Fig. 2 calculates the similarity between a plurality of sentences included in an output sentence from a data set of input information and output sentences constructed as learning data. At this time, the learning data may be acquired from a database such as a corpus, or may be constructed using the acquired sentences. At this time, the output sentence may be composed of a summary sentence that summarizes the input information. Also, the input information may be composed of multimodal information including image information and character information.

[0029] When the calculation unit 2 acquires the learning data, the calculation unit 2 may calculate a word matching rate between a plurality of sentences for each output sentence included in the learning data, and calculate a similarity between the plurality of sentences based on the calculated matching rate. Specifically, the calculation unit 2 may calculate a word matching rate between a plurality of sentences based on Equation 1. For example, as shown in FIG. 5, the calculation unit 2 may perform morphological analysis on each of a plurality of sentences (first sentence, second sentence, third sentence, etc.) included in the output sentence 8 to perform word division. For example, the calculation unit 2 may divide the first sentence, “I will apply for a patent,” into seven words, “I,” “is,” “patent,” “for,” “apply,” “do,” and “will.” The calculation unit 2 may also divide the second sentence, “This is a patent for machine learning,” into five words, “machine,” “learning,” “for,” “patent,” and “is.” The calculation unit 2 may also divide the third sentence, “I will apply for an invention,” into seven words, “I,” “is,” “invention,” “for,” “apply,” “do,” and “will.” The calculation unit 2 may perform word division using an analysis tool such as nagisa.

[0030] Next, the calculation unit 2 may sequentially select two sentences from the output sentence, and calculate the matching rate of the two sentences based on the formula 1. That is, the calculation unit 2 may calculate the matching rate for all combinations of a plurality of sentences. For example, the calculation unit 2 searches for a word "patent" that matches between the first sentence and the second sentence based on the word-divided first sentence and the second sentence, and calculates the number of matching words to be 1 word. Furthermore, the calculation unit 2 calculates the number of words included in the first sentence to be 7 words based on the word-divided first sentence. Then, the calculation unit 2 calculates the matching rate (forward order) of the first sentence and the second sentence to be 1 / 7 = 0.14 based on the formula 1. Furthermore, the calculation unit 2 may calculate the matching rate (reverse order) of the first sentence and the second sentence using the number of words included in the second sentence as the denominator instead of the number of words included in the first sentence. Specifically, the calculation unit 2 calculates the number of words included in the second sentence to be 5 words based on the word-divided second sentence. Similarly to the forward order, the calculation unit 2 calculates the number of matching words "patent" between the first sentence and the second sentence as 1. The calculation unit 2 may then calculate the matching rate (in reverse order) between the first sentence and the second sentence as 1 / 5 = 0.20.

[0031] Similarly, the calculation unit 2 searches for matching words "I", "wa", "wo", "application", "shi", and "masu" between the first sentence and the third sentence based on the word-divided first sentence and the third sentence, and calculates the number of matching words to be six. Furthermore, the calculation unit 2 calculates the number of words included in the first sentence to be seven based on the word-divided first sentence. Then, the calculation unit 2 calculates the matching rate (forward order) between the first sentence and the third sentence to be 6 / 7=0.86 based on the formula 1. Furthermore, the calculation unit 2 may calculate the matching rate (reverse order) between the first sentence and the third sentence using the number of words included in the third sentence as the denominator instead of the number of words included in the first sentence. Specifically, the calculation unit 2 calculates the number of words included in the third sentence to be seven based on the word-divided third sentence. Furthermore, the calculation unit 2 calculates the number of matching words between the first sentence and the third sentence to be six, similar to the forward order. Then, the calculation unit 2 may calculate the matching rate (in reverse order) between the first sentence and the third sentence as 6 / 7=0.86.

[0032] In this way, the calculation unit 2 repeatedly calculates the matching rate for all combinations of sentences included in the output sentence 8. Then, the calculation unit 2 calculates the similarity between the multiple sentences based on the calculated matching rate. For example, as shown in FIG. 6, the calculation unit 2 may set the similarity 9 to a value 9a calculated as the word matching rate between the multiple sentences in forward order, or may set the similarity 9 to a value 9b calculated as the word matching rate between the multiple sentences in reverse order. Furthermore, the calculation unit 2 may set the similarity 9 to an average value 9c of the matching rate in forward order and the matching rate in reverse order.

[0033] In this way, the calculation unit 2 can calculate the similarity 9 with high accuracy based on the word matching rate between multiple sentences. Furthermore, since the calculation unit 2 performs calculation between two sentences selected in sequence from the output document 8, the calculation unit 2 can calculate the similarity 9 with higher accuracy.

[0034] After calculating the similarity 9, the calculation unit 2 may store the similarity 9 in the storage unit in association with each corresponding sentence. Then, the correction unit 3 selects a specific sentence based on the level of the similarity 9 calculated by the calculation unit 2, and deletes the specific sentence from the output sentence 8.

[0035] For example, in step S2, the correction unit 3 may determine whether the calculated similarity 9 is higher than a predetermined threshold. At this time, the user may qualitatively evaluate the similarity of sentences in advance using sentences of the same type as the output sentence 8, and set the threshold based on the evaluation result. For example, the predetermined threshold may be set based on the value of sentences with extremely high similarity, such as the same content (e.g., meaning, gist, etc.) or redundant expressions.

[0036] If the similarity 9 of the multiple sentences is higher than a predetermined threshold, the correction unit 3 proceeds to step S3 and deletes the specific sentence from the output sentence 8. As a result, the specific sentence with an extremely high similarity is deleted, and sentences indicating different contents are left in the output sentence 8. Specifically, the correction unit 3 refers to the similarity 9 shown in Fig. 6, and when it determines that the similarity 9 between the first sentence and the third sentence is higher than a predetermined threshold, it may delete the specific sentence of either the first sentence or the third sentence from the output sentence 8.

[0037] At this time, the correction unit 3 may delete the latter sentence of the multiple sentences as a specific sentence from the output sentence 8. For example, when it is determined that the similarity 9 between the first sentence and the third sentence is higher than a predetermined threshold, the correction unit 3 may delete the third sentence of the first sentence and the third sentence from the output sentence 8 as shown in FIG. 7. In general, when the similarity between multiple sentences is high, the latter sentence tends to be less important than the previous sentence, for example, by elaborating on the previous sentence. Therefore, the correction unit 3 can generate the output sentence 8 in which the first sentence (the sentence that appeared earlier) with high importance remains by deleting the third sentence of the latter.

[0038] In this manner, the calculation unit 2 calculates the similarity 9 of a plurality of sentences included in the output sentence 8 from the data set constructed as the learning data. Then, the correction unit 3 corrects the learning data by deleting from the output sentence 8 a specific sentence selected based on the level of the similarity 9. In this way, the learning data generation device 1 can delete excessive sentences with overlapping content and generate learning data including appropriate output sentences 8. Furthermore, the learning data generation device 1 can prevent the number of learning data from decreasing. Furthermore, since the learning data generation device 1 automatically deletes excessive sentences, the user does not need to perform the work himself, and the cost required for correcting the learning data can be reduced.

[0039] Furthermore, the correction unit 3 determines the latter sentence of the multiple sentences as a specific sentence and deletes it from the output sentence 8. This enables the learning data generation device 1 to generate learning data that includes more appropriate output sentences 8.

[0040] Furthermore, the training data generation device 1 modifies the training data using only the output sentences 8, without using any input information in the data set constructed as the training data. For this reason, the training data may include input information other than character information in the data set. For example, the input information of the training data may be composed of multimodal information including image information and character information. This allows the training data generation device 1 to generate training data including input information in various data formats.

[0041] When the correction unit 3 corrects the learning data, it ends the learning data generation process. On the other hand, when the correction unit 3 determines in step S2 that the similarity 9 between the multiple sentences is equal to or less than a predetermined threshold, it ends the learning data generation process. In this manner, the learning data generation device 1 repeats the generation process until all the learning data is corrected.

[0042] In this embodiment, the calculation unit 2 calculates the importance of words included in multiple sentences, but the present invention is not limited to this as long as the importance can be acquired. For example, the importance may be set in advance by the user for words used in multiple sentences. For example, the user may select words that are frequently used based on the type of input information and set the importance for these words in advance. At this time, the user may create a setting table in which words and importance are associated with each other, and store the setting table in the calculation unit 2. Thereby, the calculation unit 2 may refer to the setting table based on the words used in multiple sentences and acquire the importance corresponding to the words. Since the importance is set in advance in this way, the calculation unit 2 can easily acquire the importance.

[0043] According to this embodiment, the calculation unit 2 calculates a similarity 9 between a plurality of sentences included in the output sentences 8 of a data set constructed as training data. Then, the correction unit 3 corrects the training data by deleting a specific sentence selected based on the similarity 9 from the output sentences 8. This enables the training data generation device 1 to generate training data including appropriate output sentences 8.

[0044] (Embodiment 2) Hereinafter, a second embodiment of the present disclosure will be described. Here, differences from the first embodiment will be mainly described, and common reference numerals will be used for common points with the first embodiment, and detailed descriptions thereof will be omitted.

[0045] In the first embodiment, the calculation unit 2 deletes a specific sentence from the output sentence 8 based on the similarity of multiple sentences included in the output sentence 8. However, the output sentence 8 may contain a sentence with high importance even if the sentence has an extremely high similarity. In such a case, if the specific sentence is deleted based only on the similarity, the sentence with high importance may be deleted.

[0046] Therefore, in the second embodiment, the similarity 9 of the multiple sentences may be corrected so as to decrease the importance of the words included in the multiple sentences in the output sentence 8. That is, when an important word (e.g., a word with high importance according to the occurrence characteristic) that represents the entire output sentence 8 is included in the multiple sentences, the calculation unit 2 may correct the similarity of the multiple sentences so as to decrease the similarity. Here, the importance of a word may be calculated based on, for example, the occurrence rate of the word in the output sentence 8 and the rate of sentences containing the word (the rarity of the word in the output sentence 8). For example, the calculation unit 2 may calculate the importance of a word based on TF-IDF (Term Frequency - Inverse Document Frequency) which is indicated by the product of the occurrence rate of the word in the output sentence 8 and the rate of sentences containing the word. Then, the calculation unit 2 calculates the importance of each word based on the TFIDF which indicates the importance of each word as shown in the following formula 2. gram1 The similarity 9 may be corrected by using a correction coefficient that is the inverse of

[0047]

number

[0048] In this way, the calculation unit 2 corrects the similarity 9 for words included in multiple sentences so that it decreases according to the level of importance in the output sentence 8. This allows the calculation unit 2 to calculate the similarity 9 with higher accuracy. For example, it is possible to prevent a sentence that is necessary in the output sentence 8, such as one that has a high level of importance in the output sentence 8, from being deleted.

[0049] Furthermore, the calculation unit 2 calculates the importance based on the appearance rate of the word and the rate of sentences containing the word in the output sentence 8. This enables the calculation unit 2 to calculate the similarity 9 with higher accuracy.

[0050] The calculation unit 2 may also correct the similarity 9 according to the importance of a word that indicates a predetermined part of speech, excluding particles, among the words included in the multiple sentences. For example, when comparing the first sentence with the third sentence in FIG. 5, there is a high possibility that a particle such as "wa" or "wo" is used as the same word throughout the entire sentence. For this reason, even though the particle is not involved in the content of the sentence, the importance of the word may be calculated to be high based on the TF-IDF. In other words, if the similarity 9 is corrected using the correction coefficient of the particle calculated based on the TF-IDF, there is a possibility that an excessive number of sentences will be left without being deleted. Therefore, the calculation unit 2 may analyze the parts of speech of the words included in the multiple sentences (for example, morphological analysis, etc.) and select words that indicate a predetermined part of speech. The calculation unit 2 may select a part of speech that is involved in the content of the sentence (a part of speech used by different words between sentences), such as a noun (for example, a proper noun, etc.) or a verb. The calculation unit 2 may then correct the similarity 9 by weighting only a specific word that corresponds to the predetermined part of speech with a correction coefficient according to the importance.

[0051] In this way, the calculation unit 2 corrects the similarity 9 according to the importance of words indicating a predetermined part of speech excluding particles among the words included in the multiple sentences. This allows the calculation unit 2 to more accurately calculate the similarity 9. Note that the calculation unit 2 may correct the similarity 9 according to the importance of words indicating a predetermined part of speech excluding articles among the words included in the multiple sentences.

[0052] According to this embodiment, the calculation unit 2 corrects the similarity 9 for words included in multiple sentences so as to decrease according to the importance in the output sentence 8. This allows the calculation unit 2 to calculate the similarity 9 with high accuracy.

[0053] (Embodiment 3) Hereinafter, a third embodiment of the present disclosure will be described. Here, the differences from the first and second embodiments will be mainly described, and common reference numerals will be used for common points with the first and second embodiments, and detailed descriptions thereof will be omitted.

[0054] 8, in this embodiment, when the model generation device 10 acquires the learning data 11 generated by the learning data generation device 1, the model generation device 10 trains the base model 12 using the learning data 11 to generate a sentence output model 13. The model generation device 10 provides the generated sentence output model 13 to the sentence output device 14. On the other hand, when the sentence output device 14 acquires input information from a user, for example, the sentence output device 14 inputs the input information to the sentence output model 13 and provides the sentence output from the sentence output model 13 to the user.

[0055] FIG. 9 shows a functional configuration of the model generating device 10. The model generating device 10 includes an acquisition unit 15 and a generation unit 16. The acquisition unit 15 acquires the learning data 11 generated by the learning data generating device 1. The generation unit 16 trains the base model 12 using the learning data 11 acquired by the acquisition unit 15 to generate a sentence output model 13. Here, the base model 12 is a machine learning model trained for a specific task (e.g., sentence output, etc.), and may be realized by, for example, a convolutional neural network. Like the learning data generating device 1, the model generating device 10 may be composed of hardware such as a storage device, a processor, a user interface device, and a communication device.

[0056] The process of generating the sentence output model 13 by the model generating device 10 will be described with reference to the flowchart shown in FIG.

[0057] First, in step S21, the acquisition unit 15 acquires the training data 11 generated by the training data generation device 1. For example, the acquisition unit 15 may acquire the training data 11, which is a data set of input information and an output sentence summarizing the input information, from the training data generation device 1. The acquisition unit 15 provides the acquired training data 11 to the generation unit 16.

[0058] In step S22, the acquisition unit 15 trains the base model 12 using the learning data 11 provided from the acquisition unit 15 to generate the sentence output model 13. Specifically, the generation unit 16 inputs the input information of the learning data 11 to the base model 12 and acquires an output result from the base model 12. Then, the generation unit 16 adjusts the parameters of the base model 12 according to the error between the acquired output result and the output sentence 8 of the learning data 11. For example, when the base model 12 is a neural network model, the generation unit 16 may adjust the parameters of the base model 12 according to the backpropagation method. The generation unit 16 generates the sentence output model 13 by repeating the above-mentioned training process until a predetermined termination condition is satisfied, such as using all the learning data 11.

[0059] After generating the sentence output model 13, the generation unit 16 provides the sentence output model 13 to the sentence output device 14. The generation unit 16 may generate the sentence output model 13 that outputs output sentences 8 that summarize input information, for example.

[0060] Next, the sentence output device 14 will be described. Fig. 11 shows the functional configuration of the sentence output device 14. The sentence output device 14 has an acquisition unit 17 and an output unit 18. The acquisition unit 17 accepts input information from a user or the like and acquires the input information. The output unit 18 inputs the input information acquired by the acquisition unit 17 to the sentence output model 13 generated by the model generation device 10, and outputs output sentence 8 from the sentence output model 13. Like the learning data generation device 1, the sentence output device 14 may be composed of hardware such as a storage device, a processor, a user interface device, and a communication device.

[0061] The text output process performed by the text output device 14 will be described with reference to the flow chart shown in FIG.

[0062] First, in step S31, the acquisition unit 17 receives input information from a user or the like, and acquires the input information in response to the input of the input information. For example, if the sentence output model 13 has a function of outputting an output sentence 8 that summarizes the input information, the acquisition unit 17 may receive sentence data that the user wishes to summarize as input information. The acquisition unit 17 outputs the acquired input information to the output unit 18.

[0063] When the output unit 18 acquires the input information from the acquisition unit 17, in step S32, the output unit 18 inputs the input information to the sentence output model 13 and outputs the output sentence 8 from the sentence output model 13 to the user. For example, the output unit 18 may output the output sentence 8 summarizing the input information to the user. For example, the output unit 18 may transmit the output sentence 8 to the user via a communication device, or may display the output sentence 8 on a display.

[0064] According to this embodiment, the acquisition unit 17 acquires input information. Then, the output unit 18 inputs the input information to the sentence output model 13 trained using the learning data generated by the learning data generation method, and outputs the output sentence 8 from the sentence output model 13. Therefore, the sentence output device 14 can output the output sentence 8 appropriate to the input information.

[0065] Although specific examples of the present disclosure have been described in detail above, these are merely examples and do not limit the scope of the claims. The technology described in the claims includes various modifications and alterations of the specific examples exemplified above. [Industrial Applicability]

[0066] The learning data generation method according to the present disclosure can be used as a method for generating learning data used in a machine learning model that outputs sentences. [Explanation of symbols]

[0067] 1. Learning data generation device 2 Calculation section 3 Correction section 4 Storage device 5 Processors 6 User Interface Devices 7. Communications Equipment 8 Output text 9,9a~9c Similarity 10. Model Generator 11 Training Data 12 Base model 13 Text output model 14 Text output device

Claims

1. Calculating a similarity between a plurality of sentences included in an output sentence from a dataset constructed as learning data used in a machine learning model that outputs sentences; removing from the output sentences certain sentences selected based on the similarity; A computer-implemented method for generating training data, comprising:

2. The method for generating training data according to claim 1 , wherein the similarity is corrected so as to decrease in accordance with the importance of words contained in the plurality of sentences in the output text.

3. The method for generating training data according to claim 2 , wherein the importance is calculated based on an appearance rate of the word in the output text and a rate of sentences containing the word.

4. The method for generating training data according to claim 2 , wherein the similarity is corrected in accordance with the importance of a word indicating a predetermined part of speech, excluding a particle, among the words included in the plurality of sentences.

5. The method for generating training data according to claim 2 , wherein the importance level is preset for each of the words by a user.

6. The method for generating training data according to claim 1 , wherein the similarity is calculated based on a word matching rate between the plurality of sentences.

7. The method for generating training data according to claim 1 , wherein the training data is composed of a data set of input information including image information or character information and the output sentence.

8. The method for generating training data according to claim 1 , wherein the similarity is calculated between two sentences selected in sequence from the output sentences.

9. The method for generating training data according to claim 1 , wherein the specific sentence indicates a later sentence among the plurality of sentences.

10. Calculating a similarity between a plurality of sentences included in an output sentence from a dataset constructed as learning data used in a machine learning model that outputs sentences; removing from the output sentences certain sentences selected based on the similarity; A learning data generation program that causes a computer to execute the above.

11. A calculation unit that calculates a similarity between a plurality of sentences included in an output sentence from a dataset constructed as learning data used in a machine learning model that outputs a sentence; a correction unit for deleting from the output sentence a particular sentence selected based on the similarity; A learning data generating device comprising:

12. An acquisition unit for acquiring input information; an output unit that inputs the input information to a sentence output model trained using the learning data generated by the learning data generation method according to claim 1, and outputs an output sentence from the sentence output model; A text output device comprising:

Images