Sentence generation device, learning method, sentence generation method, and program

Abstract

To generate a sentence by using words corresponding to the age or experience of an object person in the case of generating the sentence by machine translation or the like.SOLUTION: According to an embodiment of the present disclosure, there is provided a sentence generation device including a sentence generation unit that receives a first text, a second text related to the content of the first text, and a first word included in the second text, and generates a third text by converting the first word in the second text into an expression with a lower difficulty level than the first word based on a trained model.SELECTED DRAWING: Figure 3

Description

[Technical Field]

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

[0002] In recent years, advances in neural machine translation technology have made machine translation accessible to many people. For example, when children use machine translation, it is preferable to generate simplified translations so that children can properly understand the meaning of the translated text.

[0003] A technique has been proposed in the past for performing machine translation according to a specified level of difficulty (Non-Patent Document 1). In this conventional method, the difficulty level of the entire sentence can be specified using a grade from 2 to 12 (higher grades are more difficult). [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Kazuki Tani, Ryoya Yuasa, Kazuki Takikawa, Akihiro Tamura, Tomoyuki Kajiwara, Takashi Ninomiya, and Tsuneo Kato. Abenchmark dataset for multi-level complexity controllable machine translation. In Proceedings of the Thirteenth Language Resources and Evaluation Conference (LREC), pp. 6744.6752, 2022. Summary of the Invention [Problem to be solved by the invention]

[0005] One of the reasons why the meaning of a translated text cannot be properly understood is the use of difficult words. However, while conventional methods can take into account the difficulty of the entire sentence, they have the problem of not being able to take into account the difficulty of expressions in predetermined units such as words (hereinafter referred to as predetermined unit expressions).

[0006] The present disclosure has been made in consideration of the above points, and aims to make it possible to take into account the difficulty level of a predetermined unit expression when generating a sentence using machine translation or the like. [Means for solving the problem]

[0007] In order to solve the above problem, the present disclosure provides a sentence generation device having a sentence generation unit that takes as input a first text, a second text related to the content of the first text, and a first word contained in the second text, and generates a third text by converting the first word in the second text so that it is expressed with a lower level of difficulty than the first word, based on a trained model. [Effects of the Invention]

[0008] As described above, the present disclosure provides the advantage that, when generating sentences by machine translation or the like, it is possible to take into account the level of difficulty of predetermined unit expressions. [Brief explanation of the drawings]

[0009] [Figure 1] FIG. 2 is a functional configuration diagram of the sentence generation device in the learning phase. [Figure 2] FIG. 2 is a functional configuration diagram of a learning data creation unit. [Figure 3] FIG. 2 is a functional configuration diagram of the sentence generation device in the inference phase. [Figure 4] 10 is a diagram showing a specific example in which a sentence generation unit repeatedly generates sentences under the control of a control unit. FIG. [Figure 5] FIG. 2 is a diagram illustrating the electrical hardware configuration of the sentence generation device. [Figure 6] 10 is a flowchart showing a process in a learning phase. [Figure 7] 10 is a flowchart showing the processing of a learning data creation unit. [Figure 8] 10 is a flowchart showing the processing of the inference phase. [Figure 9] FIG. 10 is a diagram showing experimental results. DETAILED DESCRIPTION OF THE INVENTION

[0010] Hereinafter, embodiments of the present invention will be described with reference to the drawings. The present invention is not limited to the embodiments shown below, and various modifications are possible within the scope of the technical concept of the present invention. Since the drawings are intended to conceptually explain the present invention, dimensions, ratios, or numbers may be exaggerated or simplified as necessary to facilitate understanding.

[0011] In Figures 1 to 4, 7, and 8, the "Japanese sentences," "English sentences," "English words," "acquisition age," and "target age" shown in [ ] are merely examples and are not limited to these.

[0012] Although the description will be given taking the case where a word is used as the predetermined unit expression, the present invention is not limited to a word, and the predetermined unit expression may be an expression or phrase consisting of multiple words.

[0013] [Configuration of the sentence generation device in the learning phase] First, the functional configuration of a sentence generation device 10 in the learning phase according to this embodiment will be described with reference to Fig. 1. Fig. 1 is a diagram showing the functional configuration of the sentence generation device in the learning phase.

[0014] As shown in Fig. 1, the sentence generation device 10 of this embodiment has a training data creation unit 11, a sentence generation unit 14, and an update unit 15. Each of these units has a function realized by an instruction from a processor 1004 in Fig. 5 based on a program. Furthermore, a training data storage unit 21 is realized by an auxiliary storage device 1002 or a memory device 1003 in Fig. 5.

[0015] Furthermore, the sentence generation device 10 uses a difficulty level management DB 20 when creating training data. The difficulty level management DB 20 is constructed in a database server external to the sentence generation device 10. Note that the difficulty level management DB 20 may also be constructed in the auxiliary storage device 1002 or the memory device 1003 within the sentence generation device 10. The difficulty level management DB 20 is an example of a difficulty level management unit.

[0016] <Difficulty level management DB> One of the reasons why a target user may not be able to properly understand the meaning of a sentence is the use of difficult words. One indicator of word difficulty is the Age of Acquisition (AoA) (Reference 1). By converting words into simpler ones according to the AoA, it is possible to simplify the sentence by lowering the difficulty level to a level appropriate for the target user's age. (Reference 1) Victor Kuperman, Hans Stadthagen-Gonzalez, and Marc Brysbaert. Age-of-acquisition ratings for 30,000 english words. Behavior Research Methods, Vol. 44, pp. 978.990, 2012. In this embodiment, a difficulty level management DB 20 is used that takes the above circumstances into consideration. The difficulty level management DB 20 manages word difficulty (the difficulty of understanding the vocabulary of a word) for each word. The word difficulty indicates the estimated average period for a subject to understand the vocabulary, and includes, for example, the age at which the subject acquired the vocabulary, which is the subject's actual age, the number of years of employment, etc. Furthermore, the estimated period does not necessarily have to be a year, but may include any period such as a week, a month, or a season. However, the unit for registering the difficulty level is not necessarily limited to a word, and may be any unit that matches a predetermined unit expression. For example, phrases consisting of multiple words, such as idiomatic expressions that convey a specific meaning when used as a set, may also be included. In this case, the difficulty level of each expression unit for phrases consisting of multiple words may also be registered in the difficulty level management DB 20.

[0017] Here are some examples of words and word difficulty: · Word: this ; Word difficulty: 4.93 [age of acquisition] · Word: vote ; Word difficulty: 7.79 [age of acquisition] Word: erratum; Word difficulty: 15.77 [age of acquisition] The word difficulty can be determined by defining the frequency of occurrence of a word in a monolingual corpus, since words that occur infrequently are considered to have a high word difficulty.

[0018] Furthermore, instead of using the difficulty level management DB 20, the word ease for each word may be registered in an ease management DB, and when the word ease is equal to or less than a threshold value, the reciprocal of the word ease may be used as the difficulty level.

[0019] <Learning Data Creation Department> Since a data set suitable for this type of machine learning does not exist or is difficult to obtain, the learning data creation unit 11 creates learning data by using the difficulty management DB 20 and translating and reverse-translating original text in a single language such as English.

[0020] The learning data creation unit 11 will now be described with reference to Fig. 2. Fig. 2 is a functional configuration diagram of the learning data creation unit.

[0021] As shown in FIG. 2, the training data creation unit 11 includes a translation unit 111, a reverse translation unit 112, a difficulty level extraction unit 113, and a training data identification unit 114.

[0022] (Translation Department) The translation unit 111 creates a translated text (Japanese text) by translating an original text (English text) using a translation application or an LLM (Large Language Model), etc. When the translation unit 111 uses an LLM, it may share the LLM with the sentence generation unit 14 described below.

[0023] The translation application can be any general translation device. When using LLM, it is sufficient to have the same functionality as a translation application by specifying the translation language using a prompt, for example.

[0024] (Reverse Translation Department) The reverse translation unit 112 uses a translation application or LLM to reverse-translate the translation text [Japanese text] created by the translation unit 111 into the language before translation, thereby creating a reverse-translated text [English text]. When the reverse translation unit 112 uses an LLM, it may share it with the sentence generation unit 14, which will be described later. Furthermore, the translation unit 111 and the reverse translation unit 112 may be a single translation unit that translates into a target language in response to changes in the settings of the input language and the output language.

[0025] (Difficulty level extraction part) The difficulty level extraction unit 113 searches the difficulty level management DB 20 using each word contained in the original text and the reverse-translated text as a search key to extract the word difficulty level [age of acquisition] of each corresponding word. As mentioned above, the processing unit used as a search key is not limited to a single word. For example, not only a single word such as "America" ​​but also a combination of multiple words such as "United States of America" ​​can be processed as a predetermined unit expression in the same way as a word.

[0026] Furthermore, the difficulty extraction unit 113 identifies the most difficult word of each of the words contained in the original text and the backward-translated text, respectively. Then, the difficulty extraction unit 113 compares the word difficulties corresponding to each of the most difficult words in the original text and the backward-translated text, and if there is a difference in the difficulty levels, identifies the most difficult word with the higher difficulty as the input data (word to be corrected).

[0027] As an example of a case where there is a difference in difficulty level, if the difference in maximum difficulty level between the original text and the backward-translated text is equal to or greater than a predetermined value, this pair may be selected as a candidate for training data.

[0028] Furthermore, to realize machine translation for a target age of n, the learning data creation unit 11 may select a dataset by specifying the target age n. In this case, for example, the word difficulty of the words to be corrected is set to n or higher.

[0029] (Learning data identification part) The training data identification unit 114 identifies the translated text as input data (first text). The training data identification unit 114 also compares the maximum difficulty levels of the original text and the backward-translated text, and identifies the one containing words with a higher maximum difficulty level as input data (second text), and the one containing words with a lower maximum difficulty level as correct answer data. That is, as a method that provides specific improvements over conventional methods, the original text and the backward-translated text have different roles depending on the maximum difficulty levels contained in each. Note that, as described above, if the difficulty level extraction unit 113 compares the maximum difficulty levels of the original text and the backward-translated text, the training data identification unit 114 may identify the one containing the correction target words extracted by the difficulty level extraction unit 113 as input data (second text), and the one containing words with a lower maximum difficulty level as correct answer data, without comparing the maximum difficulties again. Furthermore, to achieve machine translation for a target age of n, the training data creation unit 11 may select a data set by specifying the target age n. In this case, for example, the training data is set to be data containing correct answer data with a word difficulty level of less than n.

[0030] Here, an example of the original text and training data is shown. Original text: This term is often used to denote certain songs on the album by number. · Primary Text: This term is often used to refer to a specific song on an album by number. Text 2: This term is often used to denote certain songs on the album by number. · Modified word: denote (age of acquisition: 11.24) Correct answer: The term is often used to mean a specific song on the album by number. The training data creation unit 11 may create training data by translating a pre-existing pair of an original text and a first text to generate a second text. For example, in the case of creating training data for a summarization task rather than a translation task, a pair of a source text and a summary text is prepared, and the training data creation unit 11 generates another summary text as the second text, using the summary text as the source text and the source text as the first text. Specifically, if the source text that is the summary text is "A referendum was held," and the first text that is the source text is "A referendum on ... was held on ... month ... day," the other summary text that is the second text is "A referendum was held."

[0031] As a result of the above, the training data creation unit 11 stores the data set of input data (first text, second text, and correction target word) and correct answer data in the training data storage unit 21. By repeating the above process, a large amount of training data can be prepared.

[0032] <Sentence generation section 14> The sentence generation unit 14 has a machine learning model Ma to be learned, which is read from the auxiliary storage device 1002 or the memory device 1003. In this embodiment, an LLM is used as the machine learning model Ma.

[0033] Here are some specific examples of prompts to use when giving instructions to an LLM: Instructions: Translate the following Japanese based on the hypothetical sentence without using the words enclosed in <edit> ### Japanese: Japanese sentences ### Hypothesis sentence: English sentence ### Translation: Plain text Note that LLM refers to the general name for a general-purpose language model and does not imply any limitation on parameter size, etc., being "large." In other words, the model must be able to generate a third text based on the machine learning model Ma using the first text, the second text, and the word to be corrected as input. The above description of the machine learning model Ma also applies to the trained machine learning model Mb described below.

[0034] The sentence generation unit 14 receives, as input data, the first text [Japanese sentence], the second text [original sentence or reverse-translated sentence], and the word to be corrected [English word] from the learning data storage unit 21.

[0035] The sentence generation unit 14 uses the machine learning model Ma (a model whose parameters are being trained) to convert (correct) at least the words to be corrected that are included in the second text based on the input data, thereby generating a third text by converting the second text, and outputs the third text to the update unit 15. Note that the sentence generation unit 14 may also convert words other than the words to be corrected.

[0036] The update unit 15 obtains an error E between the output data (third text) obtained from the machine learning model Ma and the correct answer data obtained from the training data storage unit 21, and calculates a loss L representing the error E using a loss function. Furthermore, the update unit 15 learns the machine learning model Ma by updating the connection weighting coefficients between the nodes of the neural network based on the loss L so that the loss L becomes smaller (so that the loss L approaches 0).

[0037] The update unit 15 updates the connection weighting coefficients between nodes of the neural network, for example, using the backpropagation algorithm. Regarding machine learning, any common learning method for machine learning may be used, as long as input data is converted based on the machine learning model Ma and learning is performed so that the output data approaches the correct output for a predetermined purpose. When an LLM is used as the machine learning model Ma, fine tuning may be performed using the learning data.

[0038] By repeating the above learning, the sentence generation device 10 can generate a trained machine learning model (trained model) Mb. Note that the trained machine learning model Mb is a neural network model that has been trained in advance.

[0039] [Configuration of sentence generation device in the inference phase] First, the functional configuration of the sentence generation device 30 in the inference phase according to this embodiment will be described with reference to Fig. 3. Fig. 3 is a diagram showing the functional configuration of the sentence generation device in the inference phase.

[0040] The sentence generation device 30, for example, takes as input the target age of the first text and the third text, and generates the third text (final text) based on a difficulty management DB 20 in which the acquisition age for each specified unit expression in the domain of the third text is set, and a trained machine learning model Mb, so that the word difficulty of each word contained in the first text satisfies specified conditions for difficulty related to the target age.

[0041] Here, in this embodiment, "domain" refers to the domain to which the text belongs. For example, in a translation task, it refers to the language (Japanese, English, Chinese, etc.). In the case of Japanese-to-English translation, the domain of the third text is English. Also, in conversion processing within the same language, it may be considered that the domain is different before and after conversion. For example, in a conversion task from spoken language to written language, the domain of the third text is the difficulty of the written language. In the case of a summarization task, the domain generally does not change before and after summarization, so the domain of the first text becomes the domain of the third text.

[0042] 3, the sentence generation device 30 of this embodiment has a sentence generation unit 31, a control unit 32, and a sentence generation unit 34. These units are functions realized by instructions from the processor 1004 in FIG. 5 based on a program. The control unit 32 also has a correction target word identification unit 33.

[0043] Unlike the sentence generation unit 14, the sentence generation unit 34 has a trained machine learning model Mb, but other than that, it has the same functions as the sentence generation unit 14, so a description thereof will be omitted.

[0044] Furthermore, the sentence generation device 30 uses the above-mentioned difficulty level management DB 20 when inputting the first text and outputting the final text. The difficulty level management DB 20 is constructed in a database server external to the sentence generation device 30. Note that the difficulty level management DB 20 may also be constructed in the auxiliary storage device 1002 or the memory device 1003 within the sentence generation device 30.

[0045] <Sentence generation section 31> The sentence generation unit 31 uses, for example, a machine translator that translates an input first text [Japanese sentence] and outputs a second text [English sentence]. In this case, the second text is a text related to the content of the first text. In the following description, a translation task using a machine translator will be described as an example. However, the sentence generation unit 31 may handle tasks other than translation, and the second text may be any text related to the content of the first text. For example, if the first text is a "question sentence," the second text may include a "response sentence." Also, if the first text is an "original text," the second text may include a "summary sentence."

[0046] For example, if the task is not translation, the sentence generation unit 31 may use any sentence generator appropriate for the task. For example, if the task is a dialogue, the sentence generation unit 31 of a chatbot or the like generates a second text [response sentence] from the input first text [question sentence]. In this case, the correction target word identification unit 33 identifies words to be corrected in the response sentence.

[0047] Furthermore, when the task is summarization, the sentence generation unit 31 generates a second text [summary sentence] from the input first text [original sentence]. In this case, the correction target word identification unit 33 identifies words to be corrected in the summary sentence.

[0048] <Control unit> The control unit 32 inputs the first text [Japanese sentence] from outside the sentence generation device 30, inputs the second text [English sentence] from the sentence generation unit 31, and inputs, for example, a preset difficulty level [target age] according to the user's settings.

[0049] The correction target word specifying unit 33 of the control unit 32 searches the difficulty management DB 20 using each word included in the second text as a search key, and extracts the word difficulty level [acquisition age] of each corresponding word.

[0050] Furthermore, the correction target word identification unit 33 determines whether or not the extracted words include words with a difficulty level higher than the set difficulty level [target age]. If a word with a difficulty level higher than the set difficulty level [target age] is included, the correction target word identification unit 33 identifies the word as a correction target word. On the other hand, if a word with a difficulty level higher than the set difficulty level [target age] is not included, the correction target word identification unit 33 does not identify a correction target word.

[0051] Furthermore, when the correction target word identification unit 33 identifies a word as a correction target word, the control unit 32 outputs input data (first text [Japanese sentence], second text [English sentence] as the conversion target sentence, and correction target word [English word]) to the sentence generation unit 34.

[0052] On the other hand, the sentence generation unit 34 uses the trained machine learning model Mb to convert (correct) at least the words to be corrected that are included in the second text based on the input data, thereby generating a third text by converting the second text, and outputs the third text to the control unit 32. Note that the sentence generation unit 34 may also convert words other than the words to be corrected.

[0053] The method for determining the correction target word is not limited to the above. For example, the user may be prompted to select one of the words included in the second text, the selected word may be acquired, and the acquired correction target word may be used to perform processing by the sentence generation unit 34.

[0054] Furthermore, the control unit 32 receives the third text [English sentence] converted from the second text from the sentence generation unit 34.

[0055] Here, as described above, the correction target word identification unit 33 of the control unit 32 searches the difficulty management DB 20 using each word included in the third text as a search key to extract the word difficulty [acquisition age] of each corresponding word. Furthermore, the correction target word identification unit 33 determines whether or not the extracted word difficulty levels include any words with a difficulty level higher than the set difficulty level [target age]. If any words with a difficulty level higher than the set difficulty level [target age] are included, the correction target word identification unit 33 identifies the words as new correction target words. On the other hand, if any words with a difficulty level higher than the set difficulty level [target age] are not included, the correction target word identification unit 33 does not identify any correction target words.

[0056] Furthermore, when the correction target word identification unit 33 identifies a word included in the third text as a new correction target word, the control unit 32 identifies the third text as a new second text [English sentence]. Then, the control unit 32 outputs input data (the first text [Japanese sentence], the new second text [English sentence], and the new correction target word [English word]) to the sentence generation unit 34.

[0057] As a method that provides a specific improvement over conventional methods, while the correction target word identification unit 33 is identifying words to be corrected, the control unit 32 controls the control unit 32 to repeat processing by the control unit 32 and the sentence generation unit 34, and when a predetermined condition regarding difficulty is satisfied, that is, when the correction target word identification unit 33 no longer identifies words to be corrected, the control unit 32 outputs the third text (second text after conversion) at that time as the final text [English text]. However, if the third text does not exist, the control unit 32 outputs the second text as the final text [English text].

[0058] Here, a specific example will be described with reference to Fig. 4 when the sentence generation unit 34 repeatedly generates sentences under the control of the control unit 32. Fig. 4 is a diagram showing a specific example when the sentence generation unit repeatedly generates sentences under the control of the control unit.

[0059] 4, when the correction target word identification unit 33 determines that the difficulty level [16.88] of a word (an example of a first word) included in the second text [English sentence] is higher than the set difficulty level

[10] , it identifies this word as a correction target word [English word]. Then, the control unit 32 inputs the first text [Japanese sentence], the second text [English sentence], and the correction target word [English word] to the sentence generation unit 34 for the first time, thereby acquiring a third text (second text after conversion) [English sentence] from the sentence generation unit 34.

[0060] Next, if the correction target word identification unit 33 determines that the difficulty level [11.58] of a word (an example of a second word) included in the third text is still higher than the set difficulty level

[10] , it identifies this word as a correction target word [English word]. Then, the control unit 32 inputs the first text [Japanese sentence], the third text (second text after conversion) [English sentence], and the correction target word [English word] to the sentence generation unit 34 for the second time, thereby obtaining a further third text (second text after reconversion) [English sentence] from the sentence generation unit 34.

[0061] Next, when the correction target word identification unit 33 determines that the difficulty level [7.79] of a word (an example of a third word) contained in the further third text is equal to or lower than the set difficulty level

[10] , the control unit 32 outputs the third text at this point as a translation of the first text [Japanese text] and as a final text that can be understood by the vocabulary of an average 10-year-old.

[0062] The reason why not only the second text and the words to be corrected contained in the second text but also the first text is used is to prevent the third text (the second text after conversion) from being significantly different in meaning from the second text before conversion. When the control unit 32 inputs the words to be corrected to the sentence generation unit 34, the difficulty level [age of acquisition] of the words to be corrected may or may not be input.

[0063] [Electrical hardware configuration diagram of the sentence generation device] Next, the hardware configuration of the sentence generation device 10 will be described with reference to Fig. 5. Fig. 5 is a diagram showing the electrical hardware configuration of the sentence generation device.

[0064] As shown in FIG. 5, the sentence generation device 10 includes a drive device 1000, an auxiliary storage device 1002, a memory device 1003, a processor 1004, an interface device 1005, a display device 1006, an input device 1007, an output device 1008, and the like, all of which are interconnected by a bus 1010.

[0065] A program for realizing processing on the computer is provided by a recording medium 1001 such as a CD-ROM or a memory card. When the recording medium 1001 storing the program is set in the drive device 1000, the program is installed from the recording medium 1001 to the auxiliary storage device 1002 via the drive device 1000. However, the program does not necessarily have to be installed from the recording medium 1001, but may be downloaded from another computer via a communication network such as the Internet. The auxiliary storage device 1002 stores the installed program as well as necessary files, data, etc.

[0066] When an instruction to start a program is received, the memory device 1003 reads the program from the auxiliary storage device 1002 and stores it.

[0067] The processor 1004 is hardware that implements functions related to the sentence generation device 10 in accordance with a program stored in the memory device 1003. The processor 1004 may include not only a CPU (Central Processing Unit) but also a GPU (Graphics Processing Unit).

[0068] The interface device 1005 is used as an interface for connecting to a communication network, etc. The display device 1006 displays a GUI (Graphical User Interface) etc. according to a program. The input device 1007 is composed of a keyboard, mouse, buttons, a touch panel, etc., and is used to input various operation instructions. The output device 1008 outputs the calculation results to the outside.

[0069] The electrical hardware configuration of the sentence generation device 30 is the same as that of the sentence generation device 10, and therefore a description thereof will be omitted.

[0070] [Learning Phase Processing] Next, the process of the learning phase of the embodiment will be described with reference to Figures 6 and 7. Figure 6 is a flowchart showing the process of the learning phase.

[0071] S11: The learning data creation unit 11 creates learning data (input data (first text, second text, word to be corrected) and corrective data) based on the input original text, and stores the learning data in the learning data storage unit 21. Here, process S11 will be described in detail with reference to FIG. 7. FIG. 7 is a flowchart showing the processing of the learning data creation unit.

[0072] S111: The translation unit 111 translates the original text [English text] to create a translated text [Japanese text].

[0073] S112: The reverse translation unit 112 reverse-translates the translation text [Japanese text] created by the translation unit 111 into the language before translation, thereby creating a reverse-translated text [English text].

[0074] S113: The difficulty level extraction unit 113 searches the difficulty level management DB 20 based on each word contained in the original text and the backward-translated text, and extracts the word difficulty level [age of acquisition] of each corresponding word.

[0075] S114: The difficulty extraction unit 113 identifies the most difficult word of each of the words contained in the original text and the backward-translated text, respectively. Then, the difficulty extraction unit 113 compares the word difficulties corresponding to each of the most difficult words in the original text and the backward-translated text, and if there is a difference in the difficulties, identifies the most difficult word with the higher difficulty as the input data (word to be corrected).

[0076] S115: The learning data identification unit 114 identifies the translated text as input data (first text), and compares the maximum difficulty levels of the original text and the reverse-translated text, identifying the one containing the words with the higher maximum difficulty level as input data (second text), and the one containing the words with the lower maximum difficulty level as correct answer data. This completes process S11.

[0077] S12: Next, the sentence generation unit 14 uses the machine learning model Ma to generate a third text (the converted second text) based on the input data acquired from the learning data storage unit 21, and outputs the generated third text to the update unit 15.

[0078] S13: The update unit 15 obtains the error E between the output data (third text) obtained from the machine learning model Ma and the correct answer data obtained from the learning data memory unit 21, and learns the machine learning model Ma by calculating the loss L representing the error E using a loss function.

[0079] [Inference Phase Processing] Next, the processing of the inference phase of the embodiment will be described with reference to Fig. 8. Fig. 8 is a flowchart showing the processing of the inference phase.

[0080] S31: The sentence generation unit 31 generates a second text [English sentence] based on the input first text [Japanese sentence].

[0081] S32: The correction target word identification unit 33 of the control unit 32 searches the difficulty management DB 20 based on each word contained in the second text generated by the sentence generation unit 31, and extracts the word difficulty [age of acquisition] of each corresponding word.

[0082] S33: The correction target word identification unit 33 determines whether or not the extracted words include words with a difficulty level higher than the set difficulty level [target age]. That is, the correction target word identification unit 33 determines whether or not the difficulty level of the words satisfies a predetermined condition regarding the difficulty level of the words. For example, the correction target word identification unit 33 determines whether or not the difficulty level of the words is higher than a preset set difficulty level.

[0083] S34: If a word of a difficulty level higher than the set difficulty level [target age] is included (S33; YES), the correction target word identification unit 33 identifies the word as a correction target word. Then, the control unit 32 outputs input data (the first text [Japanese sentence], the second text [English sentence] as the conversion target sentence, and the correction target word [English word]) to the sentence generation unit 34. As a result, the sentence generation unit 34 converts the second text to generate a third text by converting (correcting) at least the correction target word included in the second text into, for example, simplified words, based on the input data, using the trained machine learning model Mb. Then, the sentence generation unit 34 outputs the third text (the converted second text) to the control unit 32. Next, the control unit 32 acquires the third text. Then, the process returns to step S32. From this point onwards (from the second time onwards), the control unit 32 processes the third text by treating it as the second text.

[0084] S35: If the text does not contain any words with a difficulty level higher than the set difficulty level [target age] (S33; NO), the control unit 32 outputs the second text or the third text at that time as the final text.

[0085] Through the above processing, the sentence generation device 30 uses the trained machine learning model Mb based on the first text (an example of a specified text) and the set difficulty level, which is set to indicate the difficulty level of understanding the vocabulary of words, to generate and output a final text in which the difficulty level of each word contained in the specified text is equal to or lower than the set difficulty level.

[0086] [Experimental results] Next, experimental results according to this embodiment will be described with reference to Fig. 9. Fig. 9 is a diagram showing the experimental results. Here, BLEU (Reference 2), COMET (Reference 3), SARI (Reference 4), FKGL (Reference 5), and MUSS (Reference 6) shown in Fig. 9 are described in the following documents. (Reference 2) Kishore Papineni, SalimRoukos, ToddWard, andWei-Jing Zhu. Bleu: a method for automatic evaluation of machine translation. In Proceedings of the 40th Annual Meeting of the Association for Computational Linguistics(ACL), pp. 311.318, 2002. (Reference 3) Ricardo Rei, Craig Stewart, Ana C Farinha, and Alon Lavie. COMET: A neural framework for MT evaluation. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 2685.2702, 2020. (Reference 4) Wei Xu, Courtney Napoles, Ellie Pavlick, Quanze Chen, and Chris Callison-Burch. Optimizing statistical machine translation for text simplification. Transactions of the Association for Computational Linguistics (TACL), Vol. 4, pp. 401.415, 2016. (Reference 5) Peter Kincaid, Robert P. Fishburne, Richard L. Rogers, and Brad S. Chissom. Derivation of new readability formulas (automated readability index, fog count and flesch reading ease formula) for navy enlisted personnel. 1975. (Reference 6) Louis Martin, Angela Fan, Eric de la Clergerie, Antoine Bordes, and Beno. Sagot. MUSS: Multilingual unsupervised sentence simplification by mining paraphrases. In Proceedings of the Thirteenth Language Resources and Evaluation Conference (LREC), pp. 1651.1664, 2022. We created the training data by applying machine translation from English to Japanese and back to English. The dataset we created consisted of 169,672 sentences, and experiments were conducted by dividing the training data, validation data, and test data into 8:1:1 ratios. When fine-tuning the machine learning model Ma, we used all of the training data so that a single model could be used to simplify the sentences for all ages. Furthermore, in this experiment, the target age (set difficulty) at the time of generation was set to "10 years old." In other words, the test data was generated using only data in which the maximum AoA (maximum difficulty) contained in the back-translated sentences was 10 or greater and the maximum AoA contained in the source text was less than 10. A total of 6,194 sentences were generated.

[0087] In FIG. 9, the results of MUSS (English sentences output by machine translation are simplified using a different sentence simplification model) are shown, and "proposed method" indicates the results of the method of this embodiment.

[0088] In Figure 9, we can see that the evaluation of machine translation by BLEU and COMET is significantly lower when the sentence is simplified using MUSS than when it is reverse-translated. This shows that simply simplifying the translated sentence can impair the meaning of the translated sentence.

[0089] Next, it can be seen that the performance of the first generation of the proposed method is improved in all evaluation indices compared to the reverse-translated sentence. This demonstrates that the proposed method can simplify the translated sentence while preserving its meaning, confirming the effectiveness of the proposed method.

[0090] Next, we compare the results when the proposed method is applied iteratively to sentences whose AoA exceeds the reference value. Comparisons within the proposed method show that all indicators are maintained or improved to the same extent by iteratively applying the method. This suggests that iterative simplification makes it possible to simplify a translated sentence by using words with a lower AoA while maintaining the meaning of the translated sentence.

[0091] [Major Effects of the Embodiments] As described above, according to this embodiment, the sentence generation unit 34 receives as input a first text, a second text related to the content of the first text, and a first predetermined unit expression included in the second text, and generates a third text by converting the first predetermined unit expression in the second text to an expression of lower difficulty than the first predetermined unit expression based on a trained model. This provides the advantage that the difficulty of the predetermined unit expression can be taken into consideration when generating sentences using machine translation or the like. For example, by generating sentences using predetermined unit expressions (e.g., words, phrases) appropriate for the age or experience of the subject, the subject can understand the generated sentences with vocabulary appropriate for their age or experience.

[0092] 〔supplement〕 The sentence generation devices 10 and 30 can be realized by a computer and a program, and this program can be recorded on a (non-transitory) recording medium or provided via a communication network.

[0093] The sentence generation devices 10 and 30 also include a sentence generation system in which multiple servers connected via a network execute the processing of each unit included in the sentence generation devices 10 and 30. The sentence generation devices 10 and 30 may be devices housed in the same housing.

[0094] [Additional notes] The above-described embodiment can also be expressed as the following content.

[0095] [Additional Note 1] A sentence generation device having a hardware processor, The processor: A sentence generation device that takes as input a first text, a second text related to the content of the first text, and a first word contained in the second text, and performs a sentence generation process to generate a third text by converting the first word in the second text so that it is expressed with a lower level of difficulty than the first word, based on a trained model.

[0096] [Additional note 2] A sentence generation device according to claim 1, The processor: a sentence generation device that, if any word included in the third text does not satisfy a predetermined condition regarding difficulty, sets the word as the first word, inputs the third text as the second text to the sentence generation unit, and executes a control process to repeat the processing of the sentence generation unit until a newly generated third text satisfies the predetermined condition.

[0097] [Additional note 3] The processor: 3. The sentence generation device according to claim 1, wherein a target age is acquired, and a word included in the second text that does not satisfy a predetermined condition of difficulty related to the target age is set as the first word.

[0098] [Additional Note 4] A sentence generation device having a processor as hardware, which generates, from a first text, a third text related to the content of the first text, The processor: A sentence generation device that receives as input the target ages of the first text and the third text, and generates the third text based on a difficulty management unit in which the acquisition age for each word in the domain of the third text is set, and a trained model, so that the difficulty level of each word included in the first text satisfies a predetermined difficulty condition related to the target age.

[0099] [Additional Note 5] A learning method executed by a sentence generation device having a hardware processor, The processor: A learning method that performs an update process to train a machine learning model by using a first text, a second text related to the content of the first text, and a first word contained in the second text as input data, and a third text consisting of words that are less difficult than the first word as correct answer data, and updating model parameters based on the third text as output data and the correct answer data.

[0100] [Additional Note 6] A learning method executed by a sentence generation device having a hardware processor, The processor: A learning method comprising: generating, based on an original text, a first text related to the content of the original text; and a second text related to the content of the first text; comparing the maximum difficulty of each word included in the original text with the maximum difficulty of each word included in the second text; and specifying, as the input data, one of the original text and the second text that includes the word with the highest level of difficulty, and specifying, as the correct answer data, one of the original text and the second text that does not include the word with the highest level of difficulty.

[0101] [Additional Note 7] A sentence generation method executed by a sentence generation device having a hardware processor, The processor: A sentence generation method that takes as input a first text, a second text related to the content of the first text, and a first word contained in the second text, and generates a third text by converting the first word in the second text so that it is expressed with a lower level of difficulty than the first word, based on a trained model.

[0102] [Additional Note 8] A non-transitory recording medium on which a program for causing a computer to function as the device described in any one of appendixes 1 to 4 is recorded. [Explanation of symbols]

[0103] 10 Sentence generator 30 Sentence generator 11 Training data creation section 14 Sentence generator 15 Update section 20 Difficulty Management DB (Example of Difficulty Management Section) 21 Learning data storage unit 30 Sentence generator 31 sentence generation unit (an example of a second sentence generation unit) 32 Control Unit 33 Correction target word identification section 34 sentence generation unit (an example of a first sentence generation unit) 111 Translation Department 112 Reverse Translation Department 113 Difficulty level extraction part 114 Learning Data Identification Unit Ma Machine Learning Model Mb Trained machine learning model (trained model)

Claims

1. A sentence generation device having a sentence generation unit that receives as input a first text, a second text related to the content of the first text, and a first word contained in the second text, and generates a third text by converting the first word in the second text so that it is expressed with a lower level of difficulty than the first word, based on a trained model.

2. 2. A sentence generation device according to claim 1, a control unit that, if any word contained in the third text does not satisfy a predetermined condition regarding difficulty, sets the word as the first word, inputs the third text as the second text to the sentence generation unit, and repeats the processing of the sentence generation unit until the newly generated third text satisfies the predetermined condition.

3. The sentence generation device according to claim 1 , wherein a target age is acquired, and words included in the second text that do not satisfy a predetermined condition of difficulty related to the target age are set as the first words.

4. A sentence generation device that generates, from a first text, a third text related to the content of the first text, A sentence generation device that receives as input the target ages of the first text and the third text, and generates the third text based on a difficulty management unit in which the acquisition age for each word in the domain of the third text is set, and a trained model, so that the difficulty level of each word included in the first text satisfies predetermined difficulty conditions related to the target ages.

5. A learning method including an update process for learning a machine learning model by using a first text, a second text related to the content of the first text, and a first word contained in the second text as input data, and a third text consisting of words of lower difficulty than the first word as correct answer data, and updating model parameters based on the third text as output data and the correct answer data.

6. a learning data creation process for generating, based on an original text, a first text related to the content of the original text and a second text related to the content of the first text, and comparing a maximum difficulty level of each word included in the original text with a maximum difficulty level of each word included in the second text, and specifying, as input data, one of the original text and the second text that includes the word with the highest level of difficulty, and specifying, as correct data, one of the original text and the second text that does not include the word with the highest level of difficulty.

7. A sentence generation method that takes as input a first text, a second text related to the content of the first text, and a first word contained in the second text, and generates a third text by converting the first word in the second text so that it is expressed with a lower level of difficulty than the first word, based on a trained model.

8. A program for causing a computer to function as the device according to claim 1, 2 or 4.

Images