Text processing method and text processing device

Abstract

To provide a text processing method that can easily modify text to suit a purpose.SOLUTION: A text processing method comprises the steps of: causing a computer to specify a translation into a first language to input a source text in a first language into a language model (404); causing the computer to acquire an output text of the language model for the source text (406); and causing the computer, following the acquisition step (406), to selectively perform, on the basis of an evaluation of the output text, a first step (422) of using the source text and second steps (416, 418, 420) of using a paraphrase of the source text in place of the source text.SELECTED DRAWING: Figure 8

Description

[Technical Field]

[0001] The present invention relates to natural language processing technology, and more particularly to a sentence refinement technology using a language model. [Background technology]

[0002] Machine translation is a typical example of natural language technology. A problem with machine translation is the reliability of the translation results. This is especially important when the user of the machine translation does not understand the target language, as the user cannot check the translation results.

[0003] To solve these problems, so-called reverse translation, in which the translation results of an automatic translation are translated back into the original language, is used. Patent Document 1, listed below, discloses an automatic translation device that uses reverse translation. This device has, for example, a translation engine that translates Japanese into English and a translation engine that translates English back into Japanese. When translating an original Japanese text into English, this device uses a Japanese-to-English translation engine and displays the translated English sentence. This device also reverse-translates the translated English sentence into Japanese using an English-to-Japanese translation engine and displays the Japanese obtained by the reverse translation. A user compares the original Japanese text with the reverse-translated Japanese sentence, and if they express the same meaning (match), the user determines that the English translation is also correct. If they express different meanings, the user determines that the translation into English was incorrect, corrects the Japanese to a different expression, and attempts retranslation. [Prior art documents] [Patent documents]

[0004] [Patent Document 1] Patent Publication No. 2021-117557 Summary of the Invention [Problem to be solved by the invention]

[0005] The technology disclosed in Patent Document 1 has the advantage that even a user who is not familiar with the target language can evaluate the translation result as long as he or she understands the source language.

[0006] However, when using the technology described in Patent Document 1, a bidirectional translation engine is required for each pair of languages. In other words, two translation engines are required for each pair of languages. Furthermore, to enable a translation device to support multiple languages, the number of translation engines required is equal to twice the number of language combinations. For example, to translate between five languages, 5C2 x 2 = 20 translation engines are required. As the number of languages ​​increases, the number of translation engines required and the cost of the translation device also increase exponentially. As a result, it is difficult to provide such an automatic translation device to the consumer market.

[0007] Furthermore, for example, when translating an original text into another language, it is unclear to the average user whether a desirable translation result will be obtained. Therefore, it is difficult to obtain a reliable translation result, and there is a risk of repeating unnecessary work. This problem can also arise in tasks that do not necessarily involve translation, such as proofreading. In proofreading, the issue is whether the original text is correctly interpreted. Therefore, if the original text is correctly interpreted, there is no need to make any corrections, and if there is a possibility of misinterpretation, corrections must be made to minimize such possibility. In essence, there is a need to process text in a way that achieves the purpose while minimizing the burden on the user. The technology disclosed in Patent Document 1 cannot meet this demand.

[0008] SUMMARY OF THE INVENTION Therefore, an object of the present invention is to provide a text processing method, a text processing device, and a computer program that can easily modify text to suit a purpose. [Means for solving the problem]

[0009] A text processing method according to a first aspect of the present invention includes the steps of: inputting an original sentence in a first language into a language model while specifying a translation into a first language; obtaining an output sentence of the language model for the original sentence; and, following the obtaining step, selectively executing a first step of using the original sentence and a second step of using a paraphrase of the original sentence in place of the original sentence based on an evaluation of the output sentence.

[0010] Preferably, the executing step includes a step in which the computer presents the output sentence to a user so that the user can evaluate the output sentence; a step in which the computer receives instructions from the user regarding whether to execute the first step or the second step after the presenting step; and a step in which the computer executes the first step or the second step in accordance with the instructions.

[0011] More preferably, the executing step includes a step in which the computer calculates the degree of similarity between the original sentence and the output sentence to evaluate the output sentence, and a step in which the computer executes the first step or the second step depending on whether the degree of similarity exceeds a threshold value.

[0012] More preferably, the first step includes a step in which the computer provides the original text to the language model specifying a translation into a second language different from the first language, thereby receiving a translation of the original text in the second language at the output of the language model, and the second step includes a step in which the computer generates a paraphrase of the original text, and a step in which the computer replaces the original text with the paraphrase, and performs the steps of obtaining and selectively executing.

[0013] Preferably, the step of generating a paraphrase sentence includes a step in which the computer inputs an input sentence consisting of the output sentence or the original sentence into a paraphrase model of a first language that is different from the language model, and a step in which the computer obtains the output sentence of the paraphrase model for the input sentence as the paraphrase sentence.

[0014] A text processing device according to a second aspect of the present invention includes an output acquisition device that inputs an original text in a first language into a language model by specifying a translation into a first language and acquires an output sentence of the language model for the original text, and a processing device that selectively executes a first process using the original text and a second process using a paraphrase sentence in place of the original text based on an evaluation of the output sentence acquired by the output acquisition device.

[0015] A computer program according to a third aspect of the present invention causes a computer to perform the steps of inputting an original sentence in a first language into a language model while specifying a translation into the first language, obtaining an output sentence of the language model for the original sentence, and, following the obtaining step, selectively performing a first step of using the original sentence and a second step of using a paraphrase of the original sentence in place of the original sentence based on an evaluation of the output sentence.

[0016] The above and other objects, features, aspects and advantages of the present invention will become apparent from the following detailed description of the invention taken in conjunction with the accompanying drawings. [Brief explanation of the drawings]

[0017] [Figure 1] FIG. 1 is a schematic diagram showing input and output to a large-scale language model during training and estimation processes of the large-scale language model. [Figure 2] FIG. 2 is a schematic diagram showing the configuration of training data when a large-scale language model is used for translation. [Figure 3] FIG. 3 is a schematic diagram showing an example of input and output to a large-scale language model when translation is performed using the large-scale language model. [Figure 4] FIG. 4 is a schematic diagram showing an example of input and output to a large-scale language model when paraphrasing using the large-scale language model. [Figure 5] FIG. 5 is a schematic diagram showing an example of input and output of a large-scale language model when a paraphrase of an input sentence is generated using the large-scale language model in the first embodiment of the present invention. [Figure 6]FIG. 6 is a schematic diagram showing an example of input and output of a large-scale language model when translation is performed using the large-scale language model in the first embodiment of the present invention. [Figure 7] FIG. 7 is a block diagram showing the functional configuration of the translation system according to the first embodiment of the present invention. [Figure 8] FIG. 8 is a flowchart showing a control structure of a program for realizing the translation system in cooperation with computer hardware in accordance with the first embodiment of the present invention. [Figure 9] FIG. 9 is a block diagram showing the functional configuration of a translation system according to the second embodiment of the present invention. [Figure 10] FIG. 10 is a flowchart showing a control structure of a program for realizing the translation system in cooperation with computer hardware in accordance with the second embodiment of the present invention. [Figure 11] FIG. 11 is an external view of a computer system for realizing a system according to each embodiment of the present invention. [Figure 12] FIG. 12 is a block diagram showing the configuration of the computer system shown in FIG. DETAILED DESCRIPTION OF THE INVENTION

[0018] In the following description and drawings, the same parts are designated by the same reference numerals, and therefore detailed description thereof will not be repeated.

[0019] 1. First embodiment 1.1 Configuration Large-scale language models have been a hot topic recently. Figure 1 shows a schematic diagram of the input and output of data to a large-scale language model during training and inference. Large-scale language models require an extremely long time and a great deal of computational resources for training. Therefore, when building a system using a large-scale language model, the large-scale language model is trained in advance using general data, and then subjected to additional training, known as fine-tuning, to suit the system's purpose.

[0020] 1, in a natural language processing system 50 using a large-scale language model 60, a pre-trained large-scale language model 60 is fine-tuned by a training program 64 using training data 62 for fine-tuning. Once fine-tuning is complete, an input token sequence 66 is input to the large-scale language model 60, and an output token sequence 68 generated in accordance with the intended purpose is obtained at the output of the large-scale language model 60.

[0021] Generally, decoder-type generative AI (Artificial Intelligence) predicts and outputs the string that follows the input string (token string). Therefore, fine-tuning is performed as follows.

[0022] An example of the format of training data 100 for fine-tuning a large-scale language model 60 for automatic translation is shown in Figure 2. Referring to Figure 2, the example shows fine-tuning data 110, 112, 114, 116, 118, and 120 for Japanese to English translation, English to Japanese translation, Japanese to Chinese translation, Chinese to Japanese translation, Japanese to Korean translation, and Korean to Japanese translation, respectively.

[0023] In the following, an example will be described in which the source language (first language) is Japanese and the target language (second language) is English. Fine-tuning data 110 includes a token 130 indicating that it is a translation into English, a Japanese string, and an English string corresponding to the Japanese string. A token " " is inserted at the end of the Japanese string and the English string to indicate the end of each string. Fine-tuning data 112, 114, 116, 118, and 120 have a similar structure, and are each preceded by a token 132, 134, 136, 138, or 140 indicating the language into which the string should be translated.

[0024] When a large-scale language model 60 fine-tuned using such data is used for translation, the following process is performed. For example, referring to Fig. 3, when the large-scale language model 60 is made to translate from Japanese to English, an input token sequence 150 is prepared as follows. That is, the input token sequence 150 is a concatenation of a token 160 indicating a translation into English, a Japanese character string 162 to be translated, and a token 164 indicating the end of the input character string.

[0025] When an input token sequence 150 is input to the large-scale language model 60, an output token sequence 152 is obtained at the output of the large-scale language model 60 in response to the input. The output token sequence 152 includes an English string 170 corresponding to the Japanese string 162 and a token 172 at the end of the English string 170.

[0026] On the other hand, it is known that a large-scale language model 60 fine-tuned using data such as that shown in Figure 2 can translate between languages ​​that are not included in the training data without any additional training. For example, translation between Chinese and English, or Korean and Chinese, which are not included in Figure 2, is also possible.

[0027] For example, consider inputting an input token sequence 200 such as that shown in FIG. 4 into the large-scale language model 60. The input token sequence 200 includes a token 210 indicating that it is to be translated into Japanese, a Japanese character string 212, and a token 214 indicating the end of the input. From the large-scale language model 60, a Japanese character string 202 is obtained for the input token sequence 200. The Japanese character string 202 includes a Japanese character string 220, which is a character string obtained by "translating" the Japanese character string 212 into Japanese, and a token 222 indicating the end of the Japanese character string 202. In short, the Japanese character string 220 can be said to be a paraphrase of the Japanese character string 212. Hereinafter, paraphrasing a sentence within the same language in this way will be referred to as a paraphrase, and a sentence obtained by paraphrasing will be referred to as a paraphrased sentence.

[0028] Now, consider the problems of the prior art. For example, referring to FIG. 6 , assume that an input token sequence 300 is provided to a large-scale language model 60, and an output token sequence 302 is obtained from the large-scale language model 60. The input token sequence 300 includes a token 310 indicating that a translation into English is to be performed, a Japanese character string 312, and a token 314 indicating the end of the input. Therefore, the output token sequence 302 includes an English character string 320 corresponding to the Japanese character string 312, and a token 322 indicating the end. In this case, the English character string 320 is the output sentence of the large-scale language model 60. In this system, it is important to determine whether the English character string 320 in the output token sequence 302 is a correct translation of the Japanese character string 312 in the input token sequence 300.

[0029] If a user of large-scale language model 60 can understand English, it would be easy to determine whether English string 320 is a correct translation of Japanese string 312. However, users of translation systems such as large-scale language model 60 generally do not understand the target language. Therefore, such users may not know whether English string 320 is correct or not, and may end up using an incorrect translation.

[0030] The first embodiment is intended to enable even a user who does not understand the target language to predict whether the translation obtained by the large-scale language model 60 is correct or not.

[0031] For this purpose, in this embodiment, Japanese paraphrasing using a large-scale language model 60 is used.

[0032] 5, consider inputting an input token sequence 250 into a large-scale language model 60. The input token sequence 250 includes a token 260 indicating a translation into Japanese, a Japanese character string 262, and a token 264 indicating an end. In this case, an output token sequence 252 obtained at the output of the large-scale language model 60 includes a Japanese character string 270 and a token 272 indicating an end. The output sentence of the large-scale language model 60 in this case is the Japanese character string 270, which is a paraphrase of the original sentence.

[0033] On the other hand, the only difference between the input token sequence 300 shown in Fig. 6 and the input token sequence 250 shown in Fig. 5 is that the first token is either token 310 (Fig. 6) or token 260. Therefore, the internal state of the large-scale language model 60 when the input of the input token sequence 250 shown in Fig. 5 is complete and the internal state of the large-scale language model 60 when the input of the input token sequence 300 shown in Fig. 6 is complete are considered to be almost identical, with only the output language differing.

[0034] That is, if Japanese string 270, which is a paraphrase of input Japanese string 262, is correct as shown in Figure 5, then English string 320 output by large-scale language model 60 when input token string 300, which differs from input token string 250 only in token 310, is considered to be a correct translation of Japanese string 312, as shown in Figure 6. Conversely, if Japanese string 270 is incorrect as a paraphrase of Japanese string 262, then English string 320 obtained for input token string 300 is also considered to be an incorrect translation.

[0035] In accordance with this concept, the first embodiment makes it possible to accurately predict whether a translation obtained by the large-scale language model 60 is correct, even when the translation is into a language that the user does not understand.

[0036] FIG. 7 is a block diagram showing the functional configuration of a translation system 350 according to the first embodiment, which implements a text processing method. Referring to FIG. 7 , the translation system 350 performs translation while communicating with a large-scale language model 60, which can communicate via a network (not shown), and a large-scale language model for paraphrasing 352, which is different from the large-scale language model 60 and can also communicate via a network (not shown). The large-scale language model 60 and the large-scale language model for paraphrasing 352 are preferably separate models. As will be described later, the large-scale language model for paraphrasing 352 is used only for paraphrasing input strings. Therefore, the large-scale language model for paraphrasing 352 may be a dedicated paraphrasing model. In this embodiment, both the large-scale language model 60 and the large-scale language model for paraphrasing 352 exist outside the translation system 350. However, the present invention is not limited to such an embodiment. One or both of the large-scale language model 60 and the large-scale language model for paraphrasing 352 may be located within the translation system 350.

[0037] The translation system 350 is further connected to an external display device 354 and an input device 356. The display device 354 is typically a monitor directly connected to the translation system 350. The input device 356 typically includes a keyboard, mouse, or touch panel directly connected to the translation system 350. The display device 354 and the input device 356 are not limited to those connected to the translation system 350. For example, they may be portable terminal devices (e.g., smartphones), other computers, or the like that are connected to the translation system 350 by communication via a network.

[0038] The translation system 350 includes a control unit 380 connected to a display device 354 and an input device 356 for controlling each unit of the translation system 350, and an original text input unit 374 for receiving an input string to be translated and a target language specification via the display device 354 and the input device 356, assembling an input token sequence for the large-scale language model 60, and inputting the resulting token sequence into the large-scale language model 60. The translation system 350 is capable of both translating the input normally and predicting the accuracy of the translation using the paraphrased sentences described above. In this embodiment, when an original text is input, the original text input unit 374 has the function of specifying both an input token sequence for generating a paraphrase of the original text and an input token sequence for generating a specified translation of the target text, and inputting these tokens into the large-scale language model 60 in order. Therefore, the large-scale language model 60 also outputs two output token sequences: an output token sequence containing a paraphrase of the original text, and an output token sequence containing a translation of the original text.

[0039] Note that this embodiment is premised on the premise that both translation and prediction of translation accuracy are performed simultaneously. Therefore, it is necessary to specify the target language when predicting translation accuracy. However, this invention is not limited to such an embodiment. As will become clear from the following explanation, prediction of translation accuracy can be performed regardless of the target language. Therefore, when only predicting translation accuracy, it may be possible to perform this without specifying the target language.

[0040] The translation system 350 further includes an output acquisition unit 378 for acquiring two output token sequences from the large-scale language model 60 in response to an input token sequence from the original text input unit 374, and a storage device 376 for storing both the input original text received by the original text input unit 374 and the sentences (paraphrased sentences and translated sentences) included in the two output token sequences output by the large-scale language model 60.

[0041] The translation system 350 further includes a paraphrase processor 370 that displays the original text, paraphrased text, and translated text stored in the storage device 376 on the display device 354, receives instructions from the user regarding whether to approve the translation and proceed to processing the next input, and performs the next processing in accordance with the instructions. The original text and paraphrase processor 370 simultaneously present the original text and paraphrase text to the user. The user compares the original text and paraphrase text and determines whether the paraphrase text matches the original text. If the original text and paraphrase text match, the user determines that the translation is correct and instructs the system to process the next sentence. If the original text and paraphrase text do not match, i.e., if the paraphrase text is incorrect, the user determines that the translation is also incorrect and instructs the system to perform a retranslation. If a retranslation is required, the paraphrase processor 370 generates a new paraphrase text from the original text and then processes the new paraphrase text as the original text to perform another translation using the large-scale language model 60. To generate this new paraphrased sentence, the paraphrasing processor 370 uses a large-scale language model for paraphrasing 352, which is separate from the large-scale language model 60. Note that in this embodiment, the large-scale language model for paraphrasing 352 is used only for performing paraphrasing.

[0042] Note that for a paraphrase to match the original text, it does not necessarily mean that the two are exactly the same. It is rather natural that the paraphrase will differ from the original text. Considering the purpose of translation, for a paraphrase to match the original text means that the paraphrase can be understood to mean the same thing as the original text, or to convey substantially the same thing as the original text. Therefore, even if the paraphrase and original texts are significantly different in appearance, there may be cases where they can be determined to match.

[0043] The translation system 350 further includes an original text replacement unit 372 that replaces the original text used in the most recently executed translation when a retranslation is specified by the user with a new paraphrased text obtained by the paraphrasing processing unit 370 from the large-scale language model for paraphrasing 352, provides the new paraphrased text to the original text input unit 374, and executes translation again using the large-scale language model 60 with the new paraphrased text as the original text.

[0044] FIG. 8 shows, in flowchart form, the control structure of a program for implementing the translation system 350 shown in FIG. 7 in cooperation with computer hardware. Referring to FIG. 8, this program includes step 400 of reading from an auxiliary storage device (not shown) a configuration file that records, for example, the user's identification number and password for using the translation service using the large-scale language model 60, the user's primary language, and a list of target languages ​​available for use in the translation service using the large-scale language model 60; and step 402 of displaying, on the display device 354 (see FIG. 7), a pre-prepared source text input form for inputting source text for translation, and receiving the source text and other input. The source text input form displays a text input field for the source text, a user interface (UI) for selecting the target language, such as a drop-down list of languages, an execute button for instructing the execution of translation, and an end button for instructing the end of the program. The processing of step 402 ends when all necessary information has been entered and the execute button is pressed, and control proceeds to the next step.

[0045] This program further includes, following step 402, step 404 in which a token specifying the language used by the user (this token indicates paraphrasing, and therefore will be referred to as a "paraphrase token" hereafter for ease of explanation) is added to the beginning of the original text input in step 402, and a token indicating the end is added to the end, thereby generating an input token string and inputting the input token string into the large-scale language model 60.

[0046] The program further includes a step 406 of obtaining an output token sequence output by the large-scale language model 60 in response to the input token sequence in step 404. Because the input token sequence in step 404 includes paraphrase tokens, the token sequence obtained in step 406 includes a paraphrase of the original sentence.

[0047] This program further includes step 408 of inputting an input token string created by adding a translation token indicating the target language specified in step 402 and an end token indicating the end to the source text input in step 402 into the large-scale language model 60, and step 410 of acquiring an output of the large-scale language model 60 in response to this input. The output token string acquired in step 410 includes a translation of the source text into the target language.

[0048] This program further includes step 412, which displays the original text, the paraphrased text obtained in step 406, and the translation of the original text obtained in step 410 on the screen and receives instructions for the next process from the user; and step 414, which branches the control flow in accordance with the instructions received in step 412. In step 412, the user compares the original text with the paraphrased text and determines whether they match. If the user determines that they match, the user executes step 422, which uses the translation as a first process. That is, the user uses the translation by, for example, showing the translation to a conversation partner or copying it into another program. The user then requests the translation system 350 to display an original text input form to translate the next sentence. Therefore, in this case, control returns to step 402. On the other hand, if the user determines that the original text and the paraphrased text do not match, the user instructs retranslation as a second process. In this case, control proceeds to step 416, which will be described later.

[0049] In addition to presenting the original text, paraphrased text, and translated text, step 412 can also display a text input field for the original text to be translated, a UI for specifying the target language, and a UI such as a button for instructing to translate the next sentence. In this case, the user can enter the next original text in the text input field for the original text, specify the target language, and instruct to translate the next sentence, thereby performing translation processing on the next original text. In this case, if the determination in step 414 is positive, control can be returned to step 404.

[0050] This program further includes step 416, which is a second process, in which, when the determination in step 414 is negative, a paraphrased sentence of the original sentence is generated using the large-scale language model for paraphrasing 352 (see FIG. 7 ) and displayed on display device 354. A UI is displayed on display device 354, prompting the user to select whether or not to retranslate using this paraphrased sentence. When the user selects to retranslate using this paraphrased sentence, control proceeds to step 420, which will be described later. When the user selects not to use this paraphrased sentence, control returns to step 402, in which the user inputs a new original sentence and attempts the next translation.

[0051] This program further includes step 420, in which, when the determination in step 418 is affirmative, i.e., when the user has instructed that translation using paraphrased text be performed, the previously input source text is replaced with the new paraphrased text obtained in step 416, and control is returned to step 404. Thereafter, the above-mentioned processing is repeated using this new paraphrased text as the source text. In other words, this program proceeds with processing by using the new paraphrased text as the source text in the next processing, instead of the original source text.

[0052] 1.2 Operation The translation system 350 according to the first embodiment operates as follows. Referring to FIG. 8, a user launches the program shown in the figure. In step 400, this program reads information necessary for program execution from a configuration file and stores it in memory. In step 402, the source text input unit 374 of the translation system 350 displays a source text input form on the display device 354. When the user inputs the source text to be translated and the target language into this source text input form and instructs translation, in step 404 of FIG. 8, the source text input unit 374 of the translation system 350 creates an input token string by prefixing a paraphrase token and appending an end token to the input source text, and inputs the input token string to the large-scale language model 60. In step 406, the source text input unit 374 stores the source text in the storage device 376.

[0053] In response to this input token string, the large-scale language model 60 generates and outputs a paraphrased sentence of the original sentence. The output acquisition unit 378 of the translation system 350 acquires this paraphrased sentence and stores it in the storage device 376.

[0054] 8, the original text input unit 374 creates an input token sequence by prefixing the original text with a translation token indicating that the original text is to be translated into a language selected by the user and appending an end token to the original text, and inputs this input token sequence to the large-scale language model 60. In response to this input token sequence, the large-scale language model 60 generates and outputs an output token sequence including a translation of the original text into the specified language. In step 410, the output acquisition unit 378 acquires the translation from this output token sequence and stores it in the storage device 376.

[0055] In step 412, the paraphrase processing unit 370 reads the original text, the paraphrase text, and the translation from the storage device 376. The paraphrase processing unit 370 presents these on the display device 354. At this time, the paraphrase processing unit 370 displays a UI for receiving an instruction as to whether to approve the translation and proceed to processing for the next input. The user compares the displayed original text with the paraphrase text. If the comparison shows that the paraphrase text matches the original text, the user considers that the translation of this original text is also correct. As a result, the user selects to translate the next sentence. In response to this selection, the paraphrase processing unit 370 clears the display on the screen. The paraphrase processing unit 370 further instructs the original text input unit 374 to start inputting the next original text. The processing after the next original text has been input is as already described.

[0056] When the user instructs the paraphrase processor 370 to retranslate the original text (when the determination in step 414 is negative), the paraphrase processor 370 generates a new paraphrased text by inputting the original text into the large-scale paraphrasing language model 352. The paraphrase processor 370 displays the paraphrased text on the display device 354 (step 416) and prompts the user to select whether to retranslate using the paraphrased text (step 418). If the user selects retranslation (YES in step 418), the original text is replaced with the paraphrased text generated in step 416 in step 420 of FIG. 8, and control returns to step 404. As a result, the translation process is re-executed using the new paraphrased text as the original text. On the other hand, if the user does not select retranslation (NO in step 418), control returns to step 402. As a result, the user is prompted to select whether to rewrite the original text or to start translating a new sentence.

[0057] 1.3 Variations In this first embodiment, if the user selects not to retranslate using the paraphrased sentence in step 418 of FIG. 8 , control returns to step 402. That is, the user either rewrites the original sentence and performs the translation again, or performs translation of an entirely new original sentence. However, this first embodiment is not limited to such an embodiment. If the determination in step 418 is negative, control may return to step 416. In this case, a new paraphrased sentence is generated by the large-scale language model for paraphrasing 352. For example, when generating a paraphrase from the original sentence, the large-scale language model for paraphrasing 352 may create a list of the top N paraphrase sentence candidates, and each time step 416 is repeated, the candidates on the list may be selected in order. Furthermore, during this repetition, the user may have the option of returning control to step 402 of FIG. 8 to input another original sentence.

[0058] 7 shows only one large-scale language model for paraphrasing 352. However, the first embodiment is not limited to such an embodiment. A paraphrasing service using a large-scale language model different from the large-scale language model for paraphrasing 352 may be used, and a predetermined number of paraphrased sentences may be obtained from each large-scale language model, and these may be used in turn in the processing of step 416.

[0059] Furthermore, the translation system 350 according to the above embodiment uses external services, namely, the large-scale language model 60 and the large-scale language model for paraphrasing 352. However, this first embodiment is not limited to such an embodiment. For example, the large-scale language model 60, the large-scale language model for paraphrasing 352, or both may be provided inside the translation system 350. In this case, the large-scale language model 60, the large-scale language model for paraphrasing 352, the display device 354, and the input device 356 shown in FIG. 7 may communicate with the translation system 350 via the same communication interface. In other words, not only the large-scale language model 60 and the translation system 350, but also the translation system 350 may provide translation services to the outside world as a web service.

[0060] In the above embodiment, the paraphrase sentence and the translation sentence are simultaneously generated and displayed. However, the present invention is not limited to such an embodiment. For example, a process may be performed in which only the paraphrase sentence is generated without translation and only the paraphrase sentence is displayed. In this case, when the user confirms that the original sentence and the paraphrase sentence match, the target language may be specified and the original sentence may be translated. Furthermore, as in the above embodiment, the paraphrase sentence and the translation sentence may be generated simultaneously, and only the paraphrase sentence may be displayed so that the user can compare it with the original sentence. When the user determines that the paraphrase sentence matches the original sentence, the user may perform a predetermined operation to display the translation sentence.

[0061] Furthermore, multiple languages ​​may be specified as the translation destination language. In this case, it is particularly preferable to display only the paraphrase text first, and then display the translation into each language once it has been confirmed that the two match. Furthermore, the translation destination language may be selected before confirming that the paraphrase text and the original text match, or after confirmation has been made, the translation destination language may be selected. This also applies when there is only one translation destination language.

[0062] 1.4 Effects of the First Embodiment As described above, according to the first embodiment, when translating an original text using the large-scale language model 60, the translation system 350 also uses the large-scale language model 60 to generate a paraphrase from the original text in the same language and presents it to the user along with the original text. The token sequence input to the large-scale language model 60 during paraphrase and the token sequence input to the large-scale language model 60 during translation are identical except for the initial paraphrase token and translation token. Therefore, the internal states of the large-scale language model 60 during paraphrase and translation are substantially the same. Therefore, if the paraphrase is a correct paraphrase of the input text, the translation is also considered to be a correct translation of the input text. Even a user who does not understand the translation target language can easily determine whether the translation of the original text is correct by comparing the original text in their native language with the paraphrase.

[0063] If the user believes that the paraphrase is a correct paraphrase of the original text, the user can proceed with the next step (e.g., translating the next sentence) assuming that the proposed translation is also a correct translation of the original text.

[0064] If the user believes that the paraphrase is not a correct paraphrase of the original text, the translation system 350 generates a new paraphrase of the original text, replaces the original text with the new paraphrase, and repeats the process described above for the new original text. By comparing the new original text with the paraphrase generated by the large-scale language model 60, the user can determine whether the paraphrase is a correct paraphrase of the new original text, and therefore whether the translation is a correct translation of the new original text. In some cases, the user can also perform a similar process by modifying the original text themselves.

[0065] Furthermore, in this first embodiment, if the user believes that the paraphrased text obtained from the original text does not match the original text, the original text itself is replaced in step 420 (FIG. 8) with the new paraphrased text obtained in step 416. If the paraphrased text obtained from the new original text matches the new original text, the new original text is considered to be more suitable for translation than the original text. Therefore, the user can proofread the original text to make it more suitable for translation, for example, by replacing the original text with the new original text.

[0066] As a result, the text can be adapted to suit the purpose of translation.

[0067] Second Embodiment 2.1 Configuration In the first embodiment, the user compares the original text with the paraphrased text. This allows the user to predict the accuracy of the translation based on their own judgment. However, the present invention is not limited to such an embodiment. For example, the comparison between the original text and the paraphrased text can also be performed automatically. This second embodiment is such an embodiment.

[0068] To automatically determine whether an original sentence and a paraphrase sentence match, methods such as scoring the degree of match between words used in the original sentence and words used in the paraphrase sentence, and a model for determining whether the two match, etc. In the embodiment described below, a match determination model is used in advance to determine whether two sentences in the same language match semantically.

[0069] Fig. 9 shows a schematic configuration of a translation system 450 according to the second embodiment. Referring to Fig. 9, the translation system 450 is connected to the large-scale paraphrasing language model 352 and the large-scale language model 60, which are similar to those used in the first embodiment, as well as the above-mentioned match determination model 452, and uses the services provided by these models.

[0070] 9, the translation system 450 is for performing translation while communicating with the large-scale language model 60 and the large-scale language model for paraphrasing 352. The translation system 450 is further connected to an external display device 354 and an input device 356.

[0071] The translation system 450 includes a control unit 468 connected to the display device 354 and the input device 356 for controlling each unit of the translation system 450, and an original text input unit 466 for receiving an input string to be translated and a specification of a target language via the display device 354 and the input device 356, assembling an input token sequence for the large-scale language model 60, and inputting the token sequence into the large-scale language model 60. In this embodiment as well, when an original text is input, the original text input unit 466 has the function of specifying both an input token sequence for generating a paraphrase of the original text and an input token sequence for generating a translation of the specified target language, and inputting them into the large-scale language model 60 in that order.

[0072] The translation system 450 further includes an output acquisition unit 378 and a storage device 376 .

[0073] The translation system 450 further includes an evaluation unit 460 that provides the original sentence and the paraphrase sentence stored in the storage device 376 to a match determination model 452 and receives the output of the match determination model 452. The match determination model 452 outputs a score indicating the degree to which the meanings of the two input sentences match. By comparing the output of the match determination model 452 with a predetermined threshold, the evaluation unit 460 can evaluate whether the original sentence and the paraphrase sentence match.

[0074] The match determination model 452 can be trained using learning data consisting of sentence pairs in the same language and labels indicating whether the two sentences that make up the sentence pair match each other.

[0075] The translation system 450 further includes a paraphrase processing unit 462 that has the function of selectively executing, based on the evaluation by the evaluation unit 460, a process of outputting a translation of the original text as is, or a process of automatically starting a re-translation using a new paraphrased sentence for the original text by the large-scale paraphrasing language model 352. When starting a re-translation, the paraphrase processing unit 462 provides the original text to the large-scale paraphrasing language model 352, and uses the new paraphrased sentence of the original text output from the large-scale paraphrasing language model 352 in the re-translation. The paraphrase processing unit 462 executes this process without displaying the original text, the paraphrased sentence, or the new paraphrased sentence.

[0076] The translation system 450 further includes an original text replacement unit 464 that, in response to the paraphrase processing unit 462 selecting to start retranslation, replaces the previous original text with a new paraphrase from the translation system 450 and provides the new paraphrase to the original text input unit 466, thereby restarting translation of the new paraphrase text. The original text input unit 466, the output acquisition unit 378, the storage device 376, the evaluation unit 460, and the original text replacement unit 464 perform the above-described processing on the new paraphrase text. As a result, the above processing is repeated until the score output by the match determination model 452 exceeds the threshold. However, in this embodiment, the control unit 468 counts the number of repetitions and terminates the processing when the maximum number of repetitions is reached. In this case, a correct translation was not obtained from the input original text. Therefore, the control unit 468 preferably has a function to notify the user to rewrite the original text itself or translate a different sentence.

[0077] Fig. 10 is a flowchart showing the control structure of a computer program for causing a computer to function as translation system 450 shown in Fig. 9 in cooperation with computer hardware. Referring to Fig. 10, this program includes steps 400 to 406, which are the same as those shown in Fig. 8.

[0078] The program further includes step 500 of evaluating the paraphrased sentence using match determination model 452 and calculating a score, and step 502 of branching the control flow depending on whether the score is greater than a predetermined threshold. The paraphrased sentence is obtained by applying large-scale language model 60 (see FIG. 9) to the original sentence.

[0079] The program further includes step 418, when the determination in step 502 is positive, of assembling an input token sequence by prefixing a translation token to the source text and appending an end token to the source text, and inputting the resulting input token sequence to the large-scale language model 60, and step 420, receiving an output token sequence output by the large-scale language model 60 in response to the input token sequence. Since the input in this process is a translation token, the output token sequence includes a translation of the source text. In step 420, a translation of the source text is obtained from the output token sequence. The program further includes step 504, which displays an original text input form having the original text, the translation, and a text input field for new original text, and returns control to step 404 in response to receiving an instruction from the user.

[0080] This program further includes step 505, in which, if the determination in step 502 is negative, i.e., if the score output by match determination model 452 for the paraphrased sentence of the original text is equal to or less than a threshold value, assigning 0 to variable N, which represents the number of repetitions; step 506, following step 505, branching the control flow depending on whether variable N is greater than the maximum number of repetitions (maximum number of times); and step 520, in which, if the determination in step 506 is positive, i.e., if the value of variable N is greater than the maximum value, displaying an error message indicating that a high-accuracy translation was not possible, and then returning control to step 402. There is no guarantee that the score output by match determination model 452 will always exceed the threshold value during the repetitions. Therefore, in this embodiment, if the score output by match determination model 452 remains equal to or less than the threshold value and the number of repetitions exceeds the maximum value, an error message is displayed, prompting the user to either rewrite the original text or start translation processing for an entirely different original text.

[0081] The program further includes step 508 of generating a new paraphrased sentence from the original sentence using the large-scale language model for paraphrasing 352 when the determination in step 506 is negative; step 510 of creating an input token sequence by prefixing a paraphrase token to the new paraphrased sentence and appending an end token to the new paraphrased sentence, and inputting the input token sequence to the large-scale language model 60; and step 512 of receiving an output token sequence output by the large-scale language model 60 in response to the input token sequence. The input token sequence has the paraphrase token appended to it. Therefore, the output token sequence output by the large-scale language model 60 includes a paraphrase for the new paraphrased sentence (such a paraphrase obtained from the large-scale language model 60 for the new paraphrased sentence obtained from the large-scale language model for paraphrasing 352 is referred to as a "re-paraphrased sentence" to distinguish it from the new paraphrased sentence obtained from the large-scale language model 352). In step 512, a re-paraphrased sentence is obtained from the output token sequence of the large-scale language model 60.

[0082] This program further includes step 512 of evaluating the re-paraphrased sentence obtained in step 510 using the match determination model 452 shown in Fig. 9. As described above, in step 512, the paraphrased sentence generated in step 506 and the re-paraphrased sentence obtained in step 510 are provided to the match determination model 452, and the match determination model 452 outputs a score indicating the degree of match between the two.

[0083] This program further includes step 514 of evaluating the re-paraphrased sentence obtained in step 512 using the match determination model 452, step 516 of branching the control flow depending on whether the score indicating the evaluation of the re-paraphrased sentence obtained in step 514 is greater than a threshold, and step 518 of, when the determination in step 516 is affirmative, creating an input token string by prefixing a translation token to the paraphrased sentence obtained in step 508 and appending an end token to the end, inputting the input token string to the large-scale language model 60, and then proceeding to step 420. When the determination in step 516 is negative, control proceeds to step 519, where 1 is added to the value of the variable N. After step 519, control returns to step 506, and the processes from step 506 to step 519 are repeated until the determination in step 516 or step 506 becomes affirmative.

[0084] In this embodiment, the paraphrased sentences generated by the large-scale language model for paraphrasing 352 in step 508 are different for each iteration. For this purpose, as described above, the large-scale language model for paraphrasing 352 may create a list of candidate paraphrased sentences and select a paraphrased sentence from the list for each iteration. Alternatively, multiple large-scale language models for paraphrasing 352 may be prepared and a different large-scale language model for paraphrasing 352 may be used to generate a paraphrased sentence for each iteration.

[0085] 2.2 Operation The translation system 450 according to the second embodiment operates as follows. When the program shown in Fig. 10 is started, steps 400, 402, 404, and 406 are executed in this order, as in the first embodiment. Furthermore, in step 500, the paraphrased sentence is evaluated. In this embodiment, the paraphrased sentence and the original sentence are input to a match determination model 452, and a score of the paraphrased sentence is obtained as an output of the match determination model 452. This score is a value indicating the evaluation of the paraphrased sentence.

[0086] Further, following step 500, in step 502, it is determined whether the score indicating the evaluation is greater than a threshold value. If the score is greater than the evaluation, in steps 418 and 420, a translation for the original text is obtained. In the following step 504, an original text input form is displayed for inputting the original text, the translation, and the original text for the next translation. As a result, the user can obtain the translation for the previous original text. Since the paraphrase text matches the original text, the translation is also considered to be a correct translation of the original text.

[0087] On the other hand, if it is determined in step 502 that the score is not greater than the threshold, control proceeds to step 505, where 0 is assigned to variable N. In the following step 506, the value of variable N is compared with the maximum value. At the beginning of the iteration, the determination in step 506 is negative. Therefore, in step 508, a new paraphrased sentence is generated from the source sentence using the large-scale language model for paraphrasing 352. Steps 510, 512, and 514 are executed for this new paraphrased sentence, and it is evaluated whether the resulting re-paraphrased sentence matches the new paraphrased sentence obtained in step 508. That is, the source sentence is replaced with the new paraphrased sentence obtained in step 508, and the same processing as in steps 404 to 502 is executed. If the evaluation is higher than the threshold, in steps 518 and 420, a translation is generated using the large-scale language model 60 with the paraphrased sentence obtained in step 508 as the source sentence. Then, step 504 is executed, and the translation process for the next source sentence is performed.

[0088] On the other hand, if the determination in step 516 is negative, 1 is added to the value of variable N in step 519, and control returns to step 506. Thereafter, the processing from step 506 to step 519 is repeatedly executed until the determination in step 506 or step 516 becomes positive.

[0089] If the determination in step 519 is positive, the processing is as described above. However, if the determination in step 506 is positive, computer system 600 displays a message that a reliable translation could not be made, and then displays a form for inputting the original text (step 402). As a result, the user must either rewrite the original text or translate a different original text.

[0090] 2.3 Effects of the second embodiment As described above, according to the second embodiment, the translation system 450, not the user, determines whether the original text and the paraphrased text match. If the determination result is negative, a new paraphrased text of the original text is generated using the large-scale language model for paraphrasing 352. The original text is replaced with this new paraphrased text, and the same process as that performed on the original text is performed. This process is basically performed until the new paraphrased text matches the new paraphrased text. If the new paraphrased text matches the new paraphrased text, a translation of the new paraphrased text is performed. The translation obtained by this translation, like the new paraphrased text, is considered to be a correct translation that expresses the same meaning as the new paraphrased text. Therefore, the original text is replaced with a new paraphrased text that matches the new paraphrased text, and a translation of the new paraphrased text is obtained. As a result, a correct translation of the original text is obtained without the user having to determine whether the original text and the paraphrased text match. Furthermore, in the second embodiment, the translation is a direct translation of the paraphrased text last generated using the new large-scale language model for paraphrasing 352. Therefore, compared to the original text, the last generated paraphrase is considered to have the same meaning as the original text and to have a higher probability of producing a correct translation. Therefore, by replacing the original text that was initially input as the translation target with the last generated paraphrase, the original text can be proofread.

[0091] Third Variation In the above embodiment, the translation and paraphrase of the original text are presented simultaneously. However, the present invention is not limited to such an embodiment. The paraphrase of the original text may be presented first, and the original text may be translated only after it is determined that the paraphrase matches the original text.

[0092] Furthermore, the present invention can also be used to proofread the original text as described in the second embodiment. Furthermore, the present invention can also be used to determine whether or not the original text needs to be proofread prior to processing using a neural network, as described below.

[0093] For example, in the case of a speech recognition device, a string of characters that should be output as kanji may be output as hiragana, or punctuation marks may be omitted. Normally, proofreading or formatting processes are required to correct such imperfections. If the imperfections are not corrected, subsequent processing (e.g., translation into a foreign language) may not proceed smoothly. If such proofreading processes were to be performed solely by hand, the meaning of the original text would need to be confirmed and then appropriately corrected, which requires manpower and time.

[0094] A similar problem occurs when the output of OCR (Optical Character Recognition) is input to subsequent processing. For example, the string "Yes, and dolphins so beautiful" may be output as a continuous string such as "Yes, and dolphins so beautiful" because the spaces between the characters are ignored. Similarly, the string "Please press the buttons in the following order: up, down, left, right, right, right, down, up, down" may be output as "Please press the buttons in the following order: up, down, down, down, left, right, right, right, down, up, down." because the number of consecutive identical characters is incorrectly recognized. Naturally, if such errors remain in the string, they increase the likelihood of errors in subsequent processing. Therefore, the input must be properly proofread. However, proofreading in such cases can be surprisingly time-consuming.

[0095] A system such as the translation system 350 of the above embodiment, which converts input into an internal format and then converts it into another form based on that internal format, can eliminate or mitigate these problems for the following reasons.

[0096] In the translation system 350, during translation, a translation and a paraphrase of an original text are generated from the same internal state of the translation engine and presented to the user. If the paraphrase matches the meaning of the original text and is appropriately expressed, the original text should be translated correctly without proofreading. In other words, proofreading of an original text that does not require proofreading can be omitted. Of course, it is still necessary to correctly understand the meaning of the original text, but because the paraphrase is presented as a comparison, it is believed that the meaning of the original text can be understood much more easily than when only the original text is used. Furthermore, even if the meaning of the paraphrase differs slightly from that of the original text, it is believed that proofreading the paraphrase is easier than proofreading the original text. Therefore, using a system such as the translation system 350 has the effect of eliminating or reducing problems associated with proofreading output from speech recognition processing, character recognition processing, and the like.

[0097] 4. Computer implementation Fig. 11 is an external view of a computer system 600 that realizes the translation system 350 (Fig. 8) according to the first embodiment or the translation system 450 (Fig. 9) according to the second embodiment. Fig. 12 is a hardware block diagram of the computer system 600. The hardware configuration of the computer system 600 will be described below.

[0098] 11, this computer system 600 includes a computer 650 having a DVD (Digital Versatile Disc) drive 662, and a keyboard 654, a mouse 656, a monitor 652, a microphone 660, and a pair of speakers 658 for interacting with a user, all of which are connected to the computer 650. These are examples of devices for interaction, and any general hardware and software (e.g., a touch panel, voice input, or a general pointing device) that can be used for interacting with a user can also be used.

[0099] 11 and 12 , the computer 650 includes, in addition to a DVD drive 662, a CPU (Central Processing Unit) 710, a GPU (Graphics Processing Unit) 712, and a bus 720 connected to the CPU 710, the GPU 712, and the DVD drive 662. The computer 650 further includes a ROM (Read-Only Memory) 714 connected to the bus 720 and storing a boot-up program and the like of the computer 650, a RAM (Random Access Memory) 716 connected to the bus 720 and storing instructions constituting a program, a system program, working data, and the like, and an SSD (Solid State Drive) 718, which is nonvolatile memory, connected to the bus 720. The SSD 718 is used to store programs executed by the CPU 710 and the GPU 712, as well as data used by the programs executed by the CPU 710 and the GPU 712. The computer 650 further includes a network I / F (Interface) 726 that provides connection to a network enabling communication with other terminals, and a USB port 664 to which a USB (Universal Serial Bus) memory 702 can be attached / detached and that provides communication between the USB memory 702 and each part within the computer 650.

[0100] The computer 650 further includes an audio I / F 722 that is connected to the microphone 660, the speaker 658, and the bus 720, and has the function of reading out audio signals, image signals, and text data generated by the CPU 710 and stored in the RAM 716 or the SSD 718 according to instructions from the CPU 710, converting them to analog, amplifying them, and driving the speaker 658, and digitizing the analog audio signal from the microphone 660 and storing it at any address in the RAM 716 or the SSD 718 specified by the CPU 710.

[0101] In the above embodiments, the programs and the like that realize the functions of the translation system 350 according to the first embodiment and the translation system 450 according to the second embodiment are stored in, for example, the ROM 714, SSD 718, DVD 700, or USB memory 702 shown in Fig. 12, or in a storage medium of an external device (not shown) connected via the network I / F 726 and the network 704. Typically, these data and parameters are written to the SSD 718 from the outside, for example, and loaded into the RAM 716 when the computer 650 is executed.

[0102] Computer programs for operating this computer system 600 to realize the functions of each component of translation system 350 or translation system 450 are stored on a DVD 700 inserted into a DVD drive 662 and transferred from the DVD drive 662 to SSD 718. Alternatively, these programs may be stored in a USB memory 702, which may be inserted into a USB port 664 and the programs transferred to SSD 718. Alternatively, these programs may be transmitted to computer 650 via network 704 and stored in SSD 718.

[0103] The program is loaded into RAM 716 when executed. Machine learning models such as deep neural networks are used for the large-scale language model 60 shown in Figures 7 and 9, the large-scale paraphrasing language model 352, and the match determination model 452 shown in Figure 9. In the computer system 600, a machine learning model that has already been trained in another device may be used, or the computer system 600 may be used as a training device to train a machine learning model.

[0104] The CPU 710 reads a program from the RAM 716 according to an address indicated by an internal register called a program counter (not shown) and interprets the instructions. The CPU 710 reads data required to execute the instructions from the RAM 716, the SSD 718, or another device according to the address specified by the instruction, and executes the processing specified by the instruction. The CPU 710 stores the execution result data at an address specified by the program, such as in the RAM 716, the SSD 718, or a register within the CPU 710. Depending on the address, the execution result data is output from the computer to an external device via, for example, the network I / F 726. At this time, the program counter value is also updated by the program. The computer program may be loaded directly into the RAM 716 from the DVD 700, the USB memory 702, or via the network 704. Note that some tasks (mainly numerical calculations) of the program executed by the CPU 710 are issued to the GPU 712 according to instructions included in the program or according to the analysis results obtained when the CPU 710 executes the instructions.

[0105] The program that enables the computer 650 to realize the functions of each of the above-described translation systems 350 and 450 includes a plurality of instructions written and arranged to cause the computer 650 to operate to realize those functions. Some of the basic functions required to execute these instructions may be provided by the OS (Operating System) or third-party programs running on the computer 650, various toolkit modules installed on the computer 650, or the program execution environment. Therefore, the program does not necessarily include all of the functions required to realize the system and method of this embodiment. The program may include only instructions that execute the operations of the above-described devices and their components by statically linking appropriate functions or modules at compile time or by dynamically calling them at runtime in a controlled manner to achieve the desired results. The method for operating the computer 650 for this purpose is well known. Therefore, a description of the method for operating the computer 650 will not be repeated here.

[0106] The embodiments disclosed herein are merely examples, and the present invention is not limited to the above-described embodiments. The scope of the present invention is defined by the claims in the appended claims, taking into consideration the detailed description of the invention, and includes all modifications within the meaning and scope equivalent to the wordings described therein. [Explanation of symbols]

[0107] 50 Natural Language Processing System 60 Large-scale Language Models 62,100 training data 64 Study Programs 66, 150, 200, 250, 300 Input token sequence 68, 152, 252, 302 output token sequence 162, 202, 212, 220, 262, 270, 312 Japanese strings 170, 320 English string 350, 450 translation system 352 Large-scale language model for paraphrasing 354 Display device 356 Input Device 370, 462 Paraphrase processing unit 372, 464 Text Replacement Section 374, 466 Original text input section 376 Storage device 378 Output Acquisition Unit 380, 468 Control section 452 Matching Judgment Model 460 Evaluation Department 600 Computer Systems 650 Computer 710 CPU 712 GPU 714 ROM 716 RAM 718 SSD

Claims

1. inputting a source text in a first language into a language model, the source text specifying a translation into the first language; A step in which a computer obtains an output sentence of the language model for the source sentence; A text processing method including, following the obtaining step, a step in which a computer selectively executes a first step of using the original sentence and a second step of using a paraphrase of the original sentence in place of the original sentence based on an evaluation of the output sentence.

2. The performing step includes: a step of the computer presenting the output sentence to a user so that the user can evaluate the output sentence; a step of receiving, by the computer after the step of presenting, an instruction from a user regarding whether the first step or the second step is to be executed; The text processing method according to claim 1 , further comprising a step of causing a computer to execute the first step or the second step in accordance with the instruction.

3. The performing step includes: A step in which a computer calculates a degree of agreement between the original sentence and the output sentence in order to evaluate the output sentence; The text processing method according to claim 1 , further comprising a step of: a computer executing the first step or the second step depending on whether the degree of match exceeds a threshold value.

4. the first step includes a step of providing the source text to the language model specifying a translation into a second language different from the first language, and receiving a translation of the source text in the second language at the output of the language model; The second step is A step in which a computer generates a paraphrase of the original sentence; The text processing method according to claim 3 , further comprising the step of: a computer replacing the original sentence with the paraphrase sentence to perform the obtaining step and the selectively executing step.

5. The step of generating paraphrased sentences includes: a step of inputting the output sentence or the input sentence consisting of the source sentence into a paraphrase model of the first language, the paraphrase model being different from the language model; The text processing method according to claim 4 , further comprising: a step of: a computer acquiring, as the paraphrase sentence, an output sentence of the paraphrase model for the input sentence.

6. an output acquisition device that inputs an original sentence in a first language into a language model by specifying a translation into the first language, and acquires an output sentence of the language model for the original sentence; A text processing device including a processing device that selectively executes a first process using the original sentence and a second process using the paraphrase sentence instead of the original sentence based on an evaluation of the output sentence acquired by the output acquisition device.

Images