Data structure of adjustment data of generative ai, dialog system, adjustment method, and program

Abstract

To provide a data structure of adjustment data of generative AI, an interaction system, an adjustment method, and a program for achieving smooth interaction by a low-resource language by using an interaction system by a large-scale language model.SOLUTION: In a dialogue system 1, adjustment data of an input / output unit 3 includes: bilingual example sentence text data including a first conversational sentence described in standard Japanese which is a standard language, and a second conversational sentence having the same meaning as the first conversational sentence and described in Kagoshima dialect which is a low-resource language classified into a language homologous to standard Japanese and having less language resources for learning of generative AI than standard Japanese, in a comparable manner; vocabulary collection text data including a first word used in standard Japanese and a second word used in Kagoshima dialect and having the same meaning as the first word in a comparable manner; and grammar text data 10 in which a grammar of Kagoshima dialect is described in a comparable manner to a grammar of standard Japanese. A data structure of adjustment data of generative AI is used for adjustment processing of adjusting generative AI2 using a large-scale language model supporting a dialogue in standard Japanese.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a data structure of adjustment data for generative artificial intelligence (AI), an interactive system, an adjustment method, and a program. [Background technology]

[0002] Interactive generative AI that generates sentences and other information through dialogue with users is becoming more common. Such generative AI is based on a large-scale language model (LLM), and the degree of completion, or accuracy, of the output of the generative AI depends on the amount of language data used for training. Therefore, while generative AI can accurately generate sentences written in standard Japanese, which requires a relatively large amount of data, it does not perform as well on sentences written in low-resource languages, which require a small amount of data.

[0003] On the other hand, research is being conducted on comprehensively generating bilingual dictionaries for low-resource languages ​​(see, for example, Non-Patent Document 1). Also, research is being conducted on improving the accuracy of machine translation between foreign languages ​​and standard Japanese by utilizing linguistic knowledge (see, for example, Non-Patent Document 2). [Prior art documents] [Non-patent literature]

[0004] [Non-Patent Document 1] Yohei Murakami et al., "Language Resources," [online], Ritsumeikan University Social Intelligence Laboratory, [Retrieved February 16, 2024], Internet<URL:https: / / www.si-lab.org / research_language-ja.html> [Non-patent document 2] Zhuoyuan Mao, Raj Dabre, Fabien Cromieres, Haiyue Song, Ryota Nakao, and Yoshio Kurohashi, "Linguistic Knowledge-Based Multitask Pre-Training for Neural Machine Translation," 26th Annual Conference of the Association for Natural Language Processing, Ibaraki, pp. 1061-1064, March 2020 Summary of the Invention [Problem to be solved by the invention]

[0005] Dialects are a typical example of such low-resource languages, and it is precisely dialects that generative AI is expected to be able to solve regional issues. In the past, communication between people was closely tied to local communities, and it was sufficient for local residents to converse with each other using the local dialect. However, since the modern era, with the spread of standard Japanese, younger generations in local areas are increasingly unable to understand the local dialect. Furthermore, with the rise of inter-regional exchange, opportunities for contact with people from other regions are increasing. As a result of this trend, in modern society, the "dialect barrier" has become apparent, where miscommunication occurs when speakers and non-speakers of a local dialect come into contact, and this has become a problem.

[0006] The "dialect barrier" problem is particularly serious in medical settings, as it makes communication between patients and doctors difficult, and this has become a regional issue. Furthermore, some dialects are in danger of disappearing, and their preservation has become a regional issue. Against this background, there are hopes for solving regional issues, such as developing tools to eliminate the "dialect barrier" or educational materials to support the preservation of dialects in danger of disappearing.

[0007] The present invention has been made in light of the above-mentioned circumstances, and aims to provide a data structure for adjustment data for a generation AI, a dialogue system, an adjustment method, and a program that can realize smooth dialogue in a low-resource language using a dialogue system with a large-scale language model. [Means for solving the problem]

[0008] In order to achieve the above object, the data structure of the adjustment data for the generation AI according to the first aspect of the present invention is as follows: bilingual example text data including a first conversational sentence written in a first language that is a standard language, and a second conversational sentence that has the same meaning as the first conversational sentence and is written in a second language that is a low-resource language that is classified as a dialect related to the first language and has fewer language resources for training a generation AI than the first language, so that they can be compared; lexicon text data including a first word used in the first language and a second word used in the second language and having the same meaning as the first word, in a comparable manner; grammar text data in which the grammar of the second language is described so as to be comparable to the grammar of the first language; It is used in an adjustment process to adjust a generative AI that uses a large-scale language model that supports dialogue in the first language so that it can understand sentences written in the second language and create sentences written in the second language.

[0009] the training data is used for fine-tuning the large-scale language model; This may also be the case.

[0010] The adjustment data is included in the prompt together with the task to be assigned to the generation AI. This may also be the case.

[0011] The grammar text data includes: The sentences include descriptions of correspondences between the first language and the second language in cases, adjectives, verb conjugations, sentence endings, and conjunctions, as well as grammatical prohibitions in the second language. This may also be the case.

[0012] A dialogue system according to a second aspect of the present invention comprises: It is equipped with generative AI that uses a large-scale language model to support dialogue in the first language, which is standard Japanese. The generated AI is bilingual example text data including a first conversational sentence written in a first language and a second conversational sentence having the same meaning as the first conversational sentence, the second conversational sentence being written in a second language that is classified as a cognate language of the first language and has fewer language resources for learning by a generation AI than the first language, and that can be compared; lexicon text data including a first word used in the first language and a second word used in the second language and having the same meaning as the first word, in a comparable manner; and grammar text data in which the grammar of the second language is described so as to be comparable to the grammar of the first language.

[0013] The adjustment method according to the third aspect of the present invention comprises: A method for adjusting a generative AI using a large-scale language model that supports dialogue in a first language that is a standard language, comprising: bilingual example text data including a first conversational sentence written in a first language and a second conversational sentence having the same meaning as the first conversational sentence, the second conversational sentence being written in a second language that is a dialect classified as a cognate language of the first language and has fewer language resources for training a generation AI than the first language, and which can be compared with the first conversational sentence; lexicon text data including a first word used in the first language and a second word used in the second language and having the same meaning as the first word, in a comparable manner; The generation AI is provided with adjustment data including grammar text data in which the grammar of the second language is described so as to be comparable to the grammar of the first language, and an adjustment process is performed to adjust the generation AI so that it can create sentences written in the second language.

[0014] A program according to a fourth aspect of the present invention comprises: bilingual example text data including a first conversational sentence written in a first language that is a standard language, and a second conversational sentence that has the same meaning as the first conversational sentence and is written in a second language that is a low-resource language that is classified as a cognate language of the first language and has fewer language resources for training a generation AI than the first language, and that can be compared; lexicon text data including a first word used in the first language and a second word used in the second language and having the same meaning as the first word, in a comparable manner; an adjustment process for adjusting a generative AI using a large-scale language model that supports dialogue in the first language, using adjustment data including grammar text data in which the grammar of the second language is described so as to be comparable to the grammar of the first language; Have the computer run it. [Effects of the Invention]

[0015] According to the present invention, a dialogue system based on a large-scale language model can be used to realize smooth dialogue in a low-resource language. [Brief explanation of the drawings]

[0016] [Figure 1] 1 is a block diagram showing a data structure of adjustment data and a functional configuration of a dialogue system according to a first embodiment of the present invention. FIG. [Figure 2] 2 is a block diagram showing the hardware configuration of a user terminal and a server computer that realize the interactive system of FIG. 1. FIG. [Figure 3] 1A is a block diagram showing the flow of a dialogue in standard Japanese in the dialogue system, and FIG. 1B is a diagram showing an example of a chat screen in standard Japanese in the dialogue system. [Figure 4] This figure shows an example of a task given to a generation AI and an answer it generates. [Figure 5] FIG. 2 is a schematic diagram showing an example of bilingual example sentence text data; [Figure 6] FIG. 2 is a schematic diagram illustrating an example of vocabulary text data. [Figure 7] FIG. 2 is a schematic diagram illustrating an example of grammar text data. [Figure 8] 10 is a flowchart of an interaction process. [Figure 9] This figure shows an example of a task and an answer generated by a generation AI given the task. [Figure 10] 1A is a block diagram showing the flow of a dialogue in the Kagoshima dialect in the dialogue system, and FIG. 1B is a diagram showing an example of a chat screen in the Kagoshima dialect in the dialogue system. [Figure 11] FIG. 10 is a block diagram showing the data structure of adjustment data and the functional configuration of a dialogue system according to a second embodiment of the present invention. [Figure 12] 10 is a flowchart of an interaction process. DETAILED DESCRIPTION OF THE INVENTION

[0017] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In each drawing, the same or equivalent parts are denoted by the same reference numerals. In the following embodiments, the terms "have," "include," or "contain" also mean "consist of" or "consist of."

[0018] Embodiment 1 First, a first embodiment of the present invention will be described. As shown in Fig. 1, a dialogue system 1 according to this embodiment includes a generation AI 2 and an input / output unit 3. The generation AI 2 engages in a dialogue with a user by generating an answer to a task (question sentence, instruction sentence) according to the task included in an input prompt. The generation AI 2 is implemented, for example, in a server computer 20 shown in Fig. 2, and the input / output unit 3 is implemented, for example, in a user terminal 30 shown in Fig. 2.

[0019] [Hardware configuration] The generated AI2 shown in Fig. 1 is realized, for example, by a server computer 20 having the hardware configuration shown in Fig. 2 executing a software program. Specifically, the server computer 20 includes a CPU (Central Processing Unit) 21, which is a processor that controls the entire device, a main memory 22 such as a RAM (Random Access Memory), an external memory 23 configured from a non-volatile memory such as a flash memory or a hard disk, a communication interface 26 that communicates data with an external computer, and an internal bus 28 that connects these.

[0020] The program 29 is loaded from the external memory 23 into the main memory 22 and executed by the CPU 21. This realizes the functions of the server computer 20. When executing the program 29, the CPU 21 performs data communication with an external computer via the communication interface 26 as necessary. In this embodiment, the program 29 executed by the CPU 21 includes a program for generation AI2.

[0021] The functions of the generation AI2 can be implemented in a computer system consisting of one or more computers, each including one or more processors and one or more storage devices, including a non-transitory storage medium. The multiple computers communicate with each other via a communications network, thereby realizing the functions of the generation AI2. For example, some of the functions of the generation AI2 may be implemented in one computer, and others may be implemented in another computer. The functions of the generation AI2 may also be realized by a cloud computer.

[0022] The input / output unit 3 shown in Fig. 1 is realized, for example, by a user terminal 30 having the hardware configuration shown in Fig. 2 executing a software program. Specifically, the user terminal 30 includes a CPU (Central Processing Unit) 31, a main memory 32 of the CPU 31, an external memory 33 that stores a program 39, an operation unit 34 that is a device such as a keyboard and a mouse, a display 35 that is a display device such as a CRT (Cathode Ray Tube) or a liquid crystal monitor, a communication interface 36 that performs data communication with other computers, and an internal bus 38 that connects these.

[0023] The program 39 is loaded from the external memory 33 into the main memory 32 and executed by the CPU 31. This realizes the user terminal 30. When executing the program 39, the CPU 31 performs data communication with an external computer via the communication interface 36 as necessary. In this embodiment, the program 39 executed by the CPU 31 includes a program for the input / output unit 3.

[0024] Returning to FIG. 1, the generation AI2 can be, for example, chatGPT (Chat Generative Pre-trained Transformer), but is not limited to this. The task assigned to the generation AI2 and the answer generated by the generation AI2 are text data of sentences written in a natural language used by humans. The generation AI2 understands the content of the input task using a large-scale language model (LLM) and generates an answer to the task. The LLM is a natural language processing model trained on a large amount of natural language text data (language data). In this embodiment, the natural language supported by the LLM is standard Japanese (hereinafter simply referred to as "standard Japanese") as the first language. In other words, the first language is standard Japanese.

[0025] For example, as shown in Figure 3(A), when the input / output unit 3 receives text data of the sentence "My head hurts" as a task, the generation AI2 uses natural language processing to generate text data such as "Let's go to the hospital" as a response to "My head hurts" and sends it to the input / output unit 3.

[0026] The input / output unit 3 displays, for example, a chat screen 50 shown in FIG. 3(B) on the display 35 (see FIG. 2) of the user terminal 30. While viewing the chat screen 50 displayed on the display 35, the user operates the operation unit 34 (see FIG. 2) to input a sentence corresponding to a question or instruction. The input sentence (task) is also called a prompt. The prompt is displayed on the chat screen 50. On the chat screen 50 shown in FIG. 3(A), "My head hurts" is displayed in the upper left. The sentence "Let's go to the hospital," indicated by the text data of the answer sent from the generation AI2, is displayed in the chat screen 50 shown in FIG. 2 at the lower right of the prompt. In this way, in the dialogue system 1, a dialogue between the user and the generation AI2 is carried out by displaying the task (prompt) and answer between the generation AI2 and the input / output unit 3 on the input / output unit 3.

[0027] As mentioned above, Generator AI2 supports standard Japanese. However, as shown in Figure 4, when we used Zero-Shot Prompting to give Generator AI2 the task, "Please speak in the Kagoshima dialect. Please explain Tokyo Skytree in about 200 characters." The answer to the task, as shown in Figure 4, contained no Kagoshima dialect elements, was dominated by standard Japanese, and included dialect elements from other dialects such as Chugoku-Shikoku (jaken, jakedo) and Kyushu-Hizen (batten, bai). It also included dialect elements that did not fit the context, such as "batten boasts of its height (although it boasts of its height)." Without learning anything about the Kagoshima dialect, i.e., using Zero-Shot Prompting, Generator AI2 would generate answers that contained hallucinations and were therefore of low accuracy.

[0028] Therefore, in this embodiment, generation AI2 is made to correspond to the Kagoshima dialect, which is classified as a language (dialect) of the same family as standard Japanese. In other words, the Kagoshima dialect corresponds to the second language. The Kagoshima dialect has fewer language resources for training generation AI2 than standard Japanese. For this reason, generation AI2 is unable to learn the Kagoshima dialect. In this embodiment, an adjustment process is performed to adjust generation AI2 using adjustment data 10, which is text data showing the differences between standard Japanese and the Kagoshima dialect, as shown in Figure 1.

[0029] 1, the adjustment data 10 has a data structure that includes bilingual example sentence text data 10A, vocabulary text data 10B, and grammar text data 10C. The adjustment data 10 is used in an adjustment process that adjusts the generation AI 2 that uses a large-scale language model that supports dialogue in standard Japanese so that it can understand sentences written in the Kagoshima dialect and create sentences written in the Kagoshima dialect.

[0030] The bilingual example sentence text data 10A includes a first conversation written in standard Japanese and a second conversation written in a second language, which is a dialect of standard Japanese and has the same meaning as the first conversation, in a manner that allows comparison. FIG. 5 shows an example of the bilingual example sentence text data 10A. As shown in FIG. 5, the bilingual example sentence text data 10A first indicates that the text data is a collection of example sentences in the Kagoshima dialect. Separated by an "=", the first conversation in standard Japanese is written on the left side, and the second conversation in the Kagoshima dialect, which has the same meaning as the first conversation, is written on the right side. Because the generation AI2 can understand standard Japanese, the rules for describing the text data can also be written in a natural language format that humans can understand. As an example, FIG. 5 shows "It's Saigo" as the first conversation and "It's Saigo tomorrow" as the second conversation. The bilingual example sentence text data 10A includes multiple pairs of the first conversation and the second conversation. The number of sets is determined depending on the variety of variations between the example sentences in the first language and the example sentences in the second language.

[0031] The lexicon text data 10B is bilingual dictionary data that includes a first word used in standard Japanese and a second word used in the Kagoshima dialect that has the same meaning as the first word, in a comparable manner. As shown in FIG. 6, the lexicon text data 10B first indicates that it is a dictionary of Kagoshima dialect vocabulary, and shows that, separated by an equal sign (=), a first word in standard Japanese is written on the left side, and a second word in the Kagoshima dialect that has the same meaning as the sentence is written on the right side. In FIG. 6, as an example, "fumoto ni" (at the base) is shown as the first word, and "fumote" (at the base) is shown as the second conversation sentence. The lexicon text data 10B includes not only pairs of nouns but also pairs of verbs. The lexicon text data 10B includes multiple pairs of first and second words, both nouns and verbs. The number of pairs is determined depending on the variety of variations between the first and second words.

[0032] The grammar text data 10C is a text that explains the grammar of the Kagoshima dialect. The grammar text data 10C describes the grammar of the Kagoshima dialect in a manner that allows comparison with the grammar of standard Japanese. As shown in FIG. 7, the grammar text data 10C first indicates that it is a text that explains the grammar of the Kagoshima dialect. Next, the grammar text data 10C includes text that describes the differences in case, adjectives, verb conjugations, sentence endings, conjunctions, and prohibited actions between standard Japanese and the Kagoshima dialect. Note that the number of grammar items to be included in the grammar text data 10C varies depending on the type of dialect. Therefore, various items can be added without being limited to these items.

[0033] 1. case The grammar text data 10C can include sentences describing the correspondence between particles or particles, that is, the correspondence between cases, between standard Japanese and the Kagoshima dialect. For example, the grammar text data 10C can include information that the case particle "ga" in standard Japanese is also "ga" in Kagoshima, and example sentences in the Kagoshima dialect using this case particle. In addition, correspondences such as "wo" → "woba," "made" → "zui," "kara" → "kara," and "niwa" → "nya" and example sentences can be described. In this way, the grammar text data 10C is written so that the grammar of standard Japanese and the grammar of the Kagoshima dialect can be compared.

[0034] 2.Adjective The grammar text data 10C can contain sentences relating to the correspondence between adjective endings in standard Japanese and the Kagoshima dialect. For example, a sentence such as "The adjective ending in standard Japanese is 'i', but the adjective ending in the Kagoshima dialect is 'ka'" can be included in the grammar text data 10C.

[0035] 3.verb 3.1. Basics of verb conjugation The grammar text data 10C can contain sentences about the basics of verb conjugation for standard Japanese and the Kagoshima dialect. The grammar text data 10C first describes that the following explanation is a sentence about the basics of verb conjugation, such as, "In the Kagoshima dialect, when a verb is conjugated, it is conjugated into the following forms depending on the meaning." Following this sentence, explanations of various verb conjugations, such as negative, polite, compound, ta, sentence-terminating, base, conditional, and imperative forms, conjugation methods, and comparisons of conjugations between standard Japanese and the Kagoshima dialect are described as follows: (Example) Negative form: A negative expression. Ends with "n." Example (<Standard Japanese> = [Kagoshima dialect]: <MiNai> = [Min]) Polite form: A polite expression. Ends with "mosu." Example: = [Mimos] Compound form: The form that forms the front element of a compound verb. The compound form is in the form of "っ". Example: <mik>kita = [mik]kita Ta form: Same as standard Japanese. Indicates the past. Example: <saw> = [saw] · Sentence termination form: A form used when a sentence is terminated with a comma (,) and another sentence follows. Add "ssee,". Example: I'm going to watch a movie and then go to dinner. = I'm going to watch a movie and then go to dinner. · Basic shape: The basic shape. Example: <see> = [see] ·Conditional form: Expression of a hypothesis. Example: <If you look> = [If you look] Imperative form: An expression used when giving an order. One-step and two-step verb stems have the prefix "re" attached. Multi-step verb stems end in the e-row, just like standard Japanese. Example: <look> = [see]

[0036] 3.2. Types of conjugation The grammar text data 10C can contain sentences about the types of verb conjugation for standard Japanese and the Kagoshima dialect. The grammar text data 10C first contains a sentence about the types of verb conjugation, such as "Verb stems in the Kagoshima dialect are classified into one-stage, two-stage, and multi-stage types." 3.2.1. Single stage system Following that sentence, you can write that examples of one-step systems will follow, such as, "For example, one-step systems are conjugated as follows," and below that you can write examples of one-step systems as follows: (Example) See... Stem: mi-, dialect base form [mi] Negative form: min (not to see) Polite form: mi-mosu Compound form: mi-t <look at me> ta-form mi-ta <saw> ·Sentence-stopping form: Mi-ssee, <look,> Basic form: mi-t <look> ·Conditional form: mi-reba <if you look> · Imperative form: mi-re <look>

[0037] 3.2.2.Two-stage system Following the one-stage system, a description of the two-stage system can be written as follows. (Example) In the two-stage system, two types of stems alternate. Stem 2 is used in the base and subjunctive forms, and stem 1 is used otherwise. (Example) To get up... Stem 1: oki- / Stem 2: oku-, dialect base form "okut" Negative stem 1 oki-n (does not happen) Polite form Stem 1 oki-mosu <get up> Compound stem 1 oki-t <wake up> ·Ta form Stem 1 oki-ta <got up> ·Sentence-stopping form Stem 1 oki-sse, <wake up,> ·Basic form Stem 2 Oku-t <Owake> ·Conditional form Stem 2 Oku-reba <Okireba> Imperative stem 1 oki-re <wake up>

[0038] 3.2.3.Multi-stage utilization Following the two-stage system, a description of the multi-stage system can be written as follows. (Example) Multi-step verbs are easier to understand if you imagine them as five-step verbs in standard Japanese. The stem-final sounds are set back. The stem-final sounds of negative forms are in the a-row, the stem-final sounds of polite, compound, and past tense forms are in the i-row, and the stem-final sounds of conditional and imperative forms are in the e-row. The sentence-stopping form is made by adding "-se," to the stem, which has the same form as the basic form. (Example) Stab... Stem: Sa row multi-stage sasa- / sashi- / sase- / sa, dialect base form "sa" Negative form: sasan (do not stab) Polite form: sashimosu (to stab) Compound form: sashi <stab> TA form sashi-ta <stabbed> ·Sentence-stop form sass-see, <stab,> Basic form: Sa-<stab> ·Conditional form: sase-ba <sashi-ba> Imperative form: sase - <stab> The ta form is very simple. It is the same form as in standard Japanese, and the phonetic form is also the same. Example: <kaku> = the ta form of [kat] → <kakatta> = [kaita] Example: <kamu> = the ta form of [kan] → <kamitta> = [kanda] Example: <Iku> = the ta form of [ichi] → <Ichita> = [itta]

[0039] 4. End of sentence In the grammar text data 10C, the correspondence between the sentence endings of standard Japanese and the Kagoshima dialect can be described, for example, as follows: (Example) The standard Japanese "da" and "desu" are "ja" in the Kagoshima dialect. (Example) Taro da. = Taro ja. The past tense affirmative form of "ja" is "jatta." (Example) Sensei datta. = Sensei datta. The negative present form of "ja" is "janaka." (Example) Taro janai. = Taro janaka. The negative past tense of "ja" is "janatana." (Example) It was not Taro. The sentence-final particle "yo" in standard Japanese is "do" in the Kagoshima dialect. (Example) It's Taro. = Taro jadd. In the Kagoshima dialect, "~nandayo" at the end of a sentence is translated as "~jattaddo." (Example) Taro is a teacher. = Taro is a teacher. "~dayo" at the end of a sentence is pronounced "~tattdo" in Kagoshima dialect. (Example) "It's going to be fun." = "sutattdo." / "It's wide." = "hirokatatattdo." The past tense at the end of a sentence, "~datta da yo" is pronounced "~jatta tada" in the Kagoshima dialect. (Example) Taro was a teacher. = Taro was a teacher. The standard Japanese "masu" is "mosu" in the Kagoshima dialect. (Example) Lower it. = Lower it too. The past tense affirmative form of "mosu" is "moshita." (Example) I lowered it. = I also lowered it. The negative present form of "mosu" is "mohan." (Example) I won't lower it. = I won't lower it. The "so" in hearsay is "gena" in the Kagoshima dialect. (Example) It seems there was a mountain. = Yama ata gena.

[0040] 5. Conjunctions In the grammar text data 10C, the correspondence between conjunctions in standard Japanese and the Kagoshima dialect can be described, for example, as follows: (Example) In the Kagoshima dialect, the "kara" (because there is a reason) is changed to "de." (Example) I went because I had business to attend to. = I went because I had business to attend to. In the Kagoshima dialect, "therefore" is used to state a reason. (Example) Taro is a teacher, so he is serious. = Taro is a teacher, so he is serious. The adversative conjunction "kedo" is "don" in the Kagoshima dialect. (Example) It's a little far, but I'll go. The adversative conjunction "but" is "jaddon" in the Kagoshima dialect. (Example) Hanako is a girl, but she is strong. = Hanako is a girl, but she is strong.

[0041] 6.Prohibitions In the grammar text data 10C, grammatical prohibitions in the Kagoshima dialect can be described, for example, as follows: (Example) Do not use "bai" or "tai" at the end of a sentence. These are not Kagoshima dialects. Examples of prohibited words: "Ikubai" and "Ikutai." -Do not use "batten" as an adversative conjunction. "Batten" is not a Kagoshima dialect word. The adversative conjunction in the Kagoshima dialect is "jaddon."

[0042] In this way, the grammar text data 10C includes sentences describing the correspondence between the cases, adjectives, verbs, sentence endings, and conjunctions of the second language and the first language, as well as grammatical prohibitions in the Kagoshima dialect.

[0043] The comparison of the example sentences may be written as follows, for example: How would you directly translate the Japanese phrase "It's Saigo" into Kagoshima dialect? It would be "It's Saigo tomorrow." The adjustment data 10 may be generated in the form of a text file or a spreadsheet. If the adjustment data 10 is generated in the form of a spreadsheet, the Kagoshima dialect "Saigo de goasu" (It's Saigo desu) can be inserted into the cell to the right of the Japanese cell "Saigo desu" (It's Saigo desu) to instruct the generation AI 2 how to read the spreadsheet.

[0044] Returning to FIG. 1 , the input / output unit 3 includes an adjustment unit 5, a prompt generation unit 6, and an answer reception unit 7. In this embodiment, the adjustment data 10 is used for fine-tuning the LLM by the adjustment unit 5. Fine-tuning is a method of adding a new layer to an already trained large-scale language model and relearning a part or the entire large-scale language model using the adjustment data 10. Fine-tuning is performed by incorporating a fine-tuning API (Application Programming Interface) provided for the generation AI2 into the adjustment unit 5, and the adjustment unit 5 sending the adjustment data 10 to the generation AI2 via the API. Note that the adjustment data 10 sent by the adjustment unit 5 via the API also basically follows a prompt-and-answer procedure. The prompt generation unit 6 generates a prompt (task) based on an operation input to the operation unit 34 and sends it to the generation AI2. The answer reception unit 7 receives the answer sent from the generation AI2 and displays it on the display 35.

[0045] Dialogue processing in the Kagoshima dialect will be described below. As shown in FIG. 8, first, fine tuning is performed (step S1). This fine tuning corresponds to the adjustment process. Here, the adjustment unit 5 sends pre-generated adjustment data 10 to the generation AI 2. The generation AI 2 adds a new layer to the large-scale language model and re-learns part or all of the large-scale language model using the adjustment data 10. After the fine tuning is completed, the generation AI 2 may notify the input / output unit 3 that the adjustment is complete.

[0046] The input / output unit 3 waits until task input is started by operation input from the operation unit 34 (step S2; No). When task input is started by operation input (step S2; Yes), the input / output unit 3 waits until task input is completed (step S3; No). During this time, the prompt generation unit 6 generates a prompt (task) based on the operation input from the operation unit 34.

[0047] When the task input is completed by an operation input to end the task input (step S3; Yes), the input / output unit 3 sends a prompt (task) to the generation AI 2 (step S4). Then, the input / output unit 3 waits until it receives a response (step S5; No).

[0048] When the answer is received (Step S5; Yes), the input / output unit 3 displays the answer (Step S6). Subsequently, the input / output unit 3 determines whether or not an instruction to end the dialogue has been given (Step S7). If an instruction to end the dialogue has not been given (Step S7; No), the input / output unit 3 returns to Step S2 and waits until the task input is completed (Step S2; No).

[0049] In this dialogue processing, fine tuning to generate sentences in the Kagoshima dialect is performed before the dialogue, so it is possible to create sentences in the Kagoshima dialect, as shown in Figure 9. The prompt is shown at the top of Figure 9, and the answer sentence is shown below it. A to D displayed in the answer sentence indicate the results of a human evaluation of the dialect in that part later. The evaluation contents of A to D are as follows: A: No problem B: Insufficient (It may be used in dialect contexts, but there are more traditional ways of saying it) C: Unnatural (the meaning is understandable, but it feels strange in the dialect) D: Illegible (grammatically and semantically impossible) Even one sentence containing a D grade can hinder comprehension, so if an answer sentence containing a D grade is generated, it will be considered a failure. The answer sentence shown in Figure 9 does not contain any D grades, meaning that Generation AI2 has successfully generated an answer sentence written in the Kagoshima dialect.

[0050] Furthermore, by fine-tuning the adjustment unit 5 of the generation AI2 using the adjustment data 10, the generation AI2 becomes capable of conversing in the Kagoshima dialect. As shown in FIG. 10(A), when the Kagoshima dialect (bintagai taka) is input as a task by the prompt generation unit 6 of the input / output unit 3, the generation AI2 converts "bintagai taka" to "my head hurts." Furthermore, the generation AI2 generates the answer "Let's go to the hospital," converts this sentence to the Kagoshima dialect "ishadon niigga," and outputs it to the input / output unit 3. The answer reception unit 7 of the input / output unit 3 displays "ishadon niigga." In this way, a conversation using sentences in the Kagoshima dialect becomes possible.

[0051] Furthermore, the input / output unit 3 is not limited to a normal user terminal. For example, it may be implemented in a chatbot that converses with a user. In this case, fine tuning can be performed so that the chatbot can use the dialect used by the user, allowing the user to converse with the chatbot in that dialect. In this case, the chatbot is equipped with a microphone and a speaker, and the user can converse by voice. Voice input through the microphone is converted into text data and sent to the generation AI 2, and the text data sent from the generation AI 2 is converted into voice by the chatbot and output from the speaker.

[0052] Embodiment 2 Next, a second embodiment of the present invention will be described. In the first embodiment, the generated AI2 is adjusted by fine tuning using the adjustment data 10. In this embodiment, the adjustment data 10 is included in a prompt together with the original main task and sent to the generated AI2, thereby adjusting the generated AI2.

[0053] 11, the prompt generation unit 6 adds the adjustment data 10 to the main task, which is the question and answer to the generation AI 2, and sends it as a prompt to the generation AI 2. In this case, the adjustment data 10 is written in standard Japanese, and the main task may be written in standard Japanese or in the Kagoshima dialect.

[0054] The generation AI2 uses a large-scale language model to understand the input training data 10 and prompts including the primary task, and interacts with the user by generating an answer to the primary task using the Kagoshima dialect grammar specified in the training data 10. In this embodiment, the generation AI2 is input with the primary task so that it can understand a sentence written in the second language or create a sentence written in the second language.

[0055] The following describes dialogue processing in the Kagoshima dialect. As shown in Fig. 12, the input / output unit 3 waits until task input is started by operation input from the operation unit 34 (step S2; No). When task input is started by operation input (step S2; Yes), the input / output unit 3 waits until task input is completed (step S3; No). During this time, the prompt generation unit 6 generates a main task based on the operation input from the operation unit 34.

[0056] When the end of task input is instructed by an operation input and task input is completed (step S3; Yes), the prompt generator 6 generates a prompt by adding adjustment data to the generated main task (step S10) and sends the prompt (task) to the generator AI 2 (step S4). Subsequently, the input / output unit 3 waits until it receives a response (step S5; No).

[0057] When the answer is received (Step S5; Yes), the input / output unit 3 displays the answer (Step S6). Subsequently, the input / output unit 3 determines whether or not an instruction to end the dialogue has been given (Step S7). If an instruction to end the dialogue has not been given (Step S7; No), the input / output unit 3 returns to Step S2 and waits until task input is started (Step S2; No).

[0058] In this way, in this embodiment, by including information about the grammar of the Kagoshima dialect along with the main task in the prompt, the AI2 is prompted to accurately generate sentences in the Kagoshima dialect. In this way, accurate sentences in the Kagoshima dialect can be generated.

[0059] As explained in detail above, according to the data structure of the adjustment data 10, the dialogue system, the adjustment method, and the program of this embodiment, a large-scale language model that supports standard Japanese can be used as a base to train the large-scale language model using the adjustment data 10 relating to bilingual example sentences between standard Japanese and the Kagoshima dialect, a bilingual dictionary, and grammatical changes.Therefore, by using a dialogue system based on a large-scale language model, smooth dialogue in the Kagoshima dialect can be realized with a small amount of data, even if the Kagoshima dialect is a low-resource language with little training data.

[0060] In the above embodiment, the first language is standard Japanese and the second language is the Kagoshima dialect. However, this is not limiting. The first language may be a language other than standard Japanese, such as English, and the second language may be another dialect, such as the Kansai dialect or the Tohoku dialect. Furthermore, the second language may be narrowed down to a dialect used only in Kagoshima City, rather than a dialect used in Kagoshima Prefecture. In either case, the second language is derived from the first language.

[0061] As the generation AI2, various large-scale language models can be used, including chatGPT, Transformer-based large-scale language models such as BERT (Bidirectional Encoder Representations from Transformers), CNN (Convolutional Neural Network), RNN (Recurrent Neural Network), and LSTM (Long Short Term Memory).

[0062] The hardware and software configurations of the dialogue system 1 are merely examples and can be changed and modified as desired.

[0063] The core processing portion of the server computer 20, which is composed of CPUs 21, 31, main memories 22, 32, external memories 23, 33, operation unit 34, display 35, communication interfaces 26, 36, and internal buses 28, 38, etc., can be realized using an ordinary computer system rather than a dedicated system. The same applies to the user terminal 30. For example, a computer program for executing the above operations may be stored on a computer-readable recording medium (such as a flexible disk, CD-ROM, or DVD-ROM) and distributed, and the computer program may be installed on a computer to configure the interactive system 1 that executes the above processing. Alternatively, the computer program may be stored in a storage device of a server device on a communication network such as the Internet, and the interactive system 1 may be configured by downloading the computer program to an ordinary computer system.

[0064] When the functions of the interactive system 1 are realized by sharing the functions between an OS (operating system) and an application program, or by cooperation between the OS and the application program, only the application program portion may be stored in a recording medium or storage device.

[0065] It is also possible to superimpose a computer program on a carrier wave and distribute it over a communications network. For example, the computer program may be posted on a bulletin board system (BBS) on the communications network and distributed over the network. The computer program may then be started and executed under the control of an operating system in the same way as any other application program, thereby enabling the above-mentioned processing to be performed.

[0066] This invention allows various embodiments and modifications without departing from the broad spirit and scope of this invention. Furthermore, the above-described embodiments are intended to explain this invention and do not limit the scope of this invention. That is, the scope of this invention is defined by the claims, not the embodiments. Various modifications made within the scope of the claims and the meaning of the invention equivalent thereto are considered to be within the scope of this invention. [Industrial Applicability]

[0067] The present invention can be applied to generative AI that uses large-scale language models. [Explanation of symbols]

[0068] 1 Dialogue system, 2 Generative AI, 3 Input / output unit, 5 Adjustment unit, 6 Prompt generation unit, 7 Answer acceptance unit, 8 Main task, 10 Adjustment data, 10A Bilingual example sentence text data, 10B Lexicography text data, 10C Grammar text data, 20 Server computer, 21 CPU, 22 Main memory, 23 External memory, 26 Communication interface, 28 Internal bus, 29 Program, 30 User terminal, 31 CPU, 32 Main memory, 33 External memory, 34 Operation unit, 35 Display, 36 Communication interface, 38 Internal bus, 39 Program, 50 Chat screen

Claims

1. bilingual example text data including, in a comparable manner, a first conversational sentence written in a first language that is a standard language, and a second conversational sentence that has the same meaning as the first conversational sentence and is written in a second language that is a low-resource language that is classified as a cognate language of the first language and has fewer language resources for learning generative artificial intelligence (AI) than the first language; lexicon text data including a first word used in the first language and a second word used in the second language and having the same meaning as the first word, in a comparable manner; grammar text data in which the grammar of the second language is described so as to be comparable to the grammar of the first language; The AI ​​is used in an adjustment process to adjust a generation AI using a large-scale language model that supports dialogue in the first language so that the AI ​​can understand sentences written in the second language and create sentences written in the second language. Data structure of data for adjusting generated AI.

2. the training data is used for fine-tuning the large-scale language model; A data structure of adjustment data for the generated AI according to claim 1.

3. The adjustment data is included in a prompt together with a task to be assigned to the generated AI. A data structure of adjustment data for the generated AI according to claim 1.

4. The grammar text data includes: The sentences include descriptions of correspondences between the first language and the second language in cases, adjectives, verb conjugations, sentence endings, and conjunctions, as well as grammatical prohibitions in the second language. A data structure of adjustment data for the generated AI according to claim 1.

5. Equipped with a generative AI using a large-scale language model that supports dialogue in a first language, which is standard English; The generated AI is bilingual example text data including a first conversational sentence written in a first language and a second conversational sentence having the same meaning as the first conversational sentence, the second conversational sentence being written in a second language that is classified as a cognate language of the first language and has fewer language resources for training a generation AI than the first language, and that can be compared; lexicon text data including a first word used in the first language and a second word used in the second language and having the same meaning as the first word, in a comparable manner; and grammar text data in which the grammar of the second language is described so as to be comparable to the grammar of the first language. Dialogue system.

6. A method for tuning a generative AI using a large-scale language model that supports dialogue in a first language that is a standard language, comprising: bilingual example text data including a first conversational sentence written in a first language and a second conversational sentence having the same meaning as the first conversational sentence, the second conversational sentence being written in a second language that is a dialect classified as a cognate language of the first language and has fewer language resources for training the generation AI than the first language, and which can be compared; lexicon text data including a first word used in the first language and a second word used in the second language and having the same meaning as the first word, in a comparable manner; and providing the generation AI with adjustment data including grammar text data in which the grammar of the second language is described so as to be comparable to the grammar of the first language, and executing an adjustment process to adjust the generation AI so that the generation AI can create sentences written in the second language. Adjustment method.

7. bilingual example text data including a first conversational sentence written in a first language that is a standard language, and a second conversational sentence that has the same meaning as the first conversational sentence and is written in a second language that is classified as a language of the same family as the first language and has fewer language resources for learning the generation AI than the first language, and that can be compared; lexicon text data including a first word used in the first language and a second word used in the second language and having the same meaning as the first word, in a comparable manner; an adjustment process for adjusting a generative AI using a large-scale language model that supports dialogue in the first language, using adjustment data including grammar text data in which the grammar of the second language is described so as to be comparable to the grammar of the first language; A program that a computer runs.

Images