Information processing system and information processing method

Abstract

Provided is a novel information processing system that is excellent in convenience, usefulness, and reliability. The information processing system includes a first component which has: a function of receiving a first original sentence; a function of receiving first vector data into which an embedding model converted the first original sentence; a function of receiving second vector data into which the embedding model converted a first translation sentence generated by a translation model from the first original sentence; a function of searching a first database by using the first vector data to extract a first similar sentence and first similar vector data; a function of searching the first database by using the second vector data to extract a second similar sentence and second similar vector data; a function of generating an instruction sentence; a function of receiving an answer sentence generated by a large language model on the basis of the instruction sentence; a function of generating a score on the basis of the answer sentence; and a function of updating parameters of the translation model on the basis of the score.

Description

Information processing system and information processing method

[0001] One aspect of the present invention relates to an information processing system and an information processing method.

[0002] Note that one embodiment of the present invention is not limited to the above technical field. The technical field of one embodiment of the invention disclosed in this specification and the like relates to an object, a method, or a manufacturing method. Alternatively, one embodiment of the present invention relates to a process, a machine, manufacture, or a composition of matter. Therefore, more specifically, examples of the technical field of one embodiment of the present invention disclosed in this specification include a semiconductor device, a display device, a light-emitting device, a power storage device, a memory device, a driving method thereof, or a manufacturing method thereof.

[0003] There has been active research and development into machine translation, which uses computers to translate one natural language into another. Machine translation can be categorized into rule-based machine translation, statistical machine translation, and neural machine translation, which uses neural networks.

[0004] In recent years, there has been active development of language models using neural networks, with large-scale language models (LLMs) attracting particular attention. A large-scale language model is a natural language processing model trained using a large amount of data. A large-scale language model can realize, for example, a dialogue model that responds to user instructions. Non-Patent Document 1 discloses GPT-4 (Generative Pre-trained Transformer 4) (registered trademark) as a large-scale language model, and ChatGPT as a dialogue model.

[0005] Summary of ChatGPT / GPT-4 Research and Perspective Towards the Future of Large Language Models, Yiheng Liu et al. (Submitted on 4 Apr 2023, [online], Internet <URL: https: / / arxiv.org / abs / 2304.01852>

[0006] When translating using a translation model provided by an external service, it can be difficult to replace the original text with a sentence in another language without losing the subtle nuances inherent in the original text. For example, consider a case where the user of the translation model is a person engaged in specialized work (e.g., a writer of specifications related to intellectual property). The user wants to create a translation of an original text containing content related to their work. In such a case, using a translation model provided by an external service may result in a generic translation that is not specific to the genre, and may not be able to convert the text into a sentence that is appropriate for the user's work.

[0007] Furthermore, many of the translation models provided by external services are provided in cloud format, so using an external service to translate text containing confidential company information may not be desirable from the perspective of information leakage.

[0008] Therefore, translation models used by professionals in the above-mentioned specialized fields are required to be able to generate translations that are as close to the specific field as possible.In addition, assuming cases where highly confidential documents such as internal company information are handled, translation models are required to be in-house developed for each organization to which the user belongs.

[0009] In view of the above, an object of one embodiment of the present invention is to provide an information processing system that can generate a translation that matches a user's field of expertise. Another object of one embodiment of the present invention is to provide a novel information processing system that is highly convenient, useful, or reliable. Another object of one embodiment of the present invention is to provide an information processing method that can generate a translation that matches a user's field of expertise. Another object of one embodiment of the present invention is to provide a novel information processing method that is highly convenient, useful, or reliable.

[0010] Note that the description of these problems does not preclude the existence of other problems. Note that one embodiment of the present invention does not necessarily need to solve all of these problems. Note that problems other than these will become apparent from the description in the specification, drawings, claims, etc., and it is possible to extract other problems from the description in the specification, drawings, claims, etc.

[0011] One aspect of the present invention includes a first component having a function of accepting a first original sentence, a function of inputting the first original sentence into an embedded model and accepting first vector data converted by the embedded model, a function of inputting the first original sentence into a translation model and accepting second vector data converted by the embedded model from a first translation sentence generated by the translation model, a function of searching a first database using the first vector data to extract first similar sentences and first similar vector data, and a function of searching the first database using the second vector data to extract second similar sentences and second similar vector data. The information processing system has the following functions: a function of generating an instruction sentence asking whether a first original sentence has been translated into a first translation sentence so as to correspond to the relationship between a first similar sentence and a second similar sentence; a function of inputting the instruction sentence into a large-scale language model and accepting an answer sentence generated by the large-scale language model; a function of generating a first score and a second score based on the answer sentence; a function of updating parameters of the translation model based on a total score obtained by adding up the first score and the second score; a function of calculating a score change amount and updating the score history number based on the total score; and a function of determining whether the score change amount and the score history number meet standard values.

[0012] Furthermore, in the above, it is preferable that the first component has a function of accepting a first document and a function of dividing the first document into paragraphs and inputting the paragraphs into an embedded model, and that the embedded model has a function of converting the paragraphs into third vector data and a function of linking the paragraphs and the third vector data and storing them in the first database.

[0013] In the above, the first document is preferably a document related to the content of the first original text.

[0014] Furthermore, in the above, the first score and the second score are each a reward source used in reinforcement learning of the translation model, the first score is a value whose magnitude varies depending on the content of the answer sentence, the second score is a value corresponding to the distance between the centers of gravity of the first group of similar vector data and the second group of similar vector data, the parameter is a weighting coefficient of the neural network used in the translation model, the score change amount is the amount of change since the last update of the score total, the score history count is a value corresponding to the number of learning times in reinforcement learning, and the reference value is preferably a value that can be set arbitrarily by the user.

[0015] Another aspect of the present invention has first to twentieth steps, in which in the first step, a first component receives a first original text and then passes the first original text to a second component, in the second step, the second component passes the first original text to a third component, in the third step, the third component generates a first translation based on the first original text and then passes the first translation to a fourth component, and in the fourth step, the fourth component converts the first translation into first vector data. After converting the first original sentence into the first vector data, the first vector data is transferred to the second component; in a fifth step, the second component searches the first database for the first vector data; in a sixth step, the second component extracts the first similar sentence and the first similar vector data linked to the first similar sentence from the first database, and the first similar vector data has the highest similarity to the first vector data; in a seventh step, the second component transfers the first original sentence to the fourth component; and in an eighth step, In step 10, the second component converts the first original sentence into second vector data and then passes the second vector data to the second component; in step 9, the second component searches the first database for the second vector data; in step 10, the second component extracts, from the first database, a second similar sentence and second similar vector data linked to the second similar sentence, and the second similar vector data has the highest similarity to the second vector data; and in step 11, The second component generates an instruction sentence asking whether the first original sentence has been translated into the first translation sentence so as to correspond to the relationship between the first similar sentence and the second similar sentence, and then passes the instruction sentence to the fifth component; in a twelfth step, the fifth component generates an answer sentence based on the instruction sentence, and then passes the answer sentence to the second component; in a thirteenth step, the second component generates a first score based on the answer sentence, and then passes the first score to the third component; and in a fourteenth step, the third component:The parameter is updated based on the first score, and in a fifteenth step, the second component calculates the score change amount and updates the score history number based on the first score, and in a sixteenth step, the second component determines whether the score change amount and the score history number each satisfy a reference value, and if it is determined that the score change amount and the score history number each satisfy a reference value, in a seventeenth step, the second component passes the first original sentence to the third component, and calculates the score change amount and the score history number each satisfy a reference value. If it is determined that the reference value is not cleared, the second component repeats the processes of steps 2 to 16 until the reference value is cleared; in step 18, the third component generates a second translation based on the first original text and then passes the second translation to the second component; in step 19, the second component accepts the second translation and then passes the second translation to the first component; and in step 20, the first component accepts the second translation.

[0016] Furthermore, in the above, it is preferable that steps 21 to 23 are included before step 1, and that in step 21, the first component receives the first document and then passes the first document to the second component, and in step 22, the second component divides the first document into paragraphs and passes the paragraphs to the fourth component, and in step 23, the fourth component converts the paragraphs into third vector data, and then links the paragraphs to the third vector data and stores them in the first database.

[0017] In the above, the first document is preferably a document related to the content of the first original text.

[0018] Also, in the above, in the second step to the sixteenth step, the fourth component performs reinforcement learning using a translation model, the first score is a reward source used in the reinforcement learning, the first score is the sum of the second score and the third score, the second score is a value that varies in magnitude depending on the content of the answer sentence, the third score is a value corresponding to the distance between the centers of gravity of the first group of similar vector data and the second group of vector data, the parameter is a weighting coefficient of the neural network used in the translation model, the score change amount is the amount of change since the last update of the score total, the score history count is a value corresponding to the number of learning times in the reinforcement learning, and the reference value is a value that can be set arbitrarily by the user.

[0019] According to one aspect of the present invention, it is possible to provide an information processing system capable of generating a translation that conforms to a user's field of expertise. Alternatively, according to one aspect of the present invention, it is possible to provide a novel information processing system that is highly convenient, useful, or reliable. Alternatively, according to one aspect of the present invention, it is possible to provide an information processing method that is capable of generating a translation that conforms to a user's field of expertise. Alternatively, according to one aspect of the present invention, it is possible to provide a novel information processing method that is highly convenient, useful, or reliable.

[0020] Note that the description of these effects does not preclude the existence of other effects. Note that one embodiment of the present invention does not necessarily have all of these effects. Note that effects other than these will become apparent from the description in the specification, drawings, claims, etc., and it is possible to extract other effects from the description in the specification, drawings, claims, etc.

[0021] FIG. 1 is a diagram illustrating the configuration of an information processing system. FIG. 2 is a block diagram illustrating the configuration of an information processing system. FIG. 3 is a block diagram illustrating the configuration of an information processing system. FIG. 4 is a block diagram illustrating the configuration of an information processing system. FIG. 5 is a block diagram illustrating the configuration of an information processing system. FIG. 6 is a block diagram illustrating the configuration of an information processing system. FIG. 7 is a block diagram illustrating the configuration of an information processing system. FIG. 8 is a block diagram illustrating the configuration of an information processing device used in the information processing system. FIG. 9 is a flowchart illustrating an information processing method. FIG. 10 is a flowchart illustrating an information processing method. FIG. 11 is a flowchart illustrating an information processing method. FIG. 12 is a flowchart illustrating an information processing method. FIG. 13 is a flowchart illustrating an information processing method. FIG. 14 is a flowchart illustrating an information processing method. FIG. 15 is a diagram illustrating a document input to an information processing system. FIG. 16 is a diagram illustrating how a document input to an information processing system is divided into paragraphs. FIG. 17 is a diagram illustrating vector data generated by an embedding model. FIG. 18A is a diagram illustrating an original text input to a translation model. FIG. 18B is a diagram explaining a translation generated by a translation model. FIG. 19 is a diagram explaining a directive sentence input to a large-scale language model. FIG. 20 is a diagram explaining how a score and a total score acquired by an information processing system are generated from an answer sentence generated by a large-scale language model. FIG. 21 is a diagram explaining how a score and a total score acquired by an information processing system are generated from an answer sentence generated by a large-scale language model. FIG. 22A is a diagram explaining the correlation between parameters of a translation model and the amount of score change of an information processing system. FIG. 22B is a diagram explaining the dependency of the amount of score change of an information processing system on the number of score histories. FIG. 23A is a diagram explaining an original sentence input to a translation model. FIG. 23B is a diagram explaining a translation generated by a translation model.

[0022] The embodiments will be described in detail with reference to the drawings. However, the present invention is not limited to the following description, and it will be readily understood by those skilled in the art that various changes in form and details can be made without departing from the spirit and scope of the present invention. Therefore, the present invention should not be interpreted as being limited to the description of the embodiments shown below. In the configuration of the invention described below, the same parts or parts having similar functions will be denoted by the same reference numerals in different drawings, and repeated description thereof will be omitted.

[0023] In this specification, the ordinal numbers such as "first" and "second" are used for convenience and do not limit the number of components or the order of the components (for example, the order of processes or the order of stacking). Furthermore, the ordinal numbers assigned to components in one part of this specification may not match the ordinal numbers assigned to the same components in other parts of this specification or in the claims.

[0024] (Embodiment 1) An information processing system according to one embodiment of the present invention has a function of constructing a user-specific database containing data based on content related to a text that the user wishes to translate by inputting literature containing that content.

[0025] An information processing system according to one embodiment of the present invention has a function of translating a sentence to be translated (original sentence) input by a user into a desired language using a translation model.

[0026] Furthermore, the information processing system according to one embodiment of the present invention has a function of converting an original sentence input by a user or a translation sentence generated by a translation model into vector data using an embedded model.

[0027] Furthermore, the information processing system according to one embodiment of the present invention has a function of performing a similarity search within the database using the vector data generated by the embedding model.

[0028] The information processing system according to one embodiment of the present invention has a function of generating a response sentence corresponding to an input instruction sentence by using a large-scale language model.

[0029] Furthermore, the information processing system according to one aspect of the present invention has a function of generating a score according to the answer sentence and performing reinforcement learning of the translation model using the score as a reward.

[0030] The information processing system according to one embodiment of the present invention has the above-described various functions, allowing it to build a unique database specialized for the user's field of expertise. Furthermore, the translation model, embedding model, and large-scale language model perform processing in conjunction with each other via the database, allowing the information processing system to generate translations that are tailored to the user's field of expertise.

[0031] In this specification, "vector data" refers to data consisting of a multidimensional array of numerical values, as opposed to text data consisting of character strings (natural language) such as sentences. Vector data can also be said to be data in a format that allows for arithmetic processing. Text data can be converted into vector data by digitizing it based on the characteristics, regularity, etc., of the text data. Known specific methods for converting text data into vector data include Bag of Words, distributed representation, and embedded representation.

[0032] Furthermore, in this specification, the term "similarity search" refers to a search that determines the degree of similarity between the above-mentioned vector data. One index that indicates the degree of similarity between vector data is cosine similarity. The distance between vector data can also be used. In this case, the closer the distance between vector data, the higher the similarity between the vector data. Methods for determining the distance between vector data include Euclidean distance, standard (normalized, average) Euclidean distance, Mahalanobis distance, Manhattan distance, Chebyshev distance, and Minkowski distance.

[0033] An information processing system according to one embodiment of the present invention will be described below with reference to FIGS. 1 to 8 and 15 to 23B.

[0034] 1 illustrates an example of the configuration of each component of an information processing system according to one embodiment of the present invention, a database, and a network connecting these components. It also illustrates an example of the flow of data exchanged between each component of the information processing system and the database.

[0035] 2 to 7 show more detailed examples of data flows exchanged between the components of the information processing system shown in FIG. 1 and with a database. FIG. 2 is a diagram summarizing on one page the data flows (processes T1 to T19) exchanged between the components of an information processing system according to one embodiment of the present invention or with a database. To facilitate understanding of the data flows between the components or with a database and the details of each process, FIG. 3 shows the data flows, process details, and components related to each process for processes T1 to T3, FIG. 4 shows processes T4 to T9, FIG. 5 shows processes T10 to T13, FIG. 6 shows processes T14 and T15, and FIG. 7 shows processes T16 to T19.

[0036] In Figures 1 to 7, the components are classified by function and shown as independent components or blocks, but in reality it is difficult to completely separate the components by function, and one component may be involved in multiple functions.

[0037] 1, an information processing system according to one aspect of the present invention includes a component 10, a component 20, a component 30, a component 40, and a component 50. The information processing system also exchanges predetermined data with a database 60 via a network 70.

[0038] The following describes the functions of an information processing system according to one embodiment of the present invention, broken down into components that make up the information processing system, with reference to Figures 1 to 8. Note that some of the descriptions of data exchange between the components may overlap.

[0039] <<Configuration Example of Component 10>> As the component 10, for example, a desktop computer can be used.

[0040] The component 10 can accept data input by a user, and can provide data output by the component 30 to the user.

[0041] For example, dedicated application software or a web browser may be operated. A user can access the information processing system via either of these, thereby enjoying services using the information processing system according to one aspect of the present invention.

[0042] The component 10 has a function of accepting information input by the user (document DOC shown in FIG. 1) and transferring it to the component 20 in a process T1 indicated by an arrow in FIGS.

[0043] A literature DOC is a document related to the content of a sentence (original text) that a user wishes to translate into another language. For example, if the original text contains keywords such as "secondary battery," "reliability," and "carrying capacity," the literature DOC can be a document containing one or more of the aforementioned keywords. A literature DOC can be a commercially available book, a published document (e.g., a paper, patent specification, etc.), or a private document (e.g., an internal company document, etc.). Note that the language of the literature DOC is not limited to the language of the original text. A document written in a language other than the language of the original text can also be used as a literature DOC. Furthermore, there is no particular limit to the number of literature DOCs that a user can input into component 10, and it can be one or two or more.

[0044] The document DOC can be, for example, a document (e.g., PDF) downloaded by the user from a search destination. It can also be, for example, a user-specified text file stored in an internal data source (e.g., folder, database, etc.). The document DOC can be in any file format as long as it is text data written in natural language.

[0045] FIG. 15 shows, as an example of a document DOC, a part of a patent specification (including paragraphs AAAA, BBBB, CCCC, and DDDD) that includes the above-mentioned three terms, "secondary battery," "reliability," and "carrying amount."

[0046] Component 10 also has a function of accepting a translation target sentence (original sentence ORG shown in FIG. 1) input by the user and transferring it to component 20 in process T4 indicated by an arrow in FIGS.

[0047] The original ORG is written in a natural language. FIGS. 18A and 23A show, as an example of the original ORG, a sentence such as "The relationship between reliability and loading amount in a secondary battery is shown below." The original ORG is a sentence that a user intends to translate into another language using an information processing system according to one embodiment of the present invention. Note that the original ORG does not necessarily have to be a single sentence, and may be composed of two or more sentences.

[0048] 2 and 7, the component 10 has a function of accepting a translation (translation NTRA shown in FIG. 1) from the component 20 and presenting it to the user.

[0049] The translation NTRA is written in a natural language. FIG. 23B shows an example of the translation NTRA, which is an English sentence: "The relationship between reliability and amount in secondary batteries is shown below." The translation NTRA is a sentence that is finally generated by an information processing system according to one embodiment of the present invention as an appropriate translation of the original text ORG input by the user. Note that the language of the translation is not limited to English, and the user can arbitrarily select a language from among all languages ​​that the information processing system can support.

[0050] <<Configuration Example of Component 20>> As the component 20, for example, a workstation, a server computer, a supercomputer, or the like can be used.

[0051] Furthermore, the component 20 preferably has a function as a parallel computer. By using the component 20 as a parallel computer, it is possible to perform large-scale calculations necessary for learning and inference in artificial intelligence (AI), for example.

[0052] Furthermore, the component 20 can perform processing using a natural language processing model that uses AI.

[0053] For example, processing can be performed using natural language models such as BERT (Bidirectional Encoder Representations from Transformers), T5 (Text-to-Text Transfer Transformer), GPT-3, GPT-3.5, GPT-4, LaMDA (Language Model for Dialogue Applications), PaLM (Pathways Language Model), and Llama2.

[0054] The component 20 has a function of receiving a document DOC from the component 10 in a process T1 indicated by an arrow in FIGS.

[0055] Component 20 has a function of dividing the written content of the document DOC received from component 10 into paragraphs (paragraphs PAR shown in FIG. 1). Fig. 16 shows an example of dividing the document DOC shown in Fig. 15 into paragraphs. Fig. 16 shows an example in which the section of the document DOC corresponding to paragraph number AAAA is divided into paragraphs PAR_A, the section of the document DOC corresponding to paragraph number BBBB is divided into paragraphs PAR_B, and the section of the document DOC corresponding to paragraph number CCCC is divided into paragraphs PAR_C.

[0056] 16 shows an example in which the contents of a document DOC are divided into paragraphs, but this is not limiting. The component 20 may also be configured to have a function to divide the contents of a document DOC into parts, chapters, sections, or paragraphs. The component 20 may also be configured to have a function to divide the contents of a document DOC into sentences, for example.

[0057] 2 and 3, the component 20 has a function of transferring paragraphs PAR (paragraphs PAR_A, PAR_B, PAR_C, etc. shown in FIG. 16) to the component 40.

[0058] The component 20 also has a function of receiving original text ORG from the component 10 in process T4 indicated by an arrow in FIGS.

[0059] Component 20 also has the function of transferring original text ORG to component 30 in process T5 indicated by an arrow in FIGS.

[0060] 2 and 4, the component 20 has a function of receiving vector data (vector data TVEC shown in FIG. 1) from the component 40. The vector data TVEC is obtained by converting a sentence (translated sentence TRA shown in FIG. 1; details will be explained in "Configuration Example of Component 30") obtained by translating an original sentence ORG into another language by the component 30 into vector data.

[0061] 2 and 4, the component 20 has a function of performing a similarity search for the vector data TVEC in the database 60. In Fig. 4, text data such as the above-mentioned paragraphs PAR_A, PAR_B, and PAR_C and vector data such as vector data PVEC_A, PVEC_B, and PVEC_C are linked and stored in the database 60, and the dashed arrows indicate how the component 20 evaluates the similarity between each vector data and the vector data TVEC. Details of the vector data PVEC_A, PVEC_B, and PVEC_C will be described below in "Example of Configuration of Component 40."

[0062] Cosine similarity is one index that indicates the degree of similarity between vector data. Cosine similarity is a commonly used index because it can be easily calculated by simply taking the dot product of two vector data. Cosine similarity is expressed as a real number between -1 and 1, and the closer the cosine similarity is to 1, the higher the similarity can be determined to be.

[0063] 2 and 4, the component 20 has a function of extracting and accepting from the database 60 similar sentences (similar sentences STRA shown in FIG. 1) determined to have a high degree of similarity in the similarity search. In FIG. 4, the dashed arrows indicate how the component 20 extracts similar vector data STVEC having a high degree of similarity to the vector data TVEC and similar sentences STRA linked to the similar vector data STVEC from the database 60. The similar vector data STVEC can be, for example, the vector data that is most similar to the vector data TVEC in the similarity search.

[0064] 4 shows only one similar vector data (the similar vector data STVEC) determined to have a high degree of similarity through the similarity search, but this is not limited to this. For example, it is also possible to extract the top 10 similar vector data determined to have a high degree of similarity through the similarity search, along with the similar sentences linked to each of the similar vector data. The number of similar vector data and similar sentences to be extracted from the database 60 can be freely set by the user.

[0065] Alternatively, when using cosine similarity as an index representing the level of similarity in the similarity search, for example, a cosine similarity of 0.8 or more can be considered to be high similarity, and vector data corresponding to this can be considered as similar vector data, and the similar vector data and similar sentences linked to the similar vector data can be extracted. The cosine similarity value that is considered to be high similarity can be freely set by the user.

[0066] Component 20 also has the function of transferring original text ORG to component 40 in process T10 indicated by an arrow in FIGS.

[0067] 2 and 5, the component 20 has a function of receiving vector data (vector data OVEC shown in FIG. 1) from the component 40. The vector data OVEC is obtained by converting the original text ORG by the component 40.

[0068] 2 and 5, the component 20 has a function of performing a similarity search for the vector data OVEC in the database 60. In Fig. 5, text data such as the above-mentioned paragraphs PAR_A, PAR_B, and PAR_C and vector data such as vector data PVEC_A, vector data PVEC_B, and vector data PVEC_C are linked and stored in the database 60, and the dashed arrows indicate how the component 20 evaluates the similarity between each vector data and the vector data OVEC.

[0069] 2 and 5, the component 20 has a function of extracting and accepting from the database 60 similar sentences (similar sentences SORG shown in FIG. 1) determined to have a high degree of similarity in the similarity search. In FIG. 5, the dashed arrows indicate how the component 20 extracts similar vector data SOVEC having a high degree of similarity to the vector data OVEC and similar sentences SORG linked to the similar vector data SOVEC from the database 60. The similar vector data SOVEC can be, for example, the vector data that has the highest degree of similarity to the vector data OVEC in the similarity search.

[0070] 5 shows only one similar vector data (the similar vector data SOVEC) determined to have a high degree of similarity by the similarity search, but this is not limited to this. As with the similar vector data SVEC described above, it is also possible to extract the top multiple similar vector data determined to have a high degree of similarity and the similar sentences linked to each of the similar vector data.

[0071] 2 and 6, component 20 has a function of transferring a directive (directive PT shown in FIG. 1) to component 50 and causing component 50 to execute the instructions described in the directive PT or to respond to a question. The directive PT is a statement to be transferred to component 50, including original text ORG transferred from component 10, translation text TRA transferred from component 30, etc.

[0072] The instruction statement PT describes in natural language the specific processing content to be executed by the component 50. Fig. 19 shows an example of an instruction statement PT in which three items, "1. Reference Information," "2. Japanese," and "3. English," are listed, and below that, a question is written: "Is the above English (3) a translation of Japanese (2) based on the reference information (1)?"

[0073] In the instruction sentence PT shown in Figure 19, the sentence written in the "2. Japanese" field corresponds to the sentence that the user intends to translate using the information processing system of one aspect of the present invention, i.e., the original sentence ORG (see Figure 18A). The sentence written in the "3. English" field corresponds to the translation sentence TRA (see Figure 18B) generated by the component 30 for the original sentence ORG. The English sentence written in the "1. Reference Information" field corresponds to a list of sentences extracted from the database 60 as being highly similar to the translation sentence TRA, i.e., a list of similar sentences STRA. The Japanese sentence written in the "1. Reference Information" field corresponds to a list of sentences extracted from the database 60 as being highly similar to the original sentence ORG, i.e., a list of similar sentences SORG.

[0074] In other words, the instruction sentence PT is a sentence asking whether the original sentence ORG written in the "2. Japanese" section has been translated so that the translation sentence TRA written in the "3. English" section corresponds to the relationship (which can also be called the degree of similarity) between the similar sentence SORG and the similar sentence STRA written in the "1. Reference Information" section.

[0075] The component 20 can automatically generate the instruction sentence PT based on the original sentence ORG, the translation sentence TRA, the similar sentences STRA, and the similar sentences SORG. Alternatively, a template containing only a question sentence, such as that shown in Figure 19, can be prepared in advance, and the instruction sentence PT can be created by inserting sentences into the three fields "1. Reference information," "2. Japanese," and "3. English" of the template.

[0076] Component 20 also has a function of accepting an answer sentence (answer sentence ANS shown in FIG. 1) generated by component 50 based on the instruction sentence PT in process T15 indicated by an arrow in FIGS. 2 and 6. Examples of answer sentences ANS are shown in FIGS. 20 and 21. FIG. 20 is an example of an answer sentence ANS when component 50 answers "Yes" to the question written in the instruction sentence PT shown in FIG. 19. FIG. 21 is an example of an answer sentence ANS when component 50 answers "No" to the question written in the instruction sentence PT shown in FIG. 19.

[0077] Furthermore, the component 20 has a function of generating a first score (score SC1) based on the content of the answer sentence ANS, which serves as a reward source necessary for reinforcement learning of the translation model used in the component 30. The score SC1 has a different value when a "Yes" answer sentence ANS ( FIG. 20 ) is generated in response to the question written in the instruction sentence PT than when a "No" answer sentence ANS ( FIG. 21 ) is generated. FIG. 20 shows an example in which 1 point is assigned as the score SC1 in response to the answer sentence ANS. Also, FIG. 21 shows an example in which 0 point is assigned as the score SC1 in response to the answer sentence ANS. Note that the values ​​of the score SC1 shown in FIGS. 20 and 21 are merely examples, and values ​​other than those shown in the examples may be used.

[0078] The component 20 also has a function of calculating the distance between the centers of gravity of each group of vector data (similar vector data SVEC) in the list of English sentences (list of similar sentences STRA) listed in the "1. Reference Information" section of the instruction sentence PT and each group of vector data (similar vector data SVEC) in the list of Japanese sentences (list of similar sentences SORG). The component 20 also has a function of generating a second score (score SC2) based on the calculation result, which serves as a reward source necessary for reinforcement learning of the translation model used in the component 30. For example, the closeness (or similarity) between the two vector data groups can be calculated using an index such as the aforementioned cosine similarity, and the calculation result can be used as the score SC2. FIG. 20 shows an example in which a score of 0.8 points is assigned as the score SC2 based on the cosine similarity between the group of similar vector data SVEC and the group of similar vector data SVEC. Similarly, FIG. 21 shows an example in which a score of -0.2 points is assigned as the score SC2.

[0079] In this way, even if the text data are written in different languages ​​(similar sentence SORG and similar sentence STRA), the similarity between them can be quantitatively expressed by performing calculations (such as cosine similarity calculations) using the vector data (similar vector data SOVEC and similar vector data SVEC) linked to the text data.

[0080] Component 20 also has a function of adding up the scores SC1 and SC2 described above to calculate a total score value (the score sum SUM shown in FIG. 1). FIG. 20 shows an example in which score SC1 (1 point) and score SC2 (0.8 points) are added up to give a score sum SUM of 1.8 points. FIG. 21 shows an example in which score SC1 (0 point) and score SC2 (-0.2 points) are added up to give a score sum SUM of -0.2 points.

[0081] 2 and 7, component 20 has a function of transferring the total score SUM to component 30 and causing component 30 to update a parameter (parameter PWT) based on the total score SUM. As described in "Configuration Example of Component 30," component 30 can have a function of performing translation using a neural network. Therefore, for example, the weighting coefficient of the neural network used in component 30 can be regarded as the parameter PWT.

[0082] The component 20 also has a function of calculating the amount of change in the total score SUM since the previous update (amount of score change DSC) and updating the cumulative number of updates to the total score SUM (number of score histories NUMSC).

[0083] The score change DSC is a value that changes each time the parameter PWT is updated. That is, depending on the value of the parameter PWT, the score change DSC can either increase or decrease. Furthermore, when reinforcement learning of the translation model used in component 30 progresses and the parameter PWT is optimized, the score change DSC will take on a minimal value.

[0084] 22A is an image diagram illustrating the correlation between the parameter PWT and the score change DSC. Fig. 22A shows three different magnitudes of parameters PWT (parameter PWT1, parameter PWT2, and parameter PWT3) and their corresponding score change DSC (score change DSC1, score change DSC2, and score change DSC3).

[0085] 22A shows that the score change DSC2 is smallest when component 30 is set to the value of parameter PWT2, and that whether the value of parameter PWT is set smaller (parameter PWT1) or larger (parameter PWT3) than parameter PWT2, the score change DSC increases more than the score change DSC2. Therefore, in the case of Figure 22A, when component 30 is set to the value of parameter PWT2, it can be said that the translation model used in component 30 is in an optimized state (i.e., the state in which the translation accuracy of component 30 is highest).

[0086] The score history number NUMSC can be said to be a value equivalent to the number of times of learning in reinforcement learning of the translation model used in component 30 (i.e., the number of times the parameter PWT is updated). As the number of times of learning increases, the translation model used in component 30 becomes more optimized, and therefore, as the score history number NUMSC increases, the score change amount DSC continues to decrease and converges to a constant value. FIG. 22B shows an image illustrating the dependency of the score change amount DSC on the score history number NUMSC. Note that, for ease of understanding, the image shown in FIG. 22B shows how the score change amount DSC decreases along a smooth curve as the score history number NUMSC increases, but in reality, it decreases overall with repeated small increases and decreases.

[0087] Furthermore, component 20 has a function of determining whether the number of score histories NUMSC and the score change DSC each meet a reference value. As described above, the number of score histories NUMSC corresponds to the number of times of learning in reinforcement learning of the translation model used in component 30. Therefore, from the viewpoint of ensuring the accuracy of the output result of component 30 (specifically, the translated sentence TRA), it is preferable that the number of score histories NUMSC be a certain value or more. For example, it is preferable that the number of score histories NUMSC be 100 or more. Furthermore, as described above, it is preferable that the score change DSC be as small as possible. For example, it is preferable that the score change DSC be as close to 0 as possible.

[0088] In the image diagram shown in FIG. 22B, a hatched pattern is applied to the area where the score history number is equal to or greater than NUMSC0 and the score change amount is equal to or less than DSC0. For example, if the score history number is NUMSC0 and the score change amount is DSC0, and both the score history number and the score change amount are within the above-mentioned area, it is determined that the standard value is met. Furthermore, if either or both of the score history number and the score change amount are not within the above-mentioned area, it is determined that the standard value is not met. The standard value can be set by the user as desired.

[0089] 2 and 7. If the component 20 determines that the above-mentioned reference value is satisfied, the component 20 has the function of transferring the original text ORG to the component 30 in process T17 indicated by the arrow in FIG.

[0090] The process in which component 20 passes original text ORG to component 30 is also performed in process T5 described above. However, there is a difference in that process T5 is performed when the translation model used by component 30 is in a state before optimization, whereas process T17 is performed when the translation model has been optimized.

[0091] The component 20 also has a function of accepting the translation NTRA generated by the component 30 in a process T18 indicated by an arrow in FIGS.

[0092] Component 20 also has the function of transferring translation text NTRA to component 10 in process T19 indicated by an arrow in FIGS.

[0093] <<Configuration Example of Component 30>> For example, a workstation, a server computer, a supercomputer, etc. can be used as the component 30. Note that the component 30 is larger in scale than the component 20 and has higher calculation capabilities.

[0094] Furthermore, the component 30 preferably has a function as a parallel computer, which enables the component 30 to perform large-scale calculations required for AI learning and inference, for example.

[0095] Furthermore, the component 30 can perform processing using a translation model that uses AI, i.e., a translation model that uses a neural network.

[0096] For example, processing can be performed using natural language models such as BERT, T5, GPT-3, GPT-3.5, GPT-4, LaMDA, PaLM, Llama2, ALBERT, and XLNet.

[0097] For example, the component 30 can translate a document written in language A into a document written in language B. The component 30 can also translate a document written in language A into a document written in language B according to a given directive. Constraints can also be included in the directive. This allows the degree of freedom in translation to be controlled by providing constraints.

[0098] The component 30 can also perform processing using a general-purpose language processing model that can perform a variety of natural language processing tasks.

[0099] Note that a person who provides a service using an information processing system according to one embodiment of the present invention does not necessarily have to own the component 30. For example, a service provider can use part of a service provided by another business or the like as the component 30.

[0100] The component 30 has a function of receiving the original text ORG from the component 20 in a process T5 indicated by an arrow in FIGS.

[0101] Component 30 also has a function of generating a translation TRA based on the original ORG passed from component 20. The translation TRA is a sentence obtained by translating the original ORG into a language specified by the user and supported by component 30. The translation TRA is written in a natural language. FIG. 18B shows an example of an English translation TRA generated based on the Japanese original ORG shown in FIG. 18A. FIG. 18B shows an example of an English translation TRA generated from the Japanese original ORG shown in FIG. 18A, which is "The relationship between reliability and capacity in secondary batteries is shown below."

[0102] Component 30 also has the function of transferring translation text TRA to component 40 in process T6 indicated by an arrow in FIGS.

[0103] The component 30 also has a function of receiving the score sum SUM from the component 20 and updating the parameter PWT in process T16 indicated by an arrow in FIGS.

[0104] Component 30 also has a function of receiving original text ORG from component 20 in process T17 indicated by an arrow in FIGS.

[0105] Component 30 also has a function of generating a translation NTRA based on original text ORG passed from component 20. The translation NTRA is a sentence generated after the translation model used in component 30 is optimized. FIG. 23B shows an example in which an English translation NTRA, "The relationship between reliability and amount in secondary batteries is shown below," is generated from the Japanese original text ORG shown in FIG. 23A , "The relationship between reliability and amount in secondary batteries is shown below." It can be seen that some words have been replaced with different ones compared to the translation TRA (see FIG. 18B ) generated before the translation model used in component 30 was optimized. The translation NTRA is a sentence presented to a user as the final version of the translation of the original text ORG.

[0106] Component 30 also has the function of transferring translation text NTRA to component 20 in process T18 indicated by an arrow in FIGS.

[0107] <<Configuration Example of Component 40>> For example, a workstation, a server computer, a supercomputer, etc. can be used as the component 40. Note that the component 40 is larger in scale than the component 20 and has higher calculation capabilities.

[0108] Furthermore, the component 40 preferably has a function as a parallel computer. By using the component 40 as a parallel computer, it is possible to perform large-scale calculations required for AI learning and inference, for example.

[0109] Furthermore, the component 40 can perform processing using an embedded model that uses AI.

[0110] For example, processing can be performed using natural language models such as BERT, T5, GPT-3, GPT-3.5, GPT-4, LaMDA, PaLM, Llama2, ALBERT, and XLNet.

[0111] For example, the component 40 can convert text data into vector data by performing processing using an embedded model.

[0112] Note that a person who provides a service using an information processing system according to one embodiment of the present invention does not necessarily have to own the component 40. For example, a service provider can use part of a service provided by another business or the like as the component 40.

[0113] The component 40 has a function of receiving paragraphs PAR (paragraphs PAR_A, PAR_B, PAR_C, etc. shown in FIG. 16) from the component 20 in a process T2 indicated by an arrow in FIGS.

[0114] The component 40 also has a function of converting the paragraph PAR passed from the component 20 into vector data PVEC. FIG. 17 shows, as an example of the vector data PVEC, vector data PVEC_A, vector data PVEC_B, and vector data PVEC_C obtained by converting the paragraphs PAR_A, PAR_B, and PAR_C shown in FIG. 16 , respectively. In FIG. 17 , the vector data PVEC_A is shown as an n-dimensional vector consisting of terms a1 through an (n is an integer of 2 or greater). The vector data PVEC_B is shown as an n-dimensional vector consisting of terms b1 through bn. The vector data PVEC_C is shown as an n-dimensional vector consisting of terms c1 through cn. The terms a1 through an, b1 through bn, and c1 through cn are each expressed as a positive number (integer or real number), a negative number (integer or real number), or 0.

[0115] 2 and 3, the component 40 has a function of linking a paragraph PAR with the corresponding vector data PVEC and storing the linked data in the database 60. Fig. 3 shows how combinations such as paragraph PAR_A and vector data PVEC_A, paragraph PAR_B and vector data PVEC_B, paragraph PAR_C and vector data PVEC_C are stored in the database 60.

[0116] In this way, by configuring the database 60 to be constructed based on the contents of the document DOC entered in advance by the user, the user can construct their own database 60 specialized for the contents related to the original ORG to be translated. Note that the database 60 can also be called a "vector database" because it has the function of storing vector data.

[0117] The component 40 also has a function of accepting the translation TRA generated by the component 30 in a process T6 indicated by an arrow in FIGS.

[0118] The component 40 also has a function of converting the translation text TRA passed from the component 30 into vector data TVEC. The vector data TVEC is also expressed as an n-dimensional vector, similar to the vector data PVEC described above.

[0119] The component 40 also has a function of transferring the vector data TVEC to the component 20 in process T7 indicated by an arrow in FIGS.

[0120] The component 40 also has a function of receiving original text ORG from the component 20 in a process T10 indicated by an arrow in FIGS.

[0121] The component 40 also has a function of converting the original text ORG passed from the component 20 into vector data OVEC. The vector data OVEC is also expressed as an n-dimensional vector, similar to the vector data PVEC described above.

[0122] The component 40 also has a function of transferring the vector data OVEC to the component 20 in a process T11 indicated by an arrow in FIGS.

[0123] <<Configuration Example of Component 50>> A large computer such as a server computer or a supercomputer can be used as the component 50. Note that the component 50 is larger in scale and has higher computing power than the component 20.

[0124] Furthermore, the component 50 preferably has a function as a parallel computer. By using the component 50 as a parallel computer, it is possible to perform large-scale calculations required for AI learning and inference, for example.

[0125] Furthermore, the component 50 can perform processing using a natural language processing model that uses AI. In particular, the component 50 can perform processing using a general-purpose language processing model that can perform various natural language processing tasks.

[0126] For example, processing can be performed using natural language models such as BERT, T5, GPT-3, GPT-3.5, GPT-4, LaMDA, PaLM, Llama2, ALBERT, and XLNet. In particular, processing using GPT-4 is preferable. This allows for more natural sentence generation or dialogue.

[0127] Note that a person who provides a service using an information processing system according to one embodiment of the present invention does not necessarily have to own the component 50. For example, a service provider can use part of a service provided by another business or the like as the component 50.

[0128] The component 50 has a function of receiving a directive PT from the component 20 in a process T14 indicated by an arrow in FIGS.

[0129] The component 50 also has the function of generating a reply sentence ANS based on the directive sentence PT, and transferring the reply sentence ANS to the component 20 in a process T15 indicated by an arrow in FIGS.

[0130] The component 50 has a function of performing processing using a large-scale language model. The large-scale language model has already learned a data set. As a result, as described above, the component 50 can automatically generate an answer sentence ANS based on a directive sentence PT passed from the component 20.

[0131] By having the various functions described above, an information processing system according to one embodiment of the present invention can build a unique database specialized for a user's field of expertise. Furthermore, by linking the translation model, embedding model, and large-scale language model through the database, the information processing system can generate translations tailored to the user's field of expertise. As a result, a novel information processing system with excellent convenience, usefulness, and reliability can be provided.

[0132] <Configuration Example of Information Processing Device> An information processing system according to one embodiment of the present invention includes an information processing device that performs the functions of the above-described components.

[0133] For example, an information processing system according to an embodiment of the present invention can be configured with an information processing device that performs the functions of component 10, an information processing device that performs the functions of component 20, an information processing device that performs the functions of component 30, an information processing device that performs the functions of component 40, and an information processing device that performs the functions of component 50. Note that the number of information processing devices that make up the information processing system according to an embodiment of the present invention is one or more. Furthermore, for example, the information processing system according to an embodiment of the present invention can be configured by connecting a plurality of information processing devices using a network 70.

[0134] When the number of information processing devices is small relative to the number of components, one information processing device can be shared by multiple components. When the number of information processing devices is large relative to the number of components, the functions of one component can be shared by multiple information processing devices.

[0135] By configuring an information processing system according to one embodiment of the present invention using a plurality of information processing devices, the load related to information processing can be distributed.

[0136] Hereinafter, a detailed configuration example of an information processing device (here, an information processing device that performs the functions of the component 20) that can be used in the information processing system of one embodiment of the present invention will be described.

[0137] <<Configuration Example of Information Processing Device>> An information processing device that can be used in an information processing system according to one embodiment of the present invention includes, for example, an input unit 110, a storage unit 120, a processing unit 130, an output unit 140, and a transmission path 150 (see FIG. 8).

[0138] In the drawings accompanying this specification, the components are classified by function and shown as independent blocks in the block diagrams. However, in reality, it is difficult to completely separate the components by function, and one component may be involved in multiple functions. For example, part of the processing unit 130 may function as the input unit 110. Also, one function may be involved in multiple components. For example, the processing performed by the processing unit 130 may be executed by different servers depending on the processing.

[0139] [Input Unit 110] The input unit 110 can receive data from outside the information processing device. For example, the input unit 110 receives data via the network 70.

[0140] The input unit 110 supplies the received data to one or both of the storage unit 120 and the processing unit 130 via the transmission path 150 .

[0141] [Storage Unit 120] The storage unit 120 has a function of storing a program executed by the processing unit 130. The storage unit 120 can also have a function of storing data generated by the processing unit 130 (e.g., calculation results, analysis results, inference results), data accepted by the input unit 110, etc.

[0142] The storage unit 120 may have a database. Furthermore, the information processing device may have a database separate from the storage unit 120. The information processing device may have a function to retrieve data from a database that exists outside the storage unit 120, outside the information processing device, or outside the information processing system. Furthermore, the information processing device may have a function to retrieve data from both its own database and an external database.

[0143] Either or both of a storage and a file server can be used as the memory unit 120. Also, a database that records paths of files stored in a file server can be used as the memory unit 120.

[0144] The storage unit 120 includes at least one of a volatile memory and a non-volatile memory. Examples of the volatile memory include a dynamic random access memory (DRAM) and a static random access memory (SRAM). Examples of the non-volatile memory include a resistive random access memory (ReRAM), a phase change random access memory (PRAM), a ferroelectric random access memory (FeRAM), a magnetoresistive random access memory (MRAM), and a flash memory. The storage unit 120 may include at least one of NOSRAM (registered trademark) and DOSRAM (registered trademark). The storage unit 120 may include a recording media drive. Examples of the recording media drive include a hard disk drive (HDD) and a solid state drive (SSD).

[0145] NOSRAM is an abbreviation for "Nonvolatile Oxide Semiconductor Random Access Memory (RAM)." NOSRAM refers to a memory in which memory cells are two-transistor (2T) or three-transistor (3T) gain cells and transistors (also referred to as OS transistors) that use metal oxide in their channel formation regions. OS transistors have an extremely small leakage current, i.e., a current that flows between the source and drain in an off state. NOSRAM can be used as a nonvolatile memory by retaining a charge corresponding to data in the memory cell using its extremely small leakage current characteristic. In particular, NOSRAM can read stored data without destroying it (nondestructive read), making it suitable for arithmetic processing in which only data read operations are repeated a large number of times. NOSRAM can increase its data capacity by stacking layers, and therefore can be used as a large-scale cache memory, main memory, or storage memory to improve the performance of semiconductor devices.

[0146] DOSRAM is an abbreviation for "Dynamic Oxide Semiconductor RAM" and refers to a RAM having 1T (transistor) 1C (capacitor) type memory cells. DOSRAM is a DRAM formed using OS transistors, and is a memory that temporarily stores information sent from an external device. DOSRAM is a memory that takes advantage of the low off-state current of OS transistors.

[0147] In this specification and the like, a metal oxide refers to an oxide of a metal in a broad sense. Metal oxides are classified into oxide insulators, oxide conductors (including transparent oxide conductors), oxide semiconductors (also referred to as oxide semiconductors or simply as OSs), and the like. For example, when a metal oxide is used for a semiconductor layer of a transistor, the metal oxide may be referred to as an oxide semiconductor.

[0148] The metal oxide included in the channel formation region preferably contains indium (In). When the metal oxide included in the channel formation region contains indium, the carrier mobility (electron mobility) of the OS transistor is increased. Furthermore, the metal oxide included in the channel formation region is preferably an oxide semiconductor containing element M. The element M is preferably at least one of aluminum (Al), gallium (Ga), and tin (Sn). Other elements applicable to element M include boron (B), silicon (Si), titanium (Ti), iron (Fe), nickel (Ni), germanium (Ge), yttrium (Y), zirconium (Zr), molybdenum (Mo), lanthanum (La), cerium (Ce), neodymium (Nd), hafnium (Hf), tantalum (Ta), and tungsten (W). However, a combination of two or more of the above elements may be used as element M. The element M is, for example, an element having a high bond energy with oxygen. For example, it is an element having a higher bond energy with oxygen than indium. The metal oxide contained in the channel formation region is preferably a metal oxide containing zinc (Zn), since zinc-containing metal oxides may be easily crystallized.

[0149] The metal oxide contained in the channel formation region is not limited to a metal oxide containing indium, but may be, for example, a metal oxide containing zinc but not indium, such as zinc tin oxide or gallium tin oxide, a metal oxide containing gallium, or a metal oxide containing tin.

[0150] [Processing Unit 130] The processing unit 130 has a function of performing processes such as calculation, analysis, and inference using data supplied from one or both of the input unit 110 and the storage unit 120. The processing unit 130 can supply generated data (e.g., calculation results, analysis results, and inference results) to one or both of the storage unit 120 and the output unit 140.

[0151] The processing unit 130 has a function of acquiring data from the storage unit 120. The processing unit 130 can also have a function of recording or registering data in the storage unit 120.

[0152] The processing unit 130 may include, for example, an arithmetic circuit. The processing unit 130 may include, for example, a central processing unit (CPU). The processing unit 130 may also include a graphics processing unit (GPU). The processing unit 130 may also include a neural processing unit (NPU).

[0153] The processing unit 130 may include a microprocessor such as a DSP (Digital Signal Processor). The microprocessor may be implemented by a PLD (Programmable Logic Device) such as an FPGA (Field Programmable Gate Array) or an FPAA (Field Programmable Analog Array). The processing unit 130 may also include a quantum processor. The processing unit 130 can perform various data processing and program control by interpreting and executing instructions from various programs using the processor. Programs that can be executed by the processor are stored in at least one of the memory area of ​​the processor and the storage unit 120.

[0154] The processing unit 130 may include a main memory. The main memory may include at least one of a volatile memory such as a RAM and a non-volatile memory such as a ROM (Read Only Memory). The main memory may also include at least one of the above-mentioned NOSRAM and DOSRAM.

[0155] The RAM may be, for example, a DRAM or an SRAM, and a virtual memory space is allocated and used as a working space for the processing unit 130. The operating system, application programs, program modules, program data, lookup tables, etc. stored in the storage unit 120 are loaded into the RAM for execution. The data, programs, and program modules loaded into the RAM are each directly accessed and operated by the processing unit 130.

[0156] The ROM can store BIOS (Basic Input / Output System), firmware, etc., which do not require rewriting. Examples of ROM include mask ROM, OTPROM (One Time Programmable Read Only Memory), and EPROM (Erasable Programmable Read Only Memory). Examples of EPROMs include UV-EPROMs (Ultra-Violet Erasable Programmable Read Only Memories), which allow stored data to be erased by exposure to ultraviolet light, EEPROMs (Electrically Erasable Programmable Read Only Memories), and flash memories.

[0157] The processing section 130 can include one or both of an OS transistor and a transistor having silicon in a channel formation region (Si transistor).

[0158] The processing unit 130 preferably includes an OS transistor. Because an OS transistor has an extremely small off-state current, using the OS transistor as a switch for retaining charge (data) flowing into a capacitor functioning as a memory element can ensure a long data retention period. By using this characteristic in at least one of the register and cache memory of the processing unit, the processing unit can be operated only when necessary, and can be turned off in other cases by saving information from the previous processing in the memory element. In other words, normally-off computing is possible, and the power consumption of the information processing system can be reduced.

[0159] It is preferable that the information processing device uses AI for at least some of its processing.

[0160] It is particularly preferable that the information processing device uses an artificial neural network (ANN, hereinafter also simply referred to as a neural network). A neural network is realized by a circuit (hardware) or a program (software).

[0161] In this specification, a neural network refers to a general model that mimics the neural circuit network of a living organism, determines the connection strength between neurons through learning, and has problem-solving capabilities. A neural network has an input layer, an intermediate layer (hidden layer), and an output layer.

[0162] In this specification and the like, when discussing neural networks, determining the connection strengths (also called weighting coefficients) between neurons from existing information may be referred to as "learning."

[0163] In this specification and the like, the act of constructing a neural network using connection strengths obtained by learning and deriving a new conclusion from it may be referred to as "inference."

[0164] [Output Unit 140] The output unit 140 can output at least one of the calculation result, analysis result, and inference result in the processing unit 130 to the outside of the information processing device. For example, the output unit 140 can transmit data via the network 70.

[0165] [Transmission Path 150] The transmission path 150 has a function of transmitting data. Data can be transmitted and received between the input unit 110, the storage unit 120, the processing unit 130, and the output unit 140 via the transmission path 150.

[0166] The network 70 that can be used in the information processing system of one embodiment of the present invention can connect multiple information processing devices. This allows the connected multiple information processing devices to transmit and receive data to and from each other. Furthermore, the load related to information processing can be distributed.

[0167] When wireless communication is performed, communication standards such as the fourth generation mobile communication system (4G), fifth generation mobile communication system (5G), and sixth generation mobile communication system (6G), or specifications standardized by the IEEE such as Wi-Fi (registered trademark) and Bluetooth (registered trademark), can be used as communication protocols or communication technologies.

[0168] For example, a local network can be used for the network 70. Also, an intranet or an extranet can be used for the network 70. Also, a personal area network (PAN), a local area network (LAN), a campus area network (CAN), a metropolitan area network (MAN), a wide area network (WAN), a global area network (GAN), etc. can be used for the network 70.

[0169] Furthermore, for example, a global network can be used for the network 70. Specifically, the Internet, which is the foundation of the World Wide Web (WWW), can be used.

[0170] Furthermore, a person who provides a service using the information processing system according to one aspect of the present invention can provide the service using the information processing method according to one aspect of the present invention via a network 70, for example.

[0171] When the information processing system according to an embodiment of the present invention is built within a local network, the possibility of confidential information leaking can be reduced, for example, compared to when the Internet is used.

[0172] The network 70 is not limited to using either a local network or a global network. For example, the network 70 may be configured to include both a local network and a global network.

[0173] This embodiment mode can be appropriately combined with other embodiment modes described in this specification.

[0174] Embodiment 2 In this embodiment, an information processing method according to one embodiment of the present invention will be described with reference to FIGS.

[0175] 9 to 14 are diagrams illustrating an information processing method according to one embodiment of the present invention.

[0176] In the following, detailed explanations may be omitted for the specific contents of the original text ORG, translation text TRA, vector data PVEC, instruction text PT, answer text ANS, translation text NTRA, etc., and the specific functions of component 10, component 20, component 30, component 40, and component 50, as reference may be made to the description in embodiment 1.

[0177] In the following description, the same symbols as those used in the first embodiment can be applied to those described in the first embodiment. Therefore, detailed descriptions of the meanings and definitions of symbols may be omitted.

[0178] <Example of Information Processing Method> An information processing method according to one embodiment of the present invention includes steps S1 to S29 (see FIGS. 9 to 14).

[0179] [Step S1] In step S1 (see FIG. 9), the component 10 receives a document DOC (see FIG. 15) input by the user, and then passes the document DOC to (inputs) the component 20.

[0180] [Step S2] In step S2 (see FIG. 9), the component 20 receives the document DOC from the component 10, and then divides the document DOC into paragraphs PAR (see FIG. 16).

[0181] [Step S3] In step S3 (see FIG. 9), the component 40 receives the paragraph PAR from the component 20, and then converts the paragraph PAR into vector data PVEC (see FIG. 17). The component 40 also associates the paragraph PAR with the corresponding vector data PVEC and stores them in the database 60.

[0182] [Step S4] In step S4 (see FIG. 9), the component 10 receives the original text ORG (see FIG. 18A) input by the user, and then passes (inputs) the original text ORG to the component 20.

[0183] [Step S5] In step S5 (see FIG. 9), the component 20 receives the original text ORG from the component 10 and then passes (inputs) the original text ORG to the component 30.

[0184] [Step S6] In step S6 (see FIG. 9), the component 30 receives the original text ORG from the component 20, and then generates a translation text TRA (see FIG. 18B) based on the original text ORG. The component 30 also passes the translation text TRA to the component 40.

[0185] [Step S7] In step S7 (see FIG. 9), the component 40 receives the translation text TRA from the component 30, and then converts the translation text TRA into vector data TVEC.

[0186] [Step S8] In step S8 (see FIG. 9), the component 20 receives (obtains) the vector data TVEC from the component 40.

[0187] [Step S9] In step S9 (see FIG. 10), the component 20 performs a similarity search of the database 60 for the vector data TVEC.

[0188] [Step S10] In step S10 (see FIG. 10), the component 20 extracts the similar sentences STRA and the similar vector data STVEC from the database 60 and accepts them.

[0189] [Step S11] In step S11 (see FIG. 10), the component 20 passes (inputs) the original text ORG to the component 40.

[0190] [Step S12] In step S12 (see FIG. 10), the component 40 receives the original text ORG from the component 20, and then converts the original text ORG into vector data OVEC.

[0191] [Step S13] In step S13 (see FIG. 11), the component 20 receives (obtains) the vector data OVEC from the component 40.

[0192] [Step S14] In step S14 (see FIG. 11), the component 20 performs a similarity search of the database 60 for the vector data OVEC.

[0193] [Step S15] In step S15 (see FIG. 11), the component 20 extracts the similar sentences SORG and the similar vector data SOVEC from the database 60 and accepts them.

[0194] In the present embodiment, an example is shown in which the processes of steps S5 to S10 are performed and then the processes of steps S11 to S15 are performed, but this is not limiting. In an information processing method according to one aspect of the present invention, the processes of steps S11 to S15 may be performed after the processes of step S4, and then the processes of steps S5 to S10 may be performed.

[0195] [Step S16] In step S16 (see Figure 12), component 20 passes (inputs) to component 50 an instruction statement PT (see Figure 19) asking whether the original sentence ORG has been translated into the translation sentence TRA so as to correspond to the relationship between the similar sentence SORG and the similar sentence STRA.

[0196] [Step S17] In step S17 (see FIG. 12), the component 50 receives the instruction statement PT from the component 20, and then generates a reply statement ANS (see FIGS. 20 and 21) based on the instruction statement PT. The component 50 also passes the reply statement ANS to the component 20.

[0197] [Step S18] In step S18 (see FIG. 12), the component 20 receives (obtains) the answer sentence ANS from the component 50.

[0198] [Step S19] In step S19 (see FIG. 12), the component 20 generates a score SC1 (see FIGS. 20 and 21) based on the content of the answer sentence ANS.

[0199] [Step S20] In step S20 (see FIG. 12), the component 20 generates a score SC2 (see FIGS. 20 and 21) based on the similar sentences STRA and SORG described in the directive sentence PT.

[0200] [Step S21] In step S21 (see FIG. 13), the component 20 adds up the scores SC1 and SC2 to calculate the total score SUM (see FIGS. 20 and 21).

[0201] [Step S22] In step S22 (see FIG. 13), the component 20 passes (inputs) the score sum SUM to the component 30.

[0202] [Step S23] In step S23 (see FIG. 13), the component 30 receives the score sum SUM from the component 20, and then performs processing to update the parameter PWT based on the score sum SUM.

[0203] [Step S24] In step S24 (see FIG. 13), the component 20 calculates the score change amount DSC based on the score total SUM, and updates the score history number NUMSC.

[0204] [Step S25] In step S25 (see FIG. 14), the component 20 determines whether the score history number NUMSC and the score change amount DSC each satisfy a reference value.

[0205] [Step S26] If the determination in step S25 is "Yes," the component 20 passes (inputs) the original text ORG (see FIG. 23A) to the component 30 in step S26 (see FIG. 14).

[0206] If the determination in step S25 is "No," the component 20 repeats the processes from step S5 to step S25 until the component 20 determines "Yes" in step S25.

[0207] [Step S27] After receiving the original text ORG from the component 20 in step S26, the component 30 generates a translation text NTRA (see FIG. 23B) based on the original text ORG in step S27 (see FIG. 14). The component 30 also passes the translation text NTRA to the component 20.

[0208] [Step S28] In step S28 (see FIG. 14), the component 20 receives (obtains) the translation NTRA from the component 30.

[0209] [Step S29] In step S29 (see FIG. 14), the component 10 receives the translation NTRA from the component 20, and then presents the translation NTRA to the user.

[0210] By going through the above series of processes, the user can check the final translation NTRA of the original text ORG via the component 10.

[0211] By applying the above-described information processing method, an information processing system according to one embodiment of the present invention can build a unique database specialized for a user's field of expertise. Furthermore, by linking the above-described translation model, embedding model, and large-scale language model through the database, the information processing system can generate translations tailored to the user's field of expertise. As a result, a novel information processing system with excellent convenience, usefulness, and reliability can be provided.

[0212] This embodiment mode can be appropriately combined with other embodiment modes described in this specification.

[0213] 10: component, 20: component, 30: component, 40: component, 50: component, 60: database, 70: network, 110: input unit, 120: storage unit, 130: processing unit, 140: output unit, 150: transmission path

Claims

a first component; The first component has a function of accepting a first original sentence, a function of inputting the first original sentence into an embedded model and accepting first vector data converted by the embedded model, a function of inputting the first original sentence into a translation model and accepting second vector data converted by the embedded model from a first translation sentence generated by the translation model, a function of searching a first database using the first vector data to extract first similar sentences and first similar vector data, a function of searching the first database using the second vector data to extract second similar sentences and second similar vector data, and a function of extracting the first similar sentences and the translation model has a function of generating an instruction sentence inquiring whether the first original sentence has been successfully translated into the first translation sentence so as to correspond to a relationship with the second similar sentence; a function of inputting the instruction sentence into a large-scale language model and accepting an answer sentence generated by the large-scale language model; a function of generating a first score and a second score based on the answer sentence; a function of updating parameters of the translation model based on a total score obtained by adding up the first score and the second score; a function of calculating a score change amount and updating a score history number based on the total score; and a function of determining whether the score change amount and the score history number satisfy a reference value. Information processing system. 　 In claim 1, the first component has a function of receiving a first document and a function of dividing the first document into paragraphs and inputting the paragraphs into the embedding model; the embedding model has a function of converting the paragraph into third vector data, and a function of linking the paragraph and the third vector data and storing them in the first database; Information processing system. 　 In claim 2, The first document is a document related to the content of the first original text. Information processing system. 　 In any one of claims 1 to 3, the first score and the second score are each a reward source used in reinforcement learning of the translation model; the first score is a value that varies depending on the content of the answer sentence, the second score is a value corresponding to a distance between centers of gravity of the first group of similar vector data and the second group of similar vector data; the parameters are weighting coefficients of a neural network used in the translation model, the score change amount is the amount of change since the last update of the score total; the score history number is a value corresponding to the number of times of learning in the reinforcement learning, The reference value is a value that can be arbitrarily set by the user. Information processing system. 　 The method includes steps 1 to 20, In the first step, the first component receives a first original sentence and then passes the first original sentence to the second component; In the second step, the second component passes the first text to a third component; In the third step, the third component generates a first translation based on the first original sentence, and then passes the first translation to a fourth component; In the fourth step, the fourth component converts the first translation into first vector data, and then transfers the first vector data to the second component; In the fifth step, the second component searches a first database for the first vector data; In the sixth step, the second component extracts, from the first database, a first similar sentence and a first similarity vector data associated with the first similar sentence; the first similar vector data has the highest similarity to the first vector data; In the seventh step, the second component passes the first text to the fourth component; In the eighth step, the fourth component converts the first original text into second vector data, and then transfers the second vector data to the second component; In the ninth step, the second component searches the first database for the second vector data; In the tenth step, the second component extracts second similar sentences and second similarity vector data associated with the second similar sentences from the first database; the second similar vector data has the highest similarity to the second vector data; In the eleventh step, the second component generates an instruction statement inquiring whether the first original sentence has been translated into the first translation sentence so as to correspond to the relationship between the first similar sentence and the second similar sentence, and then passes the instruction statement to a fifth component; In the twelfth step, the fifth component generates a response sentence based on the instruction sentence, and then passes the response sentence to the second component; In the thirteenth step, the second component generates a first score based on the answer sentence, and then passes the first score to the third component; In the fourteenth step, the third component updates a parameter based on the first score; In the fifteenth step, the second component calculates a score change amount and updates a score history number based on the first score; In the sixteenth step, the second component determines whether the score change amount and the score history number each satisfy a reference value; If it is determined that the score change amount and the score history number each satisfy the reference value, in the seventeenth step, the second component passes the first original sentence to the third component; When it is determined that the score change amount and the score history number do not satisfy the reference values, the second component repeats the processes of the second step to the sixteenth step until the reference values ​​are satisfied; In the eighteenth step, the third component generates a second translation based on the first original sentence, and then passes the second translation to the second component; In the nineteenth step, after receiving the second translation, the second component passes the second translation to the first component; In the twentieth step, the first component accepts the second translation. Information processing methods. 　 In claim 5, The method includes steps 21 to 23 before the first step, In the twenty-first step, after receiving the first document, the first component transfers the first document to the second component; In the twenty-second step, the second component divides the first document into paragraphs and passes the paragraphs to the fourth component; In the 23rd step, the fourth component converts the paragraph into third vector data, and then associates the paragraph with the third vector data and stores the third vector data in the first database. Information processing methods. 　 In claim 6, The first document is a document related to the content of the first original text. Information processing methods. 　 In any one of claims 5 to 7, In the second step to the sixteenth step, the fourth component performs reinforcement learning using a translation model; the first score is a reward source used in the reinforcement learning; the first score is the sum of the second score and the third score; the second score is a value that varies depending on the content of the answer sentence, the third score is a value corresponding to a distance between centers of gravity of the first similar vector data group and the second similar vector data group, the parameters are weighting coefficients of a neural network used in the translation model, the score change amount is the amount of change since the last update of the score total; the score history number is a value corresponding to the number of times of learning in the reinforcement learning, The reference value is a value that can be arbitrarily set by the user. Information processing methods.

Images