Specification creation support device, specification creation support system, specification creation support method, computer program, and non-transitory computer-readable medium

Abstract

A specification creation support device (1200) comprises: an input accepting unit (1210) into which invention information specifying an invention is input; and a prompt generating unit (1260) which generates, on the basis of the invention information, prompts for causing a large-scale language model that has been trained in advance on a large amount of text data to generate reference text that can be used in the specification of the invention.

Description

Specification writing support device, specification writing support system, specification writing support method, computer program, and non-transitory computer-readable medium

[0001] The present disclosure relates to a specification writing support device, a specification writing support system, a specification writing support method, a computer program, and a non-transitory computer-readable medium.

[0002] Patent document 1 discloses a patent document generation device that extracts from a database multiple documents similar to a received invention document, and generates columns and text for background art, means for solving the problem, and form for implementing the invention based on a document selected by the user from the multiple documents.

[0003] Japanese Patent Application Publication No. 2023-007268

[0004] The text generated by conventional specification writing support software contains text that appears strange to the user, requiring the user to make corrections that are burdensome for the user, and is therefore inconvenient.

[0005] The present disclosure aims to provide a specification creation support device, specification creation support method, computer program, and non-transitory computer-readable medium that can reduce the burden of correction work performed by users and improve convenience.

[0006] To achieve the above object, one aspect of the specification writing support device comprises an input receiving unit into which invention information specifying an invention is input, and a prompt generating unit that generates, based on the invention information, a prompt that causes a large-scale language model that has previously trained on a large amount of text data to generate reference sentences that can be used in the specification of the invention.

[0007] In addition, one aspect of a specification writing support method for achieving the above-mentioned object is a specification writing support method executed by a specification writing support device, which includes the steps of receiving invention information that identifies an invention, and generating, based on the invention information, a prompt that causes a large-scale language model that has previously trained on a large amount of text data to generate reference sentences that can be used in the specification of the invention.

[0008] Furthermore, one embodiment of a computer program for achieving the above object is a computer program including at least one instruction executed by a processor of a specification writing support device, wherein when the at least one instruction is executed by the processor, the specification writing support device receives invention information that identifies an invention, and generates, based on the invention information, a prompt that causes a large-scale language model that has previously trained on a large amount of text data to generate reference sentences that can be used in the specification of the invention.

[0009] The specification drafting support device having the above configuration can generate reference sentences in a large-scale language model in response to prompts generated by a prompt generation unit based on invention information. Unlike conventional technology that uses similarity calculations based on a certain number of sentences, these reference sentences are well-connected word-by-word and consistent across sentences. This allows users to easily access high-quality reference sentences when drafting specifications, reducing the burden of revision work and improving convenience. Similar effects can also be achieved with the specification drafting support method and computer program having the above configuration.

[0010] In addition, a specification drafting support device according to one embodiment for achieving the above object includes: a fine-tuning instruction unit that instructs a general-purpose large-scale language model, which has been previously trained on a large amount of text data, to fine-tune predetermined parameters of the large-scale language model by additionally training the model using training data consisting of patent documents specified by a user; an input unit into which an invention sentence that specifies the invention is input; and a display unit that displays a generated sentence generated by the fine-tuned large-scale language model in response to the invention sentence.

[0011] In addition, a specification drafting support system according to one embodiment for achieving the above object includes: a fine-tuning instruction unit that instructs a general-purpose large-scale language model, which has been previously trained on a large amount of text data, to fine-tune predetermined parameters of the large-scale language model by additionally training the model using training data consisting of patent documents specified by a user; an input unit into which an invention sentence that specifies the invention is input; and a display unit that displays a generated sentence generated by the fine-tuned large-scale language model in response to the invention sentence.

[0012] Furthermore, a specification writing support method according to one aspect for achieving the above object is a specification writing support method executed by a specification writing support device, and includes the steps of: issuing an instruction to fine-tune predetermined parameters of a general-purpose large-scale language model that has been previously trained on a large amount of text data by additionally training the large-scale language model using training data consisting of patent documents specified by a user; and displaying a generated sentence generated by the fine-tuned large-scale language model for an invention sentence that specifies an invention input to an input unit provided in the specification writing support device.

[0013] Furthermore, a computer program according to one embodiment for achieving the above object is a computer program including at least one instruction executed by a processor of a specification writing support device, wherein when the at least one instruction is executed by the processor, the specification writing support device is instructed to fine-tune predetermined parameters of the large-scale language model by additionally training a general-purpose large-scale language model that has been trained in advance on a large amount of text data using training data consisting of patent documents specified by a user, and to display a generated sentence generated by the fine-tuned large-scale language model in response to an invention sentence that specifies an invention input to an input unit provided in the specification writing support device.

[0014] In addition, according to one aspect of the present invention, a non-transitory computer-readable medium stores the computer program.

[0015] According to the specification drafting support device having the above configuration, a general-purpose large-scale language model is additionally trained using training data consisting of patent documents specified by a user, thereby fine-tuning certain parameters of the general-purpose large-scale language model. This allows the general-purpose large-scale language model to learn user preferences contained in the patent documents of the training data, for example. Note that user preferences may include, for example, sentence length, dependency relationships, writing style, and the order of explaining technical causal relationships. As a result, the generated text generated by the fine-tuned large-scale language model for a newly input invention text by the user reflects the user's preferences. This reduces the burden felt by the user when proofreading and correcting the generated text displayed on the display unit. Furthermore, the specification drafting support system, specification drafting support method, computer program, and non-transitory computer-readable medium having the above configuration can also achieve similar effects.

[0016] According to the present disclosure, it is possible to provide a specification creation support device, specification creation support system, specification creation support method, computer program, and non-transitory computer-readable medium that can reduce the burden of correction work performed by the user.

[0017] Furthermore, according to the present disclosure, it is possible to provide a specification preparation support device, a specification preparation support method, and a computer program that can improve convenience.

[0018] FIG. 1 is a schematic diagram illustrating a specification drafting support system and a server according to an embodiment of the present disclosure (hereinafter referred to as the present embodiment). FIG. 2 is a flowchart of a specification drafting support method according to a first embodiment. FIG. 3 is a diagram illustrating an example of training data consisting of patent documents specified by a user. FIG. 4 is a diagram for explaining fine tuning performed in the first embodiment. FIG. 5 is a diagram illustrating an example of an invention sentence input to an input unit in the first embodiment. FIG. 6 is a diagram illustrating an example of a generated sentence displayed on a display unit in the first embodiment. FIG. 7 is a flowchart of a specification drafting support method according to a second embodiment. FIG. 8 is a diagram illustrating an example of field-specific training data. FIG. 9 is a diagram for explaining fine tuning performed in the second embodiment. FIG. 10 is a diagram illustrating an example of an invention sentence input to an input unit. FIG. 11 is a diagram illustrating an example of a detailed independent claim. FIG. 12 is a diagram illustrating an example of an invention sentence input to an input unit. FIG. 13 is a diagram illustrating an example of an invention sentence input to an input unit. FIG. 14 is a diagram illustrating an example of a generated sentence displayed on a display unit. FIG. 15 is a schematic diagram illustrating a specification drafting support system according to another embodiment of the present disclosure. FIG. 16 is a block diagram of a specification drafting support device according to another embodiment of the present disclosure. FIG. 17 is a diagram showing an example of invention information. FIG. 18 is a diagram showing an example of pre-set rules used when converting drawings into text. FIG. 19 is a sequence diagram illustrating a series of processes until reference sentences generated by a large-scale language model are displayed on the display unit of a user terminal device. FIG. 20 is a flowchart illustrating the flow from uploading claims and drawings to generating a design section. FIG. 21 is a diagram showing an example of an operation screen. FIG. 22 is a diagram showing an example of an operation screen. FIG. 23 is a flowchart illustrating the flow from uploading claims and drawings to generating a design section. FIG. 24 is a diagram showing an example of an operation screen. FIG. 25 is a diagram showing an example of an operation screen. FIG. 26 is a diagram showing an example of an operation screen when generating reference sentences for modified examples and specific examples. FIG. 27 is a diagram showing an example of an operation screen on which no reference sentences are displayed in a first example of generating reference sentences using the free input function.FIG. 28 is a diagram showing an example of character information. FIG. 29 is a diagram showing an example of an operation screen displaying reference text in a first example of generating reference text using the free input function. FIG. 30 is a diagram showing an example of an operation screen displaying reference text in a second example of generating reference text using the free input function. FIG. 31 is a diagram showing an example of an operation screen displaying no reference text in a third example of generating reference text using the free input function. FIG. 32 is a diagram showing an example of character information. FIG. 33 is a diagram showing an example of an operation screen displaying reference text in a third example of generating reference text using the free input function. FIG. 34 is a diagram showing an example of a screen displayed on the display unit 1302 of the user terminal device 1300 when the drawing number of a drawing displayed on the operation screen is double-clicked. FIG. 35 is a diagram showing an example of an operation screen displaying no reference text when multiple drawing numbers and multiple claim item numbers are selected. FIG. 36 is a diagram showing an example of an operation screen displaying reference text when multiple drawing numbers and multiple claim item numbers are selected.

[0019] Hereinafter, an example of an embodiment of the present disclosure will be described with reference to the drawings.

[0020] (Overall Configuration) A specification creation support system 1 and a server 2 according to this embodiment will be described with reference to FIG. 1. FIG. 1 is a schematic diagram illustrating a specification creation support system 1 and a server 2 according to an embodiment of the present disclosure. As illustrated in FIG. 1, the specification creation support system 1 includes a specification creation support device 3 and a user terminal device 4. The server 2 and the specification creation support device 3 can communicate with each other via a network N, for example. The network N is, for example, a cloud network. A cloud network is a computer network that exists within a cloud computing infrastructure. The specification creation support device 3 and the user terminal device 4 can communicate with each other via wired or wireless communication.

[0021] The server 2 is, for example, a server computer that exists on a cloud network. The server 2 includes a storage unit 21 and a control unit 22.

[0022] The storage unit 21 may be configured with a storage device such as a hard disk drive (HDD), a solid state drive (SSD), or a flash memory. A large-scale language model is stored in the storage unit 21. The large-scale language model is a large-scale network trained with a large amount of text data. The large-scale language model is constructed by a neural network.

[0023] A neural network is a mathematical model that represents the characteristics of brain function through computer simulation. A neural network is a model in which, for example, artificial neurons (nodes) that form a network through synaptic connections change the strength of the synaptic connections through learning, thereby acquiring problem-solving capabilities. A neural network includes, for example, an input layer, an intermediate layer, and an output layer. A neural network acquires problem-solving capabilities through, for example, deep learning. Deep learning is performed using, for example, a Transformer, a Recurrent Neural Network (RNN), a Long Short-Term Memory (LSTM), or the like. The large-scale language model according to this embodiment is constructed by performing deep learning using a Transformer. In other words, the large-scale language model according to this embodiment is a language model based on a Transformer.

[0024] The Transformer is mainly used in the field of natural language processing. The Transformer has more than 100 billion parameters, allowing for highly accurate natural language processing. The Transformer is an encoder-decoder model that uses only attention. Because the Transformer does not require sequential processing of time-series data, it can be parallelized to a greater extent than a recurrent neural network, enabling shorter training times.

[0025] In this embodiment, the large-scale language model includes a general-purpose large-scale language model M1 and a fine-tuned large-scale language model M2. The general-purpose large-scale language model M1 has previously trained on a large amount of text data and can solve various language processing tasks. Examples of language processing tasks include sentence generation, machine translation, and question answering. For example, when performing sentence generation, the general-purpose large-scale language model M1 tokenizes an input sentence and converts it into a continuous numerical representation (multidimensional vector). Next, an encoder of the general-purpose large-scale language model M1 analyzes the converted numerical representation and converts it into an intermediate representation. Then, a decoder of the general-purpose large-scale language model M1 converts the intermediate representation back into a numerical representation, converts the tokens back into a sentence, and outputs the resulting representation. In this way, the general-purpose large-scale language model M1 generates sentences based on the input sentences.

[0026] The fine-tuned large-scale language model M2 is a large-scale language model created by fine-tuning a general-purpose large-scale language model M1. Fine-tuning is a method of developing a new neural network while utilizing a trained neural network by using part of the trained neural network as is and modifying or adding the rest as necessary and performing additional training. In fine-tuning a large-scale language model, for example, an additional layer for a new language processing task (e.g., text generation such as writing a specification) is added to the end of the neural network of the trained model, and then the parameters of some of the trained output layers are fine-tuned. In other words, layers other than the additional layer at the end are also adjusted. Note that fine-tuning is an example of additional training. Additional training is a technique for generating another trained model (additional training model) using a different dataset for an existing trained model (pre-trained model). The fine-tuned large-scale language model M2 can perform language processing tasks (e.g., sentence generation) using the same principles as the general-purpose large-scale language model M1, but can solve language processing tasks in specific fields with greater accuracy than the general-purpose large-scale language model M1.

[0027] The control unit 22 includes, as its hardware configuration, a memory and a processor. The memory may be, for example, a read-only memory (ROM) in which various computer programs are stored, or a random access memory (RAM) having multiple work areas in which various computer programs executed by the processor are stored. The processor may be, for example, a central processing unit (CPU), and is configured to load a computer program specified by the various computer programs stored in the ROM onto the RAM and execute various processes in cooperation with the RAM.

[0028] The control unit 22 includes a fine-tuning unit 221 and a generation instruction unit 222. The fine-tuning unit 221 performs fine tuning (additional learning) based on instructions from the specification creation support device 3, which will be described later. As a result, predetermined parameters of the general-purpose large-scale language model M1 are fine-tuned, and a fine-tuned large-scale language model M2 is created. Details of the fine-tuning will be described later. The predetermined parameters of the general-purpose large-scale language model M1 are, for example, at least some of the parameters (weight coefficients) of the output layer in the entire neural network. The created fine-tuned large-scale language model M2 is transmitted to the storage unit 21 and stored in the storage unit 21.

[0029] The generation instruction unit 222 causes the fine-tuned large-scale language model M2 to generate new data different from the predetermined data based on the predetermined data input to the server 2. The predetermined data and the new data are, for example, text data. The generated new data is transmitted from the server 2 to the specification creation support device 3 via the network N.

[0030] The specification preparation support device 3 includes a storage unit 31, an input unit 32, a control unit 33, and a display unit 34.

[0031] The storage unit 31 may be configured, for example, by a storage device such as an HDD, SSD, or flash memory. The storage unit 31 stores training data and the like. The training data is training data in supervised learning, and is a data group configured by a combination of prompts and completions. A prompt is an input by a user to give a desired instruction to a large-scale language model. A completion is an output generated by the large-scale language model based on a prompt.

[0032] In this embodiment, the training data includes field-specific training data composed of patent documents in a specified technical field and training data composed of patent documents specified by the user. Patent documents include, for example, published patent gazettes, patent publications, and utility model registration publications. The specified technical field is, for example, a technical field identified by the International Patent Classification (IPC), FI (File Index), F-terms (File Forming Terms), etc. The IPC is a patent classification divided into several categories based on the technical content of patent documents and is an internationally standardized patent classification. The IPC is composed of sections, classes, subclasses, main groups, and subgroups. The FI is a patent classification that is more specific than the IPC and is used exclusively by the Japan Patent Office. The F-term is a patent classification based on the technical features of the inventions described in each patent document and is used exclusively by the Japan Patent Office. For example, when the technical field is identified by the International Patent Classification, the International Patent Classification used to identify the technical field may include all of the sections, classes, subclasses, main groups and subgroups, or may include only some of these.

[0033] The input unit 32 is configured to accept user input operations to the specification drafting support device 3 and to generate a request signal corresponding to the input operation. The input unit 32 is, for example, an input device such as a touch panel overlaid on the display unit 34 or a keyboard electrically connected to the specification drafting support device 3. The user inputs, for example, instructions for fine-tuning the fine-tuning unit 221 of the server 2 and invention text specifying the invention into the input unit 32. The invention text may be, for example, an invention disclosure statement, a claim statement specifying the invention, or a detailed independent claim statement formed by combining the invention-specifying matters of multiple claims. The invention text may also include a description of the effects achieved by the invention specified in the claim and a description of prior art literature explaining the background art.

[0034] When an instruction to cause the fine-tuning unit 221 of the server 2 to perform fine tuning is input to the input unit 32, the input unit 32 transmits instruction content information indicating the content of the instruction to the control unit 33. On the other hand, when an invention sentence is input to the input unit 32, the input unit 32 transmits the input invention sentence to the server 2 via the network N.

[0035] The control unit 33 may have the same hardware configuration as the control unit 22. That is, the control unit 33 may include a memory and a processor similar to those included in the control unit 22. The control unit 33 includes a fine adjustment instruction unit 331.

[0036] In this embodiment, a non-transitory computer-readable medium may be used as a memory. A non-transitory computer-readable medium refers to any type of physical memory (such as a RAM or a ROM) that can store information or data that can be read by a processor. A non-transitory computer-readable medium may store instructions for execution by one or more processors. The term "non-transitory computer-readable medium" includes tangible items and excludes carrier waves and transitory signals (i.e., non-transitory). Examples of non-transitory computer-readable media include magnetic recording media (e.g., flexible disks, magnetic tapes, hard disk drives), magneto-optical recording media (e.g., magneto-optical disks), CD-ROMs, CD-Rs, CD-R / Ws, and semiconductor memories (e.g., mask ROMs, programmable ROMs (PROMs), erasable PROMs (EPROMs), and flash ROMs).

[0037] The fine-tuning instruction unit 331 generates a transmission request signal for transmitting the learning data consisting of the patent document specified by the user to the server 2 via the network N based on the instruction content information received from the input unit 32. The fine-tuning instruction unit 331 transmits the generated transmission request signal to the storage unit 31.

[0038] At this time, the fine-tuning instruction unit 331 generates an instruction signal for causing the fine-tuning unit 221 of the server 2 to perform fine tuning based on the instruction content information received from the input unit 32. To this end, the storage unit 31 transmits training data composed of patent documents specified by the user, and the fine-tuning instruction unit 331 transmits an instruction signal for causing the fine-tuning unit 221 to perform fine tuning, to the fine-tuning unit 221. The fine-tuning unit 221 performs fine tuning based on the instruction signal received from the fine-tuning instruction unit 331 and the training data composed of patent documents specified by the user received from the storage unit 31. In this way, the fine-tuning instruction unit 331 issues an instruction to fine-tune predetermined parameters of the general-purpose large-scale language model M1 by additionally training the general-purpose large-scale language model M1 using the training data composed of patent documents specified by the user. As a result, the predetermined parameters of the general-purpose large-scale language model M1 are fine-tuned, and a fine-tuned large-scale language model M2 is created.

[0039] The display unit 34 is, for example, a touch screen display such as a liquid crystal display or an organic EL display. The display unit 34 outputs generated sentences generated by the fine-tuned large-scale language model M2 in response to the invention sentences input to the input unit 32. The generated sentences include a description including a description of a mode for carrying out the invention. Note that the generated sentences may include, in addition to the description, drawings based on the invention sentences input to the input unit 32. When the display unit 34 outputs the generated sentences generated by the fine-tuned large-scale language model M2, the generated sentences are displayed on the display unit 34.

[0040] The user terminal device 4 is a terminal device operated by a user of the specification creation support device 3. The user terminal device 4 is, for example, an electronic device such as a desktop PC, a notebook PC, a tablet terminal, or a smartphone. The user terminal device 4 is equipped with a display unit 41. The display unit 41 is, for example, a liquid crystal display. The display unit 41 is configured, for example, so that the user terminal device 4 displays generated text output from the display unit 34 of the specification creation support device 3.

[0041] (First embodiment) Next, a description will be given of a specification drafting support method executed in a first embodiment with reference to Figures 2 to 6. Figure 2 is a flowchart of the specification drafting support method according to the first embodiment. Figure 3 is a diagram showing an example of training data consisting of patent documents specified by a user. Figure 4 is a diagram for explaining fine tuning executed in the first embodiment. Figure 5 is a diagram showing an example of invention text input to the input unit 32 in the first embodiment. Figure 6 is a diagram showing an example of generated text displayed on the display unit 34 in the first embodiment.

[0042] In this embodiment, the general-purpose large-scale language model M1 is a language model based on a Transformer. In this embodiment, the storage unit 31 of the specification drafting support device 3 stores training data consisting of patent documents specified by a user. The patent documents specified by the user are unexamined patent publications, etc., in which at least one of the applicant, agent, and specification writer is the same. In this embodiment, the patent documents specified by the user are unexamined patent publications, etc., in which the applicant is the same. In this embodiment, the user is the applicant.

[0043] 2 , the user inputs an instruction to the input unit 32 to cause the fine-tuning unit 221 of the server 2 to perform fine tuning (STEP 01). When an instruction to cause the fine-tuning unit 221 of the server 2 to perform fine tuning is input to the input unit 32, the input unit 32 transmits instruction content information indicating the content of the instruction to the fine-tuning instruction unit 331.

[0044] Based on the instruction content information received from the input unit 32, the fine-tuning instruction unit 331 sends a transmission request signal to the storage unit 31 to transmit the learning data consisting of the patent document specified by the user to the server 2 via the network N (STEP 02).

[0045] In this embodiment, the training data consisting of patent documents specified by the user is the training data exemplified in Figure 3. As exemplified in Figure 3, the training data consisting of patent documents specified by the user is stored in the storage unit 31 in a state where claims (corresponding to the scope of claims in Figure 3) previously created by the user and specifications are linked together. In this embodiment, claims previously created by the user correspond to input data in supervised learning, and specifications previously created by the user correspond to correct answer data in supervised learning. Note that while Figure 3 shows only two patent documents, the number of patent documents specified by the user in this embodiment is, for example, 100.

[0046] 2 , the specification drafting support device 3 transmits to the fine-tuning unit 221 of the server 2 an instruction signal for causing the fine-tuning unit 221 to perform fine tuning, and training data consisting of patent documents specified by the user (STEP 03). The instruction signal for causing the fine-tuning unit 221 to perform fine tuning is transmitted from the fine-tuning instruction unit 331 to the fine-tuning unit 221. The training data consisting of the patent documents specified by the user is transmitted from the storage unit 31 to the fine-tuning unit 221. In this way, the fine-tuning instruction unit 331 issues an instruction to fine-tune predetermined parameters of the general-purpose large-scale language model M1 by causing the general-purpose large-scale language model M1 to perform additional training using the training data consisting of the patent documents specified by the user.

[0047] The fine-tuning unit 221 performs fine tuning based on the received instruction signal and learning data consisting of patent documents specified by the user. When fine-tuning is performed by the fine-tuning unit 221, predetermined parameters of the general-purpose large-scale language model M1 are fine-tuned, and a fine-tuned large-scale language model M2 is created (STEP 04). Note that since the general-purpose large-scale language model M1 is a language model based on a Transformer, the fine-tuned large-scale language model M2, which is based on the general-purpose large-scale language model M1, is also a language model based on a Transformer. The created fine-tuned large-scale language model M2 is transmitted to the storage unit 21 and stored therein.

[0048] Here, with reference to FIG. 4 , the fine-tuning performed by the fine-tuning unit 221 will be described in detail. As illustrated in FIG. 4 , the fine-tuned large-scale language model M2 uses the network structure of the general-purpose large-scale language model M1 as is. Furthermore, the fine-tuned large-scale language model M2 uses the parameters (weighting coefficients) set in the general-purpose large-scale language model M1 as initial values. In this embodiment, the fine-tuning unit 221 performs fine-tuning based on the received instruction signal, thereby fine-tuning at least some of the parameters (weighting coefficients) of the output layer in the entire neural network. At this time, a first additional layer (an example of an additional layer) is added between the intermediate layer and the output layer. Note that at this time, the parameters (weighting coefficients) of the input layer and intermediate layer in the entire neural network are not fine-tuned.

[0049] As shown in Fig. 2, the user inputs invention text into the input unit 32 (STEP 05). As shown in Fig. 5, in this embodiment, claim descriptions (claim 1 description and claim 2 description) are input into the input unit 32. Note that although Fig. 5 shows only two claims, the number of claims in this embodiment is not limited to two. The input unit 32 transmits the input invention text (claim descriptions) to the server 2 via the network N.

[0050] As illustrated in FIG. 2 , the generation instruction unit 222 of the server 2 causes the fine-tuned large-scale language model M2 to generate a generated sentence based on the invention sentence received from the input unit 32 (STEP 06). In this embodiment, the generation instruction unit 222 causes the fine-tuned large-scale language model M2 to generate a specification based on the claim recitation received from the input unit 32. Specifically, the fine-tuned large-scale language model M2 represents each word included in the claim recitation received from the input unit 32 as a multidimensional vector and analyzes the relationships between words and the structure of the sentence. Next, the fine-tuned large-scale language model M2 uses Attention to infer noteworthy words from among the words included in the claim recitation received from the input unit 32 and generates words one after another to generate a generated sentence such as a specification. The generated sentence is transmitted from the server 2 to the specification drafting support device 3 via the network N (STEP 07).

[0051] The generated text received from the server 2 is displayed on the display unit 34 of the specification drafting support device 3 (STEP 08). At this time, the display unit 34 displays the generated text shown in FIG. 6. As shown in FIG. 6, in this embodiment, the generated text displayed on the display unit 34 is a specification. The generated text includes a description of the title of the invention, a description of the technical field, a description of the background art, a description of prior art documents (not shown), a description of the problem to be solved by the invention (not shown), a description of the means for solving the problem (not shown), a description of the effects of the invention (not shown), a brief description of the drawings (not shown), and a description of the mode for carrying out the invention. For example, the description of the title of the invention is generated based on a description of the category of the invention in the claim (in this embodiment, the intellectual property rights support system). The description of the mode for carrying out the invention is also generated based on the claim, but because it is generated using the fine-tuned large-scale language model M2, it reflects user preferences such as sentence length, dependency, writing style, and the order of explanation of technical causal relationships. For example, the user's writing style in this embodiment is one that frequently uses terms such as "disclose" and "exemplify." The generated sentences shown in FIG. 6 reflect such user preferences.

[0052] However, it is difficult to prepare a large number of specifications previously created by a specific user. Therefore, it is difficult to create a large-scale language model from scratch that can accurately generate a specification (an example of a generated sentence) that reflects the user's preferences using only patent documents specified by the user. Furthermore, if a claim (an example of an invention sentence) is input into a general-purpose large-scale language model M1 to create a specification, the specification generated by the general-purpose large-scale language model M1 will not reflect the user's preferences. Therefore, the specification generated by the general-purpose large-scale language model M1 is likely to contain parts that the user finds strange, and may require revisions that are burdensome for the user.

[0053] The inventors then came to the conclusion that it might be possible to generate a specification that reflects the user's preferences by inputting the claim description into a fine-tuned large-scale language model M2, which was created by fine-tuning a general-purpose large-scale language model M1.

[0054] According to the specification drafting support device 3 having the above configuration, the general-purpose large-scale language model M1 is additionally trained using training data consisting of patent documents specified by the user, thereby fine-tuning predetermined parameters (output layer parameters (weight coefficients)) of the general-purpose large-scale language model M1. This allows the general-purpose large-scale language model M1 to learn the user's preferences contained in the patent documents of the training data. As a result, the generated sentences generated by the fine-tuned large-scale language model M2 for newly input invention sentences by the user reflect the user's preferences. This reduces the burden felt by the user when proofreading and correcting the generated sentences displayed on the display unit 34. Furthermore, the specification drafting support system, specification drafting support method, computer program, and non-transitory computer-readable medium having the above configuration can also achieve similar effects.

[0055] Furthermore, with the specification writing support device 3 having the above configuration, the invention text includes a claim that specifies the invention, and the generated text includes a description of a mode for carrying out the invention. The claim that specifies the invention and the description of the mode for carrying out the invention are both expressed using an expression method used in the patent field. Because the invention text includes a claim that specifies the invention, the generated text including the description of the mode for carrying out the invention is expressed with high accuracy using an expression method used in the patent field. Therefore, with the specification writing support device 3 having the above configuration, the accuracy of the generated text can be further improved from the user's perspective.

[0056] Second Embodiment Next, a description will be given of a specification drafting support method executed in a second embodiment with reference to FIGS. 3 and 7 to 9. FIG. 7 is a flowchart of the specification drafting support method according to the second embodiment. FIG. 8 is a diagram showing an example of field-specific learning data. FIG. 9 is a diagram for explaining fine tuning executed in the second embodiment. Note that in this embodiment, parts similar to those in the first embodiment will be described using the same reference numerals as in the first embodiment, and descriptions of overlapping parts will be omitted as appropriate. Also in this embodiment, the patent document specified by the user is a published patent publication or the like with the same applicant. Furthermore, also in this embodiment, the user is the applicant.

[0057] This embodiment differs from the first embodiment in that the storage unit 31 stores field-specific learning data composed of patent documents in a specified technical field and learning data composed of patent documents specified by the user, and in that the fine-tuning instruction unit 331 is configured to instruct fine-tuning in stages.

[0058] The storage unit 31 stores field-specific training data composed of patent documents in a predetermined technical field and training data composed of patent documents specified by the user. In this embodiment, the field-specific training data composed of patent documents in a predetermined technical field is a published patent application including the International Patent Classification G06Q50 / 18. Note that the International Patent Classification G06Q50 / 18 is a patent classification related to the legal profession and the handling of legal documents. Therefore, the field-specific training data in this embodiment is training data composed of patent documents in a technical field related to the legal profession and the handling of legal documents.

[0059] The fine-tuning instruction unit 331 is configured to instruct fine-tuning in stages. In this embodiment, the fine-tuning instruction unit 331 is configured to instruct fine-tuning in two stages. In the first stage, the fine-tuning instruction unit 331 issues an instruction to fine-tune the general-purpose large-scale language model M1 using field-specific training data composed of patent documents in a predetermined technical field. In this specification, the large-scale language model created by the first-stage fine-tuning is referred to as the intermediate large-scale language model M1A (see FIG. 9 ). In the second stage, the fine-tuning instruction unit 331 issues an instruction to further fine-tune the large-scale language model M1 (the intermediate large-scale language model M1A) that has been fine-tuned in the first stage using training data composed of patent documents specified by the user. As such, in this embodiment, the instruction signal generated by the fine-tuning instruction unit 331 includes an instruction signal to cause the fine-tuning unit 221 of the server 2 to perform the first-stage fine-tuning and an instruction signal to cause the fine-tuning unit 221 of the server 2 to perform the second-stage fine-tuning.

[0060] Next, a method for supporting specification preparation according to the second embodiment will be described with reference to Fig. 7. STEP 11 is the same as STEP 01 in the first embodiment.

[0061] The fine-tuning instruction unit 331 transmits a transmission request signal based on the instruction content information received from the input unit 32 (STEP 12). The transmission request signal is a signal for transmitting the learning data, which is composed of the field-specific learning data and the patent document specified by the user, from the learning data stored in the storage unit 31 to the server 2 via the network N.

[0062] In this embodiment, the field-specific training data is a published patent publication or the like including the International Patent Classification G06Q50 / 18. As illustrated in FIG. 8 , the field-specific training data is stored in the storage unit 31 in a state in which the claims (corresponding to the scope of the claims in FIG. 8 ) and the specifications are linked to the claims described in the published patent publications or the like. Therefore, unlike training data consisting of patent documents specified by the user, the field-specific training data is composed of patent documents including patent documents previously created by the user by a party other than the user. In this embodiment, the claims described in the published patent publications or the like correspond to input data in supervised learning, and the specifications described in the published patent publications or the like correspond to correct answer data in supervised learning. Note that while FIG. 8 illustrates only two patent documents, the number of patent documents in a specified technical field in this embodiment (in this embodiment, the technical field related to the International Patent Classification G06Q50 / 18) is, for example, 100,000. As such, the number of patent documents in a specified technical field is significantly greater than the number of patent documents specified by the user. The learning data consisting of patent documents designated by the user and transmitted by the storage unit 31 to the fine-tuning unit 221 of the server 2 is the learning data exemplified in FIG. 3, as in the first embodiment.

[0063] 7 , the specification preparation support device 3 transmits an instruction signal to the fine-tuning unit 221 of the server 2 to cause the fine-tuning unit 221 to perform a first-stage fine-tuning, and field-specific learning data to the fine-tuning unit 221 of the server 2 (STEP 13). The instruction signal to cause the fine-tuning unit 221 of the server 2 to perform the first-stage fine-tuning is transmitted from the fine-tuning instruction unit 331 to the fine-tuning unit 221. The field-specific learning data is transmitted from the storage unit 31 to the fine-tuning unit 221.

[0064] The fine-tuning unit 221 performs a first stage of fine-tuning based on the received instruction signal and the domain-specific training data (STEP 14). The fine-tuning unit 221 performs the first stage of fine-tuning based on the domain-specific training data. As a result, predetermined parameters (output layer parameters (weighting coefficients)) of the general-purpose large-scale language model M1 are fine-tuned, and an intermediate large-scale language model M1A is created.

[0065] Here, with reference to FIG. 9 , the fine-tuning performed by the fine-tuning unit 221 in the first stage will be described in detail. As illustrated in FIG. 9 , the intermediate large-scale language model M1A uses the network structure of the general-purpose large-scale language model M1 as is. Furthermore, the intermediate large-scale language model M1A uses the parameters (weighting coefficients) set in the general-purpose large-scale language model M1 as initial values. In this embodiment, the fine-tuning unit 221 performs the first stage of fine-tuning based on the received instruction signal, thereby fine-tuning at least some of the parameters (weighting coefficients) of the output layer in the entire neural network. At this time, a first additional layer (an example of an additional layer) is added between the intermediate layer and the output layer. Note that in this specification, the parameters (weighting coefficients) of the first additional layer and the output layer fine-tuned in the first stage are referred to as intermediate parameters (intermediate weighting coefficients). At this time, the parameters (weighting coefficients) of the input layer and intermediate layer in the entire neural network are not fine-tuned.

[0066] 7 , the specification drafting support device 3 then transmits to the fine-tuning unit 221 of the server 2 an instruction signal for causing the fine-tuning unit 221 of the server 2 to perform the second-stage fine-tuning and learning data consisting of the patent documents specified by the user (STEP 15). The instruction signal for causing the fine-tuning unit 221 of the server 2 to perform the second-stage fine-tuning is transmitted from the fine-tuning instruction unit 331 to the fine-tuning unit 221. The learning data consisting of the patent documents specified by the user is transmitted from the storage unit 31 to the fine-tuning unit 221.

[0067] The fine-tuning unit 221 performs a second stage of fine-tuning based on the received instruction signal and the training data consisting of patent documents specified by the user (STEP 16). The fine-tuning unit 221 performs a second stage of fine-tuning based on the training data consisting of patent documents specified by the user. As a result, predetermined parameters of the intermediate large-scale language model M1A (at least some intermediate parameters (intermediate weighting coefficients) of the output layer) are fine-tuned, and a fine-tuned large-scale language model M2 (see FIG. 9) is created (STEP 17). The created fine-tuned large-scale language model M2 is transmitted to the storage unit 21 and stored therein.

[0068] Here, with reference to FIG. 9 , the fine-tuning performed by the fine-tuning unit 221 in the second stage will be described in detail. As illustrated in FIG. 9 , the fine-tuned large-scale language model M2 directly utilizes the network structures of the general-purpose large-scale language model M1 and the intermediate large-scale language model M1A. Furthermore, the fine-tuned large-scale language model M2 utilizes the parameters (weighting coefficients) in the input layer and the intermediate layer set in the intermediate large-scale language model M1A, and the intermediate parameters (intermediate weighting coefficients) in the first additional layer and the output layer. In this embodiment, the fine-tuning unit 221 performs the second stage of fine-tuning based on the received instruction signal, thereby fine-tuning at least some of the intermediate parameters (intermediate weighting coefficients) in the output layer of the entire neural network. At this time, a second additional layer (an example of an additional layer) is added between the first additional layer and the output layer. Note that at this time, the parameters (weighting coefficients) in the input layer and the intermediate layer of the entire neural network are not fine-tuned.

[0069] STEP 18 to STEP 21 are the same as STEP 05 to STEP 08 in the first embodiment.

[0070] In this embodiment, the specification preparation support device 3 can also achieve the same effects as in the first embodiment.

[0071] Furthermore, while the cost of preparing training data is generally a real issue when performing additional training, the specification drafting assistance device 3 configured as described above uses field-specific training data composed of patent documents in a specific technical field in the first stage of fine-tuning. This field-specific training data can be identified using the International Patent Classification or the like, and it is relatively easy to obtain a certain amount of training data, thereby reducing the cost of collecting training data. Furthermore, because fine-tuning is performed using field-specific training data in the first stage, the amount of training data used in additional training in the second stage can be kept relatively small. In this way, by instructing fine-tuning in stages on the general-purpose large-scale language model M1, it becomes easy to adjust the balance between the cost of collecting training data and the accuracy of the learning results.

[0072] The above-described embodiments are provided to facilitate understanding of the present disclosure, and are not intended to limit the present disclosure, which may be modified or improved without departing from the spirit and scope of the present disclosure.

[0073] In the above embodiment, the description of the claims is input to the input unit 32, but a description of an invention disclosure document or the like may also be input. For example, when a description of an invention disclosure document is input to the input unit 32, the description of the invention disclosure document as shown in Fig. 10 is input to the input unit 32. In this way, the invention text input to the input unit 32 is not limited to the description of the claims, so the user can use the specification drafting support device 3 even when no claims have been created.

[0074] In the above embodiment, multiple claims, including so-called independent claims and dependent claims, are input to the input unit 32. In this case, the control unit 33 may generate a single detailed independent claim by combining the invention-specifying features of the multiple claims input to the input unit 32. For example, when the claim description illustrated in FIG. 5 is input to the input unit 32, the control unit 33 generates the detailed independent claim illustrated in FIG. 11 based on the claim description illustrated in FIG. 5. Note that the detailed independent claim is a so-called picture claim. A picture claim is a claim that directly expresses the configuration of an embodiment or example. Because the detailed independent claim combines the invention-specifying features of multiple claims, the generated sentences generated by the large-scale language model M2 fine-tuned based on the detailed independent claim are more accurate than the generated sentences generated by the large-scale language model M2 fine-tuned based on a general independent claim description.

[0075] In the above embodiment, only the description of the claim is input to the input unit 32, but for example, a description of the effect achieved by the invention specified in the claim may also be input in addition to the description of the claim, as illustrated in Fig. 12. This allows the fine-tuned large-scale language model M2 to generate generated sentences that also take into account the effect achieved by the invention specified in the claim, thereby improving the accuracy of the generated sentences generated by the fine-tuned large-scale language model M2.

[0076] In the above embodiment, only the claims are input to the input unit 32, but for example, as illustrated in Fig. 13, a description of prior art documents explaining the background art may also be input in addition to the claims. This allows the fine-tuned large-scale language model M2 to generate generated sentences that also take into account the prior art documents explaining the background art, thereby improving the accuracy of the generated sentences generated by the fine-tuned large-scale language model M2.

[0077] In the above embodiment, the generated text generated by the fine-tuned large-scale language model M2 is a specification, but for example, as illustrated in Figure 14, the generated text generated by the fine-tuned large-scale language model M2 may be a specification and drawings. Note that the drawings are generated by the fine-tuned large-scale language model M2 based on the invention text input to the input unit 32. In the example shown in Figure 14, the fine-tuned large-scale language model M2 generates drawings including a block diagram based on the invention text input to the input unit 32.

[0078] In the above embodiment, the generated sentences generated by the fine-tuned large-scale language model M2 are displayed on the display unit 34, but they may also be displayed on the display unit 41 of the user terminal device 4, for example. In this case, the display unit 34 outputs the generated sentences to the user terminal device 4.

[0079] In the above embodiment, the specification preparation support system 1 includes the user terminal device 4, but it does not necessarily have to include the user terminal device 4.

[0080] In the second embodiment, the field-specific learning data composed of patent documents in a specified technical field is a published patent publication or the like including the International Patent Classification G06Q50 / 18, but may also be, for example, a published patent publication or the like including a desired FI or F-term (File Forming Term).

[0081] The international patent classification used in the second embodiment includes all of the sections, classes, subclasses, main groups, and subgroups, but may include only some of these.

[0082] In the above embodiment, at least some of the parameters (weighting coefficients) of the output layer are fine-tuned, but for example, at least some of the parameters (weighting coefficients) of the intermediate layer and additional layer may also be fine-tuned.

[0083] In the above embodiment, an additional layer (first additional layer or second additional layer) may be added after the output layer.

[0084] In the above embodiment, the patent document designated by the user is a patent publication or the like with the same applicant, but it may also be, for example, a patent publication or the like with the same agent.

[0085] In the above embodiment, the user is the applicant of the patent document designated by the user, but may be a person different from the applicant.

[0086] Hereinafter, an example of another embodiment of the present disclosure will be described with reference to the drawings.

[0087] (Overall Configuration) A specification creation support system 1001 according to this embodiment will be described with reference to FIG. 15 . FIG. 15 is a schematic diagram illustrating the specification creation support system 1001 according to this embodiment. As illustrated in FIG. 15 , the specification creation support system 1001 includes a large-scale language model 1100, a specification creation support device 1200, and a user terminal device 1300. The large-scale language model 1100 and the specification creation support device 1200 can communicate with each other via a network, for example. The network is, for example, a cloud network. A cloud network is a computer network that exists within a cloud computing infrastructure. The specification creation support device 1200 and the user terminal device 1300 can communicate with each other via wired or wireless communication. The specification creation support device 1200 can communicate with multiple user terminal devices 1300.

[0088] The large-scale language model 1100 has previously learned a large amount of text data and is capable of solving various language processing tasks. Examples of language processing tasks include sentence generation and question answering. For example, when generating sentences, the large-scale language model 1100 tokenizes the input sentence and converts it into a continuous numerical representation (multidimensional vector). Next, an encoder of the large-scale language model 1100 analyzes the converted numerical representation and converts it into an intermediate representation. Then, a decoder of the large-scale language model 1100 converts the intermediate representation back into a numerical representation, converts the tokens back into sentences, and outputs the resulting representation. In this way, the large-scale language model 1100 generates sentences based on the input sentences.

[0089] In this embodiment, the large-scale language model 1100 generates reference sentences that can be used in the specification of the invention based on prompts received from the specification drafting support device 1200. A prompt is an input sentence that a user uses to give desired instructions to the large-scale language model 1100. Note that the reference sentences generated by the large-scale language model 1100 based on the prompts correspond to completions. The large-scale language model 1100 transmits the generated reference sentences to the specification drafting support device 1200.

[0090] The specification drafting support device 1200 generates a prompt and transmits the generated prompt to the large-scale language model 1100. The method of generating a prompt performed by the specification drafting support device 1200 will be described later.

[0091] The user terminal device 1300 is a terminal device operated by a user of the specification creation support device 1200. The user terminal device 1300 is, for example, an electronic device such as a desktop PC, a notebook PC, a tablet terminal, a smartphone, etc. The user terminal device 1300 includes an input unit 1301 and a display unit 1302.

[0092] The input unit 1301 is configured to, for example, accept a user's input operation to the user terminal device 1300 and generate an operation signal corresponding to the input operation. The input unit 1301 is, for example, an input device such as a touch panel arranged overlaid on the display unit 1302 or a keyboard electrically connected to the user terminal device 1300.

[0093] A user inputs, for example, invention information into the input unit 1301. The invention information is information for identifying an invention. The invention information includes, for example, claims that identify the invention, information on the effects of the invention corresponding to the claims, an invention disclosure document, drawings, etc. The drawings may include image data of the drawings, drawing information in which drawings depicting embodiments of the invention are converted into text, etc. The drawing information includes text explaining the contents of the drawings and symbolized configurations. The symbolized configurations are information including symbols included in the drawings and the configurations indicated by the symbols, and correspond to the "symbol explanations" in the specification. Note that each symbol and the configuration indicated by the symbol are linked in a one-to-one correspondence. The input unit 1301 transmits the input invention information and an operation signal to the specification drafting support device 1200. Note that the operation signal transmitted by the input unit 1301 may be, for example, a signal for causing the specification drafting support device 1200 to generate drawing information, a signal for causing the specification drafting support device 1200 to design a prompt and then causing the large-scale language model 1100 to generate reference text based on the prompt, etc.

[0094] The display unit 1302 is, for example, a liquid crystal display. The display unit 1302 is configured to display, for example, an operation screen based on operation screen data described below. Note that the operation screen may display reference sentences generated by the large-scale language model 1100.

[0095] Fig. 16 is a block diagram of a specification creation support device 1200 according to this embodiment. As illustrated in Fig. 16, the specification creation support device 1200 includes an input receiving unit 1210, an operation screen data generating unit 1220, a transmitting / receiving unit 1230, a storage unit 1240, a drawing information generating unit 1250, and a prompt generating unit 1260. These units are connected to each other via a bus 1270 so as to be able to communicate with each other.

[0096] The input accepting unit 1210 receives as input the invention information transmitted from the user terminal device 1300. The input accepting unit 1210 transmits the input invention information to the storage unit 1240. If the invention information input to the input accepting unit 1210 includes a drawing, the input accepting unit 1210 transmits the drawing input to the input accepting unit 1210 to the storage unit 1240 and the drawing information generating unit 1250.

[0097] The operation screen data generating unit 1220 generates operation screen data to be displayed on the user terminal device 1300. The operation screen data generating unit 1220 transmits the generated operation screen data to the transmitting / receiving unit 1230. The operation screen data may include invention information transmitted from the user terminal device 1300 and reference sentences generated by the large-scale language model 1100. For example, when new reference sentences are added to the operation screen data, this specification will express this as "reflecting the reference sentences in the operation screen data."

[0098] The transmitting / receiving unit 1230 is configured to transmit prompts generated by the prompt generation unit 1260 to the large-scale language model 1100. The transmitting / receiving unit 1230 is configured to transmit operation screen data to the user terminal device 1300. The transmitting / receiving unit 1230 is configured to transmit operation signals from the user terminal device 1300 to the drawing information generation unit 1250 and the prompt generation unit 1260. The transmitting / receiving unit 1230 is configured to receive, for example, various types of information and operation signals transmitted from the user terminal device 1300, and reference sentences and the like transmitted from the large-scale language model 1100.

[0099] The storage unit 1240 may be configured with a storage device (storage) such as an HDD, SSD, or flash memory. The storage unit 1240 stores, for example, invention information received from the input receiving unit 1210. The invention information is stored in the storage unit 1240 in, for example, itemized format, for the purpose of improving the accuracy of completions in response to prompts. In this embodiment, the storage unit 1240 stores invention information such as that shown in FIG. 17 . As shown in FIG. 17 , the storage unit 1240 stores, as invention information, sentences explaining the contents of drawings for each drawing. Furthermore, the storage unit 1240 also stores, as invention information, sentences explaining the contents indicated by symbols.

[0100] Returning to Fig. 16, the drawing information generation unit 1250 will be described. The drawing information generation unit 1250 has a hardware configuration including a memory and a processor. The memory is composed of, for example, a ROM storing various computer programs and a RAM having multiple work areas storing various computer programs executed by the processor. The processor is, for example, a CPU, and is configured to load a computer program specified by the various computer programs stored in the ROM onto the RAM and execute various processes in cooperation with the RAM.

[0101] The drawing information generation unit 1250 generates drawing information by converting the drawing input to the input acceptance unit 1210 into text based on the operation signal received from the transmission / reception unit 1230. In this case, the drawing input to the input acceptance unit 1210 is converted into text in the form of an itemized list. The drawing information generation unit 1250, for example, reads the drawing input to the input acceptance unit 1210 and converts the drawing into text based on a pre-set rule. The pre-set rule is, for example, the rule shown in FIG. 18 . For example, if the drawing input to the input acceptance unit 1210 is a sequence diagram, the drawing information generation unit 1250 converts the sequence diagram input to the input acceptance unit 1210 into text based on the rule, "Use the sequence diagram to explain data exchange between each device." The drawing information generated by the drawing information generation unit 1250 is transmitted to the storage unit 1240 and stored therein.

[0102] For example, the drawing information generation unit 1250 may prepare a trained model in advance that has learned patterns of combinations of drawing types and drawing descriptions, and use the trained model to convert drawings into text. Examples of drawing types include system configuration diagrams, block diagrams, data configuration diagrams, sequence diagrams, flowcharts, and diagrams showing display screens. A trained model is a machine learning model that has been trained in advance using a large dataset. In this case, the trained model prepared in advance may be, for example, a deep learning model constructed using a neural network. In this case, the drawing information generated by the drawing information generation unit 1250 is also transmitted to the storage unit 1240 and stored therein.

[0103] The drawing information generating unit 1250 can generate a referenced configuration from the drawing input to the input receiving unit 1210 based on the operation signal received from the transmitting / receiving unit 1230. The drawing information generating unit 1250 performs image analysis of the drawing input to the input receiving unit 1210 and generates a referenced configuration by linking the symbols included in the drawing with the configuration indicated by the symbols.

[0104] Returning to FIG. 16 , the prompt generation unit 1260 will be described. The prompt generation unit 1260 may have the same hardware configuration as the drawing information generation unit 1250, for example. The prompt generation unit 1260 acquires invention information stored in the storage unit 1240 based on an operation signal received from the transmission / reception unit 1230. The prompt generation unit 1260 generates a prompt, based on the invention information acquired from the storage unit 1240, that causes the large-scale language model 1100 to generate reference sentences that can be used in the specification of the invention. The prompt generation unit 1260 includes a design unit 1261 and an instruction sentence generation unit 1262.

[0105] The design unit 1261 is generated by the prompt generation unit 1260 based on the invention information. The design unit 1261 can design the content of the prompt. The content of the prompt is designed based on the drawing corresponding to the drawing number selected by the user and the claim corresponding to the item number selected by the user. How the user selects the drawing number, the claim item number, etc. will be described later. The instruction statement generation unit 1262 generates a prompt to be input to the large-scale language model 1100 based on the content of the prompt designed by the design unit 1261. The design unit 1261 includes a drawing selection unit 2611, a claim selection unit 2612, a signed configuration selection unit 2613, and a free input unit 2614.

[0106] The drawing selection unit 2611 identifies a drawing stored in the storage unit 1240 that corresponds to the drawing number selected by the user, based on an operation signal received from the transmission / reception unit 1230. The drawing selection unit 2611 acquires drawing information related to the identified drawing from the storage unit 1240. In this manner, the design unit 1261 can select drawing information on a drawing number basis. If the user selects multiple drawing numbers, the drawing selection unit 2611 identifies multiple drawings corresponding to the multiple drawing numbers selected by the user, based on the operation signal. Then, the drawing selection unit 2611 acquires drawing information related to the identified multiple drawings for each drawing number from the storage unit 1240. The instruction statement generation unit 1262 generates a prompt for each drawing number based on the drawing information acquired from the storage unit 1240. In this manner, if the user selects multiple drawing numbers, the prompt generation unit 1260 generates a prompt for each drawing number.

[0107] The claim selection unit 2612 identifies a claim stored in the storage unit 1240 that corresponds to the item number selected by the user, based on an operation signal received from the transmission / reception unit 1230. The claim selection unit 2612 retrieves the identified claim from the storage unit 1240. In this manner, the design unit 1261 can select claims by item number. If the user selects multiple item numbers, the drawing selection unit 2611 retrieves the identified claims for each item number from the storage unit 1240. The instruction statement generation unit 1262 generates a prompt for each item number based on the claims retrieved from the storage unit 1240. In this manner, if the user selects multiple item numbers, the prompt generation unit 1260 generates a prompt for each item number.

[0108] The signed configuration selection unit 2613 identifies a signed configuration selected by the user from among the signed configurations stored in the storage unit 1240, based on the operation signal received from the transmission / reception unit 1230. The signed configuration selection unit 2613 acquires the identified signed configuration from the storage unit 1240.

[0109] Character information can be input into the free input section 2614. Note that the character information is text data indicating, for example, definitions of terms included in claims, follow-up questions for the large-scale language model 1100, problems that the invention is intended to solve, effects of the invention, etc. When character information is input into the free input section 2614, the instruction sentence generation section 1262 combines the invention information selected by the design section 1261 with the character information input into the free input section 2614 to generate a prompt.

[0110] (Overview of a series of processes until a reference sentence is displayed on the display unit of a user terminal device) Next, a series of processes until a reference sentence generated by the large-scale language model 1100 is displayed on the display unit 1302 of the user terminal device 1300 will be outlined with reference to Fig. 19. Fig. 19 is a sequence diagram illustrating a series of processes until a reference sentence generated by the large-scale language model 1100 is displayed on the display unit 1302 of the user terminal device 1300.

[0111] First, the user performs an input operation on the input unit 1301 of the user terminal device 1300 to upload claims and drawings to the specification drafting support device 1200 (STEP 1001). That is, the user inputs invention information into the input unit 1301. When the user performs an input operation on the input unit 1301 to upload claims and drawings to the specification drafting support device 1200, the claims and drawings are uploaded to the specification drafting support device 1200. As a result, the claims and drawings input into the input unit 1301 are transmitted to the storage unit 1240 of the specification drafting support device 1200. Note that in this embodiment, the drawing information generation unit 1250 of the specification drafting support device 1200 generates symbol-added configurations based on the drawings input into the input unit 1301 of the user terminal device 1300 and transmits the generated symbol-added configurations to the storage unit 1240. Therefore, in this embodiment, the claims, drawings, and symbol-added configurations are stored in the storage unit 1240.

[0112] The specification drafting support device 1200 reflects the uploaded claims and drawings and the generated symbolized configuration in operation screen data, and transmits the operation screen data to the user terminal device 1300 (STEP 1002). At this time, the display unit 1302 of the user terminal device 1300 displays an operation screen based on the operation screen data including the claims and drawings uploaded to the specification drafting support device 1200 and the symbolized configuration generated by the specification drafting support device 1200.

[0113] When the user inputs an operation signal to the prompt generating unit 1260 into the input unit 1301 of the user terminal device 1300, the operation signal is transmitted to the specification preparation support device 1200 (STEP 1003).

[0114] When the specification creation support device 1200 receives an operation signal from the user terminal device 1300, it generates a prompt and transmits the generated prompt to the large-scale language model 1100 (STEP 1004).

[0115] When the large-scale language model 1100 receives the prompt from the specification creation support device 1200, it generates a reference sentence and transmits the generated reference sentence to the specification creation support device 1200. The specification creation support device 1200 receives the reference sentence generated by the large-scale language model 1100 (STEP 1005).

[0116] When the specification drafting support device 1200 receives the reference sentence from the large-scale language model 1100, it reflects the received reference sentence in the operation screen data. The specification drafting support device 1200 transmits the operation screen data reflecting the reference sentence received from the large-scale language model 1100 to the user terminal device 1300 (STEP 1006). At this time, the display unit 1302 of the user terminal device 1300 displays an operation screen based on the operation screen data including the above-mentioned claims, drawings, and configuration with symbols, and the reference sentence received from the large-scale language model 1100.

[0117] (Detailed Description from Uploading Claims and Drawings to Generating Design Section) Next, with reference to Figures 20 to 22, a description will be given of a specification drafting support method executed by the specification drafting support device 1200 from uploading claims and drawings to generating the design section 1261. Figure 20 is a flowchart illustrating an example of the flow from uploading claims and drawings to generating the design section 1261. Note that before STEP 1011 illustrated in Figure 20 is executed, the display unit 1302 of the user terminal device 1300 displays the operation screen illustrated in Figure 21. In other words, before STEP 1011 illustrated in Figure 20 is executed, no invention information is displayed on the operation screen displayed on the display unit 1302.

[0118] 20 , the user performs an input operation on the input unit 1301 of the user terminal device 1300 to upload claims and drawings to the specification drafting support device 1200 (STEP 1011). Note that the drawings input to the input unit 1301 contain image data of the drawings but do not contain drawing information. The claims and drawings input to the input unit 1301 are stored in the storage unit 1240.

[0119] The drawing information generating unit 1250 of the specification drafting support device 1200 converts the uploaded drawings into text to generate drawing information (STEP 1012). The drawing information generating unit 1250 generates drawing information including a reference numeral configuration for each uploaded drawing. The generated drawing information is stored in the storage unit 1240.

[0120] The prompt generation unit 1260 of the specification drafting support device 1200 generates a design unit 1261 based on the invention information (in this embodiment, claims and drawings) (STEP 1013).

[0121] The specification drafting support device 1200 reflects the uploaded claims and drawings and the generated reference numerals in the operation screen data (STEP 1014). The operation screen data is transmitted to the user terminal device 1300, and as a result, the operation screen shown in Fig. 22 is displayed on the display unit 1302 of the user terminal device 1300. As shown in Fig. 22, the uploaded claims and drawings (image data) and the reference numerals are displayed on the display unit 1302.

[0122] (Detailed Description of the Process from Creation of the Design Section to Display of the Reference Text) Next, with reference to Figures 23 and 25, a description will be given of the specification creation support method executed by the specification creation support device 1200 from creation of the design section 1261 to display of the reference text. Figure 23 is a flowchart illustrating the process from uploading claims and drawings to creation of the design section 1261.

[0123] The user performs an input operation on the input unit 1301 of the user terminal device 1300 to have the specification drafting support device 1200 design a prompt and then have the large-scale language model 1100 generate a reference sentence based on the prompt (STEP 1021). In this embodiment, the user selects FIG. 1, claim 1, and the reference sentence "30 User's Terminal," as indicated by vertical hatching in FIG. 24 , and then selects the "Generate Reference Sentence" button. This selection by the user is an input operation on the input unit 1301 of the user terminal device 1300 in this embodiment. When this input operation is performed on the input unit 1301, the input unit 1301 transmits an operation signal based on the input operation to the specification drafting support device 1200.

[0124] Returning to FIG. 23 , STEP 1022 will be described. The prompt generation unit 1260 of the specification drafting support device 1200 generates a prompt based on an operation signal received from the user terminal device 1300 (STEP 1022). In this embodiment, FIG. 1, claim 1, and the numbered configuration "30 user's terminal" are selected. Therefore, the drawing selection unit 2611 acquires drawing information related to FIG. 1 from the storage unit 1240. The claim selection unit 2612 acquires the claim related to claim 1 from the storage unit 1240. The numbered configuration selection unit 2613 acquires the numbered configuration related to "30 user's terminal" from the storage unit 1240. The instruction statement generation unit 1262 generates a prompt based on the information acquired by the drawing selection unit 2611, the claim selection unit 2612, and the numbered configuration selection unit 2613. In this embodiment, the instruction sentence generation unit 1262 generates a prompt to cause the large-scale language model 1100 to generate a reference sentence that serves as a reference when creating a sentence that explains the user terminal linked to the reference symbol 30. The instruction sentence generation unit 1262 transmits the generated prompt to the transmission / reception unit 1230.

[0125] The transmitting / receiving unit 1230 transmits the prompt received from the instruction sentence generating unit 1262 to the large-scale language model 1100 (STEP 1023).

[0126] The large-scale language model 1100 generates reference sentences based on the prompts received from the specification creation support device 1200, and transmits the generated reference sentences to the specification creation support device 1200. As a result, the specification creation support device 1200 receives the reference sentences generated by the large-scale language model 1100 (STEP 1024).

[0127] Upon receiving the reference sentence from the large-scale language model 1100, the specification creation support device 1200 reflects the received reference sentence in the operation screen data (STEP 1025). The specification creation support device 1200 transmits the operation screen data reflecting the reference sentence received from the large-scale language model 1100 to the user terminal device 1300. As a result, the operation screen shown in FIG. 25 is displayed on the display unit 1302 of the user terminal device 1300.

[0128] Region R1 located at the bottom right of the operation screen shown in Figure 25 displays reference text generated by the large-scale language model 1100. Region R2 located to the left of region R1 displays text data related to the specification being created or edited by the user. In this way, the reference text and the text data related to the specification are displayed in adjacent regions, which shortens the distance the user's gaze needs to move when creating or editing a specification while referring to the reference text. This allows the user to create or edit a specification efficiently.

[0129] However, the text generated by conventional specification preparation support software contains text that appears strange to the user, and requires correction work that is burdensome for the user, resulting in low convenience.

[0130] The specification drafting support device 1200 having the above configuration can cause the large-scale language model 1100 to generate reference sentences in response to prompts generated by the prompt generation unit 1260 based on invention information. The reference sentences generated by the large-scale language model 1100 based on prompts generated by the prompt generation unit 1260 differ from conventional reference sentences that are generated using similarities calculated by comparing a certain number of multiple sentences. These reference sentences are well-connected word-by-word and have consistency between the sentences. This makes it easier for users to use high-quality reference sentences when drafting specifications, reducing the burden of revision work and improving convenience. Furthermore, the specification drafting support method and computer program having the above configuration can also achieve similar effects.

[0131] Furthermore, according to the specification creation support device 1200 having the above configuration, the user can select the drawing information of the invention information on a drawing-by-drawing basis, and therefore the prompt generation unit 1260 can easily generate detailed and accurate prompts.

[0132] Furthermore, with the specification creation support device 1200 having the above configuration, the user can select claims by item number, and therefore can easily cause the prompt generation unit 1260 to generate detailed and accurate prompts.

[0133] Furthermore, with the specification creation support device 1200 having the above configuration, the drawing information generation unit 1250 converts drawings into text and generates drawing information, eliminating the need for the user to convert drawings into text. Therefore, with the specification creation support device 1200, the burden on the user is reduced and convenience for the user is improved.

[0134] (Generation of Reference Texts Regarding Modified Examples and Specific Examples) Next, with reference to FIG. 26 , the function of the specification drafting support device 1200 for generating reference texts regarding modified examples and specific examples will be described. FIG. 26 is a diagram showing an example of an operation screen when generating reference texts regarding modified examples and specific examples. Note that, even in this case, the input operations performed by the user on the input unit 1301 of the user terminal device 1300 and the processing executed by the specification drafting support device 1200 are basically the same as the input operations and processing described above, and therefore, overlapping portions will be omitted as appropriate. Also in this case, it is assumed that the above-described invention information is stored in the memory unit 1240 of the specification drafting support device 1200. Furthermore, even in this case, it is assumed that the user selects FIG. 1 , claim 1, and the referenced configuration "30 User's Terminal." Note that, in the following description, the above assumptions apply unless otherwise specified.

[0135] When a user wants the large-scale language model 1100 to generate reference sentences related to variations and specific examples, the user selects FIG. 1, claim 1, and the referenced configuration "30 User's Terminal" as indicated by vertical hatching in Fig. 26, and then selects the "Variations and specific examples" button. Note that this selection by the user is an input operation to the input unit 1301 of the user terminal device 1300 in this case. When this input operation is performed on the input unit 1301, the input unit 1301 transmits an operation signal based on this input operation to the specification drafting support device 1200.

[0136] The instruction sentence generation unit 1262 generates a prompt for causing the large-scale language model 1100 to generate reference sentences regarding variations and specific examples related to the user's terminal linked to reference numeral 30 and the input device and output device provided in the user's terminal, based on an operation signal received from the input unit 1301 of the user terminal device 1300. The instruction sentence generation unit 1262 transmits the generated prompt to the transmission / reception unit 1230.

[0137] The large-scale language model 1100 generates reference sentences based on the prompts received from the specification creation support device 1200, and transmits the generated reference sentences to the specification creation support device 1200. Upon receiving the reference sentences from the large-scale language model 1100, the specification creation support device 1200 reflects the received reference sentences in operation screen data. The specification creation support device 1200 transmits operation screen data reflecting the reference sentences received from the large-scale language model 1100 to the user terminal device 1300. As a result, the operation screen shown in FIG. 26 is displayed on the display unit 1302 of the user terminal device 1300.

[0138] 26 , region R1 displays reference text relating to variations and specific examples relating to the user's terminal linked to reference numeral 30 and the input device and output device provided on the user's terminal. In this way, according to specification creation support device 1200 having the above configuration, when a user wants to include variations and specific examples relating to an arbitrary component in a specification, the user can efficiently create a specification by referring to the reference text relating to the variations and specific examples displayed in region R1.

[0139] (First Example of Generating Reference Sentences Using Free Input Function) Next, a first example of generating reference sentences using the free input function will be described with reference to Figs. 27 to 29. Fig. 27 is a diagram showing an example of an operation screen on which no reference sentence is displayed in the first example of generating reference sentences using the free input function. Fig. 28 is a diagram showing an example of character information. Fig. 29 is a diagram showing an example of an operation screen on which the reference sentence is also displayed in the first example of generating reference sentences using the free input function. Note that in this example as well, the specification creation support device 1200 executes the specification creation support method exemplified in Fig. 23.

[0140] In this example, the user selects FIG. 1, claim 1, and the numbered configuration "10. Recommended Filing Date Providing System," as indicated by vertical hatching in FIG. 27, and then selects the "Free Input Field" button. Note that this selection by the user is an input operation into the input unit 1301 of the user terminal device 1300 in this example (STEP 1021 in FIG. 23). Also, in this example, character information as shown in FIG. 28 is input into the "Free Input Field." When this input operation is performed into the input unit 1301, the input unit 1301 transmits an operation signal based on the input operation and the character information as shown in FIG. 28 to the specification drafting support device 1200. Note that the character information as shown in FIG. 28 is input into the free input unit 2614.

[0141] The prompt generation unit 1260 of the specification drafting support device 1200 generates a prompt by combining the invention information selected by the design unit 1261 with the character information entered in the free input unit 2614 based on an operation signal received from the user terminal device 1300 (STEP 1022 in FIG. 23 ). Specifically, the instruction sentence generation unit 1262 of the prompt generation unit 1260 generates a prompt by combining the invention information including the coded configuration of FIG. 1, claim 1, and "10. System for providing recommended filing date." with the character information exemplified in FIG. 28 . The instruction sentence generation unit 1262 transmits the generated prompt to the transmission / reception unit 1230.

[0142] Steps 1023 to 1025 in this example are the same as those described above. However, the display unit 1302 of the user terminal device 1300 displays an operation screen such as that shown in Fig. 29. As shown in Fig. 29, region R1 displays a reference sentence that is a completion for the prompt generated by combining the invention information selected in the design unit 1261 and the character information entered in the free input section 2614.

[0143] In this example as well, the specification preparation support device 1200 can achieve the effects described above.

[0144] Furthermore, the specification creation support device 1200 having the above configuration can generate a prompt by combining the invention information selected in the design unit 1261 with the character information entered in the free input unit 2614. This allows the user to easily cause the prompt generation unit 1260 to generate a flexible prompt.

[0145] (Second Example of Generating Reference Sentences Using Free Input Function) Next, a second example of generating reference sentences using the free input function will be described with reference to Fig. 30. Fig. 30 is a diagram showing an example of an operation screen in the second example of generating reference sentences using the free input function.

[0146] This example differs from the first example of generating reference text using the free input function in that text information such as "Please explain the recommended filing date with reference to claim 1 and Figure 1" is input into the free input section 2614 of the specification drafting support device 1200. Therefore, in this example, the instruction sentence generation unit 1262 of the prompt generation unit 1260 generates a prompt by combining the text information "Please explain the recommended filing date with reference to claim 1 and Figure 1" with invention information including the numbered configuration of Figure 1, claim 1, and "10 Recommended Filing Date Providing System." The large-scale language model 1100 generates reference text based on the prompt. In this example, the display unit 1302 of the user terminal device 1300 displays the operation screen shown in Figure 30.

[0147] In this example as well, the specification preparation support device 1200 can achieve the effects described above.

[0148] (Third Example of Generating Reference Sentences Using Free Input Function) Next, a third example of generating reference sentences using the free input function will be described with reference to Figs. 31 to 33. Fig. 31 is a diagram showing an example of an operation screen on which no reference sentence is displayed in the third example of generating reference sentences using the free input function. Fig. 32 is a diagram showing an example of character information. Fig. 33 is a diagram showing an example of an operation screen on which the reference sentence is also displayed in the third example of generating reference sentences using the free input function. Note that in this example as well, the specification creation support device 1200 executes the specification creation support method exemplified in Fig. 23.

[0149] In this example, the user selects the drawings from Figures 1 to 5 and claims 1 and 7, as indicated by vertical hatching in Figure 31, and then selects the "free input field" button. Note that this selection by the user is an input operation to the input unit 1301 of the user terminal device 1300 in this example (STEP 1021 in Figure 23). Also, in this example, character information as shown in Figure 32 is input into the "free input field." When this input operation is performed on the input unit 1301, the input unit 1301 transmits an operation signal based on the input operation and the character information as shown in Figure 32 to the specification drafting support device 1200. Note that the character information as shown in Figure 32 is input into the free input unit 2614. Also, the character information as shown in Figure 32 includes the effects of the invention and the problems that the invention is to solve.

[0150] The prompt generation unit 1260 of the specification drafting support device 1200 generates a prompt by combining the invention information selected by the design unit 1261 with the character information entered in the free input unit 2614 based on an operation signal received from the user terminal device 1300 (STEP 1022 in FIG. 23 ). Specifically, the instruction statement generation unit 1262 of the prompt generation unit 1260 generates a prompt by combining the invention information including the drawings in FIG. 1 to FIG. 5 and claims 1 and 7 with the character information exemplified in FIG. 32 . The instruction statement generation unit 1262 transmits the generated prompt to the transmission / reception unit 1230.

[0151] STEP 1023 to STEP 1025 in this example are the same as STEP 1023 to STEP 1025 described above. However, the display unit 1302 of the user terminal device 1300 displays an operation screen as shown in Fig. 33. As shown in Fig. 33, a reference sentence that is a completion for a prompt generated by combining the invention information selected in the design unit 1261 and the character information entered in the free input unit 2614 is displayed in region R1.

[0152] In this example as well, the specification preparation support device 1200 can achieve the effects described above.

[0153] (Help function when selecting a drawing number) Next, the help function when selecting a drawing number will be described with reference to Fig. 34. Fig. 34 is a diagram showing an example of a screen displayed on the display unit 1302 of the user terminal device 1300 when the drawing number of a drawing displayed on the operation screen is double-clicked.

[0154] For example, when a user double-clicks on the part of Fig. 3 among the drawing numbers of the drawings displayed on the operation screen, the screen shown in Fig. 34 is displayed on the display unit 1302 of the user terminal device 1300. It is assumed that Fig. 3 that the user double-clicked on is a flowchart diagram.

[0155] As shown in Figure 34, in this example, the portion of Figure 3 that the user double-clicked on corresponds to a flowchart, which is a type of drawing. This allows the user to recognize that the type of drawing corresponding to the selected figure number is a flowchart and what kind of drawing a flowchart is. Note that the specification drafting support device 1200 may be configured to add the explanatory text for each figure shown in Figure 34 to the prompt generated by the prompt generation unit 1260.

[0156] (When multiple drawing numbers and multiple claim item numbers are selected) Next, with reference to Figures 35 and 36, we will explain how to generate reference text when multiple drawing numbers and multiple claim item numbers are selected. Figure 35 is a diagram showing an example of an operation screen where no reference text is displayed when multiple drawing numbers and multiple claim item numbers are selected. Figure 36 is a diagram showing an example of an operation screen where reference text is also displayed when multiple drawing numbers and multiple claim item numbers are selected.

[0157] In this case, the user selects the drawings from Figures 1 to 29 and the claims from Claim 1 to Claim 10, as shown by the vertical hatching in Figure 35, and then selects the "Generate Reference Text" button. That is, the user selects the drawing numbers of all the drawings and the paragraph numbers of all the claims, and then selects the "Generate Reference Text" button. Note that this selection by the user is an input operation to the input unit 1301 of the user terminal device 1300 in this case (STEP 1021 in Figure 23). When this input operation is performed on the input unit 1301, the input unit 1301 transmits an operation signal based on this input operation to the specification drafting support device 1200.

[0158] The prompt generation unit 1260 of the specification creation support device 1200 generates a prompt for each drawing number and each item number in a state in which multiple drawing numbers and multiple item numbers are selected in the design unit 1261, based on an operation signal received from the user terminal device 1300 (STEP 1022 in FIG. 23 ). The instruction statement generation unit 1262 transmits the generated prompt to the transmission / reception unit 1230.

[0159] In this case, steps 1023 to 1025 are the same as steps 1023 to 1025 described above. However, the display unit 1302 of the user terminal device 1300 displays an operation screen as shown in Fig. 36. As shown in Fig. 36, a reference sentence that is a completion for the prompt generated by combining the invention information selected in the design unit 1261 and the character information entered in the free input unit 2614 is displayed in region R1.

[0160] Even in this case, the specification preparation support device 1200 can achieve the above-mentioned effects.

[0161] Furthermore, with the specification creation support device 1200 having the above configuration, if the user selects all drawing numbers and the prompt generation unit 1260 generates a prompt for each drawing number, the large-scale language model 1100 can be made to generate reference sentences sequentially for each drawing number, thereby allowing the user to check reference sentences corresponding to the entire embodiment at once.

[0162] In this embodiment, a non-transitory computer-readable medium may be used as a memory. The non-transitory computer-readable medium refers to any type of physical memory (RAM, ROM, etc.) that can store information or data that can be read by a processor. The non-transitory computer-readable medium may store instructions for execution by one or more processors. The term "non-transitory computer-readable medium" includes tangible items and excludes carrier waves and transitory signals (i.e., non-transitory).

[0163] The above-described embodiments are provided to facilitate understanding of the present disclosure, and are not intended to limit the present disclosure, which may be modified or improved without departing from the spirit and scope of the present disclosure.

[0164] In the above embodiment, the invention information input to the input receiving unit 1210 of the specification drafting support device 1200 may include information on the effects of the invention corresponding to the claims. In other words, not only claims and drawings but also, for example, text data on the effects of the invention corresponding to the claims may be uploaded. In this case, since the invention information includes information on the effects of the invention corresponding to the claims, the quality of the reference text generated from the large-scale language model 1100 is likely to be improved.

[0165] In the above embodiment, when multiple drawing numbers are selected in the design unit 1261, the prompt generation unit 1260 generates a prompt for each drawing number, but it may also generate a prompt after compiling, for example, the drawing information contained in the drawings corresponding to the selected drawing numbers.

[0166] In the above embodiment, when multiple item numbers are selected in the design unit 1261, the prompt generation unit 1260 generates a prompt for each item number, but for example, it may also generate a prompt after compiling the claims corresponding to the selected item numbers.

[0167] In the above embodiment, the operation screen data generating unit 1220, the drawing information generating unit 1250, and the prompt generating unit 1260 are separate units, but at least two of them may be integrated.

[0168] In the above embodiments, the hatched portions may be colored, for example.

[0169] As described above, this specification discloses the following: (1) A specification drafting support device comprising: a fine-tuning instruction unit that instructs a general-purpose large-scale language model, which has previously trained on a large amount of text data, to fine-tune predetermined parameters of the large-scale language model by additionally training the large-scale language model using training data consisting of patent documents specified by a user; an input unit to which an invention sentence specifying an invention is input; and a display unit that displays generated sentences generated by the large-scale language model after fine-tuning the invention sentence. (2) The fine-tuning instruction unit is configured to instruct fine-tuning in stages, in a first stage, to fine-tune the large-scale language model using field-specific training data consisting of patent documents in a predetermined technical field, and in a second stage, to further fine-tune the large-scale language model fine-tuned in the first stage using the training data consisting of patent documents specified by the user. (3) The specification drafting support device according to (1) or (2), in which the invention sentence includes a claim specifying the invention, and the generated sentences include a description of a mode for carrying out the invention. (4) The specification drafting support device according to any one of (1) to (3), which generates a detailed independent claim composed of a combination of invention-specifying matters of the multiple claims when multiple claims are input to the input unit. (5) The specification drafting support device according to any one of (1) to (4), wherein the invention sentence includes a description of an effect achieved by the invention specified in the claims. (6) The specification drafting support device according to any one of (1) to (5), wherein the invention sentence includes a description of a prior art document explaining background art. (7) A specification drafting support system comprising: a fine-tuning instruction unit that instructs a general-purpose large-scale language model, which has previously trained on a large amount of text data, to fine-tune predetermined parameters of the large-scale language model by additionally training the general-purpose large-scale language model using training data composed of patent documents specified by a user; an input unit to which an invention sentence specifying the invention is input; and a display unit that displays a generated sentence generated by the large-scale language model after fine-tuning the invention sentence.(8) A specification drafting support method executed by a specification drafting support device, comprising: a step of issuing an instruction to fine-tune predetermined parameters of a general-purpose large-scale language model that has been trained in advance on a large amount of text data by additional training using training data consisting of patent documents specified by a user, and a step of displaying a generated sentence generated by the fine-tuned large-scale language model for an invention sentence that is input to an input unit of the specification drafting support device. (9) A computer program including at least one instruction executed by a processor of the specification drafting support device, wherein, when the at least one instruction is executed by the processor, the specification drafting support device is caused to: issue an instruction to fine-tune predetermined parameters of a general-purpose large-scale language model that has been trained in advance on a large amount of text data by additional training using training data consisting of patent documents specified by a user, and display a generated sentence generated by the fine-tuned large-scale language model for an invention sentence that is input to an input unit of the specification drafting support device. (10) A non-transitory computer-readable medium on which the computer program described in (9) is stored.

[0170] The present specification also discloses the following: (11) A specification drafting support device comprising: an input receiving unit to which invention information specifying an invention is input; and a prompt generating unit that generates, based on the invention information, a prompt that causes a large-scale language model that has previously trained on a large amount of text data to generate reference sentences that can be used in the specification of the invention. (12) The specification drafting support device described in (11), wherein the invention information includes drawing information in which drawings of embodiments of the invention are converted into text, the prompt generating unit has a design unit that can design the content of the prompt, and the design unit is capable of selecting at least the drawing information on a drawing number basis. (13) The specification drafting support device described in (12), wherein, when multiple drawing numbers are selected in the design unit, the prompt generating unit is capable of generating a prompt for each drawing number. (14) The specification drafting support device described in (12) or (13), wherein the invention information includes claims specifying the invention, and the design unit is capable of selecting at least the claims on a paragraph number basis. (15) The specification drafting support device according to (14), wherein the invention information includes information on effects of the invention corresponding to the claims. (16) The specification drafting support device according to (14) or (15), wherein the design unit has a free input unit capable of inputting character information, and the prompt generation unit generates the prompt by combining the invention information selected by the design unit with the character information input to the free input unit. (17) The specification drafting support device according to any one of (11) to (16), further having a drawing information generation unit that converts drawings input to the input receiving unit into text to generate drawing information. (18) A specification drafting support method executed by the specification drafting support device, comprising: receiving invention information that identifies the invention; and generating, based on the invention information, a prompt that causes a large-scale language model that has previously trained on a large amount of text data to generate reference sentences that can be used in the specification of the invention.(19) A computer program including at least one instruction executed by a processor of a specification drafting support device, the computer program causing the specification drafting support device, when the at least one instruction is executed by the processor, to receive invention information that identifies an invention, and generate, based on the invention information, a prompt that causes a large-scale language model that has previously trained on a large amount of text data to generate reference sentences that can be used in the specification of the invention.

[0171] This application is based on a Japanese patent application filed on June 14, 2023 (Patent Application No. 2023-097826) and a Japanese patent application filed on October 13, 2023 (Patent Application No. 2023-177415), the contents of which are incorporated herein by reference.

[0172] According to the present invention, a specification writing support device, a specification writing support system, a specification writing support method, a computer program, and a non-transitory computer-readable medium are provided that can reduce the burden of correction work performed by users and improve convenience.

[0173] 1: Specification drafting support system 2: Server 3: Specification drafting support device 4: User terminal device 21: Memory unit 22: Control unit 31: Storage unit 32: Input unit 33: Control unit 34: Display unit 41: Display unit 221: Fine-tuning unit 222: Generation instruction unit 331: Fine-tuning instruction unit M1: General-purpose large-scale language model M1A: Intermediate large-scale language model M2: Fine-tuned large-scale language model N: Network 1001: Specification drafting support system 1100: Large-scale language model 1200: Specification drafting support device 1210: Input acceptance unit 1220: Operation screen data generation unit 1230: Transmission / reception unit 1240: Memory unit 1250: Drawing information generation unit 1260: Prompt generation unit 1261: Design unit 1262: Instruction statement generation unit 1270: Bus 1300: User terminal device 1301: Input section 1302: Display section 2611: Drawing selection section 2612: Claim selection section 2613: Signed structure selection section 2614: Free input section

Claims

1. A specification writing support device comprising: an input receiving unit into which invention information that specifies an invention is input; and a prompt generating unit that generates, based on the invention information, a prompt that causes a large-scale language model that has been trained in advance on a large amount of text data to generate reference sentences that can be used in the specification of the invention.

2. The specification preparation support device according to claim 1, wherein the invention information includes drawing information in which drawings depicting an embodiment of the invention are converted into text, the prompt generation unit has a design unit capable of designing the contents of the prompt, and the design unit is capable of selecting at least the drawing information on a drawing number basis.

3. The specification preparation support device according to claim 2, wherein when a plurality of drawing numbers are selected in the design section, the prompt generation section is capable of generating a prompt for each drawing number.

4. The specification preparation support device according to claim 2, wherein the invention information includes claims that specify the invention, and the design department is capable of selecting at least the claims on an item number basis.

5. The specification drafting support device according to claim 4, wherein the invention information includes information on the effects of the invention corresponding to the claim.

6. The specification preparation support device according to claim 4, wherein the design unit has a free input unit into which character information can be input, and the prompt generation unit generates the prompt by combining invention information selected by the design unit with the character information input to the free input unit.

7. The specification preparation support device according to claim 1, further comprising a drawing information generation unit which converts the drawing input to said input receiving unit into text and generates drawing information.

8. A specification writing support method executed by a specification writing support device, comprising: a step of receiving invention information that identifies an invention; and a step of generating, based on the invention information, a prompt that causes a large-scale language model that has been trained in advance on a large amount of text data to generate reference sentences that can be used in the specification of the invention.

9. A computer program comprising at least one instruction executed by a processor of a specification drafting support device, which, when the at least one instruction is executed by the processor, causes the specification drafting support device to receive invention information identifying an invention, and generate, based on the invention information, a prompt that causes a large-scale language model that has been trained in advance on a large amount of text data to generate reference sentences that can be used in the specification of the invention.

10. A specification drafting support device comprising: a fine-tuning instruction unit that instructs fine-tuning of specified parameters of a general-purpose large-scale language model by additionally training a general-purpose large-scale language model that has been pre-trained with a large amount of text data using training data consisting of patent documents specified by a user; an input unit to which an invention sentence that specifies the invention is input; and a display unit that displays a generated sentence generated by the fine-tuned large-scale language model in response to the invention sentence.

11. The specification drafting support device according to claim 10, wherein the fine-tuning instruction unit is configured to instruct fine-tuning in stages, and in a first stage issues an instruction to fine-tune the large-scale language model using field-specific training data composed of patent documents in a specified technical field, and in a second stage issues an instruction to further fine-tune the large-scale language model already fine-tuned in the first stage using the training data composed of patent documents specified by a user.

12. A specification drafting support device as described in claim 10 or 11, wherein the invention text includes a claim description that specifies the invention, and the generated text includes a description of a form for implementing the invention.

13. The specification drafting support device according to claim 12, which, when a plurality of claims are input to the input section, generates a single detailed independent claim that is constructed by combining the invention-specifying matters of the plurality of claims.

14. The specification drafting support device according to claim 13, wherein the invention text includes a description of the effects achieved by the invention specified in the claim.

15. The specification drafting support device according to claim 14, wherein the invention text includes a reference to a prior art document explaining the background art.

16. A specification drafting support system comprising: a fine-tuning instruction unit that instructs fine-tuning of predetermined parameters of a general-purpose large-scale language model by additionally training a general-purpose large-scale language model, which has been previously trained on a large amount of text data, using training data consisting of patent documents specified by a user; an input unit to which an invention sentence that specifies an invention is input; and a display unit that displays a generated sentence generated by the fine-tuned large-scale language model in response to the invention sentence.

17. A specification drafting support method executed by a specification drafting support device, comprising: a step of instructing a general-purpose large-scale language model, which has been pre-trained on a large amount of text data, to fine-tune predetermined parameters of the large-scale language model by additionally training the large-scale language model using training data consisting of patent documents specified by a user; and a step of displaying a generated sentence generated by the fine-tuned large-scale language model for an invention sentence specifying an invention input to an input unit provided in the specification drafting support device.

18. A computer program including at least one instruction executed by a processor of a specification drafting support device, which, when the at least one instruction is executed by the processor, causes the specification drafting support device to instruct the device to fine-tune certain parameters of the large-scale language model by additionally training a general-purpose large-scale language model, which has been trained in advance on a large amount of text data, using training data consisting of patent documents specified by a user, and to display a generated sentence generated by the fine-tuned large-scale language model in response to an invention sentence specifying an invention input to an input unit provided in the specification drafting support device.

19. A non-transitory computer readable medium having stored thereon a computer program as claimed in claim 18.

Images