Method for text generation, text generation program, and text generation device

Abstract

To enable a user to grasp content which is assumed not to be described in a predetermined group of documents.SOLUTION: A method for text generation causes a computer to execute a process including: vectorizing sentences described in an acquired group of documents; extracting at least one missing part assumed not to be described in the group of documents based on a group of vectors obtained by vectorizing the sentences; and generating text to be presented to a user by using a vector corresponding to at least one missing part selected from the extracted missing parts.SELECTED DRAWING: Figure 1

Description

[Technical Field]

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

[0002] Patent Document 1 discloses a technology that can make highly accurate predictions regarding the suitability of patent requirements for inventions under investigation that are in line with examination practices, thereby effectively reducing the burden of preparing application documents. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2019-40402 Summary of the Invention [Problem to be solved by the invention]

[0004] Conventionally, when holding an excavation event to discover new ideas, the users participating in the event had to read multiple publicly available documents and build up knowledge as a preliminary step, and there is room for improvement in terms of improving the convenience for the users.

[0005] Therefore, the present disclosure aims to provide a text generation method, a text generation program, and a text generation device that can enable a user to understand content that is not expected to be included in a specified group of documents. [Means for solving the problem]

[0006] In a first aspect of the text generation method, a computer executes the following process: vectorizing sentences contained in an acquired group of documents; extracting at least one gap that is assumed not to be contained in the group of documents based on the group of vectorized sentences; and generating text to be presented to a user using a vector corresponding to at least one of the extracted gaps. In this way, the text generation method allows the user to understand the content that is assumed not to be contained in the group of documents by checking the generated text.

[0007] In the text generation method of the second aspect, a computer executes a process of selecting, from the extracted gaps, the gaps related to the text content entered by the user, thereby making it easier to generate text that shows content close to the user's image.

[0008] In the text generation method of the third aspect, in the first or second aspect, a computer executes a process in which a vector whose distance to a selected group of vectors surrounding the hole is within a predetermined distance is used as the vector corresponding to the hole. As a result, according to this text generation method, the credibility of the content shown in the text can be increased compared to when text is generated using vectors whose distance to the vector exceeds the predetermined distance.

[0009] A fourth aspect of the text generation method is the same as any of the first to third aspects, and the computer executes the following process to extract the missing parts: generating data in which the number of dimensions of a vector group obtained by vectorizing the sentence is compressed; extracting at least one vector indicating coordinates included in the missing parts assumed to be not described in the document group based on the data in which the number of dimensions is compressed; and restoring the vector in a high-dimensional space. As a result, this text generation method makes it possible to extract parts assumed to be not described in the document group in an easy-to-analyze low-dimensional space while preserving the essential properties of the sentence as much as possible.

[0010] A fifth aspect of the text generation method is the fourth aspect, in which a computer executes a process of using an autoencoder to compress the number of dimensions of a group of vectors obtained by vectorizing the sentence. As a result, this text generation method can efficiently compress the data structure to a low dimension with less information loss than when the number of dimensions is compressed using methods other than an autoencoder, such as PCA (Principal Component Analysis).

[0011] A sixth aspect of the text generation method is any of the first to fifth aspects, in which a computer executes a process to extract the holes by performing topological data analysis. This makes it possible to grasp important information that would be lost in conventional statistical data analysis in detail by understanding the shape of the data.

[0012] A seventh aspect of the text generation method is any one of the first to sixth aspects, wherein a computer executes the process of vectorizing the sentence using an encoder-decoder model, and generating the text using the encoder-decoder model. This makes it possible to generate text that reflects the characteristics of sentences contained in an acquired document group.

[0013] The text generation method of an eighth aspect is any of the first to seventh aspects, in which a computer executes the process of vectorizing the sentence using an encoder model, comparing the vector corresponding to the gap with a combination of keywords included in the sentence vectorized by the encoder model, and generating the combination of keywords that satisfies a predetermined condition as the text. As a result, this text generation method can generate text using expressions contained in the acquired document group.

[0014] A ninth aspect of the text generation method is the eighth aspect, in which a computer executes a process of vectorizing, using the encoder model, the keyword combinations based on a plurality of vectors whose inter-vector distances with a group of vectors surrounding the selected hole are within a predetermined distance. This makes it possible to increase the credibility of the content shown in the text compared to a case in which a keyword combination based on a plurality of vectors whose inter-vector distances exceed the predetermined distance is vectorized.

[0015] A tenth aspect of the text generation method is the same as any of the first to ninth aspects, wherein each document constituting the document group has a description portion written from a plurality of perspectives, and a first sentence describing a first perspective and a second sentence describing a second perspective among the plurality of perspectives are vectorized, and at least one of the gaps is extracted based on the vectorized vectors of the first sentences, and the text about the second perspective is generated using a document group surrounding at least one of the gaps selected from the extracted gaps. This text generation method allows a user to grasp the content of the second perspective based on the first perspective that is assumed not to be described in the obtained document group.

[0016] An eleventh aspect of the text generation method is any one of the first to tenth aspects, wherein the document group is a patent document group. This allows the user to understand content that is not considered to be described in the patent document group by checking the generated text.

[0017] A text generation program of a twelfth aspect causes a computer to execute the following process: vectorize sentences contained in an acquired group of documents; extract at least one gap that is assumed not to be contained in the group of documents based on the group of vectorized sentences; and generate text to be presented to a user using a vector corresponding to at least one of the extracted gaps. In this way, the text generation program allows the user to understand the content that is assumed not to be contained in the group of documents by checking the generated text.

[0018] A text generator according to a thirteenth aspect of the present invention includes a processor that vectorizes sentences contained in an acquired set of documents, extracts at least one gap that is assumed not to be contained in the set of documents based on the vectorized sentences, and generates text to be presented to a user using a vector corresponding to at least one of the extracted gaps. This allows the user to understand the content that is assumed not to be contained in the set of documents by checking the generated text. [Effects of the Invention]

[0019] As described above, the text generation method, text generation program, and text generation device according to the present disclosure enable a user to understand content that is not expected to be included in a specified group of documents. [Brief explanation of the drawings]

[0020] [Figure 1] FIG. 2 is a block diagram showing the hardware configuration of the text generator. [Figure 2] FIG. 2 is a block diagram showing a storage configuration of the text generator. [Figure 3] 10 is a flowchart illustrating the flow of a text generation process. [Figure 4]4 is a flowchart showing the process of extracting a hole portion shown in step S12 of FIG. 3. [Figure 5] 10 is a first display example displayed on the display unit. [Figure 6] 10 is a second example of a display displayed on the display unit. [Figure 7] 10 is a third example of a display displayed on the display unit. [Figure 8] 10 is a fourth example of a display displayed on the display unit. DETAILED DESCRIPTION OF THE INVENTION

[0021] The text generator 50 according to this embodiment will be described below. (First embodiment) First, a first embodiment of the text generator 50 according to the present embodiment will be described.

[0022] 1 is a block diagram showing the hardware configuration of a text generator 50. As an example, the text generator 50 may be a general-purpose computer such as a server computer or a PC (Personal Computer), or a mobile terminal such as a smartphone or a tablet terminal. In this embodiment, the text generator 50 is a "PC."

[0023] 1, a text generator 50 includes a CPU (Central Processing Unit) 51, a ROM (Read Only Memory) 52, a RAM (Random Access Memory) 53, a storage 54, an input unit 55, a display unit 56, and a communication unit 57. Each component is connected to each other via a bus 58 so as to be able to communicate with each other.

[0024] The CPU 51 is a central processing unit that executes various programs and controls each part. That is, the CPU 51 reads a program from the ROM 52 or the storage 54 and executes the program using the RAM 53 as a work area. The CPU 51 controls each of the above components and performs various arithmetic processing in accordance with the program stored in the ROM 52 or the storage 54. The CPU 51 is an example of a "processor."

[0025] The ROM 52 stores various programs and various data. The RAM 53 serves as a working area for temporarily storing programs or data.

[0026] The storage 54 is configured by a storage device such as a hard disk drive (HDD), a solid state drive (SSD), or a flash memory, and stores various programs and various data.

[0027] The input unit 55 includes, for example, a pointing device such as a mouse, various buttons, a keyboard, a microphone, a camera, and the like, and is used to perform various inputs.

[0028] The display unit 56 is, for example, a liquid crystal display, and displays various information. The display unit 56 may function as the input unit 55 by adopting a touch panel system.

[0029] The communication unit 57 is an interface for communicating with other devices. For this communication, for example, a wired communication standard such as Ethernet (registered trademark) or FDDI, or a wireless communication standard such as 4G, 5G, or Wi-Fi (registered trademark) is used.

[0030] FIG. 2 is a block diagram showing the configuration of the storage 54 of the text generator 50. As shown in FIG. As shown in FIG. 2, the storage 54 stores a text generation program 54A, an algorithm storage unit 54B, and a model storage unit 54C.

[0031] The text generation program 54A is a program for causing the CPU 51 to execute a text generation process, which will be described later.

[0032] The algorithm storage unit 54B stores machine learning algorithms used in the text generation process, such as dimensionality reduction algorithms such as PCA (Principal Component Analysis), isomap, UMAP, and t-SNE, or autoencoders such as VAE (Variational Auto Encoder).

[0033] The model storage unit 54C stores machine learning models used in the text generation process. The machine learning models include, for example, encoder models such as text2vec and BERT, and encoder-decoder models such as LLM (Large Language Models).

[0034] 3 is a flowchart illustrating the flow of a text generation process as a text generation method according to the present disclosure. The text generation process is performed by CPU 51 reading text generation program 54A from storage 54, expanding it in RAM 53, and executing it. As an example, the text generation process is performed when a user operates text generator 50 and a predetermined application installed in text generator 50 is executed.

[0035] In step S10 shown in FIG. 3, the CPU 51 acquires a group of patent documents. The patent documents are, for example, unexamined patent publications and patent gazettes stored in the patent database of the Patent Information Platform provided by the National Center for Industrial Property Information and Training. Each patent document constituting the group of patent documents has a description portion from multiple perspectives. For example, unexamined patent publications and patent gazettes have a description portion such as claims and a detailed description of the invention. Furthermore, the detailed description of the invention in unexamined patent publications and patent gazettes includes a description portion from multiple perspectives, such as a problem to be solved by the invention, a means for solving the problem, and a mode for implementing the invention. In the first embodiment, the CPU 51 acquires, as the group of patent documents, a collection of description portions from one perspective (e.g., a problem to be solved by the invention) in multiple patent documents uploaded by a user to a predetermined application. Then, the CPU 51 proceeds to step S11.

[0036] The patent document group is not limited to the above-mentioned published patent applications and patent publications, and information from commercial databases may also be used. Examples of information from commercial databases include DWPI Abstracts stored in the DWPI (Derwent World Patents Index, registered trademark) patent database, concepts or key content stored in Orbit Intelligence, and problem solutions stored in BIZcluncher. DWPI Abstracts contain sections describing novelty, use, advantage, etc.

[0037] In step S11, the CPU 51 vectorizes the sentences described in the patent document group acquired in step S10. For example, if a collection of portions describing problems to be solved by inventions in a plurality of published patent publications is acquired as the patent document group in step S10, the CPU 51 vectorizes the sentences describing the problems to be solved by each invention. In the first embodiment, the CPU 51 vectorizes the sentences using an LLM, which is an encoder-decoder model. Then, the CPU 51 proceeds to step S12.

[0038] In step S12, the CPU 51 extracts at least one gap that is assumed not to be described in the patent documents acquired in step S10, based on the vector group obtained by vectorizing the sentences in step S11. For example, the CPU 51 extracts the gap by performing topological data analysis. Details of this will be described later with reference to FIG. 4. The CPU 51 then proceeds to step S13.

[0039] In step S13, the CPU 51 selects at least one specific hole portion from the hole portions extracted in step S12. For example, the CPU 51 selects a specific hole portion associated with the text content input by the user from the hole portions. This will be described in detail later with reference to Figures 5 and 6. The CPU 51 then proceeds to step S14.

[0040] In step S14, the CPU 51 generates text to be presented to the user using the vector corresponding to the specific hole selected in step S13. The text is a sentence or a group of words that indicates content that is assumed not to be described in the patent documents acquired in step S10. The vector corresponding to the specific hole is a vector whose inter-vector distance with the group of vectors surrounding the specific hole is within a predetermined distance. In the first embodiment, the CPU 51 inputs the vector corresponding to the specific hole into the LLM that vectorized the sentence in step S11, and generates the text. Then, the CPU 51 proceeds to step S15.

[0041] In step S15, the CPU 51 outputs the text generated in step S14 and displays it on the display unit 56. Then, the CPU 51 ends the text generation process.

[0042] FIG. 4 is a flowchart showing the process of extracting the hole portion shown in step S12 of FIG.

[0043] 4, the CPU 51 uses an autoencoder to generate data that compresses the number of dimensions of the vector group obtained by vectorizing the text in step S11, and then the CPU 51 proceeds to step S12-2.

[0044] In step S12-2, the CPU 51 performs persistent homology processing on the data whose dimensionality has been compressed in step S12-1, extracting vectors indicating coordinates of holes assumed to be undescribed in the patent documents acquired in step S10. Persistent homology is a type of topological data analysis, and has been described in detail in various publications. Therefore, a detailed description is omitted here. Persistent homology is a method for quantitatively extracting shape information of data as features by focusing on structural elements, such as connections, holes, and voids, of a figure in a compressed dimensional space. In this persistent homology processing, a virtual sphere is set in a vector space, with each point data representing the above data at its center, and the radius of the sphere is gradually increased. Then, in this persistent homology processing, the CPU 51 extracts holes formed between adjacent spheres as the above-mentioned hole portions when the radius of each sphere is increased. The CPU 51 then proceeds to step S12-3.

[0045] In step S12-3, the CPU 51 uses an autoencoder to restore the vectors indicating the coordinates included in the hole portions extracted in step S12-2 to a higher dimension, for example, the number of dimensions before reduction, and extracts at least one hole portion.The CPU 51 then proceeds to step S13.

[0046] Fig. 5 is a first example of a display displayed on the display unit 56. Specifically, Fig. 5 is a first example of a selection screen for selecting a specific hole portion from among a plurality of extracted hole portions.

[0047] The selection screen shown in FIG. 5 displays point cloud data 60, a text input section 62, a hole display section 64, and an OK button 66.

[0048] The point cloud data 60 shows a plurality of point data located around a hole portion. As an example, the point cloud data 60 shows point data as black circles, and also shows each sphere (which is a circle here because it is a two-dimensional diagram) virtually set at each point data. Each point data in the point cloud data 60 has a vector corresponding to a sentence described in each patent document obtained in the text generation process. Three hole portions, hole portion A, hole portion B, and hole portion C, are extracted from the point cloud data 60.

[0049] The text input section 62 is a section where the user inputs text. The text input section 62 shown in Figure 5 displays "ABC," indicating that the user has input the word "ABC."

[0050] The hole display unit 64 displays a hole portion associated with the text content input by the user. In the case shown in FIG. 5, the CPU 51 selects a hole portion whose periphery is surrounded by point data having a vector corresponding to a sentence including the word input by the user as a hole portion associated with the text content. For example, assume that the point data P1 shown in FIG. 5 has a vector corresponding to a sentence including the word "ABC" input by the user. In this case, the CPU 51 selects a hole portion A whose periphery is surrounded by point data P1 and displays it in the hole display unit 64. When the OK button 66 is operated with the hole portion associated with the text content displayed in the hole display unit 64, the CPU 51 causes the display unit 56 to display a generation screen for generating text to be presented to the user.

[0051] Fig. 6 is a second example of a display displayed on the display unit 56. Specifically, Fig. 6 is a second example of a selection screen.

[0052] The selection screen shown in FIG. 6 displays point cloud data 60, a text input section 62, a hole display section 64, and an OK button 66, similar to FIG.

[0053] The text input section 62 shown in FIG. 6 displays "D includes E and F," indicating that the user has input the sentence "D includes E and F."

[0054] In the case shown in FIG. 6, the CPU 51 selects a hole portion located at a position close in similarity to the sentence input by the user as a hole portion associated with the text content. In this case, the CPU 51 vectorizes the sentence input by the user using the above-described LLM. The CPU 51 then arranges the vector corresponding to the sentence as point data in a vector space and selects a hole portion having the closest distance to the point data. For example, suppose that the user inputs the sentence "D includes E and F" into the text input unit 62, and point data P2 is newly arranged in the point cloud data 60 shown in FIG. 6. In this case, the CPU 51 selects hole portion B having the closest distance (similarity) to point data P2 in the vector space and displays it in the hole display unit 64.

[0055] Fig. 7 is a third example of a display displayed on the display unit 56. Specifically, Fig. 7 is a first example of a generation screen for generating text to be presented to the user.

[0056] The generation screen shown in FIG. 7 displays point cloud data 70, a vector display section 72, an OK button 74, and a text display section 76.

[0057] The point cloud data 70 shows data of a plurality of points located around a hole portion selected on the selection screen. As an example, the point cloud data 70 shows a case where hole portion B is selected on the selection screen.

[0058] The point cloud data 70 displays only point data having vectors whose inter-vector distances with respect to the vector group surrounding the hole portion B are within a predetermined distance. The vector group surrounding the hole portion B is the five point data connected by the line segment L1 in FIG. 7. Hereinafter, the vector group surrounding the hole portion selected on the selection screen may be referred to as "one outline vector group." Herein, "vectors whose inter-vector distances with respect to one outline vector group are within a predetermined distance" refers to vectors that constitute a vector group that can surround the hole portion selected on the selection screen when the one outline vector group is excluded. In other words, point data having vectors that cannot surround the hole portion is not displayed in the point cloud data 70. For example, the point cloud data 70 shown in FIG. 7 displays only point data having vectors that constitute one outline vector group and two outline vector groups that surround the hole portion B with the second largest number of point data after the one outline vector group. Note that the two outline vector groups in FIG. 7 are eight point data connected by the line segment L2.

[0059] The vector display unit 72 displays a vector corresponding to the hole portion selected by the user. For example, the user operates a mouse serving as the input unit 55 to select a portion within the point cloud data 70. In FIG. 7, the user has selected a square S portion within the hole portion B. In this case, the CPU 51 obtains vector values ​​in the vector space of the square S portion and displays them on the vector display unit 72.

[0060] The text display unit 76 displays text to be presented to the user. In the first embodiment, when the OK button 74 is operated while a vector corresponding to the hole portion is displayed in the vector display unit 72, the CPU 51 inputs the vector into the LLM and displays the generated text in the text display unit 76. For example, the text display unit 76 displays a sentence indicating content that is assumed not to be described in the patent document group obtained in the text generation process.

[0061] As described above, in the text generator 50, the CPU 51 vectorizes the sentences described in the patent documents acquired in the text generation process. Furthermore, the CPU 51 extracts at least one gap that is assumed to be not described in the patent documents based on the vectors obtained by vectorizing the sentences. The CPU 51 then generates text to be presented to the user using a vector corresponding to at least one gap selected from the extracted gaps. This allows the text generator 50 to allow the user to understand what is assumed to be not described in the patent documents by reviewing the generated text.

[0062] Furthermore, in the text generator 50, the CPU 51 selects, from the extracted gaps, gaps that are related to the text content entered by the user. This makes it easier for the text generator 50 to generate text that shows content close to the user's image.

[0063] Furthermore, in the text generator 50, the CPU 51 uses, as a vector corresponding to a hole, a vector whose distance to a group of vectors surrounding the selected hole is within a predetermined distance. This allows the text generator 50 to increase the credibility of the content shown in the text compared to when text is generated using vectors whose distance to the vector exceeds the predetermined distance.

[0064] In the text generator 50, the CPU 51 extracts the missing portions by compressing the number of dimensions of vectors obtained by vectorizing the text. Next, based on the compressed data, the CPU 51 extracts at least one vector indicating the coordinates of the missing portions assumed to be not described in the patent documents acquired in the text generation process. The CPU 51 then restores the vector in a higher dimension. This allows the text generator 50 to extract portions assumed to be not described in the patent documents in a low-dimensional space that is easy to analyze, while preserving the essential characteristics of the text as much as possible.

[0065] Furthermore, in the text generator 50, the CPU 51 uses an autoencoder to compress the number of dimensions of a group of vectors obtained by vectorizing a sentence. As a result, the text generator 50 can efficiently compress the data structure to a low dimension with less information loss than when the number of dimensions is compressed using methods other than an autoencoder, such as PCA.

[0066] Furthermore, in the text generator 50, the CPU 51 extracts holes by performing topological data analysis. As a result, the text generator 50 can grasp important information in detail that would be lost in conventional statistical data analysis by grasping the shape of the data.

[0067] In the text generator 50, the CPU 51 vectorizes a sentence using an encoder-decoder model. Then, the CPU 51 generates the above text using the encoder-decoder model. As a result, the text generator 50 can generate text that reflects the characteristics of the sentences described in the patent documents acquired in the text generation process.

[0068] (Second embodiment) Next, a second embodiment of the text generator 50 according to the present invention will be described while omitting or simplifying parts that overlap with the above embodiment.

[0069] The second embodiment differs from the above embodiments in that, in the text generation process shown in Fig. 3, an encoder model is used to vectorize a sentence, and the encoder model is used to generate the text. For example, the second embodiment uses BERT as the encoder model. The text generation method in the second embodiment will be described below with reference to Fig. 7.

[0070] As shown in FIG. 7, suppose that the OK button 74 is operated while the vector of the part of square S selected by the user, i.e., the vector (x, x, x) corresponding to hole portion B, is displayed on vector display unit 72. In this case, CPU 51 vectorizes, using BERT, a combination of keywords based on a plurality of vectors whose inter-vector distances from one outline vector group of hole portion B are within a predetermined distance. As in the above embodiment, the "plurality of vectors whose inter-vector distances from one outline vector group are within a predetermined distance" are vectors that constitute a vector group that can surround the hole portion selected on the selection screen. In addition, the CPU 51 extracts keywords from the sentence corresponding to the vector by executing a predetermined algorithm stored in algorithm storage unit 54B.

[0071] For example, the CPU 51 extracts keywords from the text corresponding to three vectors constituting one contour vector group of the hole portion B (e.g., configuration A, configuration B, and configuration C). Next, the CPU 51 combines the extracted keywords to generate a predetermined number of keyword combinations (e.g., combination 1 "configuration A, configuration B," combination 2 "configuration A, configuration C," etc.). Next, the CPU 51 sequentially inputs the multiple keyword combinations into BERT and vectorizes them. Next, the CPU 51 sequentially compares the vector (x, x, x) corresponding to the hole portion B with the multiple vectorized keyword combinations to determine the keyword combination with the closest distance (similarity) in the vector space. The CPU 51 then generates the determined keyword combination (e.g., "configuration A, configuration B") as text. As a result, the text display unit 76 displays a group of words indicating content that is assumed not to be described in the patent documents acquired in the text generation process.

[0072] As described above, in the text generator 50, the CPU 51 uses an encoder model to vectorize sentences described in the patent documents acquired in the text generation process. The CPU 51 then compares the vectors corresponding to the gaps with the combinations of keywords included in the sentences, vectorized by the encoder model, and generates a combination of keywords that satisfies a predetermined condition as text. In the second embodiment, as an example, the combination of keywords that satisfies the predetermined condition is the combination of keywords that has the closest distance (similarity) in vector space to the vector corresponding to the gaps. This allows the text generator 50 to generate text using expressions described in the patent documents acquired in the text generation process. Thus, in one or more embodiments, text generation is not limited to the generation of sentences, but also includes word combinations, etc.

[0073] Furthermore, in the text generator 50, the CPU 51 vectorizes, using an encoder model, a combination of keywords based on a plurality of vectors whose inter-vector distances to the group of vectors surrounding the selected hole are within a predetermined distance. This allows the text generator 50 to increase the credibility of the content shown in the text compared to a case where a combination of keywords based on a plurality of vectors whose inter-vector distances exceed a predetermined distance are vectorized.

[0074] (Third embodiment) Next, a third embodiment of the text generator 50 according to the present invention will be described while omitting or simplifying parts that overlap with the above embodiments.

[0075] The third embodiment differs from the above-described embodiments in that, in the text generation process shown in FIG. 3, the CPU 51 acquires a collection of portions of a plurality of patent documents describing multiple perspectives (e.g., a problem to be solved by the invention and a means for solving the problem) as a patent document group. In this case, in the text generation process, the CPU 51 uses an LLM, which is an encoder-decoder model, to vectorize a first sentence describing a first perspective (e.g., a problem to be solved by the invention) and a second sentence describing a second perspective (e.g., a means for solving the problem). In the third embodiment, the content of the text generated in the text generation process differs from that of the above-described embodiments. The text generation method in the third embodiment will be described below with reference to FIG. 8.

[0076] Fig. 8 is a fourth display example displayed on the display unit 56. Specifically, Fig. 8 is a second example of a generated screen.

[0077] On the generation screen shown in FIG. 8, an OK button 74, a text display section 76, point cloud data 80, point cloud data 82, and a prompt input section 84 are displayed.

[0078] The point cloud data 80 shows point data of a group of vectors surrounding a hole portion for a first viewpoint selected on the selection screen. As an example, the point cloud data 80 shows a case where hole portion B is selected on the selection screen. The point cloud data 80 is made up of five point data, point data P5 to P9.

[0079] The point cloud data 82 indicates a plurality of point data for a second viewpoint corresponding to the point cloud data 80. The point cloud data 82 is made up of five point data items P10 to P18.

[0080] Hereinafter, in the third embodiment, the above-mentioned first aspect will be described as the "problem that the invention aims to solve (hereinafter simply referred to as the "problem")" in the published patent publication, and the second aspect will be described as the "means for solving the problem (hereinafter simply referred to as the "means")" in the published patent publication.

[0081] Here, point data P5 and point data P10, point data P6 and point data P12, point data P7 and point data P14, point data P8 and point data P16, and point data P9 and point data P18 each correspond to a sentence described in a published patent application having the same publication number. For example, point data P5 corresponds to a sentence described in a specific published patent application A from the perspective of a problem to be solved, and point data P10 corresponds to a sentence described in the published patent application A from the perspective of a means to be implemented.

[0082] At this time, the point cloud data 80 and the point cloud data 82 are a collection of the same unexamined patent publications (e.g., unexamined patent publications A to E), but a hole B is formed in the point cloud data 80, and no hole is formed in the point cloud data 82. In this case, from unexamined patent publications A to E, portions that are assumed not to be described from the perspective of the problem to be solved are extracted, but portions that are assumed not to be described from the perspective of the means to be taken are not extracted. In the third embodiment, in this situation, the CPU 51 generates text regarding the perspective of the means to be taken using a group of patent documents surrounding the hole extracted from the perspective of the problem to be solved.

[0083] The prompt input unit 84 is a unit that accepts input of a prompt to be input to the LLM that vectorized the sentence in the text generation process. As an example, in FIG. 8 , the prompt "Please tell me about a means to solve the problem shown in (x, x, x)" is input to the prompt input unit 84. When the OK button 74 is operated with a prompt input to the prompt input unit 84, the CPU 51 inputs the prompt into the LLM and displays the generated text in the text display unit 76. For example, the text display unit 76 displays a sentence indicating a means to solve a new problem that is assumed to be not described in the patent documents obtained in the text generation process.

[0084] As described above, in the text generator 50, the CPU 51 vectorizes the first sentence describing the problem perspective and the second sentence describing the means perspective, among multiple perspectives in the patent document group. The CPU 51 also extracts at least one gap based on the vector group obtained by vectorizing the first sentence. The CPU 51 then generates text about the means perspective using the patent document group surrounding at least one gap selected from the extracted gaps. This allows the text generator 50 to allow the user to understand means for solving a novel problem that is assumed not to be described in the patent document group obtained by the text generation process.

[0085] (others) In the above embodiment, a group of patent documents has been described as an example of a group of documents, but the example of a group of documents is not limited to this. For example, the example of a group of documents may be a group of academic papers, a group of technical reports, a group of report documents, etc.

[0086] In the above embodiment, a case where multiple hole portions are extracted has been described as an example, but the number of hole portions extracted is not limited to multiple. For example, in the text generation process, only one hole portion may be extracted.

[0087] In the above embodiment, a specific hole portion is selected from among multiple extracted hole portions based on a user's text input. However, the selection of a specific hole portion may be performed without a user input. For example, the CPU 51 may select a specific hole portion based on a persistence diagram generated in the persistent homology processing. The persistence diagram is a diagram in which the axis represents the radius (or time) at which the hole appears and the axis represents the radius (or time) at which the hole disappears. In this case, the CPU 51 may select a hole that is closest to the 45-degree line in the persistence diagram as a specific hole portion. The CPU 51 may also select a hole whose appearance time is earlier than a predetermined standard as a specific hole portion.

[0088] In the above embodiment, an example has been described in which one hole portion is selected as a specific hole portion from among a plurality of extracted hole portions, but the specific hole portion is not limited to one hole portion. For example, in the text generation process, a plurality of hole portions may be selected as a specific hole portion.

[0089] In the above embodiment, the number of dimensions of the vector group obtained by vectorizing the text using an autoencoder is compressed in both the case where an encoder-decoder model is used for vectorizing the text and the case where an encoder model is used. However, this is not limited to this. When an encoder model is used for vectorizing the text, the number of dimensions of the vector group obtained by vectorizing the text may be compressed using an algorithm other than the autoencoder, such as PCA, without restoring the number of dimensions.

[0090] In the above embodiment, holes are extracted by executing persistent homology as the topological data analysis, but the topological data analysis to be executed is not limited to this. For example, holes may be extracted by executing Mapper instead of persistent homology as the topological data analysis.

[0091] In the above embodiment, when generating text using an encoder-decoder model, only vectors corresponding to the holes are input to the encoder-decoder model, but the input contents to the encoder-decoder model are not limited to this. For example, the encoder-decoder model may be able to input a prompt indicating an instruction sentence by the user in addition to the vector.

[0092] In the above embodiment, the text generator 50 executes the text generation process shown in Fig. 3. However, the present invention is not limited to this, and the text generation process may be executed in cooperation with another device. For example, the algorithm storage unit 54B and the model storage unit 54C may be provided in another device, and the text generator 50 may appropriately acquire processing results based on a machine learning algorithm and a machine learning model from the other device during the text generation process.

[0093] In the above embodiment, the text generation process executed by the CPU 51 after reading the software (program) may be executed by various processors other than a CPU. Examples of such processors include programmable logic devices (PLDs) such as field-programmable gate arrays (FPGAs), whose circuit configuration can be changed after fabrication, and application-specific integrated circuits (ASICs), which are dedicated electrical circuits that are processors with circuit configurations specifically designed to execute specific processes. The text generation process may be executed by one of these processors, or by a combination of two or more processors of the same or different types (e.g., multiple FPGAs, or a combination of a CPU and an FPGA). The hardware structure of these processors is, more specifically, an electrical circuit that combines circuit elements such as semiconductor devices.

[0094] In the above embodiment, the text generation program 54A is pre-stored (installed) in the storage 54, but the present invention is not limited to this. The text generation program 54A may be provided in a form recorded on a recording medium such as a CD-ROM (Compact Disk Read Only Memory), a DVD-ROM (Digital Versatile Disk Read Only Memory), or a USB (Universal Serial Bus) memory. The text generation program 54A may also be downloaded from an external device, cloud storage, or the like via a network. [Explanation of symbols]

[0095] 50 Text Generator 51 CPU (processor) 54A Text Generator

Claims

1. The sentences contained in the acquired documents are vectorized, extracting at least one gap that is assumed not to be described in the group of documents based on a group of vectors obtained by vectorizing the sentences; generating text to be presented to a user using a vector corresponding to at least one of the extracted hole portions; A text generation method in which processing is performed by a computer.

2. selecting, from the extracted gaps, the gap associated with the text content input by the user; 10. The method of claim 1, wherein the processing is performed by a computer.

3. As the vector corresponding to the hole portion, a vector whose distance from a group of vectors surrounding the selected hole portion is within a predetermined distance is used.

10. The method of claim 1, wherein the processing is performed by a computer.

4. The hole portion is extracted as follows: generating data in which the number of dimensions of a vector group obtained by vectorizing the text is compressed; extracting at least one vector indicating coordinates included in a hole portion that is assumed not to be described in the document group based on the data with the number of dimensions compressed; Reconstructing the vector in high dimensions; 10. The method of claim 1, wherein the processing is performed by a computer.

5. Using an autoencoder, the number of dimensions of the vector group obtained by vectorizing the sentence is reduced.

5. The method of claim 4, wherein the processing is performed by a computer.

6. extracting the hole portion by performing topological data analysis; 10. The method of claim 1, wherein the processing is performed by a computer.

7. vectorizing the sentence using an encoder-decoder model; generating the text using the encoder-decoder model; 10. The method of claim 1, wherein the processing is performed by a computer.

8. vectorizing the sentence using an encoder model; comparing the vector corresponding to the hole with a combination of keywords included in the sentence that has been vectorized by the encoder model, and generating the combination of keywords that satisfies a predetermined condition as the text; 10. The method of claim 1, wherein the processing is performed by a computer.

9. vectorizing, by the encoder model, a combination of the keywords based on a plurality of vectors whose inter-vector distances with the selected group of vectors surrounding the hole portion are within a predetermined distance; 9. The method of claim 8, wherein the processing is performed by a computer.

10. Each document constituting the document group has a description portion written from a plurality of viewpoints, vectorizing a first sentence describing a first viewpoint and a second sentence describing a second viewpoint from among the plurality of viewpoints; extracting at least one of the holes based on a group of vectors obtained by vectorizing the first sentence; generating the text for the second viewpoint using a group of documents surrounding at least one of the extracted hole portions; 10. The method of claim 1, wherein the processing is performed by a computer.

11. the document group is a patent document group; The text generation method of claim 1 .

12. The sentences contained in the acquired documents are vectorized, extracting at least one gap that is assumed not to be described in the group of documents based on a group of vectors obtained by vectorizing the sentences; generating text to be presented to a user using a vector corresponding to at least one of the extracted hole portions; A text generation program that causes a computer to perform processing.

13. a processor; The processor: The sentences contained in the acquired documents are vectorized, extracting at least one gap that is assumed not to be described in the group of documents based on a group of vectors obtained by vectorizing the sentences; generating text to be presented to a user using a vector corresponding to at least one of the extracted hole portions; Text generator.

Images