Information processing system and tagging method

Abstract

This invention provides an information processing system and tagging method that improve the accuracy of tagging while reducing the processing burden on workers. [Solution] The method involves dividing a text for tag generation into multiple generation chunks, inputting these chunks into a large-scale language model, the large-scale language model generating a summary sentence for each generation chunk, combining the summary sentences into multiple generation chunks to generate multiple tag generation chunks, inputting these into the large-scale language model, the large-scale language model generating a tag structure in which tags are structured at an arbitrary hierarchy for the multiple tag generation chunks, dividing the text to be tagged into multiple target chunks, inputting these multiple target chunks into the large-scale language model, generating a tag assignment chunk based on the divided target chunks, and inputting the tag assignment chunk and tag structure into the large-scale language model, thereby allowing the large-scale language model to assign tags to the tag assignment chunk according to the tag structure.

Description

Technical Field

[0001] The present disclosure relates to an information processing system and a tagging method.

Background Art

[0002] There is a technique of extracting keywords from text data and assigning tags based on the extracted keywords. For example, a server device that associates weighted keywords obtained by adding to keywords with input Japanese document text is known (see, for example, Patent Document 1). Also, a device that extracts words included in a received document and calculates a recommendation score for each tag for the received document based on a tag word co-occurrence measure for all the extracted words is known (see, for example, Patent Document 2).

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Patent Document 2

Summary of the Invention

Problems to be Solved by the Invention

[0004] However, according to the prior art, although it is possible to calculate the similarity between words included in a sentence and group them, it is difficult to assign tags that match the content of the sentence with high accuracy, and in order to assign tags with high accuracy corresponding to the content of the sentence, an operator has to do it manually, so there is a problem that the processing burden on the operator increases.

[0005] Therefore, one of the objects of the present disclosure is to provide an information processing system and a tagging method that can improve the accuracy when assigning tags and reduce the processing burden on the operator.

Means for Solving the Problems

[0006] An information processing system according to one aspect of this disclosure includes: (1) dividing a tag generation document to generate multiple generation chunks; (2) inputting the generated multiple generation chunks into a large-scale language model, thereby generating a generation chunk summary sentence for each of the multiple generation chunks, which the large-scale language model has summarized based on the multiple generation chunks; (3) combining at least one of the multiple generation chunk summary sentences for each of the multiple generation chunks to generate multiple tag generation chunks corresponding to the multiple generation chunks; and (4) inputting the multiple tag generation chunks into the large-scale language model. The process is carried out in the order of steps (1) to (6), wherein the large-scale language model generates a tag structure in which tags are structured at an arbitrary hierarchy for a plurality of tag generation chunks, (5) divides the text to be tagged to generate a plurality of target chunks, inputs the generated plurality of target chunks into the large-scale language model to generate a tag assignment chunk based on the divided plurality of target chunks, and (6) inputs the tag assignment chunk and the tag structure into the large-scale language model so that the large-scale language model assigns the tags to the tag assignment chunk according to the tag structure.

[0007] In step (1) above, when generating a plurality of the generation chunks, the generation chunk may be generated by combining the end of the preceding generation chunk with its own generation chunk.

[0008] In step (3) above, when generating multiple tag generation chunks, the tag generation chunk summary statements corresponding to the generation chunks other than itself may be combined in the order of the generation chunks to generate the tag generation chunk.

[0009] In step (5) above, the large-scale language model may input a plurality of target chunks, causing the large-scale language model to generate a target chunk summary sentence for each of the plurality of target chunks based on the plurality of target chunks, and then combine the generated plurality of target chunk summary sentences to generate the tagging chunk.

[0010] In step (4) above, a list of tags for the multiple tag generation chunks may be input into the large-scale language model to generate a list of tags for the multiple tag generation chunks, and the large-scale language model may be input into the large-scale language model to generate the tag structure in which the tags are structured, by inputting the list of tags and a template structure in which the tags are pre-structured into an arbitrary hierarchy.

[0011] Furthermore, the process may include adjusting the tags at each level so that the number of tags assigned to the tagging chunk by the large-scale language model matches the number set at each level constituting the tag structure.

[0012] Furthermore, the process may include receiving input of importance levels for the tags constituting the tag structure, modifying the tag structure based on the tags for which importance levels have been received, and inputting the text to be assigned tags and the tag structure for which importance levels have been received into the large-scale language model, thereby enabling the large-scale language model to assign tags to the text to be assigned tags according to the tag structure adjusted by the importance levels.

[0013] Furthermore, the large-scale language model may perform a process that includes the step of assigning tags to the text to be assigned, by inputting a prompt to the large-scale language model that prevents the assignment of tags that are not present in the tag list, in addition to the text to be assigned and the tag structure.

[0014] Furthermore, the system may include a step of receiving a request for modification of the tags included in the tag structure based on an operator's operation, and modifying the tag structure based on the tags that have been modified based on the operator's operation.

[0015] Furthermore, if the target chunk has a number of tokens equal to or greater than a first number, the process may include the step of dividing the target chunk using a first algorithm, and if the target chunk has a number of tokens equal to or greater than a second number, the process may include the step of dividing the target chunk using a second algorithm.

[0016] The aforementioned text for tag generation may be text data generated by transcription.

[0017] The text used for tag generation and the text used for tag assignment may be the same text.

[0018] The text used for tag generation and the text to be assigned the tag may be different texts.

[0019] A tag - assigning method according to one aspect of the present disclosure includes: (1) splitting a tag - generating sentence to generate a plurality of generated chunks; (2) inputting the plurality of generated chunks into a large - language model, so that the large - language model generates a generated - chunk summary sentence summarized based on the plurality of generated chunks for each of the plurality of generated chunks; (3) for each of the plurality of generated chunks, generating a plurality of tag - generating chunks corresponding to the plurality of generated chunks by combining at least one or more of the plurality of generated - chunk summary sentences; (4) inputting the plurality of tag - generating chunks into the large - language model, so that the large - language model generates a tag structure in which tags are structured at an arbitrary level for the plurality of tag - generating chunks; (5) splitting a sentence for an object to be tagged to generate a plurality of target chunks, and generating tag - assigning chunks based on the plurality of split target chunks by inputting the plurality of generated target chunks into the large - language model; (6) inputting the tag - assigning chunks and the tag structure into the large - language model, so that the large - language model assigns the tags according to the tag structure to the tag - assigning chunks, and performing processing in the order of the steps (1) to (6).

Advantages of the Invention

[0020] According to one aspect of the present disclosure, it is possible to improve the accuracy when assigning tags and reduce the processing burden on the operator.

Brief Description of the Drawings

[0021] [Figure 1] It is a diagram showing an example of the schematic configuration of an information processing system according to one embodiment. [Figure 2] It is a diagram showing an example of a flowchart of tag generation and assignment processing for implementing a tag - assigning method according to one embodiment. [Figure 3] It is a diagram showing an example of generating a plurality of tag - generating chunks based on a tag - generating sentence. [Figure 4] It is a diagram showing an example of generating a tag list based on a plurality of tag - generating chunks. [Figure 5] It is a diagram showing an example of the generated tag list. [Figure 6] It is a diagram showing an example of the template structure. [Figure 7] It is a diagram showing an example of the tag structure generated by the large language model. [Figure 8] It is a diagram showing an example of generating a tag - attaching chunk based on the target chunk divided by the first algorithm. [Figure 9] It is a diagram showing an example of generating a tag - attaching chunk based on the target chunk divided by the first algorithm and the target chunk divided by the second algorithm. [Figure 10] It is a diagram showing an example of the flowchart of the tag - attaching chunk generation process implemented in step S110 of FIG. 2. [Figure 11] It is a diagram showing an example of the tags attached to the text for the attachment target. [Figure 12] It is a diagram showing an example of the functional configuration of the information processing server according to an embodiment. [Figure 13] It is a diagram showing an example of the hardware configuration of an information processing server and the like according to an embodiment.

Embodiments for Carrying Out the Invention

[0024] Embodiments of this disclosure will be described in detail below with reference to the attached drawings. In the following description, identical parts will be denoted by the same reference numerals. Since identical parts have the same name, function, etc., detailed descriptions will not be repeated.

[0025] (Information Processing System) Figure 1 is a diagram showing an example of the schematic configuration of an information processing system according to one embodiment. The information processing system 1 shown in Figure 1 includes a device 10 and an information processing server 20.

[0026] Device 10 may be a fixed communication terminal such as a personal computer (PC) or a server, or it may be a mobile communication terminal such as a mobile phone, smartphone, or tablet. In other words, device 10 in this disclosure can be read as a communication device.

[0027] Device 10 uses wired and / or wireless (e.g., Long Term Evolution (LTE), 5 th Communication may be made with the information processing server 20 or external servers (not shown) via network N through Generation New Radio (5G NR), Wi-Fi (registered trademark), etc. Network N consists of the internet or dedicated lines.

[0028] The information processing server 20 may be a fixed communication terminal such as a personal computer or server. The information processing server 20 may also be a server for receiving input instructions from an operator entered into an external server. Note that the server in this disclosure may be read as an apparatus, device, circuit, etc. The information processing server 20 stores the tag list 110, template structure 120, and tag structure 130 described later. The information processing server 20 also has the function of providing an API for the large-scale language model 50.

[0029] The information processing server 20 may communicate with device 10 or external servers via the network N, either via wired and / or wireless, or Wi-Fi.

[0030] An example of the functional and hardware configurations of each device, such as device 10 and information processing server 20, will be described later.

[0031] Note that the system configuration described is merely an example and is not limited to this. For example, although Figure 1 shows one of each device, the number of each device is not limited to this and there may be multiples. The information processing system 1 may also have a configuration that omits some devices, or a configuration in which the function of one device is realized by multiple devices.

[0032] The configuration may involve multiple devices having their functions implemented by a single device. For example, device 10 and information processing server 20 may be implemented on a single server.

[0033] (How to add tags) A tagging method according to one embodiment of this disclosure will be described below. The tagging method may be applied to the information processing system 1 described above. The tagging method performed by the information processing system 1 will be described below. The tagging method is performed by a tag generation and assignment process. In this embodiment, a tag structure is generated by performing the tag generation and assignment process. Then, by performing the tag generation and assignment process based on the generated tag structure, the accuracy of tag assignment can be improved and the processing burden can be reduced. The tagging method may be performed by a process other than the tag generation and assignment process.

[0034] <Tag generation and assignment process> Figure 2 shows an example of a flowchart of the tag generation and assignment process for implementing a tag assignment method according to one embodiment.

[0035] First, the information processing server 20 transcribes existing documents (also called "materials"), text information such as documents, or audio information such as recorded data, based on input operations from the device 10 (step S101). Documents may include PDF (Portable Document Format), JPEG (Joint Photographic Experts Group), GIF (Graphics Interchange Format), TIFF (Tagged Image File Format), HTML (HyperText Markup Language) files, etc. Text may include not only text in the files exemplified above, but also text in files containing web articles, blog posts, emails, etc. Documents are not limited to PDF or JPEG, but may be any file format containing text data. By performing OCR (Optical Character Recognition) on the document, text data can be extracted from the text, transcribed, and converted into text to generate tag generation text 60. If the document has already been converted into text, the transcription process can be omitted, and the document can be used directly as tag generation text 60.

[0036] The information processing server 20 divides the tag generation text 60 into arbitrary units to generate multiple generation chunks 61 to 64 (step S102). The tag generation text 60 is a training text used to generate a tag structure 130 for assigning tags to the tag target text 260 described later. The tag generation text 60 and the tag target text 260 used in the tag assignment described later can be the same text. This makes it possible to improve the accuracy of assigning tags that are appropriate to the content of the text by performing the tag generation and assignment process on the same text from which the tag structure 130 was generated based on the previously generated tag structure 130. Furthermore, since the processing is performed almost automatically, the processing burden on the worker can be reduced.

[0037] Note that the tag generation text 60 used in tag generation and the tag assignment target text 260 used in tag assignment may be different texts. If different texts are used for the tag generation text 60 and the tag assignment target text 260, it is preferable that the tag generation text 60 and the tag assignment target text 260 belong to the same type of technical field. In this embodiment, the same type of technical field refers to the same technical field, such as medicine, pharmaceuticals, chemistry, physics, and law. By performing the tag assignment process based on the tag structure 130 generated based on the tag generation text 60 in the same type of technical field that has been generated in advance, it is possible to use consistent answers that are not influenced by random answers generated by the large-scale language model 50, thereby improving the accuracy of assigning tags that are appropriate to the content of the text and reducing the processing burden on the operator.

[0038] In this embodiment, any unit can be any number of tokens, any number of characters, any number of words, or any number of sentences. In this embodiment, the information processing server 20 divides the tag generation text 60 into any number of tokens.

[0039] In this embodiment, a token refers to a basic element that constitutes a text, such as a word, punctuation mark, or symbol. If the tag generation text 60 has fewer than any number of tokens, the information processing server 20 may use the original tag generation text 60 as is without splitting it. A number of tokens less than 10,000 can be used as the arbitrary number of tokens. Preferably, a number of tokens less than 5,000 can be used as the arbitrary number of tokens. Furthermore, a number of tokens less than 3,000 can also be used as the arbitrary number of tokens. This prevents over-summarization of the original text and enables the generation of appropriate tags. Splitting the tag generation text 60 can be performed by any program. For example, in this embodiment, Python® can be used as the program. The program is not limited to Python, and any known program can be used. The information processing server 20 splits the tag generation text 60 to generate multiple generation chunks 61 to 64. In this embodiment, the information processing server 20 generates four chunks as the multiple generation chunks, but this is not limited to this; at least two or more chunks can be generated.

[0040] Figure 3 shows an example of generating multiple tag generation chunks based on tag generation text. As shown in Figure 3, when the information processing server 20 generates multiple generation chunks 61 to 64, it may combine the end of the sentence 61A of the preceding generation chunk 61 with the beginning of its own generation chunk 62 to generate generation chunk 62. Similarly, for other chunks, it may combine the end of the sentence 62A of the preceding generation chunk 62 with the beginning of its own generation chunk 63 to generate generation chunk 63. This allows the contextual content of the generation chunks 61 to 64 before and after the division to be connected, thereby improving the accuracy of tag assignment.

[0041] The information processing server 20 inputs the generated chunks 61-64 into the large language model 50 (LLM). The generated chunks 61-64 input into the large language model 50 are generated chunks created by concatenating the end of the sentence of the preceding generated chunk with the beginning of the sentence of the generated chunk itself.

[0042] In this embodiment, the large-scale language model is an example of Generative AI. For example, ChatGPT® can be used as the large-scale language model. In this embodiment, the information processing server 20 has the function of providing an API (application programming interface) for the large-scale language model 50, such as ChatGPT, and hosting the large-scale language model. The large-scale language model is not limited to ChatGPT, but can be any computer language model that can understand and generate natural language. The information processing server 20 is not limited to providing an API for the large-scale language model, but may also execute the functions of the large-scale language model itself.

[0043] The information processing server 20 inputs multiple generated chunks 61 to 64 to the large-scale language model 50, which then generates generated chunk summary sentences 81 to 84 based on the multiple generated chunks 61 to 64 (step S103).

[0044] The information processing server 20 is expected to input the tag generation text 60 into the large-scale language model 50 to generate a tag structure 130. However, there is a limit to the number of tokens that can be input into the large-scale language model 50 when generating the tag structure 130. Therefore, in this embodiment, in order to generate the tag structure 130, the information processing server 20 pre-divides the tag generation text 60 to generate multiple generation chunks 61 to 64. The information processing server 20 then inputs each of the divided generation chunks 61 to 64 into the large-scale language model 50 and can generate generation chunk summary sentences 81 to 84 with a reduced number of tokens by summarizing (compressing) each of them.

[0045] As shown in Figure 3, the information processing server 20 inputs multiple generation chunks 61-64 into the large-scale language model 50, which then generates generation chunk summary sentences 81-84 based on each of the multiple generation chunks 61-64. The information processing server 20 then combines the generated chunk summary sentences 81-84 into the divided generation chunks 61-64 to generate tag generation chunks 71-74. This reduces the number of tokens in the tag generation text 60, allowing tag generation chunks 71-74 to be generated regardless of the limit on the number of tokens that can be input into the large-scale language model 50.

[0046] Furthermore, since generated chunks 61 to 64 are derived from fragmented tag generation text 60, there is a possibility that the meaning may be fragmented between preceding and succeeding chunks. Therefore, in this embodiment, when the information processing server 20 divides tag generation text 60 to generate the preceding generated chunk 61, and also generates its own generated chunk 62, it can combine the end of sentence 61A of the preceding generated chunk 61 with the beginning of its own generated chunk 62 to generate its own generated chunk 62. Similarly, when the information processing server 20 divides tag generation text 60 to generate the preceding generated chunk 62, and also generates its own generated chunk 63, it can combine the end of sentence 62A of the preceding generated chunk 62 with its own generated chunk 63 to generate its own generated chunk 63. Similarly, when the information processing server 20 divides tag generation text 60 to generate the preceding generated chunk 63, and also generates its own generated chunk 64, it can combine the end of sentence 63A of the preceding generated chunk 63 with its own generated chunk 64 to generate its own generated chunk 64. This allows the content of the preceding and succeeding generation chunks to overlap, preventing the semantic content from being fragmented between them.

[0047] For example, the sentence endings 61A, 62A, and 63A of the preceding generated chunks 61, 62, and 63 contain at least one word from the content described in the preceding generated chunks 61, 62, and 63. The sentence endings 61A, 62A, and 63A may be sentences, phrases, or single words. The sentence endings 61A, 62A, and 63A may represent a few percent to several tens of percent of the content described in the preceding generated chunks 61, 62, and 63. More preferably, the sentence endings 61A, 62A, and 63A may represent 10% of the content described in the preceding generated chunks 61, 62, and 63.

[0048] This allows the generation chunk summary sentences 81-84 to be generated while maintaining the context of the generation chunks 61-64 before and after the division, even when the tag generation text 60 is divided into multiple generation chunks 61-64. Therefore, the generation chunk summary sentences 81-84 can be generated while maintaining the overall content of the tag generation text 60, thereby increasing the reliability of the tag structure 130 used when assigning tags.

[0049] In the above-described embodiment, when the information processing server 20 generates the preceding generation chunk 61 and its own generation chunk 62, it generates the chunk by concatenating the end 61A of the preceding generation chunk 61 with the beginning of the chunk 62. However, this is not limited to this. For example, when the information processing server 20 generates the preceding generation chunk 61 and its own generation chunk 62, it may concatenate the beginning of the preceding generation chunk 61 with the beginning or end of the chunk 62. Alternatively, for example, when the information processing server 20 generates the preceding generation chunk 61 and its subsequent generation chunk 62, it may concatenate the end 61A and beginning of the preceding generation chunk 61 with the beginning or end of the chunk 62. This allows the content of preceding and succeeding generation chunks to overlap, preventing the semantic content from being fragmented between them.

[0050] Furthermore, if the generated generation chunk summaries 81-84 are combined with generation chunks 61-64 respectively to generate tag generation chunks 71-74, the content of tag generation text 60 may be summarized too much, resulting in the loss of too much of the content of articles and other texts included in tag generation text 60. Therefore, the information processing server 20 may use a Python (registered trademark) algorithm to specify that the number of tokens in tag generation chunks 71-74 be within an arbitrary range.

[0051] Returning to the tag generation and assignment process in Figure 2, the information processing server 20 generates multiple tag generation chunks 71 to 74 corresponding to multiple generation chunks 61 to 64 by combining at least one of the multiple generation chunk summary sentences 81 to 84 for each of the multiple generation chunks 61 to 64 (step S104).

[0052] Referring to Figure 3, the case of generating tag generation chunk 72 among multiple tag generation chunks 71 to 74 will be explained. The information processing server 20 combines its own generated chunk 62, which has the end sentence 61A of the previous generated chunk 61 concatenated to its beginning, with generated chunk summary sentences 81, 83, and 84 generated based on generated chunks 61, 63, and 64 other than its own generated chunk 62. The order in which the generated chunk summary sentences 81, 83, and 84 are combined with its own generated chunk 62 corresponds to the order of generated chunks 61 to 64.

[0053] In this embodiment, generated chunks 61 to 64 are generated based on the tag generation text 60 in the order of generated chunk 61, generated chunk 62, generated chunk 63, and generated chunk 64. Therefore, when the information processing server 20 generates a tag generation chunk 72, it combines the generated chunk summary sentence 81, which was generated based on generated chunk 61, with the beginning of its own generated chunk 62, which has the sentence end 61A of the preceding generated chunk 61 joined to it. It then combines the generated chunk summary sentence 83, which was generated based on generated chunk 63, and the generated chunk summary sentence 84, which was generated based on generated chunk 64, with the end of its own generated chunk 62, which has the sentence end 61A of the preceding generated chunk 61 joined to it, in that order. Thus, the generated tag generation chunk 72 is generated by combining the generated chunk summary sentence 81, the sentence end 61A of the preceding generated chunk 61, its own generated chunk 62, the generated chunk summary sentence 83, and the generated chunk summary sentence 84 in that order. This allows the contextual content of the generation chunks before and after the tag generation text 60 to be sequentially linked, even when the tag generation text 60 is split, thereby improving the accuracy of tag assignment.

[0054] When the information processing server 20 inputs the multiple tag generation chunks 71 to 74 generated in step S104 into the large-scale language model 50, the large-scale language model 50 generates tags (Tag) contained in the multiple tag generation chunks 71 to 74 based on the multiple tag generation chunks 71 to 74, and generates a tag list 110 based on the generated tags (step S105). The generated tag list 110 is stored in the information processing server 20. The generated tag list 110 may also be stored in an external device other than the information processing server 20.

[0055] Figure 4 shows an example of generating a tag list 110 based on multiple tag generation chunks 71-74. The information processing server 20 inputs the generated tag generation chunks 71-74 into the large-scale language model 50. When inputting the tag generation chunks 71-74 into the large-scale language model 50, a tag generation prompt is also input. An example of a tag generation prompt is the instruction, "Generate any tag from the tag generation chunks." In response to the given prompt, the large-scale language model 50 performs the process of generating a tag list 110 based on the multiple tag generation chunks 71-74.

[0056] When tag generation chunks 71-74 and a tag generation prompt are input to the large-scale language model 50, the large-scale language model 50 generates the tags Tag contained in tag generation chunks 71-74. Tag generation by the large-scale language model 50 is performed repeatedly by looping through one of the multiple tag generation chunks 71-74 multiple times. The number of loops can be 2 to 10 times, depending on parameters such as the size and type of the chunk. This allows for the generation of appropriate tags Tag. More preferably, the number of loops can be 3 to 5 times. This allows for the generation of more favorable tags Tag compared to the case of 2 to 10 loops. The process of looping through tag generation chunks to generate tags Tag generates tags independently, without referring to past generation results. This allows for the random generation of unique tags Tag. Then, by merging the randomly generated unique tags Tag into a single table, a tag list 110 can be generated. In this embodiment, the tag list 110 merges each tag into a single table, but this is not limited to this; the tag list 110 may also be generated by distributing each tag into multiple tables and merging them therein.

[0057] Figure 5 shows an example of a generated tag list 110. The tag list 110 includes multiple tags (Tag) generated based on tag generation chunks 71-74. In this embodiment, a tag (also called a "keyword") consists of characters such as important words or phrases to represent the tag generation text 60. Therefore, by visually inspecting the tag, the operator can easily grasp the content of the text containing the tag. Furthermore, the operator can use the tag as search information, using it as an index when searching for texts containing the tag. In this embodiment, the tag is composed of characters, but is not limited to this; it may also be identification information that can be arbitrarily identified by the operator or the information processing server 20, such as numbers, symbols, colors, or sounds.

[0058] The information processing server 20 inputs the tags included in the generated tag list 110 into the large-scale language model 50 and performs tag cleansing (step S106). The generated tag list 110 may contain unnecessary tags such as mere symbols, meaningless words, duplicate tags, or similar words that are merely variations in spelling. Therefore, a tag cleansing process is performed on the tags included in the tag list 110 to remove unnecessary tags and generate a tag list 110 that contains only unique and simple tags. In this embodiment, tag cleansing includes the process of removing unnecessary tags from multiple tags included in the tag list 110.

[0059] As a result of generating the tag list 110, a “complete sentence including punctuation” may be included as a tag in the tag list 110. A sentence tag like “complete sentence including punctuation” is inappropriate as a tag. Therefore, in the tag cleansing process, the information processing server 20 may input multiple tags included in the tag list 110 into the large-scale language model 50 to detect whether or not a sentence tag like “complete sentence including punctuation” is included as a tag in the tag list 110. If it is detected that a sentence tag like “complete sentence including punctuation” is included as a tag, the information processing server 20 may re-execute the rule-based process of generating the tag list 110 for “complete sentence including punctuation,” tag it as an appropriate tag, and include it in the tag list 110. This makes it possible to generate the tag list 110 while excluding inappropriate tags, and improves the accuracy of tag assignment. The process of tagging "complete sentences including punctuation" with appropriate tags does not need to be executed in a loop; it can be performed as a one-time process.

[0060] Furthermore, as a result of generating the tag list 110, "tags consisting only of symbols" may be included in the tag list 110. "Tags consisting only of symbols" are inappropriate as tags. Therefore, the information processing server 20 may use a program such as Python to perform a deletion process using regular expressions during the tag cleansing process. This makes it possible to generate the tag list 110 while eliminating inappropriate symbols, thereby improving the accuracy of tag assignment.

[0061] In this embodiment, tag cleansing is performed by inputting the data into the large-scale language model 50, but this is not limited to this; the information processing server 20 itself may perform the cleansing without inputting the data into the large-scale language model 50. By performing tag cleansing, it is possible to prevent the assignment of unnecessary tags and improve the accuracy of tag assignment.

[0062] The information processing server 20 inputs the generated tag list 110 and a template structure 120 in which the tags are pre-structured into an arbitrary hierarchy into the large-scale language model 50. The template structure 120 is tabular information in which the tags are pre-structured into an arbitrary hierarchy.

[0063] Figure 6 shows an example of a template structure 120. In the template structure 120, each tag is stored structured in an arbitrary hierarchy. In the embodiment shown in Figure 6, each tag is stored structured in three hierarchies: major category, medium category, and minor category. In the embodiment shown in Figure 6, the major category represents a higher hierarchy than the medium category, and the medium category represents a higher hierarchy than the minor category. In the embodiment of Figure 6, the template structure is structured in three hierarchies: major category, medium category, and minor category, but this is not limited to this; it is sufficient to be structured in at least two or more hierarchies. The template structure 120 is information that serves as a guideline for the large-scale language model 50 to generate a tag structure 130 by structuring the tags included in the tag list 110 into multiple hierarchies. Tags do not need to be included in all hierarchies that constitute the template structure 120. The template structure 120 only needs to include a tag hierarchy structure that serves as a minimum guideline for the large-scale language model 50 to generate the tag structure 130.

[0064] The template structure 120 is the prototype structure for the tag tree structure. The template structure 120 is also called the seed structure. The template structure 120 is stored in the information processing server 20. The template structure may also be stored in a device other than the information processing server 20.

[0065] When the information processing server 20 inputs the tag list 110 generated and cleansed in step S105, and the template structure 120 in which the tags are pre-structured into an arbitrary hierarchy, into the large-scale language model 50, the large-scale language model 50 generates a tag structure 130 in which the tags are structured (step S107). The tag structuring is performed based on multiple tags Tag included in the tag list 110. The generated tag structure 130 is stored in the information processing server 20. The generated tag structure 130 may also be stored in an external device other than the information processing server 20.

[0066] Figure 7 shows an example of a tag structure 130 generated by the large-scale language model 50. In the tag structure 130, each tag assigned to each hierarchy according to the template structure 120 is structured and stored in an arbitrary hierarchy. In the embodiment shown in Figure 7, each tag is assigned to and structured and stored in three hierarchies: major category, medium category, and minor category. In the embodiment shown in Figure 7, the major category represents a higher hierarchy than the medium category, and the medium category represents a higher hierarchy than the minor category. In the embodiment of Figure 7, the tag structure 130 is structured in three hierarchies, but is not limited to this; it may be structured in at least two or more hierarchies. Since the tag structure 130 is generated based on the template structure 120, it is preferable that it is composed of the same number of hierarchies as the hierarchies that make up the template structure 120. In this embodiment, since the template structure 120 is composed of three hierarchies: major category, medium category, and minor category, the tag structure 130 is also composed of three hierarchies: major category, medium category, and minor category, similar to the hierarchies that make up the template structure 120. The tag structure 130 is not limited to the same number of layers as the template structure 120; it may consist of more layers or fewer layers than the template structure 120. The tag structure 130 is information that serves as a guideline for the large-scale language model 50 when assigning tags to the target document. As in this embodiment, by structuring the tag structure 130, which classifies tags by semantic level of abstraction, with multiple layers, the number of characters in the prompt input to the large-scale language model 50 can be reduced.

[0067] Furthermore, as shown in Figure 7, a column of error messages may be formed corresponding to each hierarchy. For example, even when a tag list 110 is input into the large-scale language model 50, some tags will fail to be structured into an arbitrary hierarchy. Such tags are temporarily placed into an arbitrary hierarchy (e.g., a sub-category hierarchy), and the large-scale language model 50 executes a program to write the tags it could not structure into the corresponding column of error messages. By visually checking whether there are any tags that could not be structured in the column of error messages, the operator of the information processing server 20 can easily and immediately identify which tags failed to structure and resulted in an error.

[0068] As an example of a prompt given when the information processing server 20 inputs the generated tag list 110 and a template structure 120 in which the tags are pre-structured into arbitrary hierarchies into the large-scale language model 50, the instruction "Refer to the template structure and assign the generated tag list to the appropriate positions" is given. In response to the given prompt, the large-scale language model 50 performs the process of assigning the tags included in the tag list 110 to each level of the tag structure 130 according to the template structure 120.

[0069] For example, the tag list 110 shown in Figure 5 includes the tag "migraine," but the tag "migraine" is not included in the template structure 120 shown in Figure 6. Therefore, as shown in Figure 7(1), the tag "migraine" is not included in the subcategory hierarchy in the initially generated tag structure 130. Accordingly, the large-scale language model 50 uses the template structure 120 shown in Figure 6 as a reference and assigns the tag "migraine" to the subcategory hierarchy, as shown in Figure 7(2), according to tags such as "diagnosis" included in the major category hierarchy, "symptoms" included in the intermediate category hierarchy, or "headache" included in the minor category hierarchy, to generate the tag structure 130.

[0070] Similarly, although the tag list 110 shown in Figure 5 includes the tag "Efficacy Evaluation," the tag "Efficacy Evaluation" is not included in the template structure 120 shown in Figure 6. Therefore, as shown in Figure 7(1), the tag "Efficacy Evaluation" is not included in the subcategory hierarchy in the initially generated tag structure 130. Accordingly, the large-scale language model 50, referring to the template structure 120 shown in Figure 6, assigns the tag "Efficacy Evaluation" to the subcategory hierarchy, as shown in Figure 7(2), according to tags such as "Product Information" included in the major category hierarchy, "Mechanism of Action" included in the subcategory hierarchy, or "Vomiting" included in the minor category hierarchy, and generates the tag structure 130. The information processing server 20 continues the process of generating the tag structure 130 by step S107 until all tags included in the tag list 110 have been assigned to the tag structure 130 by the large-scale language model. Once all tags included in the tag list 110 have been assigned to the tag structure 130, the process proceeds to step S108. In this way, even for tags not included in the template structure 120, the large-scale language model 50 can easily expand the tree structure according to the template structure 120 to generate the tag structure 130, thereby improving the processing speed when assigning tags and reducing the workload of the user.

[0071] Generating a tag structure by manually creating a tree structure based on individual tags requires effort and could lead to increased labor costs. On the other hand, if a large-scale language model 50 is made to generate a tag structure from scratch based on individual tags without any guiding information, it is difficult to obtain a uniform output because the output of the large-scale language model 50 will vary. Therefore, in this embodiment, a minimum prototype tree structure template structure 120 that serves as a guideline for the tag structure 130 is input to the large-scale language model 50 along with the tag list 110. As a result, the large-scale language model 50 can understand which hierarchy the target tag is in, assign it, and hierarchically organize individual tags according to the template structure 120, automatically generating and expanding the tree structure of the tag structure 130. This makes it possible to automatically generate a tag structure 130 that can improve the accuracy of tag assignment, reduce the processing burden on workers, and lower labor costs.

[0072] Once the generation of the tag structure 130 is complete, the information processing server 20 assigns importance to the tags by having the large-scale language model 50 determine the importance of the tags in N stages (step S108). As shown in Figure 7, the tag structure 130 has importance columns corresponding to any hierarchy. In the embodiment of Figure 7, importance columns are formed at the sub-category hierarchy. For example, for a specific tag, the importance of tags corresponding to the large-scale language model 50 may be determined at N stages. Importance is an example of a parameter for a tag. In Figure 7, the tag in the sub-category "headache" is judged to have an importance of "high" and is assigned to it. Similarly, the tag in the sub-category "MRI" is judged to have an importance of "low" and is assigned to it. In the embodiment of Figure 7, importance columns are formed only at the sub-category hierarchy, but this is not limited to this, and importance can be assigned to at least one of the major, medium, and minor categories. In this case, importance columns corresponding to the hierarchy to which importance is assigned are formed in the tag structure 130.

[0073] Furthermore, the system may accept input of importance levels from the operator for the tags that make up the tag structure 130. The information processing server 20 assigns importance levels to each tag included in the tag structure 130 based on the importance information received. The information processing server 20 modifies the tag structure 130 based on the assigned importance levels. The importance level is information that serves as a guideline when assigning tags to the target documents. Input of importance levels for tags may be accepted based on the operator's input operation to device 10. Input of importance levels for tags may also be accepted directly by the information processing server 20 based on the operator's input operation to the information processing server 20. Importance levels can be set using numbers, words, identification symbols, etc. For example, the system may be set so that the higher the importance level, the higher the probability of the tag being assigned, and the lower the importance level, the lower the probability of the tag being assigned, or it may be set so that the lower the importance level, the higher the probability of the tag being assigned, and the higher the importance level, the lower the probability of the tag being assigned. The tag structure 130, which has been assigned importance levels and modified, is stored in the information processing server 20. The tag structure 130, which has been assigned importance levels and modified, may also be stored in a device other than the information processing server 20. Once the tag structure 130 is generated, the processing from step S109 onward is executed. That is, the processing from step S109 onward is executed after the tag structure 130 has been generated.

[0074] Once the tag structure 130 is generated, the information processing server 20 divides the text 260 to be tagged into arbitrary units to generate multiple target chunks 261 to 263 (step S109).

[0075] The text to be tagged 260 is the text that is to be processed and for which tags are required. The text to be tagged 260 can be any text, and is not limited to any file format as long as it is a file containing text data, such as an HTML file, web article, blog post, or email. The text to be tagged 260 to be split is a text. If the text to be tagged 260 is not in text format, it may be transcribed beforehand and converted into text before being used as the text to be tagged 260. In addition, the text to be tagged 260 may be text information such as documents (materials) or text input for transcription in step S101 of Figure 2 above, or text information converted from audio information such as recorded data (it may be reinterpreted).

[0076] In this embodiment, any unit can be any number of tokens, any number of characters, any number of words, or any number of sentences. In this embodiment, the information processing server 20 divides the document to be assigned 260 into any number of tokens.

[0077] The information processing server 20 inputs the divided target chunks 261 to 263 to the large-scale language model 50, which then generates chunk summary sentences 281 to 283 of the text to be tagged based on the multiple target chunks 261 to 263, and generates a tagging chunk 264 based on the generated chunk summary sentences 281 to 283 (step S110).

[0078] The information processing server 20 is expected to input the text to be tagged 260 into the large-scale language model 50 and assign tags to it. However, there is a limit to the number of tokens that can be input into the large-scale language model 50 when assigning tags. Therefore, in this embodiment, in order to assign tags, the text to be tagged 260 is divided into a plurality of target chunks 261 to 263 in advance. The information processing server 20 can then input each of the divided target chunks 261 to 263 into the large-scale language model 50 and generate a tag-assignment chunk 264 with a reduced number of tokens by summarizing (compressing) each of them.

[0079] Figure 8 shows an example of generating a tagging chunk 264 based on target chunks 261-263 divided by the first algorithm (algorithm A). As shown in Figure 8, the information processing server 20 divides the text to be tagged 260 into multiple target chunks 261-263 using the first algorithm. Then, by inputting the divided target chunks 261-263 into the large-scale language model 50, the large-scale language model 50 generates chunk summary sentences 281-283 based on the multiple target chunks 261-263. The information processing server 20 then combines the generated chunk summary sentences 281-283 to generate a tagging chunk 264. This reduces the number of tokens in the text to be tagged 260, and allows the generation of a tagging chunk 264 regardless of the limit on the number of tokens that can be input to the large-scale language model 50.

[0080] Furthermore, since the target chunks 261 to 263 are derived from fragmented text 260, there is a possibility that the meaning may be fragmented between the preceding and succeeding chunks. Therefore, in this embodiment, when the information processing server 20 generates the preceding target chunk 261 and its own target chunk 262, it can combine the end of sentence 271 of the preceding target chunk 261 with the beginning of sentence 262. Similarly, when the information processing server 20 generates the preceding target chunk 262 and its own target chunk 263, it can combine the end of sentence 272 of the preceding target chunk 262 with the beginning of sentence 263. This allows the content of the preceding and succeeding target chunks to overlap, preventing the meaning from being fragmented between them.

[0081] The sentence end 271 of the target chunk 261 in the preceding paragraph is a text containing at least one word from the content described in the target chunk 261 in the preceding paragraph. Similarly, the sentence end 272 of the target chunk 262 in the preceding paragraph is a text containing at least one word from the content described in the target chunk 262 in the preceding paragraph. The sentence ends 271 and 272 may be sentences, phrases, or single words. The sentence ends 271 and 272 may represent a few percent to several tens of percent of the content described in the target chunks 261 and 262 in the preceding paragraph. More preferably, the sentence ends 271 and 272 may represent 10% of the content described in the target chunks 261 and 262 in the preceding paragraph.

[0082] This allows the tag-assigning chunk 264 to be generated while maintaining the context of the target chunks 261-263 before and after the split, even when the target text 260 is divided into multiple target chunks 261-263. Therefore, the tag-assigning chunk 264 can be generated while maintaining the overall content of the target text 260, thereby improving the accuracy of tagging based on the tag structure 130.

[0083] Furthermore, if the generated chunk summaries 281 to 283 are combined to create a tagging chunk 264, the content of the text to be tagged 260 may be summarized too much, resulting in the loss of too much of the content of articles and other texts contained within the text to be tagged 260. Therefore, if the number of tokens in the tagging chunk 264 is equal to or greater than the first number of tokens, the information processing server 20 may divide it into chunks using the first algorithm so that it falls within a first predetermined range, as shown in Figure 8.

[0084] For example, the predetermined range of the first number of tokens in the tagging chunk 264 generated by the first algorithm is adjusted to be between 3,000 and less than 10,000 tokens. This prevents excessive loss of content from the target text 260 even when a tagging chunk 264 is generated, and makes it possible to generate a tagging chunk 264 of an appropriate length for handling by the large-scale language model 50.

[0085] In the embodiment described above, the information processing server 20 adjusts the number of tokens in the tagging chunk 264 using a first algorithm so that it falls within a first predetermined range, but this is not limited to this. For example, the information processing server 20 may adjust using a second algorithm different from the first algorithm, as shown in Figure 9. As shown in Figure 9, the first algorithm and the second algorithm may be executed together, or they may be executed individually.

[0086] Figure 9 shows an example of generating a tag-assignment chunk 264 based on target chunks 261-263 divided by the first algorithm (Algorithm A) and target chunks 361-36n divided by the second algorithm (Algorithm B).

[0087] As shown in Figure 9, the information processing server 20 divides the text 260 to be assigned using a first algorithm to generate multiple target chunks 261 to 263, and inputs the generated multiple target chunks 261 to 263 into the large-scale language model 50, which then generates chunk summary sentences 281 to 283 based on the multiple target chunks 261 to 263.

[0088] The information processing server 20 divides the text 260 to be tagged using a second algorithm to generate multiple target chunks 361-36n, and inputs these multiple target chunks 361-36n into the large-scale language model 50, which then generates chunk summary sentences 381-38n based on the multiple target chunks 361-36n. The information processing server 20 then combines the chunk summary sentences 281-283 generated by the first algorithm and the chunk summary sentences 381-38n generated by the second algorithm to generate a tag-attachment chunk 264. The total number of tokens in the chunk summary sentences 381-38n generated by the second algorithm is shorter than the total number of tokens in the chunk summary sentences 281-283 generated by the first algorithm. This allows the content of the tag-assignment chunk 264 to be supplemented by combining the chunk summaries 381-38n generated by the second algorithm, even if the number of tokens in the chunk summaries 281-283 generated by the first algorithm is too short. Therefore, it is possible to prevent the loss of meaning in the text 260 to be tagged.

[0089] Furthermore, similar to the first algorithm explained in Figure 8, the target chunks 361 to 36n generated in the second algorithm are derived from a fragmented text 260, which may result in a fragmentation of meaning between preceding and succeeding chunks. Therefore, in the second algorithm, as in the first algorithm, the information processing server 20 divides the text 260 to generate the preceding target chunk 361, and when generating its own target chunk 362, it can combine the sentence ending 371 of the preceding target chunk 361 with its own target chunk 362. Similarly, when generating the preceding target chunk 362 and its own target chunk 36n, the information processing server 20 can combine the sentence ending 372 of the preceding target chunk 362 with its own target chunk 36n. This allows for overlapping content between preceding and succeeding target chunks, preventing a fragmentation of meaning between them.

[0090] The choice between the first and second algorithms can be determined by the summary generation process shown in Figure 10. Figure 10 is an example of a flowchart of the tagging chunk generation process performed in step S110 of Figure 2.

[0091] First, the information processing server 20 determines whether the total number of tokens in the target chunks 261-263 (or the text to be tagged 260) is 10,000 or more (step S231). If the total number of tokens in the target chunks 261-263 (or the text to be tagged 260) is less than 10,000 (step S231: NO), there is no need to summarize further, so the information processing server 20 does not generate a summary text using the first or second algorithm, and uses the text contained in the target chunks 261-263 (or the text to be tagged 260) as is for tagging chunk 264 (step S232).

[0092] If the total number of tokens in the target chunks 261-263 (or the text 260 to be assigned) is 10,000 or more (step S231: YES), the information processing server 20 inputs the target chunks 261-263 into the large-scale language model 50, and the first algorithm summarizes the target chunks 261-263 to generate chunk summary sentences 281-283 (step S233). The information processing server 20 determines whether the total number of tokens in the chunk summary sentences 281-283 generated by the first algorithm is equal to or greater than the first number of tokens (step S234). The first number of tokens can be set to any number of tokens. In this embodiment, for example, 3,000 tokens are set as the first number of tokens. If the total number of tokens in the chunk summaries 281 to 283 generated by the first algorithm is 3,000 or more (step S234: YES), then the information processing server 20 combines the chunk summaries 281 to 283 generated by the first algorithm to create a tagging chunk 264, since the chunk summaries have been generated with an appropriate length (3,000 to 10,000 tokens). (step S235)

[0093] If the total number of tokens in the generated chunk summary sentences 281-283 is less than 3,000 (step S234: NO), the information processing server 20 inputs the target chunks 361-36n into the large-scale language model 50, and the second algorithm summarizes the target chunks 361-36n to generate chunk summary sentences 381-38n (step S236). The information processing server 20 determines whether the total number of tokens in the chunk summary sentences 281-283 generated by the first algorithm and the total number of tokens in the chunk summary sentences 381-38n generated by the second algorithm is greater than or equal to the second number of tokens and less than the third number of tokens (step S237). The second number of tokens and the third number of tokens can be set to any number of tokens. In this embodiment, for example, 3,000 tokens are set as the second number of tokens, and for example, 10,000 tokens are set as the third number of tokens. The second number of tokens is set to a value smaller than the third number of tokens. This makes it possible to make the total number of tokens in the chunk summaries 381-38n generated by the second algorithm smaller than the total number of tokens in the chunk summaries 281-283 generated by the first algorithm.

[0094] If the total number of tokens in the chunk summary sentences 281-283 generated by the first algorithm and the total number of tokens in the chunk summary sentences 381-38n generated by the second algorithm is not between 3,000 and 10,000, i.e., the total number of tokens in the chunk summary sentences 281-283 and 381-38n is less than 3,000 (step S237: NO), then it is necessary to add more summary sentences. Therefore, the information processing server 20 generates chunk summary sentences 381-38n using chunks that have been divided by changing the maximum number of tokens per chunk based on the second algorithm (step S238). The maximum number of tokens can be changed to any value. Once step S238 is completed, the process returns to step S237, and steps S237 and S238 are repeatedly executed until the total number of tokens in the chunk summary sentences 281-283 generated by the first algorithm and the total number of tokens in the chunk summary sentences 381-38n generated by the second algorithm is between 3,000 and 10,000. Each time steps S237 and S238 are repeatedly executed, the information processing server 20 can increase the total number of tokens in the generated chunk summary sentences 381-38n. In the above embodiment, the information processing server 20 increases the total number of tokens in the generated chunk summary sentences 381-38n, but this is not limited to this. For example, each time steps S237 and S238 are repeatedly executed, the information processing server 20 may decrease the total number of tokens in the generated chunk summary sentences 381-38n.

[0095] The target chunks 361-36n of the second algorithm are divided with fewer tokens than the target chunks 261-263 of the first algorithm, and input into the large-scale language model 50. The information processing server 20 divides the target chunks 361-36n of the second algorithm into chunks of a predetermined length. For example, 1,000 tokens can be set as the predetermined length. However, the predetermined length is not limited to 1,000 tokens. The information processing server 20 can determine whether the number of tokens should be greater than (increased) or less than (decreased) the predetermined length using any algorithm. For example, if the number of tokens should be greater than (increased), and the target chunks 261-263 of the first algorithm have 5,000 tokens, the information processing server 20 divides the target chunks 361-36n of the second algorithm into chunks of 1,000-4,999 tokens, which is less than or equal to that number.

[0096] An arbitrary algorithm, after executing the second algorithm once, determines whether the number of tokens obtained by combining (summarizing) the chunk summaries 281-283 generated by the first algorithm and the chunk summaries 381-38n generated by the second algorithm should be greater than (increased) or less than (decreased) the number of tokens of a predetermined length. For example, if the information processing server 20, using an arbitrary algorithm, finds that the number of tokens obtained by combining (summarizing) the chunk summaries 281-283 generated by the first algorithm and the chunk summaries 381-38n generated by the second algorithm is less than the second number of tokens (e.g., 3,000 tokens), it will divide the text so that the number of tokens becomes less than the number of tokens of a predetermined length (e.g., 1,000 tokens). On the other hand, if the information processing server 20, using an arbitrary algorithm, combines (sums) the chunk summary sentences 281-283 generated by the first algorithm and the chunk summary sentences 381-38n generated by the second algorithm, and the resulting number of tokens is greater than or equal to the third number of tokens (for example, 10,000 tokens), it divides the text so that the number of tokens becomes greater than (increases) the number of tokens of a predetermined length (for example, 1,000 tokens). An arbitrary number of tokens can be set as the number of tokens to decrease (decrease) or the number of tokens to increase (increase). 100 can be set as the arbitrary number of tokens. However, the arbitrary number of tokens is not limited to 100. The following describes the case where 100 is set as the arbitrary number of tokens.

[0097] If the second algorithm divides the data into smaller chunks than 1,000 tokens (for example, 900 tokens), the number of "n"s in chunk summaries 381-38n increases, resulting in a larger (longer) total number of tokens. If the second algorithm divides the data into larger chunks than 1,000 tokens (for example, 1,100 tokens), the number of "n"s in chunk summaries 381-38n decreases, resulting in a smaller (shorter) total number of tokens. By using any of the above algorithms, the total number of tokens obtained by combining chunk summaries 281-283 generated by the first algorithm and chunk summaries 381-38n generated by the second algorithm can be kept between 3,000 (second number of tokens) and less than 10,000 (third number of tokens).

[0098] If the total number of tokens in the chunk summaries 281-283 generated by the first algorithm and the total number of tokens in the chunk summaries 381-38n generated by the second algorithm is between 3,000 and 10,000, that is, if the total number of tokens in the chunk summaries 281-283 and 381-38n is 3,000 or more (step S237: YES), then the chunks have been generated as of an appropriate length (3,000-10,000 tokens). Therefore, the information processing server 20 combines the chunk summaries 281-283 generated by the first algorithm and the chunk summaries 381-38n generated by the second algorithm to generate a tagging chunk 264 (step S239). Once this process is complete, the tagging chunk generation process performed in step S110 of Figure 2 is finished.

[0099] The information processing server 20 inputs the tag assignment chunk 264 generated in step S110 and the generated tag structure 130 into the large-scale language model 50. The generated tag structure 130 is the completed tag structure shown in Figure 7(2) that was generated in advance in step S107. When the information processing server 20 inputs the tag assignment chunk 264 and the tag structure 130 generated in the tag generation and assignment process into the large-scale language model 50, the large-scale language model 50 assigns tags to the target document 260 corresponding to the tag assignment chunk 264 (step S111). Figure 11 shows an example of tags assigned to the target document 260.

[0100] In the embodiment shown in Figure 11, text A and text B are exemplified as texts 260 to be tagged. The texts 260 to be tagged shown in Figure 11 are the texts before summarization of the tag-tag chunks 264. In this process, when input to the large-scale language model 50, the tag structure 130 is input in addition to the tag-tag chunks 264.

[0101] Although not shown in the diagram, an example of a prompt given when the information processing server 20 inputs the tagging chunk 264 and the tag structure 130 to the large-scale language model 50 is the instruction, "Assign tags in the hierarchical order according to the tag structure to the target document corresponding to the tagging chunk." In response to the given prompt, the large-scale language model 50 performs the process of assigning the tags contained in the tag structure 130 to the target document 260 according to the hierarchy of the tag structure 130.

[0102] For example, in the embodiment shown in Figure 11, the tag-target document 260 "Document A" corresponding to the tag-assigning chunk 264 is tagged with tags related to "diagnosis," "symptoms," "product information," "side effect information," and "dizziness."

[0103] The information processing server 20 adjusts the tags at each level of the tag structure 130 so that the number of tags assigned to the tag assignment chunks 264 by the large-scale language model 50 matches the number set at each level (step S112). In this case, the large-scale language model 50 selects a predetermined number of tags from each level of the input tag structure 130 that are included in the tag assignment chunks 264 corresponding to the text to be tagged 260, and assigns the tags. As a result, for example, the large-scale language model 50 can use the tag structure 130 as a guide, and assigns tags so that there are at least a predetermined number of unique tags among the tags included in the minor categories, the medium categories, and the major categories of the tag structure 130. The number of tags can be arbitrarily set based on the operator's input. As a predetermined number, for example, it can be set to 10 or more at each level. Therefore, tags can be assigned without bias in content at each level, and the accuracy of tag assignment can be improved. Alternatively, the number of tags included at each level can be adjusted to any number. For example, the total number of tags in the major, medium, and minor categories may be adjusted to be 10. Furthermore, the priority for assigning tags to each level may be determined based on scoring corresponding to parameters such as importance, frequency, and confidence level. Additionally, by simply inputting the tagging chunks 264 and the tag structure 130 to the large-scale language model 50, tags can be assigned to the target texts 260 of the tagging chunks 264, thus reducing the processing burden on the operator. Moreover, since tags can be assigned to the target texts 260 of the tagging chunks 264 according to the tag structure 130, consistent responses can be used for the random responses generated by the large-scale language model 50. As a result, the accuracy of tag assignment can be improved.

[0104] When the information processing server 20 inputs the tag assignment chunk 264 and the tag structure 130 generated in the tag generation and assignment process into the large-scale language model 50, the large-scale language model 50 may input a prompt to the large-scale language model 50 that prevents the assignment of tags that do not exist in the tag list 110, in addition to the text to be assigned 260 and the tag structure 130. An example of a prompt given when inputting into the large-scale language model 50 is the instruction, "Do not assign tags that do not exist in the tag list." This prevents the assignment of unnecessary tags and improves the accuracy of tag assignment. Furthermore, it prevents the operator from having to delete unnecessary tags, thereby reducing the workload on the operator.

[0105] Furthermore, if a tag has an importance parameter, the information processing server 20 inputs the text 260 to be tagged and the tag structure 130 that has received the importance into the large-scale language model 50, and the large-scale language model 50 assigns tags to the text 260 according to the tag structure 130 adjusted by importance. For example, although not assigned in the embodiment shown in Figure 11, if the importance of the tag "mechanism of operation" is set high, the information processing server 20 instructs to assign the "mechanism of operation" tag as a priority. This allows the worker to assign the tags they need as a priority and assign the tags they desire. Therefore, since the worker does not need to assign tags they desire later, the processing burden on the worker can be reduced.

[0106] (Equipment configuration) Figure 12 shows an example of the functional configuration of an information processing server according to one embodiment. As shown in this example, the information processing server 20 has a control unit 410, a storage unit 420, a communication unit 430, an input unit 440, and an output unit 450. In this example, the functional blocks of the characteristic parts of this embodiment are mainly shown, and the device 10 may also have other functional blocks necessary for other processing. Furthermore, the configuration may be one that does not include some functional blocks.

[0107] The control unit 410 performs control of the device 10. The control unit 410 may be composed of a controller, control circuit, or control device as described in accordance with common understanding in the art relating to this disclosure.

[0108] The storage unit 420 stores (holds) information used by the device 10. The storage unit 420 can be configured, for example, with memory, storage, or a device memory as described based on common understanding in the art relating to this disclosure.

[0109] The communication unit 430 communicates with other communication devices (equipment, servers, etc.) via the network. The communication unit 430 may output various received information to the control unit 410.

[0110] The communication unit 430 may consist of a transmitter / receiver, a transmitting / receiving circuit, or a transmitting / receiving device as described in accordance with common understanding in the art relating to this disclosure. The communication unit 430 may also consist of a transmitting unit and a communication unit.

[0111] The input unit 440 accepts input from an operator. The input unit 440 may also be connected to a predetermined device, storage medium, etc., to accept data input. The input unit 440 may output the input results to, for example, the control unit 410.

[0112] The input unit 440 can be configured with input devices such as keyboards, mice, and buttons, input / output terminals, and input / output circuits, as described based on common understanding in the technical field related to this disclosure. The input unit 440 may also be integrated with the display unit (for example, as a touch panel).

[0113] The output unit 450 outputs data, content, etc., in a format perceptible to the operator. For example, the output unit 450 may include a display unit for displaying images, an audio output unit for outputting sound, and so on.

[0114] The display unit can be configured with a display device such as a display or monitor, as described based on common understanding in the technical field related to this disclosure. The audio output unit can be configured with an output device such as a speaker, as described based on common understanding in the technical field related to this disclosure.

[0115] The output unit 450 may be configured to include, for example, arithmetic units, arithmetic circuits, arithmetic devices, players, image / video / audio processing circuits, image / video / audio processing devices, amplifiers, etc., as described based on common understanding in the art relating to this disclosure.

[0116] The communication unit 430, the input unit 440, and the control unit 410, or any combination thereof, may be called the reception unit. The reception unit may accept input attribute data, including attribute information relating to the worker's attributes.

[0117] The control unit 410 may perform processing based on the steps shown in Figures 2 and 10.

[0118] Device 10 may also have the same configuration as shown in Figure 12. A person skilled in the art will be able to interpret and understand the description of the information processing server 20 in the explanation of Figure 12 as appropriate.

[0119] The following sections will be explained with examples.

[0120] The communication unit 530 of the information processing server 20 may receive parameters including the importance of the tag.

[0121] When generating multiple generated chunks, the control unit 510 of the information processing server 20 may generate a generated chunk by combining the end of the sentence of the preceding generated chunk with its own generated chunk.

[0122] The control unit 510 of the information processing server 20 may, when generating multiple tag generation chunks, combine the generation chunk summary statements corresponding to generation chunks other than itself from among the multiple generation chunk summary statements in the order of the generation chunks to generate tag generation chunks.

[0123] The control unit 510 of the information processing server 20 may input multiple target chunks into a large-scale language model, causing the large-scale language model to generate a target chunk summary sentence for each of the multiple target chunks based on the multiple target chunks, and then combine the generated target chunk summary sentences to generate a tagging chunk.

[0124] The control unit 510 of the information processing server 20 may input multiple tag generation chunks into a large-scale language model to generate a tag list of multiple tag generation chunks, and may also input the tag list and a template structure in which the tags are pre-structured into an arbitrary hierarchy into the large-scale language model to generate a tag structure in which the large-scale language model structures the tags.

[0125] The control unit 510 of the information processing server 20 may adjust the tags at each level so that the number of tags assigned to the tagging chunk by the large-scale language model matches the number set at each level that constitutes the tag structure.

[0126] The control unit 510 of the information processing server 20 may receive input of importance levels for tags constituting the tag structure, modify the tag structure based on the tags for which importance levels have been received, and input the text to be tagged and the tag structure with received importance levels into the large-scale language model, thereby allowing the large-scale language model to tag the text to be tagged according to the tag structure adjusted by importance levels.

[0127] The control unit 510 of the information processing server 20 may assign tags to the text to be assigned by inputting a prompt to the large-scale language model that prevents the large-scale language model from assigning tags that are not present in the tag list, in addition to the text to be assigned and the tag structure.

[0128] The control unit 510 of the information processing server 20 may accept modifications to tags included in the tag structure based on the operator's operation, and may modify the tag structure based on the tags whose modifications have been accepted based on the operator's operation.

[0129] The control unit 510 of the information processing server 20 may divide the target chunk using the first algorithm if the target chunk is equal to or greater than the first number of tokens, and divide the target chunk using the second algorithm if the target chunk is less than the second number of tokens. Alternatively, the control unit 510 of the information processing server 20 may divide the target chunk using the first algorithm if the target chunk is equal to or greater than the first number of tokens and less than the second number of tokens, and divide the target chunk using the second algorithm if the target chunk is equal to or greater than the second number of tokens.

[0130] The text for tag generation input from device 10 to the control unit 510 of information processing server 20 may be text data generated by transcription.

[0131] The text for tag generation and the text for tag assignment, which are input from the device 10 to the control unit 510 of the information processing server 20, may be the same text.

[0132] The text for tag generation and the text to be assigned, which are input from the device 10 to the control unit 510 of the information processing server 20, may be different texts.

[0133] (Hardware configuration) The block diagrams used in the description of the above embodiments show functional units. These functional blocks (components) are realized by any combination of hardware and / or software. Furthermore, the means of realizing each functional block are not particularly limited. That is, each functional block may be realized by a single physically coupled device, or by two or more physically separated devices connected by wired or wireless means.

[0134] For example, the device in one embodiment of the present disclosure (such as the information processing server 20) may function as a computer that performs the information processing and tagging processing of the present disclosure. Figure 13 is a diagram showing an example of the hardware configuration of the information processing server, etc., according to one embodiment. The above-mentioned device 10, information processing server 20, etc. may be physically configured as a computer device including a processor 1001, memory 1002, storage 1003, communication device 1004, input device 1005, output device 1006, bus 1007, etc.

[0135] In this disclosure, terms such as apparatus, circuit, device, unit, and server may be interpreted interchangeably. The hardware configuration of device 10, information processing server 20, etc., may include one or more of the devices shown in the figure, or it may be configured to omit some of the devices.

[0136] For example, although only one processor 1001 is shown in the diagram, there may be multiple processors. Furthermore, processing may be performed by a single processor, or by two or more processors simultaneously, sequentially, or by other means. Note that processor 1001 may be implemented using one or more chips.

[0137] Each function in device 10, information processing server 20, etc., is realized by loading predetermined software (programs) onto hardware such as processor 1001 and memory 1002, causing processor 1001 to perform calculations and control communication by communication device 1004, data reading and / or writing to memory 1002 and storage 1003, etc.

[0138] The processor 1001 controls the entire computer, for example, by running the operating system. The processor 1001 may be composed of a central processing unit (CPU) that includes interfaces with peripheral devices, control devices, arithmetic units, registers, etc. Furthermore, each of the above-mentioned components, such as the control unit 410, may be implemented by the processor 1001.

[0139] Furthermore, processor 1001 may control the entire computer by using a quantum fluctuation-based process, specifically quantum annealing, to find the minimum value (global minimum) of any objective function from any set of candidate solutions. This allows for the provision of proposed information that utilizes hypotheses.

[0140] Furthermore, the processor 1001 reads programs (program code), software modules, data, etc., from at least one of the storage 1003 and the communication device 1004 into the memory 1002 and executes various processes accordingly. The program used is one that causes the computer to execute at least a part of the operations described in the above embodiment. For example, the control unit 410 may be implemented by a control program stored in the memory 1002 and running on the processor 1001, and other functional blocks may be implemented similarly.

[0141] Memory 1002 is a computer-readable recording medium and may consist of at least one of the following: ROM (Read Only Memory), EPROM (Erasable Programmable ROM), EEPROM (Electrically EPROM), RAM (Random Access Memory), or other suitable storage medium. Memory 1002 may also be called a register, cache, or main memory. Memory 1002 can store executable programs (program code), software modules, etc., for carrying out a method according to one embodiment.

[0142] The storage 1003 is a computer-readable recording medium and may consist of at least one of the following: a flexible disk, a floppy disk, a magneto-optical disk (e.g., a compact disk (CD-ROM (Compact Disc ROM)), a digital multipurpose disk, a Blu-ray disk), a removable disk, a hard disk drive, a solid-state drive, a smart card, a flash memory device (e.g., a card, stick, key drive), a magnetic stripe, a database, a server, or other suitable storage medium. The storage 1003 may also be called an auxiliary storage device. The storage unit 420 described above may be implemented by the memory 1002 and / or the storage 1003.

[0143] The communication device 1004 is hardware (transceiver / receiver device) for communicating between computers via at least one of a wired network and a wireless network, and is also referred to as a network device, network controller, network card, communication module, etc. The communication device 1004 may include a SIM card (Subscriber Identity Module Card). The communication unit 430 described above may be implemented by the communication device 1004.

[0144] The input device 1005 is an input device that accepts input from an external source (e.g., a keyboard, mouse, etc.). The output device 1006 is an output device that outputs to an external source (e.g., a display, speaker, etc.). The input device 1005 and the output device 1006 may be configured as an integrated unit (e.g., a touch panel). The input section 440 and the output section 450 described above may be implemented by the input device 1005 and the output device 1006, respectively.

[0145] Furthermore, each device, such as the processor 1001 and memory 1002, is connected by a bus 1007 for communicating information. The bus 1007 may consist of a single bus or different buses may be used for communication between devices.

[0146] Furthermore, the information processing server 20 may be configured to include hardware such as a microprocessor, a digital signal processor (DSP), an ASIC (Application Specific Integrated Circuit), a PLD (Programmable Logic Device), and an FPGA (Field Programmable Gate Array), and some or all of each functional block may be realized by such hardware. For example, the processor 1001 may be implemented using at least one of these pieces of hardware.

[0147] (modified version) In addition, terms used in this disclosure and / or terms necessary for understanding this disclosure may be replaced with terms having the same or similar meaning.

[0148] The information, parameters, etc., described in this disclosure may be expressed using absolute values, relative values ​​from a given value, or other corresponding information. Furthermore, the names used for parameters, etc., in this disclosure are not limited in any way.

[0149] The information, signals, etc. described in this disclosure may be represented using any of the various different techniques. For example, the data, instructions, commands, information, signals, bits, symbols, chips, etc. that may be referred to throughout the above description may be represented by voltage, current, electromagnetic waves, magnetic fields or magnetic particles, optical fields or photons, or any combination thereof.

[0150] Information, signals, etc., may be input and output via multiple network nodes. Input and output information, signals, etc., may be stored in a specific location (e.g., memory) or managed using a table. Input and output information, signals, etc., may be overwritten, updated, or appended to. Output information, signals, etc., may be deleted. Input information, signals, etc., may be transmitted to other devices.

[0151] Software should be broadly interpreted to mean instructions, instruction sets, code, code segments, program code, programs, subprograms, software modules, applications, software applications, software packages, routines, subroutines, objects, executable files, execution threads, procedures, functions, and so on, whether they are called software, firmware, middleware, microcode, hardware description languages, or by any other name.

[0152] Furthermore, software, instructions, information, etc., may be transmitted and received via at least one of a transmission medium and a signal waveform. For example, if software is transmitted from a website, server, or other remote source using at least one of wired technology (such as coaxial cable, fiber optic cable, twisted pair, or digital subscriber line (DSL)) and wireless technology (such as infrared or microwave), then at least one of these wired and wireless technologies is included in the definition of a transmission medium.

[0153] The terms “system” and “network” as used in this disclosure may be used interchangeably.

[0154] Each aspect / embodiment described in this disclosure may be used individually, in combination, or switched between during execution. Furthermore, the processing procedures, sequences, flowcharts, etc., of each aspect / embodiment described in this disclosure may be rearranged in order, provided they are consistent. For example, the methods described in this disclosure present various step elements in an exemplary order and are not limited to that specific order.

[0155] In this disclosure, the phrase "based on" does not mean "based solely on" unless otherwise specified. In other words, the phrase "based on" means both "based solely on" and "based at least on."

[0156] Any reference to elements using the designations “first,” “second,” etc., as used in this disclosure does not generally limit the quantity or order of those elements. These designations may be used in this disclosure as a convenient way to distinguish between two or more elements. Accordingly, the references to the first and second elements do not imply that only two elements may be employed or that the first element must precede the second element in any way.

[0157] Where the terms “include,” “including,” and variations thereof are used in this disclosure, these terms are intended to be inclusive, as is the term “comprising.” Furthermore, the term “or” as used in this disclosure is not intended to mean exclusive OR.

[0158] In this disclosure, if articles are added by translation, such as a, an, and the in English, this disclosure may include the fact that the noun following these articles is plural.

[0159] Although the invention described herein has been explained in detail above, it will be clear to those skilled in the art that the invention described herein is not limited to the embodiments described herein. The invention described herein can be implemented in modified and altered forms without departing from the spirit and scope of the invention as defined by the claims. Therefore, the descriptions herein are for illustrative purposes only and do not imply any limitation on the invention described herein. [Explanation of symbols]

[0160] 1: Information Processing System 10: Device 20: Information processing server 50: Large-scale language models 110: Tag List 120: Template Structure 130: Tag structure 410: Control Unit (Reception Unit) 420: Storage section 430: Communications Department (Reception Department) 440: Input section (reception section) 450: Output section 510: Control Unit 530: Communications Department 1001: Processor 1002: Memory 1003: Storage 1004: Communication device 1005: Input device 1006: Output device 1007: Bus 1008: Bus

Claims

1. (1) Divide the text for tag generation to generate multiple generation chunks, (2) By inputting the multiple generated chunks into a large-scale language model, the large-scale language model generates a summarized generated chunk sentence for each of the multiple generated chunks, (3) For each of the multiple generated chunks, at least one of the multiple generated chunk summary sentences is combined to generate multiple tag generation chunks corresponding to the multiple generated chunks, (4) By inputting a plurality of the tag generation chunks into the large-scale language model, the large-scale language model generates a tag structure in which tags are structured at an arbitrary hierarchy for the plurality of tag generation chunks. (5) Divide the text to be tagged to generate multiple target chunks, input the generated multiple target chunks into the large-scale language model, thereby generating a tag-tag chunk based on the divided multiple target chunks, (6) By inputting the tag-assigning chunk and the tag structure into the large-scale language model, the large-scale language model assigns the tag to the tag-assigning chunk according to the tag structure. The process is carried out in the order of steps (1) to (6) described above. An information processing system characterized by the following:

2. The information processing system according to claim 1, In step (1) above, when generating multiple generated chunks, the generated chunk is generated by combining the end of the preceding generated chunk with its own generated chunk. An information processing system characterized by the following:

3. The information processing system according to claim 2, In step (3) above, when generating multiple tag generation chunks, the tag generation chunk summary sentences corresponding to the generation chunks other than themselves are combined in the order of the generation chunks to generate the tag generation chunk. An information processing system characterized by the following:

4. The information processing system according to claim 3, In step (5) above, by inputting a plurality of target chunks into the large-scale language model, the large-scale language model generates a target chunk summary sentence for each of the plurality of target chunks based on the plurality of target chunks, and combines the generated plurality of target chunk summary sentences to generate the tagging chunk. An information processing system characterized by the following:

5. The information processing system according to claim 4, In step (4) above, a list of tags is generated by inputting a plurality of tag generation chunks into the large-scale language model, and the tag list and a template structure in which the tags are pre-structured into an arbitrary hierarchy are input into the large-scale language model, causing the large-scale language model to generate the tag structure in which the tags are structured. An information processing system characterized by the following:

6. The information processing system according to claim 5, Furthermore, the process includes adjusting the tags at each level so that the number of tags assigned to the tagging chunk by the large-scale language model matches the number set at each level constituting the tag structure. An information processing system characterized by the following:

7. The information processing system according to claim 6, Furthermore, the system accepts input of importance levels for the tags constituting the tag structure, and modifies the tag structure based on the tags for which importance levels have been received. The process includes inputting the aforementioned text to be assigned tags and the tag structure that has received the importance level into the large-scale language model, thereby enabling the large-scale language model to assign tags to the text to be assigned tags according to the tag structure adjusted by the importance level. An information processing system characterized by the following:

8. The information processing system according to claim 7, Furthermore, the process includes the step of assigning the tags to the text to be assigned by inputting a prompt to the large-scale language model that prevents it from assigning tags that are not present in the tag list, in addition to the text to be assigned and the tag structure. An information processing system characterized by the following:

9. The information processing system according to claim 8, Furthermore, the operator accepts modifications to the tags included in the tag structure, Based on the tag that has been modified based on the operator's actions, the process includes the step of modifying the tag structure. An information processing system characterized by the following:

10. The information processing system according to claim 9, Furthermore, if the target chunk has a number equal to or greater than the first token, the target chunk is divided into chunks using the first algorithm. If the target chunk has a number equal to or greater than the second number of tokens, the process includes the step of dividing the target chunk using the second algorithm. An information processing system characterized by the following:

11. An information processing system according to claim 10, The aforementioned text for tag generation is text data generated by transcription. An information processing system characterized by the following:

12. An information processing system according to any one of claims 1 to 11, The text used for tag generation and the text used for tag assignment are the same text. An information processing system characterized by the following:

13. An information processing system according to any one of claims 1 to 11, The text used for tag generation and the text used for tag assignment are different texts. An information processing system characterized by the following:

14. (1) Divide the text for tag generation to generate multiple generation chunks, (2) By inputting the multiple generated chunks into a large-scale language model, the large-scale language model generates a summarized generated chunk sentence for each of the multiple generated chunks, (3) For each of the multiple generated chunks, at least one of the multiple generated chunk summary sentences is combined to generate multiple tag generation chunks corresponding to the multiple generated chunks, (4) By inputting a plurality of the tag generation chunks into the large-scale language model, the large-scale language model generates a tag structure in which tags are structured at an arbitrary hierarchy for the plurality of tag generation chunks. (5) Divide the text to be tagged to generate multiple target chunks, input the generated multiple target chunks into the large-scale language model, thereby generating a tag-tag chunk based on the divided multiple target chunks, (6) By inputting the tag-assigning chunk and the tag structure into the large-scale language model, the large-scale language model assigns the tag to the tag-assigning chunk according to the tag structure. The process is carried out in the order of steps (1) to (6) described above. A method for assigning tags characterized by the following features.

Images