Aggregated information generation device, aggregated information generation method, aggregated information generation program, and aggregated information generation system

Abstract

To provide an aggregated information generation device, aggregated information generation method, aggregated information generation program, and aggregated information generation system, which enable generation of aggregated information in line with what a user is interested in.SOLUTION: An aggregated information generation device 5 for generating aggregated information obtained by aggregating sentences is provided, the aggregated information generation device 5 comprising an acquisition unit 51 configured to acquire the sentences and content of interest representing what a user is interested in, and a generation unit 52 configured to generate aggregated information based on the content of interest from the sentences using a trained language model 52a.SELECTED DRAWING: Figure 4

Description

[Technical Field]

[0001] The present invention relates to an aggregated information creation device, an aggregated information creation method, an aggregated information creation program, and an aggregated information creation system. [Background technology]

[0002] Since the advent of Transformer, the accuracy of natural language processing tasks such as summarizing text has improved significantly. As a result, systems that use natural language processing techniques to extract and reconstruct important information from text are being introduced in a variety of fields and businesses. For example, natural language processing techniques are used to create headlines and summaries from news articles, and to create simple descriptions from text.

[0003] As related techniques, for example, techniques described in Patent Documents 1 and 2 exist. The system described in Patent Document 1 relates to a task of summarizing product review sentences, etc. The system summarizes input sentences through a process of categorizing them and a process of clustering overlapping content, for example.

[0004] The system described in Patent Document 2 is related to the task of summarizing research papers. The system uses a trained machine learning model to classify sentences into pre-prepared labels. For example, the system collects sentences classified into the same label and summarizes them using the trained machine learning model. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] Japanese Patent Publication No. 2023-053867 [Patent Document 2] Japanese Patent Publication No. 2023-162306 Summary of the Invention [Problem to be solved by the invention]

[0006] However, conventional systems sometimes do not output results that are in line with the content that the user is interested in. For example, in the system of Patent Document 1, content with a large number of clustering results is prioritized as output content. As a result, the information that the user actually wants to know may not be output.

[0007] Furthermore, in the system of Patent Document 2, the output results are fixed to the content and format of the training data when the trained machine learning model was created. Therefore, it is not possible to aggregate data that was not focused on during the training phase but that has now become a focus of attention. When new content emerges that requires attention, re-training is required to redefine the requirements. Furthermore, when a single sentence contains multiple contents, it may be assigned multiple labels. For example, consider a case where a sentence is assigned labels A and B. In this case, it is difficult to determine which parts of the sentence correspond to labels A and B, and then split the sentence to extract only the parts that have only one label (e.g., label A). This can result in content that the user wants to exclude being included in the pre-summary text, making it impossible to create a summary that is focused on the content the user is interested in.

[0008] The present invention has been made in consideration of the above-mentioned problems, and an object of the present invention is to provide an aggregated information creation device, an aggregated information creation method, an aggregated information creation program, and an aggregated information creation system that can create aggregated information that is in line with the content that the user focuses on. [Means for solving the problem]

[0009] The above object of the present invention can be achieved by the following means.

[0010] (1) An aggregated information creation device that creates aggregated information by aggregating sentences, comprising: an acquisition unit that acquires the sentences and content of interest that indicates content that a user focuses on; and a creation unit that uses a trained language model to create the aggregated information based on the content of interest from the sentences.

[0011] (2) The aggregated information creation device according to (1) above, wherein the text includes at least a sentence containing multiple contents.

[0012] (3) The aggregated information creation device described in (1) above, wherein the trained language model is trained to output aggregated information limited to the target content when the sentence and the target content are input as input data.

[0013] (4) The learning data for learning the language model associates the sentence, the content of interest, and the aggregated information, and the aggregated information is composed only of sentences related to the content of interest of the user, with other sentences removed, in the aggregated information creation device described in (1) above.

[0014] (5) The aggregated information creation device described in (1) above further includes an extraction unit that extracts sentences related to the target content from the text, and the creation unit creates the aggregated information based on the target content from the extraction result of the extraction unit.

[0015] (6) An aggregated information creation device as described in (5) above, in which labels according to content are predetermined, and the target content is linked in advance to labels related to the target content, and the extraction unit creates extracted information by extracting sentences corresponding to the labels from the text based on the labels related to the target content.

[0016] (7) The aggregated information creation device described in (6) above, wherein the extraction unit uses a trained classification model to decompose the sentence, assign labels according to the content of each sentence, and extract sentences that have been assigned labels related to the content of interest, and the classification model is trained by supervised learning using training data in which labels are assigned to each sentence.

[0017] (8) The aggregated information creation device according to (1) above, wherein the acquisition unit acquires a word or a list of multiple words as the content of interest.

[0018] (9) The aggregated information creation device according to (6) above, wherein the acquisition unit acquires a selection result in which a label related to the content of interest is selected.

[0019] (10) The aggregated information creation device described in (1) above, wherein the document is one of an observation record, a question and answer session, and a meeting minutes.

[0020] (11) The aggregated information creation device according to (1) above, wherein the text is related to nursing care services.

[0021] (12) A method for creating aggregated information by aggregating sentences, the method comprising: an acquisition step for acquiring the sentences and content of interest indicating content that a user focuses on; and a creation step for creating the aggregated information based on the content of interest from the sentences using a trained language model.

[0022] (13) An aggregated information creation program for causing a computer to execute the aggregated information creation method described in (12) above.

[0023] (14) An aggregated information creation system comprising: a user terminal to which content of interest indicating content on which a user pays attention is input; and the aggregated information creation device described in (1) above. [Effects of the Invention]

[0024] According to the present invention, it is possible to create aggregated information that is in line with the content that the user is interested in. [Brief explanation of the drawings]

[0025] [Figure 1] 1 is a schematic configuration diagram of an aggregate information creation system according to a first embodiment of the present invention. [Figure 2] 1 is a diagram illustrating an example of a hardware configuration of an information processing device according to a first embodiment of the present invention. [Figure 3] 1 is an image of creating aggregated information in the information processing device according to the first embodiment of the present invention. [Figure 4]1A and 1B are functional configuration diagrams of an information processing device according to a first embodiment of the present invention, in which FIG. 1A shows functions related to a stage of learning aggregated information, and FIG. 1B shows functions related to a stage of creating aggregated information. [Figure 5] 1A and 1B are diagrams for explaining the creation of training data, where (a) is an image of the training data, and (b) is an example of the training data. [Figure 6] 1A and 1B are diagrams for explaining the creation of training data, where (a) is an image of the training data, and (b) is an example of the training data. [Figure 7] 1 is an example of a flowchart showing an operation at a creation stage of the aggregate information creation system according to the first embodiment of the present invention. [Figure 8] FIG. 10 is a schematic configuration diagram of an aggregate information creation system according to a second embodiment of the present invention. [Figure 9] 10 is a functional configuration diagram of an information processing device according to a second embodiment of the present invention, in which (a) shows functions related to the stage of learning labels, (b) shows functions related to the stage of learning aggregated information, and (c) shows functions related to the stage of creating aggregated information (including the stages of labeling and classifying). [Figure 10] 10 is an image of creating aggregated information in an information processing device according to a second embodiment of the present invention. [Figure 11] 10 is an example of a flowchart showing an operation at a creation stage of the aggregate information creation system according to the second embodiment of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0026] Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. Each drawing is merely a schematic illustration to allow a sufficient understanding of the present invention. Therefore, the present invention is not limited to the illustrated examples. Furthermore, in each drawing, common or similar components are given the same reference numerals, and redundant explanations thereof may be omitted. Furthermore, detailed explanations of known functions not directly related to the present invention may be omitted.

[0027] [First embodiment] <Configuration of aggregate information creation system according to the first embodiment> The configuration of the aggregated information creation system 1 according to the first embodiment will be described with reference to FIG. 1. FIG. 1 is a schematic diagram of the aggregated information creation system 1 according to the first embodiment. The aggregated information creation system 1 creates aggregated information from any text. The text may be, for example, an observation record, a question and answer session, or minutes of a meeting, and is composed of multiple sentences. The aggregated information is information that summarizes the text based on the content (referred to as "content of interest") that a user (for example, a reader of the text or an administrator) focuses on. The aggregated information is shorter (has fewer characters) than the original text, and is composed of, for example, one sentence or a small number of sentences (for example, several). The aggregated information may be in units smaller than a sentence (for example, a phrase), and is composed of at least multiple words.

[0028] As shown in FIG. 1, the aggregated information creation system 1 includes a user terminal 2 and an information processing device 3. The user terminal 2 is operated by a user (for example, a reader or administrator of the text). The information processing device 3 is a central device of the aggregated information creation system 1. The user terminal 2 and the information processing device 3 are communicatively connected via a network 9. The network 9 is, for example, a LAN (Local Area Network) or a WAN (Wide Area Network), and may be either wired or wireless.

[0029] 1 is, for example, a personal computer (PC), a tablet terminal, a smartphone, etc. A user operates the user terminal 2 to input content of interest. The user terminal 2 transmits the content of interest input by the user to the information processing device 3. In addition, the user terminal 2 receives aggregated information from the information processing device 3 in response to the transmitted content of interest.

[0030] The information processing device 3 shown in FIG. 1 is a device (computer) equipped with an information processing function. The information processing device 3 may be, for example, an application server. The information processing device 3 may also be a device constituting a cloud system. The information processing device 3 realizes various functions (including learning functions) related to the creation of aggregated information by program execution processing using a CPU (Central Processing Unit) or the like. The information processing device 3 receives content of interest from the user terminal 2 and creates aggregated information based on the received content of interest. The information processing device 3 transmits the created aggregated information to the user terminal 2 in response to the content of interest. Note that the user can also directly input the content of interest to the information processing device 3 without going through the user terminal 2.

[0031] 2 (and also FIG. 1 as needed), the hardware configuration of the information processing device 3 will be described. FIG. 2 shows an example of the hardware configuration of the information processing device 3. As shown in FIG. 2, the information processing device 3 includes a control unit 31, an input unit 32, an output unit 33, a communication unit 34, a storage unit 35, and the like, and each unit is connected by a bus.

[0032] The control unit 31 is configured with a CPU (Central Processing Unit), RAM (Random Access Memory), etc., and comprehensively controls the processing operations of each unit of the information processing device 3. Specifically, the CPU reads out various processing programs stored in the storage unit 35, expands them in the RAM, and performs various processes in cooperation with the programs.

[0033] The input unit 32 is configured by a keyboard having various function keys, a pointing device having various buttons (for example, a mouse or a touchpad), etc. The input unit 32 outputs an operation signal input by the user to the control unit 31.

[0034] The output unit 33 is configured by a display, a speaker, etc. The output unit 33 outputs various screens and sounds according to instructions of display signals and audio signals input from the control unit 31.

[0035] The communication unit 34 is configured with a network interface or the like, and transmits and receives data to and from external devices connected via the network 9 (see FIG. 1). For example, the communication unit 34 receives content of interest that the user focuses on from the user terminal 2. The communication unit 34 also transmits aggregated information created based on the content of interest to the user terminal 2.

[0036] The storage unit 35 is configured by a nonvolatile memory or the like (for example, a hard disk drive (HDD) or a solid state drive (SSD)). The storage unit 35 stores various data required for creating the aggregated information.

[0037] The information processing device 3 has a machine learning model. The machine learning model possessed by the information processing device 3 is a language model. The language model represents a probability distribution for a word string. The language model possessed by the information processing device 3 may be, for example, "Transformer." "Transformer" is a neural network model for sequence transformation tasks, and is a technology that has become the basis of recent interactive AI (Artificial Intelligence). A sequence transformation task is a task of converting a certain sequence of input into a different sequence of output, and in this embodiment, a task of creating aggregate information from arbitrary sentences is assumed.

[0038] Creation of the aggregated information in the information processing device 3 will be described with reference to Fig. 3. Fig. 3 is an image of creation of the aggregated information in the information processing device 3.

[0039] The information processing device 3 inputs any sentence and the user's focus as input data into the trained language model. Here, the input sentence is assumed to be an observation record (care record) created during care work. For example, the sentence "Urination occurred in the pad. During mealtime, the patient only ate soup and refused the rest. Choked during meal. Stopped. Appears to have stumbled and fallen" is input into the language model. Furthermore, "choking" is input into the language model as the focus. The trained language model aggregates the care record based on the focus "choking" and outputs aggregated information. For example, it outputs the aggregated information "Chokes, so service is stopped." In this way, the information processing device 3 outputs aggregated information according to the content the user wants to know. Note that the input data may not be a single word, but may be a list of multiple words (e.g., "Chokes, spits out, cannot swallow").

[0040] The following describes the functions of the information processing device 3 with reference to Fig. 4 (and also with reference to Figs. 1 to 3 as appropriate). Fig. 4 is a functional configuration diagram of the information processing device 3 according to the first embodiment, in which (a) shows functions related to the stage of learning aggregated information, and (b) shows functions related to the stage of creating aggregated information.

[0041] In this embodiment, the description will be given assuming that the processes related to learning a language model and creating aggregated information using the trained language model are performed by a single device. That is, the information processing device 3 performs language model learning and creates aggregated information using the trained language model. An information processing device 3 that focuses on the function related to language model learning is specifically referred to as an "aggregated information learning device" (see FIG. 1). An information processing device 3 that focuses on the function related to creating aggregated information using a trained language model is specifically referred to as an "aggregated information creation device" (see FIG. 1). Note that the processes related to learning a language model and the processes related to creating aggregated information using a trained language model can also be performed by separate devices. The aggregated information creation device 5 acquires the trained language model from the aggregated information learning device 4 and creates aggregated information using the acquired language model.

[0042] The aggregate information learning device 4 shown in Fig. 4 is a device that learns a language model 42a. As shown in Fig. 4, the aggregate information learning device 4 mainly includes an acquisition unit 41 and a learning unit 42. The acquisition unit 41 and the learning unit 42 are realized by program execution processing using a CPU or the like.

[0043] The acquisition unit 41 acquires learning data that associates (1) sentences, (2) content of interest, and (3) aggregated information. The content of interest is, for example, a single word (e.g., a noun, verb, adjective, adjectival verb, adverb, etc.). The aggregated information is a sentence limited to the content of interest, and is, for example, composed only of sentences related to the content of interest of the user, with other sentences removed. The aggregated information learning device 4 may store a collection of learning data to be processed in one learning session as a dataset in the storage unit 35, and the acquisition unit 41 may acquire the dataset during learning.

[0044] Creation of learning data will be described with reference to Figures 5 and 6. Figures 5 and 6 are diagrams for explaining creation of learning data, where (a) is an image of learning data and (b) is an example of learning data.

[0045] Here, the source sentence is assumed to be "Urination in the pad. During the meal, the patient ate only soup and refused the rest. The patient choked during the meal. The meal was stopped. It appears the patient stumbled and fell." When the focus content "choking" is specified for this sentence, we will explain how to train a language model to obtain aggregate information such as "Due to the patient choking, the meal was stopped." In this case, for example, the training data shown in (a) of Figure 5(b) is created.

[0046] In the training data shown in (A), sentences are separated by two [SEP] tokens. [SEP] tokens represent sentence boundaries. The part before the first [SEP] token corresponds to the original sentence. The part between the first [SEP] token and the second [SEP] token corresponds to the focus content. The part after the second [SEP] token corresponds to the aggregate information.

[0047] Instead of the training data shown in (a) of Figure 5(b), the training data shown in (c) may be created. In the training data shown in (c), sentences are separated by two [SEP] tokens. The part before the first [SEP] token corresponds to the target content. The part between the first [SEP] token and the second [SEP] token corresponds to the original sentence. The part after the second [SEP] token corresponds to the aggregated information.

[0048] Next, we will explain the case where a language model is trained to obtain aggregate information such as "the person stumbled and fell" when the target content "fall" is specified for a given sentence. In this case, for example, the training data shown in (A) of Figure 5(b) is created. In the training data shown in (A), sentences are separated by two [SEP] tokens. The part before the first [SEP] token corresponds to the original sentence. The part between the first and second [SEP] tokens corresponds to the target content. The part after the second [SEP] token corresponds to the aggregate information.

[0049] Furthermore, instead of the training data shown in FIG. 5(b), training data in the form of a question and answer session may be created as shown in FIG. 6. The training data shown in (f) of FIG. 6(b) is training data when the content of interest is "choking," and the training data shown in (g) is training data when the content of interest is "falling." In the training data shown in (f) and (g), sentences are separated by two [SEP] tokens. The part before the first [SEP] token corresponds to the original sentence. The part between the first [SEP] token and the second [SEP] token corresponds to the content of interest. The part after the second [SEP] token corresponds to the aggregated information.

[0050] The learning unit 42 shown in FIG. 4 uses training data to train a language model 42a. The language model 42a is, for example, an encoder-decoder model using a Transformer, or a decoder model. In this embodiment, a general-purpose language model is prepared and applied to the task of creating aggregate information by tuning using the training data described above. The tuning method may be full fine-tuning, which updates all parameters, or a method that updates only some parameters. Alternatively, the tuning method may be LoRA-tuning, which inserts an adapter into the model and trains only the adapter part.

[0051] The aggregated information creation device 5 shown in FIG. 4 uses a trained language model 52a to create aggregated information limited to content of interest that a user focuses on from any sentence. The trained language model 52a is trained so that when content of interest is input for any sentence, aggregated information limited to the content of interest is output. As shown in FIG. 4, the aggregated information creation device 5 mainly includes an acquisition unit 51 and a creation unit 52. The acquisition unit 51 and the creation unit 52 are realized by program execution processing using a CPU or the like.

[0052] The acquisition unit 51 acquires (1) a sentence and (2) a focused content. The acquired sentence and focused content are as described for the acquisition unit 41 of the aggregated information learning device 4. Here, the focused content used in the creation stage of the aggregated information (i.e., the focused content input to the trained language model 52a) may be content that was not used in the learning stage. For example, even if the focused content of "choking" has not been learned, the focused content of "choking" can be used in the creation stage. This allows the user to specify the focused content at the creation stage without worrying about the focused content used in the learning stage. Note that the acquisition unit 51 may limit information input via a user interface (UI) to prevent information unrelated to the focused content from being input. The acquisition unit 51 may, for example, limit the number of characters to be input to prevent information unrelated to the focused content from being input.

[0053] The creation unit 52 shown in Fig. 4 uses a trained language model 52a to create aggregated information limited to a content of interest from an arbitrary sentence. The language model 52a is, for example, an encoder-decoder model using a Transformer or a decoder model, and is tuned by the learning unit 42 of the aggregated information learning device 4. The creation unit 52 acquires the trained language model 52a in advance from the aggregated information learning device 4. The trained language model 52a is trained so that, when a content of interest is input for an arbitrary sentence, aggregated information limited to the content of interest is output.

[0054] <Operation of the aggregate information creation system according to the first embodiment> The operation of the aggregated information creation system 1 in the creation stage according to the first embodiment will be described with reference to Fig. 7 (and Figs. 1 to 6 as appropriate). Fig. 7 is an example of a flowchart showing the operation of the aggregated information creation system 1 in the creation stage.

[0055] (Entering information about the text and obtaining the text "S11") A user inputs information about a sentence (for example, identification information of the sentence) via the user terminal 2. The user terminal 2 transmits information about the input sentence to the aggregate information creation device 5, and the acquisition unit 51 acquires a sentence corresponding to the information about the input sentence from the storage unit 35. For example, when the user inputs information such as a person's name and a date, the acquisition unit 51 acquires the corresponding care record from the care records stored in the storage unit 35. The user may also input the care record directly into the user terminal 2.

[0056] (Input and acquisition of focus content "S12") Furthermore, the user inputs content of interest via the user terminal 2. The user terminal 2 transmits the input content of interest to the aggregated information creation device 5, and the acquisition unit 51 acquires the input content of interest. For example, the user inputs "choking" as content of interest related to the care record, and the acquisition unit 51 acquires the input "choking." There is no set order for steps S11 and S12; for example, the acquisition of the text and the content of interest may be performed simultaneously, or the text may be acquired after the content of interest is acquired.

[0057] (Information aggregation based on focus content "S13") Next, the creation unit 52 of the aggregate information creation device 5 uses the trained language model 52a to create aggregate information limited to the content of interest from the acquired sentence. For example, from the sentence "Urination occurred in the pad. During the meal, the patient ate only soup and refused the rest. The patient choked during the meal. The meal was stopped. It appears that the patient stumbled and fell," the creation unit 52 creates "Due to the patient choking, the provision was stopped" in line with the content of interest "choking."

[0058] The aggregated information creation system 1 according to the first embodiment of the present invention configured as above provides the following advantageous effects. That is, in the aggregated information creation system 1 according to this embodiment, the aggregated information creation device 5 includes an acquisition unit 51 and a creation unit 52. The acquisition unit 51 acquires the source text and the content of interest. The creation unit 52 uses a trained language model 52a to obtain aggregated information based on the content of interest from the acquired text. The language model 52a is trained so that when the content of interest for any text is input, aggregated information limited to the content of interest is output. Therefore, the aggregated information creation system 1 can create aggregated information that is in line with the content that the user wants to focus on.

[0059] Furthermore, in the conventional method of aggregating information using labels, it is not possible to aggregate information based on new content that was not focused on during the learning stage but has now become a focus of attention. When new content emerges that you want to focus on, it becomes necessary to redefine the requirements and re-learn. Furthermore, when a single sentence contains multiple contents, it may be assigned multiple labels. For example, consider a case where a sentence is assigned labels A and B. In this case, it is difficult to determine which parts of the sentence correspond to labels A and B, and then split the sentence to extract only the parts that have only one label (for example, label A). This can result in the inclusion of content that the user would like to exclude, making it impossible to create an information aggregation that is in line with the content the user is interested in.

[0060] [Second embodiment] In the second embodiment, before information is aggregated using a trained language model, each sentence constituting a text is labeled with a classification model. Then, sentences with specific labels are extracted and used as input data for the language model.

[0061] <Configuration of aggregate information creation system according to the second embodiment> The configuration of the aggregated information creation system 101 according to the second embodiment will be described with reference to Fig. 8. Fig. 8 is a schematic configuration diagram of the aggregated information creation system 101 according to the second embodiment. The aggregated information creation system 101 creates aggregated information from any text, similar to the first embodiment.

[0062] As shown in FIG. 8, the aggregated information creation system 101 includes a user terminal 2 and an information processing device 103. The user terminal 2 is operated by a user (for example, a reader or administrator of the text). The information processing device 103 is a central device of the aggregated information creation system 101. The user terminal 2 and the information processing device 103 are connected to each other so as to be able to communicate with each other via a network 9. The user terminal 2 is the same as in the first embodiment, and therefore a detailed description thereof will be omitted.

[0063] The information processing device 103 shown in FIG. 8 is a device (computer) equipped with an information processing function. The information processing device 103 may be, for example, an application server. The information processing device 103 may also be a device constituting a cloud system. The information processing device 103 realizes various functions (including learning-related functions) related to the creation of aggregated information by program execution processing using a CPU (Central Processing Unit) or the like. The hardware configuration of the information processing device 103 is similar to the hardware configuration of the information processing device 3 according to the first embodiment (see FIG. 2).

[0064] The functions of the information processing device 103 will be described with reference to Fig. 9. Fig. 9 is a functional configuration diagram of the information processing device 103 according to the second embodiment, in which (a) shows functions related to the stage of learning labels, (b) shows functions related to the stage of learning aggregated information, and (c) shows functions related to the stage of creating aggregated information (including the stage of assigning labels and classifying).

[0065] In this embodiment, it is assumed that (1) processing related to training of classification models, (2) processing related to training of language models, and (3) processing related to creation of aggregated information using trained classification models and language models are performed by a single device. In other words, the information processing device 103 trains classification models and language models and creates aggregated information using the trained classification models and language models. The information processing device 103 that focuses on functions related to training of classification models is particularly referred to as a "label training device" (see FIG. 8). The information processing device 103 that focuses on functions related to training of language models is particularly referred to as an "aggregated information training device" (see FIG. 8). Furthermore, the information processing device 103 that focuses on functions related to creation of aggregated information using trained classification models and language models is particularly referred to as an "aggregated information creation device" (see FIG. 8).

[0066] It is also possible to perform (1) the process related to learning the classification model, (2) the process related to learning the language model, and (3) the process related to creating aggregated information using the learned classification model and language model in separate devices. Aggregated information creation device 5 acquires the learned classification model from label learning device 6 and also acquires the learned language model from aggregated information learning device 4. Aggregated information creation device 5 creates aggregated information using the acquired classification model and language model.

[0067] The label learning device 6 shown in Fig. 9 is a device that, when an arbitrary sentence is input, breaks the sentence into sentences and learns a classification model so as to assign a label to each sentence. As shown in Fig. 9, the label learning device 6 mainly includes an acquisition unit 61 and a learning unit 62. The acquisition unit 61 and the learning unit 62 are realized by program execution processing using a CPU or the like.

[0068] The acquisition unit 61 acquires learning data for training a classification model. The learning data for training a classification model is, for example, sentences in which a label is assigned to each sentence. A label is defined in advance in the learning stage, and is, for example, a single word (e.g., a noun, a verb, an adjective, an adjectival verb, an adverb, etc.). The label learning device 6 may store a collection of learning data to be processed in one learning run as a dataset in the storage unit 35 (see FIG. 2), and the acquisition unit 61 may acquire the dataset during learning.

[0069] The learning unit 62 shown in FIG. 9 trains a classification model 62a using training data. The learning unit 62 trains the classification model by supervised learning, for example, using training data of sentences in which each sentence is assigned a label. The classification model can be configured with two models: a machine learning model that breaks down a sentence into individual sentences, and a machine learning model that assigns a label according to the content of each decomposed sentence. In this case, the learning unit 62 trains the machine learning model that breaks down a sentence into individual sentences using training data in which a sentence is broken down into individual sentences. Furthermore, the learning unit 62 trains the machine learning model that assigns a label according to the content of each decomposed sentence using training data to which a label according to the content of each decomposed sentence is assigned.

[0070] The aggregate information learning device 4 shown in Fig. 9 is a device that learns a language model 42a. The functions of the aggregate information learning device 4 are the same as those in the first embodiment, and therefore detailed description thereof will be omitted.

[0071] The aggregated information creation device 105 shown in FIG. 9 uses a trained classification model to decompose any text into sentences and assign a label to each sentence. The trained classification model is a model that has been trained so that when any text is input, it decomposes the text into sentences and assigns a label according to the content of each decomposed sentence. Furthermore, the aggregated information creation device 105 uses a trained language model to create aggregated information limited to content that the user wants to focus on (content of interest) from sentences to which a specific label has been assigned. The trained language model is a model that has been trained so that when the user inputs content of interest for any text, it outputs aggregated information limited to that content.

[0072] 9, the aggregated information creating device 105 mainly includes an acquiring unit 151, an extracting unit 152, and a creating unit 153. The acquiring unit 151 has a first acquiring unit 151A and a second acquiring unit 151B. The acquiring unit 151, the extracting unit 152, and the creating unit 153 are realized by a program execution process using a CPU or the like.

[0073] The first acquisition unit 151A acquires (1) an arbitrary sentence and (2) a label designated by the user. The first acquisition unit 151A may limit information input through a user interface so that information unrelated to the label is not input. For example, the first acquisition unit 151A may match the information with a pre-registered label or limit the number of characters input, thereby preventing information unrelated to the label from being input.

[0074] The extraction unit 152 uses the trained classification model 152a to decompose an arbitrary sentence into sentences and assign a label to each of the decomposed sentences. The classification model 152a is trained, for example, by supervised learning using training data of sentences in which each sentence is assigned a label. The trained classification model 152a is a model that has been trained so that, when an arbitrary sentence is input, the arbitrary sentence is decomposed into sentences and a label is assigned to each sentence. The extraction unit 152 acquires the trained classification model 152a in advance from the label learning device 6. Furthermore, the extraction unit 152 extracts sentences to which a label has been assigned based on a label designated by a user.

[0075] The second acquisition unit 151B acquires (1) a sentence to which a user-specified label has been assigned, and (2) a content of interest. (1) The sentence to which a user-specified label has been assigned is extracted based on the label assigned by the classification model 152a. Note that the second acquisition unit 151B may limit information input through a user interface so that information unrelated to the content of interest is not input.

[0076] The creation unit 153 uses the trained language model 153a to create aggregated information based on the content of interest from a sentence to which a user-specified label has been assigned. The language model 153a is tuned by the learning unit 42 of the aggregated information learning device 4. The creation unit 153 acquires the trained language model 153a in advance from the aggregated information learning device 4. The trained language model 153a is a model that has been trained so that, when the user inputs content of interest for any sentence, it outputs aggregated information limited to that content.

[0077] 10 (and also FIG. 9 as needed), the creation of the aggregated information in information processing device 103 will be described. FIG. 10 is an image of the creation of aggregated information in aggregated information creating device 105.

[0078] Specifically, the extraction unit 152 of the aggregate information creation device 105 inputs an arbitrary sentence as input data into the trained classification model 152a. Here, a nursing care record is assumed as the arbitrary sentence, and for example, the following sentence is input to the classification model: "Urination occurred in the pad. During mealtime, the patient only ate soup and refused the rest. The patient choked during mealtime. The meal was stopped. It appears that the patient stumbled and fell." The trained classification model 152a breaks the sentence down into sentences and assigns a label to each sentence. The labels are determined in advance during the training stage of the classification model 152a.

[0079] The extraction unit 152 also extracts sentences to which a specific label has been assigned based on the user-specified label, which is input data. In this example, "meal" is specified as the label, and the sentences to which this label has been assigned, "During the meal, he only ate soup and refused the rest," and "He choked during the meal. He stopped eating," are extracted as extracted information.

[0080] Next, the creation unit 153 of the aggregated information creation device 105 inputs the sentences extracted by the extraction unit 152 and the user's focused content into the trained language model 153a. Here, the sentence "During the meal, he only ate the soup and refused the rest" and the sentence "He choked during the meal. Stopped." are input into the language model 153a. In addition, "choking" is input into the language model 153a as the focused content. The trained language model 153a aggregates the extracted sentences based on the focused content "choking" and outputs aggregated information. For example, it outputs "He choked, so the service was stopped" as the aggregated information.

[0081] <Operation of the aggregate information creation system according to the second embodiment> The operation of the aggregated information creation system 101 according to the second embodiment at the creation stage will be described with reference to Fig. 11 (and Fig. 8 to Fig. 10 as appropriate). Fig. 11 is an example of a flowchart showing the operation of the aggregated information creation system 101 at the creation stage.

[0082] (Inputting information about sentences and labels, and obtaining sentences and labels "S21") The user inputs information about a sentence (for example, identification information of the sentence) via the user terminal 2. The user also inputs a specific label via the user terminal 2. The user terminal 2 transmits the information about the input sentence and the label to the aggregate information creation device 105. The acquisition unit 151 acquires a sentence corresponding to the information about the input sentence from the storage unit 35. For example, when the user inputs information such as a person's name or a date, the acquisition unit 151 acquires the corresponding care record from the care records stored in the storage unit 35.

[0083] (Labeled "S22") Next, the extraction unit 152 of the aggregate information creation device 105 uses the trained classification model 152a to break down any given sentence into sentences and assign a label according to the content of each broken down sentence. For example, suppose that the following sentence is acquired: "Urination occurred in the pad. During the meal, the patient ate only soup and refused the rest. Choked during the meal. Stopped. Appears to have stumbled and fallen." For example, the extraction unit 152 assigns the label "excretion" to "Urination occurred in the pad.". The extraction unit 152 also assigns the label "eating" to "During the meal, the patient ate only soup and refused the rest." The extraction unit 152 also assigns the label "eating" to "Chokes during the meal. Stopped." The extraction unit 152 also assigns the label "eating." The extraction unit 152 also assigns the label "exercise" to "Appears to have stumbled and fallen."

[0084] (Extract sentences with specific labels as aggregate elements "S23") Next, the extraction unit 152 of the aggregate information creation device 105 extracts sentences to which the label is assigned based on the label specified by the user. For example, if the label "meal" is specified by the user, the extraction unit 152 extracts "During the meal, he only ate the soup and refused the rest" and "He choked during the meal. He stopped eating."

[0085] (Enter and acquire the content of interest "S24") A user inputs content of interest via the user terminal 2. The user terminal 2 transmits the input content of interest to the aggregated information creation device 105, and the acquisition unit 151 acquires the input content of interest. For example, the user inputs "choking" as content of interest related to the care record, and the acquisition unit 151 acquires the input "choking." Note that step S24 may be performed before step S25, and for example, the acquisition of the sentence or label and the acquisition of the content of interest may be performed simultaneously.

[0086] (S25: Aggregation of information based on focus) Next, the creation unit 153 of the aggregate information creation device 105 uses the trained language model 153a to obtain aggregate information based on the content of interest from the extracted sentence. For example, from the sentence "During the meal, the patient ate only the soup and refused the rest. The patient choked during the meal. The meal was stopped," the creation unit 153 creates "The patient choked, so the meal was stopped" in accordance with the content of interest "choking."

[0087] The aggregate information creation system 101 according to the second embodiment of the present invention configured as above provides the same effects as the first embodiment. Furthermore, aggregated information creating system 101 according to the second embodiment extracts specific sentences using extraction unit 152 having classification model 152a (think of it as filtering a sentence). Then, aggregated information creating system 101 creates aggregated information from the extracted information extracted by extraction unit 152 using creation unit 153 having language model 153a. Therefore, when multiple sentences are input, aggregated information creating system 101 can create aggregated information that is in line with the content that the user wants to focus on with high accuracy.

[0088] Although the embodiments of the present invention have been described above, the present invention is not limited to the above-described embodiments and can be modified and implemented as appropriate. For example, in each embodiment, the original text (input text) of the aggregated information is assumed to be an observation record (nursing record) created in a nursing care service. However, the present invention can be used in various fields and in services other than nursing care. Furthermore, the present invention can create aggregated information from text other than observation records. [Explanation of symbols]

[0089] 1,101 Aggregate Information Creation System 2. User terminal 3. Information processing equipment 4. Aggregate Information Learning Device 5 Aggregate information creation device 6 Label learning device 31 Control Unit 32 Input section 33 Output section 34 Communications Department 35 Storage section 41 Acquisition Department 42 Learning Department 51 Acquisition Department 52 Creation Department 61 Acquisition Department 62 Learning Department 151 Acquisition Department 151A 1st acquisition part 151B 2nd acquisition part 152 Extraction part 153 Creation Department 42a, 52a, 153a Language Model 62a,152a Classification model

Claims

1. An aggregated information creation device that creates aggregated information by aggregating sentences, an acquisition unit that acquires the sentence and a content of interest that indicates a content that the user pays attention to; a creation unit that creates the aggregated information based on the content of interest from the sentence using a trained language model; An aggregate information creation device comprising:

2. The aggregated information creation device according to claim 1 , wherein the text includes at least a sentence containing a plurality of contents.

3. the trained language model is trained to output aggregate information limited to the content of interest when the sentence and the content of interest are input as input data; The aggregate information creation device according to claim 1 .

4. the learning data for learning the language model associates the sentence, the content of interest, and the aggregated information; The aggregated information is composed of only sentences related to content that the user pays attention to, and other sentences are removed. The aggregate information creation device according to claim 1 .

5. An extraction unit that extracts sentences related to the target content from the sentence, the creation unit creates the aggregated information based on the content of interest from the extraction result of the extraction unit. The aggregate information creation device according to claim 1 .

6. A label corresponding to the content is determined in advance, and the content of interest is associated with a label related to the content of interest in advance. the extraction unit creates extraction information by extracting sentences corresponding to the labels from the text based on the labels related to the content of interest; The aggregate information creating device according to claim 5 .

7. the extraction unit uses a trained classification model to decompose the sentence, assign a label according to the content of each sentence, and extract sentences to which a label related to the content of interest has been assigned; The classification model is trained by supervised learning using training data in which a label is assigned to each sentence. The aggregate information creation device according to claim 6.

8. The aggregate information creating device according to claim 1 , wherein the acquiring unit acquires a word or a list of a plurality of words as the content of interest.

9. The aggregated information creating device according to claim 6 , wherein the acquisition unit acquires a selection result in which a label related to the target content is selected.

10. The aggregated information creation device according to claim 1 , wherein the document is any one of an observation record, a question and answer session, and a meeting minutes.

11. The aggregated information creation device according to claim 1 , wherein the text is related to nursing care services.

12. An aggregated information creation method for creating aggregated information that aggregates sentences, comprising: an acquisition step of acquiring the sentence and a content of interest indicating a content of interest of the user; a creation step of creating the aggregated information based on the content of interest from the sentences using a trained language model; A method for creating aggregate information.

13. An aggregated information creating program for causing a computer to execute the aggregated information creating method according to claim 12.

14. a user terminal into which content of interest indicating content that a user focuses on is input; The aggregated information creation device according to claim 1 ; An aggregate information creation system comprising:

Images