Sentence generation device, sentence generation method, and program

Abstract

To provide a means capable of suppressing the omission of recording contents and the error of the recording contents due to a voice recognition error and a video recognition error and reducing the personal dependency of business recording.SOLUTION: A sentence generation device according to one embodiment includes a data structurization unit, a sentence creation unit, and a sentence generation unit. The data structuring unit creates, from record data including a plurality of sentences, structure data including two sentences among the plurality of sentences and a label indicating a relationship between the two sentences. The sentence generator uses the structure data to generate an input sentence including a first sentence that is one of two sentences of the structure data, a label, and a question sentence. The sentence generation unit generates a generated answer sentence obtained by causing a sentence generation model based on a neural network to generate a sentence based on an input sentence.SELECTED DRAWING: Figure 2

Description

[Technical Field]

[0001] The embodiments relate to a sentence generation device, a sentence generation method, and a program. [Background technology]

[0002] In various industries, business records are created daily. Creating such business records requires costs, such as money and time. In order to improve business efficiency and reduce costs, it is desirable to reduce the costs required for creating business records. For this reason, there is a growing demand for technology that can automatically create business records.

[0003] For example, in the nursing industry, it is desirable to reduce the cost of creating nursing records. In response to this, Non-Patent Document 1 considers an automatic nursing record creation system that utilizes voice recognition technology.

[0004] Furthermore, in factories, a business analysis system that utilizes image recognition technology is being considered. Non-Patent Document 2 proposes an AI (Artificial Intelligence) that utilizes image recognition technology to create business records. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Ayaka Suga, Yuko Ohno, Mayumi Nagayasu, Makoto Fujii, Natsuko Ando, ​​Takako Fujimaki, Haruka Kudo, Anna Tsutsui, Tetsuya Tajima, Haruka Shimizu, Makoto Yamamoto, Ritsuko Saito, Satoshi Nakatani, "Basic study on creating nursing records from nurses' oral statements during surgery," [online], IT Healthcare Magazine, Vol. 16, No. 1, 2021, pp. 13-20, [Retrieved April 15, 2024], Internet<URL : https: / / www.jstage.jst.go.jp / article / ithc / 16 / 1 / 16_13 / _article / -char / ja> [Non-patent document 2] Yuka Sugimura, Daisuke Uchida, Genta Suzuki, Toshio Endo, "Actlyzer: A Behavioral Analysis Technology for Recognizing Various Human Behaviors from Video," [online], Proceedings of the National Conference of the Japanese Society for Artificial Intelligence, 2020, [Retrieved April 15, 2024], Internet<URL :https: / / www.jstage.jst.go.jp / article / pjsai / JSAI2020 / 0 / JSAI2020_4Rin157 / _pdf / -char / ja> [Non-patent document 3] Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin, Scott Gray, Benjamin Chess, Jack Clark, Christopher Berner, Sam McCandlish, Alec Radford, Ilya Sutskever, Dario Amodei, “Language Models are Few-Shot Learners”, [online]arXiv:2005.14165, [cs.CL] 22 Jul 2020 [Retrieved April 15, 2024], Internet<URL :https: / / arxiv.org / abs / 2005.14165> [Non-patent document 4] Chin-Yew Lin and Eduard Hovy, “Automatic Evaluation of Summaries Using N-gram Co-Occurrence Statistics”, [online], In NAACL 2003, [Retrieved April 15, 2024], Internet<URL :https: / / aclanthology.org / N03-120 / > Summary of the Invention [Problem to be solved by the invention]

[0006] According to the method of Non-Patent Document 1, information that does not occur in speech is not recorded, and according to the method of Non-Patent Document 2, information that does not appear in the video is not recorded. As a result, the recorded content may be insufficient, for example, items that should be recorded may not be recorded. Furthermore, action plans and future predicted tasks cannot be recorded.

[0007] Furthermore, errors in voice recognition and video recognition may cause errors in the recorded content.

[0008] Furthermore, the content of business records is personal, meaning it depends on the person creating the record, and therefore the content may vary depending on who creates the record.

[0009] The present invention has been made in light of the above circumstances, and its purpose is to provide a means for suppressing omissions in recorded content and errors in recorded content due to voice recognition errors and video recognition errors, while also reducing the personal nature of business records. [Means for solving the problem]

[0010] A sentence generation device according to one embodiment includes a data structuring unit, a sentence creation unit, and a sentence generation unit. The data structuring unit creates structured data from record data including a plurality of sentences, the structured data including two of the plurality of sentences and a label indicating the relationship between the two sentences. The sentence creation unit uses the structured data to create an input sentence including a first sentence, which is one of the two sentences in the structured data, the label, and a question sentence. The sentence generation unit generates a generated answer sentence by causing a sentence generation model based on a neural network to generate a sentence based on the input sentence. [Effects of the Invention]

[0011] According to the embodiment, it is possible to provide a means for suppressing errors in recorded content and reducing the personal nature of recorded content. [Brief explanation of the drawings]

[0012] [Figure 1] FIG. 1 shows the hardware configuration of a sentence generation device according to the first embodiment. [Figure 2] FIG. 2 shows the functional configuration of the sentence generation device of the first embodiment. [Figure 3] FIG. 3 shows the configuration of a business record dataset in the sentence generation device of the first embodiment. [Figure 4] FIG. 4 shows an example of business record data in the sentence generation device of the first embodiment. [Figure 5] FIG. 5 shows the configuration of a structure dataset in the sentence generation device of the first embodiment. [Figure 6] FIG. 6 shows an example of structure data in the sentence generation device of the first embodiment. [Figure 7] FIG. 7 shows the structure of an input sentence in the sentence generation device of the first embodiment. [Figure 8] FIG. 8 shows the generation of data in the operation of the sentence generation device of the first embodiment. [Figure 9] FIG. 9 shows a flow of learning a sentence generation model by the sentence generation device of the first embodiment. [Figure 10]FIG. 10 shows the flow of sentence generation by the sentence generation device of the first embodiment. [Figure 11] FIG. 11 shows the functional configuration of a sentence generation device according to the second embodiment. [Figure 12] FIG. 12 shows the configuration of a structure dataset in the sentence generation device of the second embodiment. [Figure 13] FIG. 13 shows the generation of data in the operation of the sentence generation device of the second embodiment. [Figure 14] FIG. 14 shows a flow of learning a sentence generation model by the sentence generation device of the second embodiment. [Figure 15] FIG. 15 shows the flow of sentence generation by the sentence generation device of the second embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0013] The following embodiments are described with reference to the drawings. In the embodiments following a given embodiment, differences from the given embodiment are mainly described. Any description of one embodiment also applies to another embodiment unless expressly or obviously excluded.

[0014] 1. First embodiment 1.1.Configuration (Structure) 1.1.1. Structure of the sentence generator Fig. 1 shows the hardware configuration of a sentence generation device according to the first embodiment. As shown in Fig. 1, the sentence generation device 1 includes a CPU (Central Processing Unit) 11, a ROM (Read Only Memory) 12, a RAM (Random Access Memory) 13, a storage 14, an input device 15, an output device 16, and a communication device 17.

[0015] The CPU 11 is an integrated circuit capable of executing various programs. By executing the programs, the CPU 11 controls and executes the overall operation of the sentence generation device 1. By the CPU 11 executing the programs stored in the ROM 12 and loaded onto the RAM 13, the sentence generation device 1 executes various operations and functions as various functional blocks.

[0016] The ROM 12 is a non-volatile memory that stores programs and control data for implementing and controlling the sentence generation device 1. The RAM 13 is a volatile memory that temporarily stores data. The RAM 13 also functions as a work area for the CPU 11.

[0017] The storage 14 is an auxiliary storage device of the sentence generation device 1. Examples of the storage 14 include a hard disk drive (HDD), a solid state drive (SSD), and a memory card. The storage 14 stores data necessary for executing sentence generation. The storage 14 may store a program that causes the CPU 11 to execute sentence generation by the sentence generation device 1.

[0018] The input device 15 is a device that enables a user of the sentence generation device 1 to input data to the sentence generation device 1. Examples of the input device 15 are input devices such as a keyboard, a mouse, a touch panel, or a button switch.

[0019] The output device 16 is a device for presenting visual information and / or audio information to the user of the sentence generation device 1. Examples of the output device 16 include an LCD (Liquid Crystal Display), an EL (Electroluminescence) display, and a speaker. The output device 16 displays a screen for the user to operate the sentence generation device 1 and the generated sentences.

[0020] The communication device 17 is a device that enables the sentence generation device 1 to communicate with the outside. The communication device 17 is configured to be able to communicate with devices and / or networks external to the sentence generation device 1 via wire and / or wireless.

[0021] Fig. 2 shows the functional configuration (functional blocks) of the sentence generation device of the first embodiment. As shown in Fig. 2, the sentence generation device 1 includes a storage unit 21, a data structuring unit 22, an input sentence and correct answer sentence creation unit 23, a sentence generation unit 24, a parameter update unit 25, and an output unit 26. The storage unit 21, the data structuring unit 22, the input sentence and correct answer sentence creation unit 23, the sentence generation unit 24, the parameter update unit 25, and the output unit 26 are realized by some or all of the resources provided by the hardware of the sentence generation device 1 and by execution of a program.

[0022] The storage unit 21 stores various data used by the sentence generation device 1. In one example, the storage unit 21 is realized by a part of the storage capacity of the RAM 13 and the storage 14. The storage unit 21 particularly stores a business record dataset 211, a structure dataset 212, and a sentence generation model 213 as data related to sentence generation by the sentence generation device 1.

[0023] The business record dataset 211 includes a plurality of business record data. Each business record data is a document made up of one or more sentences for each of a plurality of items. The business record data will be described further below.

[0024] The structure data set 212 includes multiple structure data. Each structure data includes multiple sentences in a certain business record data and a relationship label. The relationship label is information that describes the relationship between two sentences that it targets. Examples of the form of the relationship label include characters and symbols. The structure data will be described further below.

[0025] The sentence generation model 213 is a neural network that generates another sentence based on a sentence (text) received as input. The sentence generation model 213 functions based on parameters. An example of the sentence generation model 213 includes a Generative Pre-trained Transformer (GPT)-3.

[0026] The input sentence and correct answer sentence creation unit 23 receives structure data and creates an input sentence and a correct answer sentence based on the received piece of structure data. The input sentence includes a sentence in the received structure data, a relationship label in the structure data, and a question sentence. The input sentence and correct answer sentence creation unit 23 stores a plurality of question sentences in advance. The input sentence and correct answer sentence creation unit 23 selects one of the plurality of question sentences based on a rule. In one example, the rule includes selecting all question sentences one by one in order. The correct answer sentence is a sentence different from the sentence included in the input sentence in the structure data including the input sentence used to generate the generated answer sentence. The input sentence and the correct answer sentence will be described further below.

[0027] The sentence generation unit 24 generates another sentence from the received sentence. The sentence generation unit 24 receives the input sentence from the input sentence and correct answer sentence creation unit 23, and receives the sentence generation model 213 from the storage unit 21. The sentence generation unit 24 uses the received input sentence as an input and also uses the sentence generation model 213 to generate another sentence as an output. Hereinafter, the generated sentence will be referred to as a generated answer sentence.

[0028] The parameter update unit 25 updates the parameters of the sentence generation model 213 based on the received data. The parameter update unit 25 receives the generated answer sentence from the sentence generation unit 24, and receives the input sentence and the correct answer sentence from the correct answer sentence creation unit 23. The parameter update unit 25 updates the parameters of the sentence generation model 213 based on the generated answer sentence and the correct answer sentence so that the generated answer sentence approaches the correct answer sentence. The parameter update is performed on the sentence generation model 213 in the storage unit 21. The sentence generation model 213 is updated by the parameter update.

[0029] The output unit 26 displays the generated answer sentence as characters and / or voice. The output unit 26 includes the output device 16 and has a function for displaying the generated answer sentence on the output device 16. The output unit 26 has a function for outputting the generated answer sentence to the outside of the output unit 26.

[0030] 3 shows the configuration of a business record dataset in the sentence generation device of the first embodiment. As shown in FIG. 3 and described above with reference to FIG. 2, each business record dataset 211 includes a plurality of business record data. Each business record data includes a plurality of items and sentences associated with each item and related to the items. In one example, the business record data is provided in the form of pre-created data from outside the sentence generation device 1 via the input device 15 and / or the communication device 17.

[0031] FIG. 4 shows an example of business record data in the sentence generation device of the first embodiment. FIG. 4 shows an example where the business record data is a nursing record in the nursing or care industry. As shown in FIG. 4, the nursing record is written in the form of focus charting. Examples of items used are focus, data, action, and response. The sentence for focus is "has fever." The business record data shown in FIG. 4 includes three data items that are independent of each other. The sentence for each data is either "body temperature 39.1°C," "no chills," or "has facial flushing." The sentence for action is "administered one tablet of loxoprofen." The sentence for response is "body temperature reaches 37.2°C."

[0032] FIG. 5 shows the configuration of a structure dataset in the sentence generation device of the first embodiment. As shown in FIG. 5 and described above with reference to FIG. 2, each structure dataset includes two arbitrary sentences (hereinafter, sometimes referred to as the "first sentence" and the "second sentence") in a certain piece of business record data and relationship labels for the two sentences. The relationship labels may be in any form as long as they are information indicating the relationship between the two sentences. Examples of when the relationship labels are text information include "concretization," "abstraction," "cause," "result," "continuation," and "precedence." The relationship labels may be unique symbols or numbers associated with the text information.

[0033] FIG. 6 shows an example of structure data in the sentence generation device of the first embodiment. FIG. 6 shows examples of sentences in a certain nursing record data set, including "has fever," "body temperature 39.1°C," "no chills," and "has flushed face." If two of these sentences are related from the perspective of concretization, the sentences "body temperature 39.1°C" and "has flushed face" can each be evaluated as a concretization of the sentence "has fever." On the other hand, the sentence "no chills" is not evaluated as a concretization of the sentence "has fever." Based on these evaluations, two pieces of structure data are created. The first piece of structure data includes "has fever" as the first sentence, "body temperature 39.1°C" as the second sentence, and "concretization" as the relationship label. The second piece of structure data includes "has fever" as the first sentence, "has flushed face" as the second sentence, and "concretization" as the relationship label.

[0034] 7 shows the structure of input sentences in the sentence generation device of the first embodiment. As shown in FIG. 7 and described above with reference to FIG. 2, the input sentences are based on one piece of structure data. Each input sentence includes the first sentence and relationship label in the structure data on which the input sentence is based, as well as a question sentence. The input sentences will be described further below.

[0035] 1.2.Operation Fig. 8 shows data generation in the operation of the sentence generation device of the first embodiment. The sentence generation device 1 can train a sentence generation model and generate sentences. Fig. 8 shows both data generation in training the sentence generation model and data generation in sentence generation.

[0036] In sentence generation using a selected piece of business record data as input, data is generated as follows. First, a structure data set 212 is generated by generating a plurality of structure data from the business record data. For example, sentence n is selected as the first sentence of a piece of structure data, and sentence n+1 is selected as the second sentence of this structure data. The relationship label included in the structure data indicates the relationship between the first sentence (i.e., sentence n) and the second sentence (i.e., sentence n+1). A plurality of unique structure data including various combinations of the first sentence, the second sentence, and the relationship label are generated. That is, the combination of the first sentence, the second sentence, and the relationship label of each piece of structure data is different from the combination of the first sentence, the second sentence, and the relationship label of another piece of structure data.

[0037] An input sentence is generated from one piece of structure data in the structure dataset 212. The input sentence includes a first sentence and a relationship label in the structure data, and further includes a question sentence. The question sentence includes a question that enables the sentence generation model 213 to generate a sentence having content similar to that of the second sentence in the structure data using the first sentence and the relationship label as input. Based on an example where the first sentence is "has fever," the relationship label is "concrete," and the second sentence is "body temperature 39.1°C," an example of a question sentence is, "You are a nurse. When you abstractly observe the following about a patient's condition, what specific situation is possible?"

[0038] The sentence generation model 213 generates a generated answer sentence from the input sentence. For example, the generated answer sentence includes a sentence such as "body temperature 39.1 degrees Celsius."

[0039] In sentence generation model training using a selected piece of business record data as input, in addition to generating a generated answer sentence, a correct answer sentence is obtained. The correct answer sentence is the second sentence in the structure data. As will be described later, based on the generated answer sentence and the correct answer sentence, constraints of the sentence generation model that bring the generated answer sentence closer to the correct answer sentence are determined.

[0040] FIG. 9 shows a flow of sentence generation model learning by the sentence generation device of the first embodiment. The flow in FIG. 9 shows a flow using one selected business record data. In one example, the flow in FIG. 9 is started by a user of the sentence generation device 1 operating the sentence generation device 1. In one example, the selection of business record data is performed by the user. In another example, the selection of business record data is performed by the sentence generation device 1 in accordance with predetermined rules.

[0041] As shown in FIG. 9, the data structuring unit 22 acquires one selected business record data from the memory unit 21, and creates a structured data set by creating multiple structured data from the acquired business record data (step ST1).

[0042] The input sentence and correct answer sentence creating unit 23 acquires one piece of structure data from the created structure data set, and creates an input sentence and a correct answer sentence from the acquired structure data (step ST2).

[0043] The sentence generating unit 24 uses the input sentence and the input sentence generated by the correct answer sentence generating unit 23 as inputs, and also uses the sentence generation model to generate a generated answer sentence (step ST3).

[0044] The parameter update unit 25 determines constraints to be imposed on the sentence generation model using the generated answer sentence and the correct answer sentence, and updates the parameters of the sentence generation model based on the determined constraints (step ST4). The constraints are based on a learning method that brings the generated answer sentence closer to the correct answer sentence. In one example, the constraints include changing the parameters of the sentence generation model to reduce the cross-entropy error. Specifically, the constraints are as follows.

[0045] The cross-entropy error is calculated by Equation 1.

[0046]

number

[0047] k is the index of the generated answer sentence and the correct answer sentence. k is the probability of correctness for each of the multiple generated answer sentences, i.e., the probability that it matches the correct answer sentence. k is y k is correct, it has a value of 1, and is incorrect, it has a value of 0. The higher the probability that a correct one of the multiple generated answer sentences, i.e., a generated answer sentence that matches the correct answer sentence, is generated, the smaller the absolute value of the cross-entropy error. Therefore, by adjusting the parameters of the sentence generation model based on the cross-entropy error so as to increase the probability that a generated answer sentence that matches the correct answer sentence is generated, the probability of generating a generated answer sentence that is the same as the correct answer sentence by the sentence generation model is improved. In one example, in step ST4, the adjustment of the sentence generation model can be repeated until the accuracy of sentence generation by the sentence generation model reaches a certain level.

[0048] When step ST4 is completed, the flow of FIG. 9 ends.

[0049] Fig. 10 shows a flow of sentence generation by the sentence generation device of the first embodiment. In one example, the flow of Fig. 10 starts when a user of the sentence generation device 1 instructs the sentence generation device 1 to generate a sentence.

[0050] As shown in FIG. 10, after step ST1, the input sentence and correct answer sentence creation unit 23 acquires one structure data set from the generated structure data sets and creates an input sentence from the acquired structure data (step ST11). Step ST11 continues to step ST3. After step ST3, the output unit 26 outputs the generated answer sentence (step ST12). Examples of output include displaying the generated answer sentence on the output device 16 and providing data indicating the generated answer sentence outside the output unit 26. When step ST12 ends, the flow in FIG. 10 ends.

[0051] 1.3.Advantages (Effects) The sentence generation device 1 of the first embodiment creates structured data including two sentences and labels indicating the relationship between the two sentences from business record data including multiple sentences, and uses a question sentence that uses the structured data as input to generate a generated answer sentence using a sentence generation model based on a neural network. The generated answer sentence can be used as a sentence in the business record data. Because the generated answer sentence is generated automatically, omissions and / or errors in content that can occur when video recognition and / or voice recognition are used in the generated record (new business record data) are reduced. Furthermore, because the generated answer sentence is generated by the device, variations in content that can occur when records are created by humans due to the dependency on the creator are reduced.

[0052] 2. Second embodiment The second embodiment differs from the first embodiment in the contents of the structure data and the creation of the structure data set.

[0053] FIG. 11 shows a functional configuration (functional blocks) of a sentence generation device according to the second embodiment. As shown in FIG. 11, the sentence generation device 1 further includes a data narrowing-down unit 28. The data narrowing-down unit 28 deletes one or more pieces of structure data from the structure data set 212 to generate a structure data set 212A that includes the remaining structure data. The data narrowing-down unit 28 deletes structure data selected in accordance with certain criteria from the structure data set 212. An example of the criteria includes whether the structure data is linked to an input sentence that is estimated not to be useful in improving the accuracy of the sentence generation model. In one example, the data narrowing-down unit 28 deletes structure data based on a confidence level regarding the accuracy of the relationship labels of the data structure.

[0054] FIG. 12 shows the configuration of structure data sets in the sentence generation device of the second embodiment. As shown in FIG. 12, each structure data set includes a confidence level in addition to a first sentence, a second sentence, and a relational label. Examples of the form of the confidence level include text information, numerical values, and symbols. In one example, when a numerical value is used, the numerical value is an integer between 1 and 10, and the higher the numerical value, the higher the confidence level. In one example, the confidence level is input by a user of the sentence generation device 1 via the input device 15. In one example, the confidence level is assigned by an expert who has knowledge of the business record indicated by the business record data. The confidence level may be received by communication using the communication device 17.

[0055] Fig. 13 shows data generation in the operation of the sentence generation device of the second embodiment. As shown in Fig. 13, after the structure data set 212 is created, one or more structure data are deleted from the structure data set 212 based on their confidence levels to create a structure data set 212A. In one example, in a case where the confidence level is expressed by an integer between 1 and 10, if a certain structure data has a confidence level of 5 or less, that structure data is deleted from the structure data set 212. An input sentence is created from the structure data in the structure data set 212A.

[0056] FIG. 14 shows a flow of sentence generation model learning by the sentence generation device of the second embodiment. The flow in FIG. 14 shows a flow using one selected business record data. In one example, the flow in FIG. 14 starts when a user of the sentence generation device 1 operates the sentence generation device 1. As shown in FIG. 14, after step ST1, the data narrowing-down unit 28 creates a narrowed-down structure data set 212A from the structure data set 212 (step ST21). Step ST21 continues to step ST2. The structure data used in step ST2 is the structure data in the structure data set 212A. Step ST2 and subsequent steps are the same as those in the first embodiment.

[0057] FIG. 15 shows a flow of sentence generation by the sentence generation device of the second embodiment. In one example, the flow of FIG. 15 starts when a user of the sentence generation device 1 instructs the sentence generation device 1 to execute sentence generation. As shown in FIG. 15, step ST1 continues to step ST21, and step ST21 continues to step ST11. The structure data used in step ST2 is the structure data in the structure data set 212A. Steps from ST11 onwards are the same as those in the first embodiment.

[0058] As in the first embodiment, the sentence generation device 1 of the second embodiment creates structured data including two sentences and labels indicating the relationship between the two sentences from business record data including multiple sentences, and uses a question sentence that uses the structured data as input to generate a generated answer sentence using a sentence generation model based on a neural network. Therefore, the same advantages as in the first embodiment can be obtained.

[0059] Furthermore, according to the second embodiment, the structure data further includes a degree of certainty indicating the accuracy of the relational label, and among the generated plurality of structured data, structured data with a low degree of certainty is deleted, and the remaining structured data is used. Therefore, the sentence generation model is adjusted based on the structured data with a high degree of relational accuracy, and thus a generated answer sentence with a high degree of relation is obtained.

[0060] Although several embodiments of the present invention have been described, these embodiments are presented as examples and are not intended to limit the scope of the invention. These embodiments can be implemented in various other forms, and various omissions, substitutions, and modifications can be made without departing from the spirit of the invention. These embodiments and their modifications are included within the scope and spirit of the invention, as well as within the scope of the invention described in the claims and their equivalents. [Explanation of symbols]

[0061] 1...Sentence generation device, 11...CPU, 12...ROM, 13...RAM, 14…Storage, 15...input device, 16...output device, 17...Communication devices, 21...Storage section, 22...Data structuring section, 23...input sentence and correct answer sentence creation unit, 24...Sentence generation section, 25...parameter update unit, 26...Output section

Claims

1. a data structuring unit that generates structured data from record data including a plurality of sentences, the structured data including two sentences among the plurality of sentences and a label indicating a relationship between the two sentences; a sentence creation unit that uses the structure data to create an input sentence including a first sentence that is one of the two sentences of the structure data, the label, and a question sentence; a sentence generation unit that generates a generated answer sentence by causing a sentence generation model based on a neural network to generate a sentence based on the input sentence; A sentence generation device comprising:

2. A parameter update unit is further provided, the sentence creation unit provides the parameter update unit with a second sentence, which is the other of the two sentences of the structure data; the parameter update unit updates parameters of the sentence generation model based on the second sentence and the generated answer sentence so as to impose a constraint on the sentence generation model that causes the sentence generation model to generate the generated answer sentence that is closer to the second sentence. The sentence generation device according to claim 1 .

3. the question sentence includes a question that causes the sentence generation model to generate the generated answer sentence having the same content as the second sentence, using the first sentence and the label; The sentence generation device according to claim 1 .

4. The record data is nursing record data, the label indicates one of instantiation, abstraction, cause, effect, continuation, and antecedent; The sentence generation device according to claim 1 .

5. Further comprising a data narrowing down unit, the data structuring unit creates a plurality of first structure data from the recording data, each of the plurality of first structure data includes two sentences among the plurality of sentences, a label indicating a relationship between the two sentences, and a confidence level; the data narrowing unit creates a plurality of second structure data by deleting one or more of the plurality of first structure data based on the degree of certainty; the structure data used by the sentence creation unit is one of the plurality of second structure data; The sentence generation device according to claim 1 .

6. the certainty of each of the plurality of first structure data has a value based on the accuracy of the label of the first structure data; The sentence generation device according to claim 5.

7. generating structure data from record data including a plurality of sentences, the structure data including two sentences among the plurality of sentences and a label indicating a relationship between the two sentences; using the structure data to create an input sentence including a first sentence, which is one of the two sentences of the structure data, the label, and a question sentence; generating a generated response sentence by causing a sentence generation model based on a neural network to generate a sentence based on the input sentence; A sentence generation method comprising:

8. To the device, generating structure data from record data including a plurality of sentences, the structure data including two sentences among the plurality of sentences and a label indicating a relationship between the two sentences; using the structure data to create an input sentence including a first sentence, which is one of the two sentences of the structure data, the label, and a question sentence; generating a generated response sentence by causing a sentence generation model based on a neural network to generate a sentence based on the input sentence; A program to execute.

Images