Abstract generation server, abstract generation system, abstract generation program, and abstract generation method

Abstract

Generate summary information indicating the content of the document information from the document information. 【Solution means】 A summary generation server that generates summary information from the document information to be summarized includes an extraction-type summary generation unit that generates extraction-type summary information by extracting sentences with high importance from the document information, an abstract-type summary generation unit that generates abstract-type summary information by summarizing the description content of the document information using a natural language processing model, a similarity calculation unit that calculates the similarity between the extraction-type summary information and the abstract-type summary information, and a summary output unit that outputs the extraction-type summary information or the abstract-type summary information based on the similarity.

Description

Technical Field

[0001] The present disclosure relates to a summary generation server, a summary generation system, a summary generation program, and a summary generation method for generating summary information from document information.

Background Art

[0002] Conventionally, a technique for analyzing a sentence and automatically generating a summary sentence has been used. For example, Patent Document 1 discloses an apparatus for automatically generating summary information from medical record information. This apparatus executes processes such as morphological analysis, syntactic analysis, and semantic analysis, divides each sentence included in the medical record information into elements, determines the importance of each element, and generates summary information by adopting elements whose importance is equal to or higher than a predetermined reference value.

[0003] In addition to an extraction-type summary generation method that evaluates the importance of sentences and phrases included in document information and generates summary information by extracting sentences and phrases with high importance, an abstract-type summary generation method that grasps the meaning of sentences included in document information and generates new sentences is known. With the development of natural language processing technology in recent years, it has become possible to generate natural sentences without a sense of incongruity, including phrases not present in the original document information, by the abstract-type summary generation method.

Prior Art Documents

Patent Documents

[0004]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0005] Since the extracted summary information uses words and sentences extracted from the original document information, it is less likely to have content breakdowns. However, there are cases where the generated text, which is created by cutting and pasting only important parts, becomes unnatural. On the other hand, although the abstract summary information is in a natural text that is easy to understand, there is a possibility that a text that does not exist in the original document information is generated, resulting in content that is different from the gist of the document information.

[0006] For example, when generating summary information from document information showing medical-related test results, medical-related words and test values may be determined to be important, and the extracted summary information may end up being a list of words and values. On the other hand, although the abstract summary information is in a natural text, there are cases where it becomes a text in which it is difficult to understand test results such as the findings of the examining doctor. For this reason, it may not be possible to grasp the test results from the summary information, and ultimately, it may be necessary to check the original document information.

[0007] The present disclosure has been made in view of the prior art including the above problems, and one of its objects is to provide a summary generation server, a summary generation system, a summary generation program, and a summary generation method that can generate summary information that can easily confirm the content included in the document information.

Means for Solving the Problems

[0008] The summary generation server according to the present disclosure is a summary generation server that generates summary information from document information to be summarized, and includes an extraction-type summary generation unit that generates extraction-type summary information by extracting sentences with high importance from the document information, an abstract-type summary generation unit that generates abstract-type summary information by summarizing the description content of the document information using a natural language processing model, a similarity calculation unit that calculates the similarity between the extraction-type summary information and the abstract-type summary information, and a summary output unit that outputs the extraction-type summary information or the abstract-type summary information based on the similarity.

[0009] In the above configuration, the summary output unit may output the abstract-type summary information when the similarity is equal to or greater than a predetermined threshold, and may output the extraction-type summary information when the similarity is lower than the threshold.

[0010] In the above configuration, a simplified summary generation unit may be further provided that generates simplified summary information obtained by changing, based on the information registered in the dictionary, the words and phrases in the field handled by the document information, the dictionary in which the words and phrases related to the words and phrases are registered, and the words and phrases in the field included in the summary information output by the summary output unit.

[0011] In the above configuration, the information registered in the dictionary corresponding to the words and phrases in the field may include at least one of a paraphrase of the words and phrases and a sentence explaining the words and phrases.

[0012] In the above configuration, an evaluation reception unit that receives evaluation of the abstract summary information may be further provided, and the abstract summary generation unit may perform machine learning using the document information and the abstract summary information generated from the document information as teacher data.

[0013] The summary generation system according to the present disclosure includes a summary generation server according to the above configuration, and a terminal device that transmits the document information to the summary generation server and receives summary information generated by the summary generation server from the document information.

[0014] The summary generation program according to the present disclosure is a summary generation program that generates summary information from document information to be summarized, and includes an extraction type summary information generation procedure that generates extraction type summary information by extracting sentences with high importance from the document information, an abstract type summary information generation procedure that generates abstract type summary information by summarizing the description content of the document information using a natural language processing model, a similarity calculation procedure that calculates the similarity between the extraction type summary information and the abstract type summary information, and an output procedure that outputs the extraction type summary information or the abstract type summary information based on the similarity, and causes a computer to execute them.

[0015] Executed by a computer device,The abstract generation method according to the present disclosure is an abstract generation method for generating abstract information from document information to be summarized, and includes an extraction-type abstract information generation step of generating extraction-type abstract information by extracting sentences with high importance from the document information, an abstract-type abstract information generation step of generating abstract-type abstract information by summarizing the description content of the document information using a natural language processing model, a similarity calculation step of calculating the similarity between the extraction-type abstract information and the abstract-type abstract information, and an output step of outputting the extraction-type abstract information or the abstract-type abstract information based on the similarity.

Effect of the Invention

[0016] According to the abstract generation server, abstract generation system, abstract generation program, and abstract generation method according to the present disclosure, it is possible to create abstract information that can easily confirm the content included in the document information from the text information.

Brief Description of the Drawings

[0017]

Figure 1

Figure 2

Figure 3

Figure 4

Figure 5

Mode for Carrying Out the Invention

[0018] Hereinafter, embodiments of the abstract generation server, abstract generation system, abstract generation program, and abstract generation method according to the present disclosure will be described with reference to the accompanying drawings.

[0019] <Outline of the Abstract Generation System> FIG. 1 is a diagram for explaining the outline of the summary generation server 10 according to the present embodiment. As shown in FIG. 1, the summary generation server 10 includes an extractive summary generation unit 22, an abstractive summary generation unit 23, a similarity calculation unit 24, and a summary output unit 25. The summary generation server 10 receives an input of document information and outputs summary information generated from the document information. The summary generation server 10 can generate and output summary information from one or a plurality of pieces of input document information.

[0020] The extractive summary generation unit 22 generates extractive summary information from the document information. The extractive summary generation unit 22 determines the importance of words and sentences included in the document information, extracts words and sentences with high importance, and generates extractive summary information. When one piece of document information is input, the extractive summary generation unit 22 generates extractive summary information for that document information. When a plurality of pieces of document information are input, the extractive summary generation unit 22 generates one piece of extractive summary information for all of the input document information. The method for generating the extractive summary is not particularly limited, and a conventionally known method may be used. For example, importance evaluation may be performed using a method such as TF-IDF (Term Frequency Inverse Document Frequency) to generate extractive summary information, or a machine learning model such as BERT (Bidirectional Encoder Representations from Transformers) may be used to generate extractive summary information.

[0021] The abstractive summary generation unit 23 generates abstractive summary information from the text information. The abstractive summary generation unit 23 captures the content indicated by the document information through natural language processing using an encoder-decoder model, and generates summary information indicating the content. When one piece of document information is input, the abstractive summary generation unit 23 generates abstractive summary information for that document information. When a plurality of pieces of document information are input, the abstractive summary generation unit 23 generates one piece of abstractive summary information for all of the input document information. The method for generating the abstractive summary is not particularly limited, and a conventionally known method may be used. For example, an encoder-decoder model such as a Transformer may be used to generate abstractive summary information.

[0022] The similarity calculation unit 24 calculates the similarity between the extractive summary information generated by the extractive summary generation unit 22 and the abstract summary information generated by the abstract summary generation unit 23. The method for calculating the similarity is not particularly limited, and a conventionally known method may be used. For example, the cosine similarity between the extractive summary information and the abstract summary information may be calculated and this value may be used as the similarity.

[0023] The summary output unit 25 compares the similarity calculated by the similarity calculation unit 24 with a preset threshold value, and based on the comparison result, outputs the extractive summary information or the abstract summary information. For example, when the similarity is equal to or greater than the threshold value, the abstract summary information is output, and when the similarity is lower than the threshold value, the extractive summary information is output.

[0024] In this way, the summary generation server 10 generates two types of summary information, extractive summary information and abstract summary information, from the same document information for one or more document information, and outputs the abstract summary information on the condition that the similarity between the abstract summary information and the extractive summary information satisfies a predetermined condition. When the value indicating the similarity between the content of the abstract summary information and the content of the extractive information is equal to or greater than a predetermined threshold value, the abstract summary information is output, and when it is smaller than the threshold value, the extractive summary information is output.

[0025] It is preferable to output the abstract summary information, which is more likely to obtain a natural sentence using natural language processing technology, rather than the extractive summary information that is likely to result in an unnatural sentence. However, the abstract summary information may indicate content different from the original document information. While basically outputting the abstract summary information, the summary generation server 10 outputs the extractive summary information when the content of the abstract summary information is significantly different from the content of the extractive summary information, thereby preventing the output of summary information with content different from the document information. Hereinafter, a specific example of the case where the summary generation server 10 generates summary information from one document information will be described.

[0026] <Configuration example of the summary generation system> FIG. 2 is a diagram showing a configuration example of a summary generation system including a summary generation server 10. The summary generation system includes, in addition to the summary generation server 10, an instruction terminal 50 used for instructing the generation of summary information, and browsing terminals 61 and 62 used for browsing the summary information. The summary generation server 10 can transmit and receive information to and from each terminal 50, 61, 62 via a network 2.

[0027] The instruction terminal 50 is a terminal used by a user who instructs the summary generation server 10 to generate summary information. Document information is transmitted from the instruction terminal 50 to the summary generation server 10. The browsing terminals 61 and 62 are terminals used by users who browse the summary information generated by the summary generation server 10. Summary information is transmitted from the summary generation server 10 to the browsing terminals 61 and 62. For example, a computer device including an operation unit, a display unit, a control unit, a storage unit, and a communication unit is used as the instruction terminal 50 and the browsing terminals 61 and 62. At least one of the instruction terminal 50 and the browsing terminals 61 and 62 may be realized by a portable terminal such as a smartphone or a tablet terminal. For the sake of convenience of explanation, the instruction terminal 50 and the browsing terminals 61 and 62 are distinguished, and one instruction terminal 50 and two browsing terminals 61 and 62 are shown. However, the number of instruction terminals and the number of browsing terminals using the summary generation system are not particularly limited, and an aspect in which the instruction terminal also serves as a browsing terminal may be adopted.

[0028] As shown in FIG. 2, the summary generation server 10 includes a control unit 20, a communication unit 30, and a storage unit 40. Although not shown in FIG. 2, the summary generation server 10 may be provided with an operation unit and a display unit. For example, a computer device including an operation unit, a display unit, a control unit, a storage unit, and a communication unit may be used as the summary generation server 10.

[0029] The communication unit 30 is used for the summary generation server 10 to communicate with an external device via the network 2. The storage unit 40 is a non-volatile storage device in which various data necessary for the operation of the summary generation server 10 are stored. The data stored in the storage unit 40 includes document information data 41, summary information data 42, a machine learning dataset 43, and dictionary data 44.

[0030] The control unit 20 controls the entire summarization generation server 10. The control unit 20 includes an instruction reception unit 21, an extraction-type summarization generation unit 22, an abstract-type summarization generation unit 23, a similarity calculation unit 24, a summarization output unit 25, a simplified-type summarization generation unit 26, and an evaluation reception unit 27.

[0031] Programs corresponding to these functional units 21 to 27 of the control unit 20 are stored in advance in the storage unit 40 or a dedicated storage device, and when the programs are executed by hardware such as a CPU, the functions and operations of the control unit 20 and each functional unit 21 to 27 are realized. The functions and operations of the summarization generation server 10 are realized by the control unit 20 and each functional unit 21 to 27.

[0032] The instruction reception unit 21 receives document information from the instruction terminal 50 via the communication unit 30 and receives an instruction for generating summarization information. The instruction reception unit 21 stores the document information received from the instruction terminal 50 in the document information data 41 of the storage unit 40.

[0033] The extraction-type summarization generation unit 22 generates extraction-type summarization information from the document information stored in the storage unit 40. The extraction-type summarization generation unit 22 stores the generated extraction-type summarization information in the summarization information data 42 of the storage unit 40. The extraction-type summarization generation unit 22 generates extraction-type summarization information by using, for example, a machine learning model such as BERT that has been pre-trained with a dataset for generating extraction-type summarization included in the machine learning dataset 43 of the storage unit 40. However, the method for generating extraction-type summarization information executed by the extraction-type summarization generation unit 22 is not particularly limited, and it may be performed in another manner.

[0034] The abstract summary generation unit 23 generates abstract summary information from the document information stored in the storage unit 40. The abstract summary generation unit 23 stores the generated abstract summary information in the summary information data 42 of the storage unit 40. The abstract summary generation unit 23 uses, for example, a natural language processing model such as T5 (Text-to-Text Transfer Transformer) pre-trained with a dataset for abstract summary generation included in the machine learning dataset 43 of the storage unit 40 to generate abstract summary information. However, the method for generating the abstract summary information executed by the abstract summary generation unit 23 is not particularly limited, and it may be performed in another manner.

[0035] The similarity calculation unit 24 calculates the similarity between the extractive summary information and the abstract summary information. The similarity calculation unit 24, for example, converts the extractive summary information and the abstract summary information into vectors, calculates the inner product of the two obtained vectors, and uses this value as the similarity between the extractive summary information and the abstract summary information. However, the method for calculating the similarity is not particularly limited, and the cosine similarity may be used as the similarity, or the similarity may be calculated using BERT or Sentence-BERT based on BERT.

[0036] The summary output unit 25 outputs the summary information to an external device via the communication unit 30. The summary output unit 25 outputs the extractive summary information or the abstract summary information stored in the summary information data 42 of the storage unit 40 to the external device based on the similarity calculated by the similarity calculation unit 24. When the value of the similarity is equal to or greater than a predetermined threshold, the abstract summary information is output, and when it is less than the predetermined threshold, the extractive summary information is output.

[0037] The abridged summary generation unit 26 generates abridged summary information with some words and expressions changed from the extractive summary information or the abstractive summary information output by the summary output unit 25 to an external device. The abridged summary generation unit 26 stores the generated abridged summary information in the summary information data 42 of the storage unit 40. When the value of the similarity calculated by the similarity calculation unit 24 is equal to or greater than a predetermined threshold, the abridged summary generation unit 26 generates abridged summary information from the abstractive summary information, and when it is less than the predetermined threshold, the abridged summary generation unit 26 generates abridged summary information from the extractive summary information.

[0038] The abridged summary information is generated using the dictionary data 44 prepared in advance in the storage unit 40. In the dictionary data 44, at least one of a word or phrase in the field handled by the document information, another word or phrase obtained by paraphrasing the word or phrase, and a sentence explaining the word or phrase is registered in association.

[0039] For example, when the summary generation system is used in a medical-related facility, words or phrases obtained by paraphrasing medical-related technical terms so that patients can understand them, and sentences for explaining the technical terms to patients are registered in the dictionary data 44 in association with each technical term. The abridged summary generation unit 26 replaces the technical terms included in the summary information output by the summary output unit 25 with the words or sentences for patients registered in the dictionary data 44 to generate abridged summary information. However, the method for generating the abridged summary information is not particularly limited, and the abridged summary generation unit 26 may use a machine learning model to generate the abridged summary information in the same manner as the extractive summary generation unit 22 or the abstractive summary generation unit 23 instead of or in addition to using the dictionary data 44.

[0040] The summary output unit 25 outputs summary information to an external device. The summary output unit 25 can output summary information corresponding to the type of the external device that requests the summary information from among the extractive summary information, the abstractive summary information, and the abridged summary information stored in the summary information data 42 of the storage unit 40. For example, in the example of a medical-related facility, extractive summary information or abstractive summary information is output to the instruction terminal 50 and the browsing terminal 61 used by medical staff based on the similarity, and abridged summary information is output to the browsing terminal 62 used by patients. Specific examples will be described later.

[0041] The evaluation reception unit 27 receives the evaluation result of the summary information from the instruction terminal 50 via the communication unit 30. On the display unit of the instruction terminal 50, the extraction-type summary information or the abstract-type summary information is displayed based on the similarity calculated by the similarity calculation unit 24. The summary indicator checks the summary information generated by the summary generation server 10, operates the operation unit of the instruction terminal 50, and inputs the evaluation value of the summary information. For example, within a predetermined evaluation value range, a higher evaluation value is input as the satisfaction degree of the summary indicator is higher. If the evaluation value received by the evaluation reception unit 27 is equal to or greater than a preset threshold value, the document information before summarization stored in the storage unit 40 and the summary information generated from the document information are additionally registered in the machine learning dataset 43. When the evaluation by the summary indicator is high, data is additionally registered and the machine learning dataset 43 is updated. Note that the evaluation method of the summary information is not limited to the evaluation by the evaluation value, and for example, it may be a mode in which the summary indicator accepts either an evaluation of whether the summary information is satisfactory or not.

[0042] If the summary information output to the instruction terminal 50 is extraction-type summary information, the dataset for extraction-type summary generation included in the machine learning dataset 43 is updated. If the summary information output to the instruction terminal 50 is abstract-type summary information, the dataset for abstract-type summary generation included in the machine learning dataset 43 is updated. When the dataset of the learning target included in the machine learning dataset 43 is updated, the extraction-type summary generation unit 22 and the abstract-type summary generation unit 23 execute re-learning using the updated dataset. Thereby, an improvement in the function of generating summary information by the summary generation server 10 can be expected.

[0043] <Summary Generation Process> FIG. 3 is a flowchart showing the flow of the summary generation process executed by the summary generation server 10. The summary generation server 10 is in a standby state where it can receive an instruction for summary generation (step S1; No). When the instruction reception unit 21 receives an instruction for generating summary information from the instruction terminal 50 (step S1; Yes), summary information is generated from the document information received from the instruction terminal 50. The extractive summary generation unit 22 generates extractive summary information (step S2), and the abstractive summary generation unit 23 generates abstractive summary information (step S3). FIG. 3 shows an example in which two pieces of summary information, i.e., extractive summary information and abstractive summary information, are generated in parallel processing, but a mode in which the two pieces of summary information are sequentially generated in order may also be acceptable.

[0044] After generating the summary information, the similarity calculation unit 24 calculates the similarity between the extractive summary information and the abstractive summary information (step S4), and this value is compared with a threshold value (step S5).

[0045] If the similarity is equal to or greater than the threshold value (step S5; Yes), the summary output unit 25 outputs the abstractive summary information (step S6), and if the similarity is lower than the threshold value (step S5; No), the summary output unit 25 outputs the extractive summary information (step S7). The abstractive summary information or the extractive summary information output by the summary output unit 25 is displayed on the screen of the instruction terminal 50 that instructed the generation of the summary information.

[0046] The simplified summary generation unit 26 generates simplified summary information from the abstractive summary information or the extractive summary information output to the instruction terminal 50 (step S8). The summary generation server 10 that has generated the simplified summary information returns to the standby state of step S1 where it can receive a summary generation instruction.

[0047] The abstractive summary information or the extractive summary information output to the instruction terminal 50 and the simplified summary information generated from the summary information are stored in the summary information data 42 of the storage unit 40. Users other than the summary instructing person can also view the abstractive summary information or the extractive summary information output to the instruction terminal 50 and the simplified summary information by using the browsing terminals 61 and 62. Specific examples will be described later.

[0048] <Learning Process> FIG. 4 is a flowchart showing the flow of the learning process executed by the summary generation server 10. When the abstract summary information or the extracted summary information output by the summary output unit 25 in step S6 or S7 shown in FIG. 3 is displayed on the instruction terminal 50 operated by the summary instructing person, the summary instructing person can evaluate the summary information. When the summary instructing person evaluates the summary information, the process shown in FIG. 4 is started.

[0049] When the evaluation reception unit 27 of the summary generation server 10 receives the evaluation value by the summary instructing person, it compares the evaluation value with a preset threshold value (step S11). If the evaluation value is smaller than the threshold value (step S11; Yes), the process ends without performing machine learning. That is, when the summary generation server 10 cannot generate summary information with a high evaluation by the summary instructing person, it does not execute machine learning using the summary information.

[0050] If the evaluation value is equal to or higher than the threshold value (step S11; Yes), the evaluation reception unit 27 registers and updates the summary information with a high evaluation value and the document information that is the source of this summary information in the machine learning dataset 43 of the storage unit 40 (step S12).

[0051] When the extracted summary information is highly evaluated by the instruction terminal 50, the dataset for generating the extracted summary included in the machine learning dataset 43 is updated, and the extracted summary generation unit 22 executes machine learning using the updated dataset (step S13).

[0052] When the abstract summary information is highly evaluated by the instruction terminal 50, the dataset for generating the abstract summary included in the machine learning dataset 43 is updated, and the abstract summary generation unit 23 executes machine learning using the updated dataset (step S13).

[0053] When the summary generation server 10 can generate summary information with a high degree of satisfaction of the summary instructing person from the document information created by the summary instructing person, by executing machine learning using these document information and summary information as teacher data, an improvement in the function related to summary information generation can be expected.

[0054] In addition, when the evaluation of the summary information is performed by a binary choice of whether or not the summary requester is satisfied with the summary information, if the summary requester is satisfied, the extraction-type summary generation unit 22 and the abstract-type summary generation unit 23 may execute re-learning. The timing of executing the re-learning is not particularly limited. For example, the re-learning may be executed immediately after the update of the dataset, or may be executed at a predetermined time on the day when the dataset is updated, or may be executed at a predetermined timing after the next day after the dataset update.

[0055] <Example of summary information generation> FIG. 5 is a schematic diagram for explaining an example of the summary generation process executed by the summary generation server 10 and the generated summary information. FIG. 5 shows an example in which a medical institution uses the summary generation system to automatically generate a summary from the examination results by an examining doctor.

[0056] The examining doctor includes a radiologist and a clinical laboratory doctor. For example, when a patient's radiation examination such as an X-ray examination or a CT examination is performed, a radiologist performs an image diagnosis of the radiation image and creates a document including the examination results. For example, when a specimen examination of cells, tissues, etc. collected from a patient is performed, a clinical laboratory doctor examines the specimen and creates a document including the examination results.

[0057] As shown in FIG. 5, the examining doctor operates the instruction terminal 50 to transmit document information 100 including the examination results to the summary generation server 10 and instructs the generation of summary information. The document information 100 shown in FIG. 5 shows an example of examination results created by an examining doctor who performs a specimen examination for breast cancer.

[0058] The summary generation server 10 generates extraction-type summary information 111 and abstract-type summary information 112 from the document information 100. In the example shown in FIG. 5, a lot of information has been input into the document information 100 by the examining doctor, but the abstract-type summary information 112 is a natural sentence that extracts important information from the document information 100.

[0059] Specifically, the abstract summary information 112 is written in an easy-to-understand manner, such as findings that, as a result of examining the specimen, it was determined to be a benign tumor called "fibrocystic breast disease" rather than breast cancer, or that "Congo-red staining" was performed to rule out the possibility of another disease called "amyloidosis".

[0060] On the other hand, the extracted summary information 111 in the example shown in FIG. 5 does not include information such as the finding that the examining doctor considered it to be "fibrocystic breast disease" or that "Congo-red staining" was performed to rule out the possibility of "amyloidosis". When generating extracted summary information for document information in the medical field, numerical values, technical terms, English words, symbols used for emphasis, etc. tend to be highly evaluated in terms of importance. Therefore, as a result of preferentially extracting those with high importance, some of the important information in the document information 100 may be omitted as in the example shown in FIG. 5.

[0061] If the similarity between the extracted summary information 111 and the abstract summary information 112 is equal to or greater than a predetermined threshold, the summary generation server 10 adopts the abstract summary information 112, and if the similarity is lower than the threshold, it adopts the extracted summary information 111. In the example shown in FIG. 5, since both the extracted summary information 111 and the abstract summary information 112 describe the breast and the like, the similarity is equal to or greater than the threshold, and the abstract summary information 112 is adopted.

[0062] The summary generation server 10 stores the document information 100 and the abstract summary information 112 adopted as the summary information of the document information 100 in the storage unit 40 in association with the examination number.

[0063] After that, when the patient's attending physician operates the browsing terminal 61 as shown in FIG. 5, enters the examination number, and requests to view the examination results, the summary generation server 10 identifies the document information 100 and the abstract summary information 112 based on the examination number. The summary generation server 10 outputs the identified document information 100 and abstract summary information 112 and displays them on the screen of the browsing terminal 61. The attending physician can easily confirm the examination results by the examining physician from the abstract summary information 112 displayed on the browsing terminal 61. In the example shown in FIG. 5, the attending physician can grasp the content described in the document information 100 exceeding 120 characters by the abstract summary information 112 of about 70 characters. Since the document information 100 is also displayed on the browsing terminal 61, the attending physician can also confirm the document information 100 as needed.

[0064] The summary generation server 10 generates the simplified summary information 120 from the abstract summary information 112 adopted for the summary information of the document information 100. The simplified summary generation unit 26 of the summary generation server 10 uses the dictionary data 44 in the storage unit 40 to replace the medical field-specific terms with easy-to-understand phrases and sentences for the patient.

[0065] The simplified summary generation unit 26 can also change the expressions used by medical personnel to general expressions, or omit the technical terms and sentences that are not important for the patient. The change of expression can be performed by registering the information on the expression to be changed and the expression after the change in the dictionary data 44, similar to the replacement of technical terms. Similarly, the omission of technical terms and sentences can be performed by registering the technical terms to be omitted in the dictionary data 44 and omitting the technical terms and the sentences containing the technical terms. The simplified summary generation unit 26 may also perform expression change, omission of technical terms and sentences in a manner of executing machine learning.

[0066] In the example of the simplified summary information 120 shown in FIG. 5, information indicating that "mastopathy" is a "benign tumor" is added. Also, the technical terms "Congo-red staining" and "amyloidosis" are replaced with "another test" and "another disease" respectively, and the original technical terms are shown in parentheses. The examiner's finding of "considered to be the image (of mastopathy)" is changed in the simplified summary information 120 to the expression "seems to be the image (of mastopathy)". The text regarding the histological findings by the examiner is omitted and not included in the simplified summary information 120.

[0067] The summary generation server 10 stores the simplified summary information 120 generated from the abstract summary information 112 in the storage unit 40 in association with the examination number. As a result, the document information 100, the abstract summary information 112, and the simplified summary information 120 are stored in the storage unit 40 in association with the examination number.

[0068] As shown in FIG. 5, the patient can use a smartphone, a tablet terminal, etc. as the browsing terminal 62 to view the examination results. When the patient operates the browsing terminal 62 to input the examination number and request to view the examination results, the summary generation server 10 identifies the simplified summary information 120 based on the examination number and displays it on the screen of the browsing terminal 62.

[0069] For example, the summary generation server 10 distinguishes between the browsing terminal 61 and the browsing terminal 62 based on the address information of each terminal connected to the network 2. The summary generation server 10 displays the document information 100 and the abstract summary information 112 on the browsing terminal 61 used by the doctor in charge, and displays the simplified summary information 120 on the browsing terminal 62 used by the patient. However, the method of distinguishing the terminals is not particularly limited. For example, the summary generation server 10 may execute an authentication process to distinguish between the browsing terminal 61 used by the doctor in charge and the browsing terminal 62 used by the patient.

[0070] The patient can check the test results by the simplified summary information 120 in the test results 162 displayed on the screen of the browsing terminal 62. In the example shown in FIG. 5, it can be easily understood that for a patient who has undergone a breast cancer test, the test result shows that it is not breast cancer but a benign tumor called mastopathy, and another test is performed to rule out the possibility of another disease called amyloidosis.

[0071] In this embodiment, an example has been described in which the extraction-type summary generation unit 22 and the abstract-type summary generation unit 23 re-execute machine learning in response to the evaluation of the summary information by the instructor who instructed the generation of the summary information. However, a mode in which only the abstract-type summary generation unit 23 re-executes machine learning may also be possible. For example, when the extraction-type summary generation unit 22 generates summary information without using a machine learning model, the extraction-type summary generation unit 22 does not re-execute machine learning. For example, in a case where, in principle, the abstract-type summary information is output and the extraction-type summary information is output only when the similarity with the extraction-type summary information is low, a mode in which only the abstract-type summary generation unit 23 re-executes machine learning may also be possible.

[0072] In this embodiment, an example has been described in which summary information is generated from the test results by the examining doctor. However, the summary generation system can also be used to generate summary information from other document information, such as generating summary information from the image diagnosis results by the radiologist or generating summary information from the electronic medical record.

[0073] In FIG. 5 of this embodiment, an example has been described in which the simplified summary generation unit 26 of the summary generation server 10 executes a plurality of processes such as replacement of technical terms, change of expressions, omission of technical terms and sentences, etc. However, a mode in which only a part of these plurality of processes is executed may also be possible. For example, a mode in which the simplified summary generation unit 26 executes only the replacement of technical terms based on the dictionary data 44 may also be possible.

[0074] Also, in FIG. 5, the first method of adding an explanation of a technical term in parentheses after the technical term and the second method of replacing the technical term with another phrase or sentence and then adding the technical term in parentheses are shown. However, the simplified summary generation unit 26 of the summary generation server 10 may be configured to execute only one of these methods. The simplified summary generation unit 26 may process the technical terms in a third method in which the technical terms are replaced as in the second method, but the technical terms are deleted without being described in parentheses. Among the first to third methods, the method of processing each technical term can be set in the dictionary data 44 for each technical term, and the simplified summary generation unit 26 may process the technical terms based on the settings of each technical term.

[0075] In FIGS. 2 to 5 of the present embodiment, an example of receiving input of document information from the instruction terminal 50 and generating summary information from the document information has been described. However, the device that is the input source of the document information is not particularly limited. For example, the document information may be stored in an external device different from the instruction terminal 50. In this case, the instruction reception unit 21 of the summary generation server 10 may receive information indicating the storage location of the document information, such as an external device, from the instruction terminal 50, and acquire the document information from the storage location via the network 2. The summary generation server 10 can output abstract summary information or extraction summary information by executing each of the above-described processes on the document information acquired from an external device or the like. Further, the summary generation server 10 can generate simplified summary information from the output abstract summary information or extraction summary information.

[0076] In FIGS. 2 to 5 of the present embodiment, an example in which the summary generation server 10 generates summary information for one document information has been described. However, as described with reference to FIG. 1, the summary generation server 10 can also generate summary information for a plurality of document information. In this case, the instruction reception unit 21 of the summary generation server 10 receives a designation of one or more storage locations where the document information is stored and one or more document information stored in each storage location. The designation of the storage location and the document information may be performed manually one by one, or may be performed automatically using data in which a plurality of storage locations and document information are registered in advance. The storage location may be the instruction terminal 50 or an external device different from the instruction terminal 50. The instruction reception unit 21 acquires the designated one or more document information from each designated storage location via the network 2. The summary generation server 10 executes each of the above-described processes for the designated plurality of document information, generates one abstract type summary information and one extraction type summary information obtained by summarizing all the document information, and outputs the abstract type summary information or the extraction type summary information based on the similarity. Further, the summary generation server 10 can generate simplified type summary information from the output abstract type summary information or extraction type summary information.

[0077] In the example of a medical-related facility, the instruction reception unit 21 receives a designation of a plurality of document information such as past medical records and inspection records, extracts the character part, that is, text data, from each document information, and stores it in the document information data 41 of the storage unit 40. From the text data stored in the document information data 41, the extraction type summary generation unit 22 generates extraction type summary information, and the abstract type summary generation unit 23 generates abstract type summary information, whereby the summary generation server 10 can execute each process such as the output of the above-described summary information.

[0078] For example, assume that an external device has prepared a database in which document information such as medical records and examination records of a plurality of patients is stored in association with the identification numbers of each patient. When a doctor in charge operates the instruction terminal 50 to input the identification number of a patient and instructs the generation of summary information, the instruction reception unit 21 of the summary generation server 10 that has received the instruction specifies all the document information stored in the database in association with the identification number input by the doctor in charge. The instruction reception unit 21 extracts text data from all the specified document information and stores it in the document information data 41 of the storage unit 40. From the text data stored in the document information data 41, the extraction-type summary generation unit 22 generates extraction-type summary information, and the abstract-type summary generation unit 23 generates abstract-type summary information. Based on the similarity between the abstract-type summary information and the extraction-type summary information calculated by the similarity calculation unit 24, the summary output unit 25 outputs the abstract-type summary information or the extraction-type summary information and displays it on the display unit of the instruction terminal 50. Thereby, the doctor in charge can easily confirm the information included in a plurality of document information such as the patient's past medical records and examination records from the displayed summary information.

[0079] In this embodiment, a configuration example of a summary generation system including a summary generation server 10, a terminal 50 that processes document information input to the summary generation server 10, and terminals 61 and 62 that view the summary information generated by the summary generation server 10 from the document information is shown. This configuration example is a schematic of functions, and the configuration of the summary generation system is not physically limited to this configuration. For example, in the configuration example shown in FIG. 2, the instruction terminal 50 may be an aspect that realizes some or all of the functions and operations of the above-described summary generation server 10. The form of distribution and integration of each device is not limited to the above-described example, and all or part of it can be configured by functionally or physically distributing and integrating it in any unit according to various loads and usage situations.

Industrial Applicability

[0080] As described above, the summary generation server, summary generation system, summary generation program, and summary generation method according to the present disclosure are useful for generating summary information that can easily confirm the content included in the document information from the document information.

Explanation of Symbols

[0081] 10 Summary Generation Server 20 Control Unit 21 Instruction Reception Unit 22 Extractive Summary Generation Unit 23 Abstract Summary Generation Unit 24 Similarity Calculation Unit 25 Summary Output Unit 26 Simplified Summary Generation Unit 27 Evaluation Reception Unit 30 Communication Unit 40 Memory Unit 50 Instruction Terminal 61, 62 Browsing Terminal

Claims

1. A summary generation server that generates summary information from document information to be summarized, an extraction-type summary generation unit that generates extraction-type summary information by extracting sentences with high importance from the document information, an abstract-type summary generation unit that generates abstract-type summary information by summarizing the content described in the document information using a natural language processing model, a similarity calculation unit that calculates the similarity between the extraction-type summary information and the abstract-type summary information, and a summary output unit that outputs the extraction-type summary information or the abstract-type summary information based on the similarity A summary generation server characterized by comprising the above.

2. The summary output unit outputs the abstract-type summary information when the similarity is equal to or greater than a predetermined threshold, and outputs the extraction-type summary information when the similarity is lower than the threshold. The summary generation server according to claim 1, characterized by the above.

3. a dictionary in which terms in the field handled by the document information and information related to the terms are registered, and a simplified-type summary generation unit that generates simplified-type summary information by changing terms in the field included in the summary information output by the summary output unit based on the information registered in the dictionary The summary generation server according to claim 1, further characterized by comprising the above.

4. The information registered in the dictionary corresponding to the terms in the field includes at least one of a term obtained by paraphrasing the term and a sentence explaining the term. The summary generation server according to claim 3, characterized by the above.

5. An evaluation reception unit that receives an evaluation of the abstract-type summary information further comprising, The abstract-type summary generation unit performs machine learning using the document information and the abstract-type summary information generated from the document information as teacher data. The summary generation server according to claim 1, characterized by the above.

6. A summary generation server according to any one of claims 1 to 5, and a terminal device that transmits the document information to the summary generation server and receives summary information generated by the summary generation server from the document information A summary generation system characterized by comprising the above.

7. A summary generation program that generates summary information from document information to be summarized, an extraction-type summary information generation procedure that generates extraction-type summary information by extracting sentences with high importance from the document information, an abstract-type summary information generation procedure that generates abstract-type summary information by summarizing the content described in the document information using a natural language processing model, a similarity calculation procedure that calculates the similarity between the extraction-type summary information and the abstract-type summary information, An output procedure for outputting the extracted summary information or the abstract summary information based on the similarity, and a summary generation program characterized by causing a computer to execute the same. **Claim 8**: A summary generation method for generating summary information from document information to be summarized, which is executed by a computer device, comprising: an extracted summary information generation step of generating extracted summary information by extracting sentences with high importance from the document information; an abstract summary information generation step of generating abstract summary information by summarizing the description content of the document information using a natural language processing model; a similarity calculation step of calculating the similarity between the extracted summary information and the abstract summary information; an output step of outputting the extracted summary information or the abstract summary information based on the similarity; and a summary generation method characterized by including the above steps.

Images