Summary generating apparatus, summary generating method, and recording medium

Abstract

To efficiently generate a summary in consideration of a sentence to be emphasized or a sentence to be ignored.SOLUTION: To generate a summary sentence by using an original sentence being a sentence to be summarized and a reference sentence indicating a sentence to be ignored. The original sentence is divided into each sentence and inputted to a score calculation part to calculate an importance score. For example, the importance scores of sentence 1, sentence 2, and sentence 3 obtained by dividing the original sentence are 9.351, 11.852, and 4.713, respectively. The divided sentence and the reference sentence are combined and inputted to a score calculation part, and a similarity score showing mutual similarity is calculated. Then, the importance score is corrected using the calculated similarity score. Specifically, since the reference sentence is a sentence to be ignored, 10 is subtracted from the importance score of the sentence having the maximum score in the similarity score. For example, since the maximum score in the similarity score is sentence 2, the importance score of sentence 2 is corrected to 1.852. A summary sentence is generated from the corrected importance score.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] The present invention relates to a summary generation program, a summary generation device, and a summary generation method that can efficiently generate summaries that take into account sentences that should be emphasized or ignored. [Background technology]

[0002] Conventional techniques that use the TF-IDF (Term Frequency - Inverse Document Frequency) algorithm, which uses the product of the frequency of occurrence of a word in a document and the inverse document frequency, or a text ranking algorithm or a deep learning model to extract important sentences or phrases and present a summary are known (see, for example, Patent Document 1).Technology that uses a deep learning model to understand information from the original text and generate a summary sentence is also known. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Application Publication No. 2024-098244 Summary of the Invention [Problem to be solved by the invention]

[0004] However, because the above-mentioned conventional technology is based on general text interpretation, information deemed irrelevant to the context is omitted from the summary, which can result in the omission of important information. Specifically, depending on the purpose of the summary, there are cases where it is necessary to include sentences or phrases that are not very relevant to the context and tend to be omitted in normal processing, and cases where it is necessary to omit information that is important from a contextual perspective, such as information that should not be disclosed to patients when creating a patient-oriented medical record summary.

[0005] Therefore, how to efficiently generate summaries that take into account sentences that should be emphasized or ignored is an important issue. However, using a fine-tuned pre-trained model is not efficient.

[0006] The present invention has been made to solve the problems (issues) of the above-mentioned conventional technology, and aims to provide a summary generation program, a summary generation device, and a summary generation method that can efficiently generate summaries that take into account sentences that you want to emphasize or sentences that you want to ignore. [Means for solving the problem]

[0007] In order to solve the above-mentioned problems, the present invention provides a summary generation program for generating a summary sentence from a summary target sentence that is the subject of summarization, the program causing a computer to execute the following steps: a similarity calculation step for calculating the similarity between a reference sentence to be included in the summary sentence and a plurality of sentences that form the summary target sentence; an importance score calculation step for calculating importance scores of the plurality of sentences that form the summary target sentence using a first large-scale language model; a correction step for correcting the importance scores of the plurality of sentences that form the summary target sentence calculated in the importance score calculation step, based on the similarities calculated for the plurality of sentences that form the summary target sentence; and a summary sentence generation step for generating the summary sentence based on the importance scores corrected in the correction step, using a second large-scale language model.

[0008] Furthermore, in the above invention, the present invention is characterized in that, when the similarity calculated for the plurality of sentences forming the sentence to be summarized is large, the correction procedure corrects the importance scores of the plurality of sentences forming the sentence to be summarized calculated in the importance score calculation procedure so that they are equal to or greater than a predetermined threshold value.

[0009] The present invention also provides a summary generation program for generating a summary sentence from a summary target sentence that is the subject of summarization, the program causing a computer to execute the following steps: a similarity calculation step for calculating the similarity between a reference sentence that is not included in the summary sentence and a plurality of sentences that form the summary target sentence; an importance score calculation step for calculating importance scores of the plurality of sentences that form the summary target sentence using a first large-scale language model; a correction step for correcting the importance scores of the plurality of sentences that form the summary target sentence calculated in the importance score calculation step, based on the similarities calculated for the plurality of sentences that form the summary target sentence; and a summary sentence generation step for generating the summary sentence based on the importance scores corrected in the correction step, using a second large-scale language model.

[0010] In addition, in the above invention, the present invention is characterized in that, when the similarity calculated for the multiple sentences that form the sentence to be summarized is large, the correction procedure corrects the importance scores of the multiple sentences that form the sentence to be summarized, calculated in the importance score calculation procedure, so that they are less than a predetermined threshold.

[0011] Furthermore, in the above invention, the present invention is characterized in that the summary generation procedure uses the second large-scale language model to generate a summary including sentences whose importance scores corrected in the correction procedure are greater than a predetermined threshold.

[0012] Furthermore, in the above invention, the present invention is characterized in that the similarity calculation procedure includes a reading procedure of reading from a storage unit a first vector obtained by vectorizing the reference sentence; a vectorization procedure of vectorizing each of the multiple sentences included in the text to be summarized; and a calculation procedure of inputting the multiple vectors vectorized by the vectorization procedure and the first vector into the first large-scale language model and calculating the similarity of the multiple sentences included in the text to be summarized that are output from the first large-scale language model.

[0013] Furthermore, in the above invention, the present invention is characterized in that the vectorization step uses an encoder of the first large-scale language model to vectorize each of a plurality of sentences included in the text to be summarized.

[0014] In the present invention, in the above-mentioned invention, the first large-scale language model and the second large-scale language model use the same trained model.

[0015] The present invention also provides a summary generation device that generates a summary sentence from a summary target sentence that is the subject of summarization, comprising: a similarity calculation unit that calculates a similarity between a reference sentence to be included in the summary sentence and a plurality of sentences that form the summary target sentence; an importance score calculation unit that calculates an importance score for the plurality of sentences that form the summary target sentence using a first large-scale language model; a correction unit that corrects the importance scores of the plurality of sentences that form the summary target sentence calculated by the importance score calculation unit based on the similarities calculated for the plurality of sentences that form the summary target sentence; and a summary sentence generation unit that generates the summary sentence based on the importance scores corrected by the correction unit using a second large-scale language model.

[0016] The present invention also provides a summary generation device that generates a summary sentence from a summary target sentence that is the subject of summarization, and is characterized by comprising: a similarity calculation unit that calculates a similarity between a reference sentence that is not to be included in the summary sentence and a plurality of sentences that form the summary target sentence; an importance score calculation unit that calculates an importance score for the plurality of sentences that form the summary target sentence using a first large-scale language model; a correction unit that corrects the importance scores of the plurality of sentences that form the summary target sentence calculated by the importance score calculation unit based on the similarities calculated for the plurality of sentences that form the summary target sentence; and a summary sentence generation unit that generates the summary sentence based on the importance scores corrected by the correction unit using a second large-scale language model.

[0017] The present invention also provides a summary generation method for a summary generation device that generates a summary sentence from a summary target sentence to be summarized, comprising the steps of: a similarity calculation step of calculating the similarity between a reference sentence to be included in the summary sentence and a plurality of sentences that form the summary target sentence; an importance score calculation step of calculating importance scores for the plurality of sentences that form the summary target sentence using a first large-scale language model; a correction step of correcting the importance scores of the plurality of sentences that form the summary target sentence calculated in the importance score calculation step, based on the similarities calculated for the plurality of sentences that form the summary target sentence; and a summary sentence generation step of generating the summary sentence based on the importance scores corrected in the correction step, using a second large-scale language model.

[0018] The present invention also provides a summary generation method for a summary generation device that generates a summary sentence from a summary target sentence to be summarized, comprising: a similarity calculation step of calculating the similarity between a reference sentence not to be included in the summary sentence and a plurality of sentences that form the summary target sentence; an importance score calculation step of calculating importance scores for the plurality of sentences that form the summary target sentence using a first large-scale language model; a correction step of correcting the importance scores of the plurality of sentences that form the summary target sentence calculated in the importance score calculation step, based on the similarities calculated for the plurality of sentences that form the summary target sentence; and a summary sentence generation step of generating the summary sentence based on the importance scores corrected in the correction step, using a second large-scale language model. [Effects of the Invention]

[0019] According to the present invention, it is possible to efficiently generate a summary that takes into account sentences that one wants to emphasize or ignore. [Brief explanation of the drawings]

[0020] [Figure 1] FIG. 1 is an explanatory diagram of an overview of the summary generation program. [Figure 2]FIG. 2 is a functional block diagram showing the configuration of the summary generation device. [Figure 3] FIG. 3 is a diagram showing an example of the summary setting data, sentence data, and divided sentence data shown in FIG. [Figure 4] FIG. 4 is a diagram showing an example of the importance score data, the similarity score data, the correction value data, and the correction score data shown in FIG. [Figure 5] FIG. 5 is a diagram illustrating an overview of Sentence-BERT. [Figure 6] FIG. 6 is a diagram showing an example (part 1) of a text summary that ignores specific content. [Figure 7] FIG. 7 is a diagram showing an example (part 2) of a text summary that ignores specific content. [Figure 8] FIG. 8 is a diagram showing an example (part 3) of a text summary that ignores specific content. [Figure 9] FIG. 9 is a diagram showing an example (part 4) of a text summary that ignores specific content. [Figure 10] FIG. 10 is a diagram showing an example (part 5) of a text summary that ignores specific content. [Figure 11] FIG. 11 is a diagram showing an example (part 1) of a text summary that emphasizes specific content. [Figure 12] FIG. 12 is a diagram showing an example (part 2) of a text summary that emphasizes specific content. [Figure 13] FIG. 13 is a diagram showing an example (part 3) of a text summary that emphasizes specific content. [Figure 14] FIG. 14 is a diagram showing an example (part 4) of a text summary that emphasizes specific content. [Figure 15] FIG. 15 is a diagram showing an example (part 5) of a text summary that emphasizes specific content. [Figure 16] FIG. 16 is a flowchart showing a processing procedure for generating a summary. [Figure 17] FIG. 17 is a diagram illustrating an example of a hardware configuration of a summary generating device. DETAILED DESCRIPTION OF THE INVENTION

[0021] Hereinafter, embodiments of a summary generating program, a summary generating device, and a summary generating method will be described in detail with reference to the accompanying drawings.

[0022] <Summary Generation Program Overview> First, we will explain the outline of this summary generation program. Figure 1 is an explanatory diagram of this summary generation program. This summary generation program uses a trained model that performs natural language processing to calculate sentence similarities and semantic relationships as importance scores, and uses these importance scores to perform extraction-type summaries of text.

[0023] As shown in Figure 1, our summary generation program generates a summary using the original text, which is the text to be summarized, and reference sentences that indicate sentences to ignore. These ignored sentences are sentences that contain expressions that should not be included in the summary.

[0024] The original text is divided into sentences and input to the score calculation unit, and the importance score is calculated. For example, the importance scores of sentences 1, 2, and 3 obtained by dividing the original text are 9.351, 11.852, and 4.713, respectively.

[0025] Each sentence (hereinafter referred to as "divided sentence") obtained by dividing the original sentence is combined with the reference sentence and input to the score calculation unit, and a similarity score indicating the mutual similarity is calculated. For example, the similarity scores of sentence 1, sentence 2, and sentence 3 with respect to the reference sentence are 0.304, 0.673, and 0.317, respectively.

[0026] The importance score is corrected using the calculated similarity score. Specifically, since the reference sentence is a sentence to be ignored, 10 is subtracted from the importance score of the sentence with the highest similarity score. For example, since sentence 2 has the highest similarity score, the importance score of sentence 2 is corrected to 1.852.

[0027] A summary is generated from this corrected importance score. For example, if the top two sentences based on the corrected importance score are used as the summary, the summary generated will be "Using a gastrointestinal endoscopy... No treatment is required, but..."

[0028] In this way, this summary generation program is configured to input the sentences to be summarized and reference sentences to be ignored, etc. into a trained model that performs natural language processing to calculate an importance score and a similarity score, and then generate a summary sentence using the importance score corrected by this similarity score, making it possible to efficiently generate summaries that take into account sentences to be emphasized or ignored.

[0029] <Configuration of summary generation device 10> Next, the configuration of the summary generation device 10 related to the present summary generation program will be described. Fig. 2 is a functional block diagram showing the configuration of the summary generation device 10 related to the present summary generation program. As shown in Fig. 2, the summary generation device 10 has a display unit 11, an input unit 12, a storage unit 15, and a control unit 16.

[0030] The display unit 11 is a display device such as a liquid crystal panel, a display device, etc. The input unit 12 is an input device such as a keyboard, a mouse, etc.

[0031] The memory unit 15 is a storage device such as a hard disk drive or non-volatile memory, and stores summary setting data 15a, sentence data 15b, divided sentence data 15c, importance score data 15d, similarity score data 15e, correction value data 15f, and correction score data 15g.

[0032] The summary setting data 15a indicates the number of sentences in the summary and whether reference sentences are to be emphasized or ignored. The sentence data 15b indicates the original sentence and the reference sentence. The divided sentence data 15c indicates the divided sentences.

[0033] The importance score data 15d is data indicating the importance score of a divided sentence. The similarity score data 15e is data indicating the similarity score of a divided sentence. The correction value data 15f is data indicating correction values ​​when the reference sentence is emphasized or ignored. The corrected score data 15g is data indicating the corrected importance score of a divided sentence.

[0034] The control unit 16 is a control unit that performs overall control of the summary generation device 10, and has a summary setting unit 16a, a sentence management unit 16b, a sentence segmentation unit 16c, a score calculation unit 16d, a score correction unit 16e, and a summary generation unit 16f. In practice, by loading these programs into a CPU (Central Processing Unit) and executing them, the summary setting unit 16a, the sentence management unit 16b, the sentence segmentation unit 16c, the score calculation unit 16d, the score correction unit 16e, and the summary generation unit 16f will execute their respective corresponding processes.

[0035] The summary setting unit 16a is a processing unit that manages the summary setting data 15a. When the summary setting unit 16a receives settings for the number of sentences in the summary and whether to emphasize or ignore reference sentences from the input unit 12, the summary setting unit 16a stores the received information in the summary setting data 15a.

[0036] The text management unit 16b is a processing unit that manages the text data 15b. When the text management unit 16b receives the original text and the reference text from the input unit 12, the text management unit 16b stores them in the text data 15b.

[0037] The sentence division unit 16c is a processing unit that manages the divided sentence data 15c. When the sentence management unit 16b updates the sentence data 15b, the sentence division unit 16c divides the original text of the sentence data 15b into sentences, associates the sentences with sentence numbers, and stores them in the divided sentence data 15c.

[0038] The score calculation unit 16d is a processing unit that calculates an importance score and a similarity score using a trained model that performs natural language processing. The score calculation unit 16d calculates an importance score for each segmented sentence by inputting the segmented sentences of the segmented sentence data 15c into Sentence-BERT (hereinafter referred to as "SBERT"), which is a trained model that performs natural language processing. The score calculation unit 16d then stores the calculated importance score in importance score data 15d. SBERT will be described later.

[0039] Furthermore, the score calculation unit 16d calculates a similarity score for each segmented sentence by inputting the segmented sentences of the segmented sentence data 15c and the reference sentences of the text data 15b into SBERT, and stores the calculated similarity score in similarity score data 15e.

[0040] The score correcting unit 16e is a processing unit that corrects the importance score. When the importance score data 15d is updated by the score calculating unit 16d, the score correcting unit 16e stores the importance score of the similarity score data 15e as a corrected importance score in the corrected score data 15g.

[0041] Furthermore, when the score calculation unit 16d updates the similarity score data 15e, the score correction unit 16e identifies the sentence number with the maximum similarity score in the similarity score data 15e and extracts a correction value corresponding to the importance / ignorance setting in the summary setting data 15a from the corrected score data 15g.The score correction unit 16e then adds the extracted correction value to the corrected importance score in the corrected score data 15g corresponding to the identified sentence number to update it.After that, it passes a summary instruction to the summary generation unit 16f.Note that when identifying sentence numbers using similarity scores, it is also possible to identify all sentence numbers whose similarity scores are equal to or greater than a predetermined value (e.g., 0.3).

[0042] The summary generation unit 16f is a processing unit that generates a summary sentence. Upon receiving a summary instruction from the score correction unit 16e, the summary generation unit 16f identifies sentence numbers of the divided sentences indicated by the summary sentence count in the summary setting data 15a from the data with high corrected importance scores in the correction score data 15g. The summary generation unit 16f then extracts divided sentences corresponding to the identified sentence numbers from the divided sentence data 15c and generates a summary sentence by connecting the extracted divided sentences in the order of their sentence numbers. The generated summary sentence is then displayed on the display unit 11.

[0043] Next, an example of data stored in the storage unit 15 of the summary generation device 10 shown in Fig. 2 will be described. Fig. 3 and Fig. 4 are diagrams showing examples of the summary setting data 15a, sentence data 15b, divided sentence data 15c, importance score data 15d, similarity score data 15e, correction value data 15f, and correction score data 15g shown in Fig. 2.

[0044] The summary setting data 15a shown in Fig. 3(a) indicates a state in which the number of summary sentences is "2" and the importance / ignorance setting is "ignorance." The text data 15b shown in Fig. 3(b) indicates a state in which the original text is "Due to a gastrointestinal endoscopy... there is a suspicion of early-stage cancer... No treatment is necessary..." and the reference text is "Cancer is suspected."

[0045] The divided sentence data 15c shown in Figure 3(c) shows a state in which the divided sentence with sentence number "1" is "By gastrointestinal endoscopy...", the divided sentence with sentence number "2" is "Due to suspicion of early stage cancer...", and the divided sentence with sentence number "3" is "No treatment is necessary, but...".

[0046] The importance score data 15d shown in Figure 4(a) indicates that the importance score of sentence number "1" is "9.351", the importance score of sentence number "2" is "11.852", and the importance score of sentence number "3" is "4.713".

[0047] The similarity score data 15e shown in Fig. 4(b) indicates a state where the similarity score for sentence number "1" is "0.304", the similarity score for sentence number "2" is "0.673", and the similarity score for sentence number "3" is "0.317".

[0048] The correction value data 15f shown in Fig. 4(c) indicates a state where the weighted correction value is "3.00" and the ignored correction value is "-10.00".

[0049] The corrected score data 15g shown in Fig. 4(d) indicates a state where the corrected importance score for sentence number "1" is "9.351", the corrected importance score for sentence number "2" is "1.852", and the corrected importance score for sentence number "3" is "4.713".

[0050] <Overview of Sentence - BERT> Next, an overview of Sentence - BERT, which is used as a pre - trained model for natural language processing, will be described. Fig. 5 is a diagram showing the overview of Sentence - BERT.

[0051] SBERT is a pre - trained model that utilizes the BERT (Bidirectional Encoder Representations from Transformers) architecture to accurately calculate the similarity between sentences. Using the similarity calculated by SBERT, extractive summarization of articles can be performed.

[0052] As shown in Fig. 5, in SBERT, multiple sentences (for example, SentenceA and SentenceB) are input. The input sentences are each input into BERT. In BERT, the input sentence is split into chunks such as words called tokens, and a result is output for each token.

[0053] The output of each token from BERT is pooled and output as a single vector. For example, Sentence A and Sentence B are output as vectors u and v. The similarity between Sentence A and Sentence B can be calculated by calculating the cosine similarity between vectors u and v.

[0054] <An example of a text summary that ignores specific content> Next, an example of a text summary that ignores specific content will be described. Figures 6 to 10 are diagrams showing an example of a text summary that ignores specific content.

[0055] As shown in Figure 6(a), if the original sentence is "Mild obesity is observed. Please continue treatment and follow-up.", the original sentence is divided into sentences and importance scores are calculated using SBERT.

[0056] Specifically, as shown in Figure 7, the original text is divided into divided sentences (sentence number 1 to sentence number 14), and importance scores are assigned to each. For example, the divided sentence with sentence number 1 is "Mild obesity was confirmed," and the corresponding importance score is "1.643." The divided sentence with sentence number 2 is "Blood tests also confirmed diabetes," and the corresponding importance score is "2.429."

[0057] As shown in Figure 6(b), if the sentence "There is a suspicion of cancer" is given as a reference sentence to be ignored, the similarity score is calculated by inputting this reference sentence and the divided sentences of the original sentence into SBERT.

[0058] Specifically, the calculated similarity score is associated with the sentence number as shown in Fig. 8. For example, the similarity score corresponding to sentence number 2 is "0.429", and the similarity score corresponding to sentence number 5 is "0.436".

[0059] The sentence number with the highest score among the calculated similarity scores is identified, and the importance score corresponding to this sentence number is corrected. For example, in the similarity scores shown in Figure 8, sentence number 12 has the highest similarity score of "0.671", so the importance score of this sentence number 12, "11.857", is corrected.

[0060] In this case, the setting is to ignore the reference sentence, so the correction value is "-10.00." The importance score corrected using this correction value is shown in Figure 9. Specifically, the importance score of sentence number 12 before correction is "11.857," while the importance score after correction is "1.857."

[0061] A predetermined number of sentence numbers are identified in order of the highest corrected importance score, and the identified divided sentences are connected in order of sentence numbers to generate a summary. For example, if the number of sentences to summarize is 5, the corrected importance scores shown in Figure 9 identify divided sentences with sentence numbers 9, 10, 11, 13, and 14. These five divided sentences are then connected to generate the summary shown in Figure 10(a).

[0062] On the other hand, when processing is performed without ignoring reference sentences, the uncorrected importance score is used to identify the segmented sentences from which the summary is generated. That is, the segmented sentences with sentence numbers 10, 11, 12, 13, and 14 are identified, and these five segmented sentences are joined to generate the summary shown in Figure 10(b). The major difference between the summary shown in Figure 10(a) and the summary shown in Figure 10(b) is whether or not the underlined sentence in Figure 10(b) is included. This sentence is highly similar to the reference sentence "There is a suspicion of cancer," which is to be ignored, and it can be seen that this sentence is not included in the summary shown in Figure 10(a).

[0063] <An example of a text summary that emphasizes specific content> Next, an example of a text summary that emphasizes specific content will be described. Figures 11 to 15 are diagrams showing an example of a text summary that emphasizes specific content.

[0064] As shown in Figure 11(a), if the original sentence is "Eight prostate biopsy specimens have been submitted. PSA follow-up is considered advisable," the original sentence is divided into sentences and an importance score is calculated using SBERT.

[0065] Specifically, as shown in Figure 12, the original sentence is divided into divided sentences (sentence number 1 to sentence number 6), and importance scores are associated with them. For example, the divided sentence with sentence number 1 is "Eight prostate biopsy specimens have been submitted," and the corresponding importance score is "2.0." Also, the divided sentence with sentence number 2 is "All specimens are prostate tissue," and the corresponding importance score is "1.5."

[0066] As shown in Figure 11(b), if the sentence "The specimen was submitted" is given as the reference sentence to be emphasized, the similarity score is calculated by inputting this reference sentence and the divided sentences of the original sentence into SBERT.

[0067] Specifically, the calculated similarity score is associated with the sentence number as shown in Fig. 13. For example, the similarity score corresponding to sentence number 1 is "0.717", and the similarity score corresponding to sentence number 2 is "0.542".

[0068] Here, all sentence numbers whose calculated similarity scores are equal to or greater than a predetermined value (e.g., 0.3) are identified, and the importance scores corresponding to these sentence numbers are corrected. For example, in the similarity scores shown in Figure 13, the divided sentences with similarity scores of 0.3 or greater are sentence numbers 1 and 2, so the importance scores of each are corrected.

[0069] Since the reference sentence is set to be important, the correction value is "3.00." The importance scores corrected using this correction value are shown in Figure 14. Specifically, the importance score of sentence number 1 before correction is "2.0," but the importance score after correction is "5.0." The importance score of sentence number 2 before correction is "1.5," but the importance score after correction is "4.5."

[0070] A predetermined number of sentence numbers are identified in order of the highest corrected importance score, and the identified divided sentences are connected in order of sentence numbers to generate a summary. For example, if the number of sentences to be summarized is three, the divided sentences with sentence numbers 1, 2, and 3 are identified based on the corrected importance score shown in Figure 14. These three divided sentences are then connected to generate the summary shown in Figure 15(a). Note that if the number of sentences to be summarized is two, the divided sentences with sentence numbers 1 and 3 are identified, and these two divided sentences are then connected to generate the summary shown in Figure 15(b).

[0071] On the other hand, when processing is performed without placing importance on the reference sentence, the uncorrected importance score is used to identify the segmented sentences from which the summary is generated. That is, the segmented sentences with sentence numbers 3 and 4 are identified, and these two segmented sentences are joined to generate the summary shown in Figure 15(c). The major difference between the summary shown in Figure 15(b) and the summary shown in Figure 15(c) is whether or not the underlined sentence in Figure 15(b) is included. This sentence is highly similar to the reference sentence "The specimen has been submitted," which is placed under emphasis, and it can be seen that this sentence is not included in the summary shown in Figure 15(c).

[0072] <Processing procedure for generating summary sentences> Next, the processing procedure for generating a summary will be described with reference to the flowchart of FIG.

[0073] As shown in FIG. 16, when the summary generation device 10 receives the original sentence, the reference sentence, and the summary setting (step S101; Yes), it divides the original sentence into sentences (step S102).

[0074] The divided sentences are input to SBERT to calculate importance scores (step S103), and the divided sentences and the reference sentence are input to SBERT to calculate similarity scores (step S104).

[0075] The divided sentence having the maximum score or a score equal to or greater than a predetermined value among the calculated similarity scores is identified, and the importance score of this identified divided sentence is corrected (step S105).

[0076] A predetermined number of divided sentences with high corrected importance scores are extracted and a summary is generated (step S106), and the process ends.

[0077] <Example of hardware configuration of summary generation device 10> Next, the correspondence between the main hardware configuration of the summary generation device 10 and the computer will be described. Fig. 17 is a diagram showing an example of the hardware configuration of the summary generation device 10.

[0078] Generally, a computer has a configuration in which a CPU 81, a ROM 82, a RAM 83, a non-volatile memory 84, etc. are connected via a bus 85. A hard disk drive may be provided instead of the non-volatile memory 84. For the sake of convenience of explanation, only the basic hardware configuration is shown.

[0079] Here, the ROM 82 or non-volatile memory 84 stores programs required to start the operating system (hereinafter simply referred to as "OS"), and the CPU 81 reads and executes the OS program from the ROM 82 or non-volatile memory 84 when the power is turned on.

[0080] On the other hand, various application programs executed on the OS are stored in non-volatile memory 84, and the CPU 81 executes the application programs while using RAM 83 as the main memory, thereby executing processes corresponding to the applications.

[0081] The summary generation program of the summary generation device 10 is also stored in non-volatile memory 84 or the like, like other application programs, and the CPU 81 loads and executes this program. In the case of the summary generation device 10, a summary generation program including routines corresponding to the summary setting unit 16a, text management unit 16b, text segmentation unit 16c, score calculation unit 16d, score correction unit 16e, and summary generation unit 16f shown in Fig. 2 is stored in non-volatile memory 84 or the like. When the summary generation program is loaded and executed by the CPU 81, summary generation processes corresponding to the summary setting unit 16a, text management unit 16b, text segmentation unit 16c, score calculation unit 16d, score correction unit 16e, and summary generation unit 16f are generated.

[0082] As described above, this summary generation program is configured to input the sentences to be summarized and reference sentences to be ignored, etc. into a trained model that performs natural language processing to calculate an importance score and a similarity score, and then generate a summary sentence using the importance score corrected by this similarity score.This makes it possible to efficiently generate summaries that take into account sentences to be emphasized or ignored.

[0083] In the above embodiment, a configuration in which the original sentence and the reference sentence are accepted in text form has been described, but the present invention is not limited to this, and the original sentence and the reference sentence can also be accepted in audio form.

[0084] Furthermore, the components illustrated in the above embodiments are merely functional schematics and are not necessarily physically configured as shown. In other words, the distribution and integration of each device is not limited to that illustrated, and all or part of the devices can be functionally or physically distributed and integrated in any unit depending on various loads, usage conditions, etc. [Industrial Applicability]

[0085] The summary generating program, summary generating device, and summary generating method according to the present invention are suitable for efficiently generating summaries that take into account sentences that one wants to emphasize or sentences that one wants to ignore. [Explanation of symbols]

[0086] 10 Summary generator 11 Display section 12 Input section 15 Storage section 15a Summary Configuration Data 15b Text data 15c Split sentence data 15d Importance Score Data 15e Similarity score data 15f Correction value data 15g Corrected Score Data 16 Control Unit 16a Summary Setting Section 16b Text Management Department 16c Sentence division part 16d Score calculation section 16e Score Correction Section 16f Summary generator 81 CPU 82 ROM 83 RAM 84 Non-volatile memory 85 Bus

Claims

1. A summary generation program that generates a summary sentence from a summary target sentence that is a target of summarization, a similarity calculation step of calculating a similarity between a reference sentence to be included in the summary sentence and a plurality of sentences forming the summary target sentence; an importance score calculation step of calculating importance scores of a plurality of sentences forming the text to be summarized using a first large-scale language model; a correction step of correcting the importance scores of the plurality of sentences forming the text to be summarized calculated in the importance score calculation step based on the similarities calculated for the plurality of sentences forming the text to be summarized; a summary generation step of generating the summary sentence based on the importance score corrected in the correction step, using a second large-scale language model; A summary generation program that causes a computer to execute the above.

2. The correction procedure includes: The summary generation program described in claim 1, characterized in that if the similarity calculated for the multiple sentences that make up the summary target sentence is large, the importance scores of the multiple sentences that make up the summary target sentence calculated in the importance score calculation procedure are corrected so that they are above a predetermined threshold.

3. A summary generation program that generates a summary sentence from a summary target sentence that is a target of summarization, a similarity calculation step of calculating a similarity between a reference sentence not included in the summary sentence and a plurality of sentences forming the summary target sentence; an importance score calculation step of calculating importance scores of a plurality of sentences forming the text to be summarized using a first large-scale language model; a correction step of correcting the importance scores of the plurality of sentences forming the text to be summarized calculated in the importance score calculation step based on the similarities calculated for the plurality of sentences forming the text to be summarized; a summary generation step of generating the summary sentence based on the importance score corrected in the correction step, using a second large-scale language model; A summary generation program that causes a computer to execute the above.

4. The correction procedure includes: The summary generation program described in claim 3, characterized in that if the similarity calculated for the multiple sentences that make up the summary target sentence is large, the importance scores of the multiple sentences that make up the summary target sentence calculated in the importance score calculation procedure are corrected so that they are less than a predetermined threshold.

5. The procedure for generating a summary is as follows:

4. The summary generation program according to claim 1, further comprising: generating a summary including sentences whose importance scores corrected by the correction procedure are greater than a predetermined threshold value, using the second large-scale language model.

6. The similarity calculation procedure includes: a reading step of reading a first vector obtained by vectorizing the reference sentence from a storage unit; a vectorization step of vectorizing each of a plurality of sentences included in the text to be summarized; a calculation step of inputting the plurality of vectors vectorized by the vectorization step and the first vector into the first large-scale language model, and calculating similarities between a plurality of sentences included in the text to be summarized that are output from the first large-scale language model; 4. The summary generating program according to claim 1, further comprising:

7. The vectorization procedure includes:

7. The program for generating a summary according to claim 6, wherein an encoder of the first large-scale language model is used to vectorize each of a plurality of sentences included in the text to be summarized.

8. 4. The summary generation program according to claim 1, wherein the first large-scale language model and the second large-scale language model use the same trained model.

9. A summary generation device that generates a summary sentence from a summary target sentence that is a target of summarization, a similarity calculation unit that calculates a similarity between a reference sentence to be included in the summary sentence and a plurality of sentences that form the summary target sentence; an importance score calculation unit that calculates importance scores of a plurality of sentences that form the text to be summarized using a first large-scale language model; a correction unit that corrects the importance scores of the plurality of sentences that form the summary target sentence calculated by the importance score calculation unit based on the similarities calculated for the plurality of sentences that form the summary target sentence; a summary generation unit that generates the summary based on the importance score corrected by the correction unit using a second large-scale language model; A summary generation device comprising:

10. A summary generation device that generates a summary sentence from a summary target sentence that is a target of summarization, a similarity calculation unit that calculates a similarity between a reference sentence not included in the summary sentence and a plurality of sentences that form the summary target sentence; an importance score calculation unit that calculates importance scores of a plurality of sentences that form the text to be summarized using a first large-scale language model; a correction unit that corrects the importance scores of the plurality of sentences that form the summary target sentence calculated by the importance score calculation unit based on the similarities calculated for the plurality of sentences that form the summary target sentence; a summary generation unit that generates the summary based on the importance score corrected by the correction unit using a second large-scale language model; A summary generation device comprising:

11. A summary generation method for a summary generation device that generates a summary sentence from a summary target sentence, the method comprising: a similarity calculation step of calculating a similarity between a reference sentence to be included in the summary sentence and a plurality of sentences forming the summary target sentence; an importance score calculation step of calculating importance scores of a plurality of sentences forming the text to be summarized using a first large-scale language model; a correction step of correcting the importance scores of the plurality of sentences forming the text to be summarized calculated in the importance score calculation step based on the similarities calculated for the plurality of sentences forming the text to be summarized; a summary sentence generation step of generating the summary sentence based on the importance score corrected in the correction step by using a second large-scale language model; 10. A method for generating a summary, comprising:

12. A summary generation method for a summary generation device that generates a summary sentence from a summary target sentence, the method comprising: a similarity calculation step of calculating a similarity between a reference sentence not included in the summary sentence and a plurality of sentences forming the summary target sentence; an importance score calculation step of calculating importance scores of a plurality of sentences forming the text to be summarized using a first large-scale language model; a correction step of correcting the importance scores of the plurality of sentences forming the text to be summarized calculated in the importance score calculation step based on the similarities calculated for the plurality of sentences forming the text to be summarized; a summary sentence generation step of generating the summary sentence based on the importance score corrected in the correction step by using a second large-scale language model; 10. A method for generating a summary, comprising:

Images