JP7309489B2 - Abstract creation method and abstract creation system - Google Patents

Description

The present invention relates to a summary creation method and a summary creation system.

A railway operator receives requests, opinions, complaints, etc. regarding railway operation from customers, documents them, and accumulates them. Based on this, railway operators are improving their operations. However, since the amount of documents that railway operators receive from customers is enormous, it is necessary to organize and summarize these documents so that they can be used effectively.

In this regard, Non-Patent Document 1 describes a method for creating a summary from a given document by machine learning. It is disclosed that a sentence to be adopted as a summary sentence is specified by extracting the sentence as . Further, regarding the setting of importance, Non-Patent Document 2 discloses a calculation method using a trained model of deep learning. Specifically, a CNN (Convolutional Neural Network) is used to classify sentences and create a summary. That is, the system sequentially executes the construction of a distributed representation of a sentence based on the distributed representation of words, the construction of a distributed representation of the entire sentence based on the distributed representation of the sentence, and the classification of the sentence based on the distributed representation of the entire sentence. Perform the learning calculation (machine learning) necessary for After that, the system inputs the sentences to be summarized to the trained CNN, finds the contribution of the distributed representation of the sentences to the classification result by back-calculating from the output of the CNN, and regards the calculated contribution as the importance.

https://dl.acm.org/citation.cfm?id=215333 (Accessed April 16, 2019) https://pdfs.semanticscholar.org/c52a/cb4e4143ace520166691a29faaeaf892ac47.pdf (Accessed April 16, 2019)

However, since railway companies carry out various tasks such as management and operation of rolling stock, management of timetables, management of passengers, and management and operation of various facilities, the contents of the documents are also diverse. Therefore, it was not easy to accurately grasp and summarize the main points of these sentences.

SUMMARY OF THE INVENTION The present invention has been made in view of such a situation, and an object of the present invention is to provide a summary creation method and a summary creation system capable of creating a summary that accurately grasps the main points of a document related to a railway business. is to provide

To achieve the above objects, according to one aspect of the present invention, there is provided a sentence summarizing system having a processor and a memory, comprising subject information in the railway business from a given sentence having a plurality of sentences about the railway business. annotating processing for specifying a sentence, a sentence having location information in the railway business, a sentence having time information in the railway business, and a sentence having information about the action or motion in the railway business; At least one or more sentences indicating each information of the subject information, the location information, the time information, and the action or action information are extracted from the sentence, and the sentence including the extracted sentence is executed as a summary sentence corresponding to the predetermined sentence.

According to the present invention, it is possible to create a summary that accurately grasps the gist of the text regarding the railway business.

It is a figure explaining an example of a structure and a function of a summary sentence creation system. It is a figure which shows an example of an annotation dictionary. It is a figure which shows an example of an annotation rule. It is a figure which shows an example of the hardware constitutions which an abstract creation apparatus has. FIG. 10 is a flow diagram illustrating an example of sentence group summarization processing; FIG. 10 is a flow chart for explaining the details of sentence analysis processing; FIG. 11 is a flow diagram illustrating details of importance score prediction processing; FIG. 10 is a flow diagram illustrating an example of annotation addition processing; FIG. 10 is a flow diagram illustrating an example of a summary creation process; It is a figure which shows an example of a result display screen.

The abstract writing system according to this embodiment will be described below.
FIG. 1 is a diagram for explaining an example of the configuration and functions of a summary sentence creation system according to this embodiment. The summary sentence creation system 1 includes a sentence storage device 20 storing one or more sentences 3 (sentence group) having a plurality of sentences (sentences) related to railway business, and a sentence storage device 20 via a communication network 5. and a summary writing device 10 communicably connected to the device. The communication network 5 is, for example, a wired or wireless communication network such as a LAN (Local Area Network), a WAN (Wide Area Network), the Internet, or a dedicated line.

As shown in FIG. 1, the summary sentence creation device 10 includes a sentence analysis unit 110 that analyzes the structure of each sentence in the sentence 3, an importance prediction unit 130 that predicts the importance of each sentence in the sentence 3, Functional units of an annotation adding unit 150 that adds predetermined annotation information to a sentence when a predetermined condition is satisfied, and a summary sentence creation unit 170 that creates a summary sentence corresponding to the sentence 3 based on the importance and annotation. have

The text analysis unit 110 includes a sentence division unit 111 , a morphological analysis unit 113 and a dependency analysis unit 115 . The sentence division unit 111 divides sentence 3 into one or more sentences. The morphological analysis unit 113 divides each sentence divided by the sentence division unit 111 into a plurality of morphemes. The dependency analysis unit 115 reconstructs each morpheme divided by the morpheme analysis unit 113 into clauses, and specifies the modification/modified relationship between the clauses. The text analysis unit 110 stores this information in the text analysis result 500. FIG.

The importance prediction unit 130 includes an importance prediction value vector calculation unit 131 , a sentence importance prediction unit 133 , and a trained model creation unit 135 .

The importance prediction value vector calculation unit 131 calculates a feature vector (feature amount) of each sentence based on the structure of each sentence of the sentence 3 specified by the sentence analysis unit 110 .

The sentence importance prediction unit 133 inputs the feature vector of each sentence calculated by the importance prediction value vector calculation unit 131 to the trained model 200 (learner data) described later, thereby summarizing the sentence 3. Calculate the importance of each sentence in That is, the sentence importance prediction unit 133 predicts the importance of each sentence in the sentence 3 by inputting the sentence 3 to the created trained model 200 .

The learned model creation unit 135 predicts the importance of each sentence in the sentence 3 using the structure of the sentence 3 analyzed by the sentence analysis unit 110 and the abstract sentence data 400 in which predetermined abstract sentences are registered. Machine learning is performed on the trained model 200 for. That is, the learned model creation unit 135 learns the relationship between the structure of the sentence about the railway business and the predetermined summary sentence associated with the sentence, and uses the sentence about the railway business as an input value, A trained model 200 is created that predicts the importance of the sentence as a summary.

If each sentence in sentence 3 has a phrase indicating “when”, “where”, “who”, and “what happened” (hereinafter referred to as “four elements”) related to the railway business, the annotation adding unit 150 is provided with predetermined meta information (annotation). That is, the annotation adding unit 150 extracts information on subjects (“who”) in the railway business (hereinafter simply referred to as “subject information” or “subject elements”) from a predetermined sentence 3 having a plurality of sentences related to the railway business. and sentences with location (“where”) information in railway business (hereinafter simply referred to as “location information” or “location elements”) and time (“when”) information in railway business ( hereinafter simply referred to as "time information" or "time element") and information on actions or actions ("what happened") in the railway business (hereinafter simply referred to as "action or action information" or "action or Each sentence that has an element of action") is identified. The annotation adding unit 150 stores these results in the annotation adding result 151 .

Note that the correspondence between the information of the four elements and the corresponding sentences (phrases) is registered in the annotation database 300 . The annotation database 300 has an annotation dictionary 310 recording a list of information on the four elements, and annotation rules 330 . The annotation rule 330 determines under what circumstances a sentence has the information of each of the four elements (for example, what kind of word does the character string have in the sentence? indicates whether it is judged to have the information of the element if it has

(annotation dictionary)
FIG. 2 is a diagram showing an example of the annotation dictionary 310. As shown in FIG. The annotation dictionary 310 includes an element name 311 of one of the four elements, a sub-element 313 of the element related to the element name 311, a sub-element 313 and a word (character string) associated with the element specified by the sub-element 313 and dictionary data 315 .

In this embodiment, a plurality of words are registered for each of the element of time (“when”), the element of subject (“who”), and the element of action or action (“what happened”). . In particular, the place element ("where") has sub-elements "train", "station", and "place" (other places), and for each sub-element, multiple words is registered.

(Annotation rule)
FIG. 3 is a diagram showing an example of annotation rules 330. As shown in FIG. The annotation rule 330 is information including a rule number 331 assigned to each rule and a rule content 333 that is the content of the rule number 331 . In the example shown in the figure, the annotation rule with the rule number "1" includes the sub-element "train" in the location element "where". This is a rule for adding annotation information "where_place" to the clause. The annotation rule 330 registers rules for at least four elements including the element of location.

Next, as shown in FIG. 1, the summary sentence creation unit 170 creates a summary sentence for sentence 3 based on the annotation information and the degree of importance. That is, the summary sentence creation unit 170 extracts sentences that clearly indicate subject information, location information, time information, and action or movement information from the sentences specified by the annotation adding unit 150. At least one piece of information is extracted, and a sentence containing the extracted sentence is created as a summary sentence corresponding to the sentence 3. At this time, based on the degree of importance predicted by the sentence importance degree prediction unit 133, the summary sentence creation unit 170 extracts sentences that well indicate each piece of information.

Note that, if only less than a predetermined number of sentences (sentence extraction threshold) can be extracted, the summary sentence creation unit 170 can extract sentences that show each piece of information well. A predetermined number of sentences that satisfy the sentence with a higher value) are additionally extracted, and a summary sentence corresponding to the sentence 3 is created including the additionally extracted sentences.

In addition, when the annotation adding unit 150 extracts a sentence that clearly indicates location information, the summary sentence creation unit 170 extracts a sentence that clearly indicates location information. , the sentences are preferentially extracted in this order. In the following, information on these trains, stations, and other locations will be referred to as sub-elements.

The functions of the abstract creation device 10 described above are realized as follows.
FIG. 4 is a diagram showing an example of the hardware configuration of the abstract creation device 10. As shown in FIG. The abstract creation device 10 includes a processing unit 51 such as a CPU (Central Processing Unit), a main storage device 52 such as a RAM (Random Access Memory) and a ROM (Read Only Memory), a HDD (Hard Disk Drive), an SSD ( Auxiliary storage device 53 such as Solid State Drive), a keyboard,
It comprises an input device 54 such as a mouse or a touch panel, an output device 55 such as a monitor (display) for screen display, and a communication device 56 for communicating with other devices. The functions of the abstract creation device 10 are realized by the hardware of the abstract creation device 10 or by the processing unit 51 reading and executing each program stored in the main storage device 52 or the auxiliary storage device 53. be. In addition, these programs are stored in a storage device such as a secondary storage device, a nonvolatile semiconductor memory, a hard disk drive, or an SSD, or a non-readable storage device such as an IC card, SD card, or DVD, which can be read by the abstract creation apparatus 10, for example. Stored on a temporary data storage medium. Note that the above also applies to the text accumulation device 20 .

<<Processing>>
Next, processing performed by the abstract writing system 1 will be described. The summary creation system 1 executes a sentence group summary process for creating a summary of the sentence 3 .

<Sentence Group Summary Processing>
FIG. 5 is a flow chart for explaining an example of sentence group summarization processing. This process is started, for example, when a predetermined input for receiving the sentence 3 is made by the user to the abstract creation device 10 . When the summary sentence creation device 10 receives the sentence 3 from the sentence storage device 20, the process of s11-s21 is executed for each sentence in the sentence 3. FIG.

That is, the summary sentence creation device 10 executes sentence analysis processing for analyzing each sentence of the sentence 3 (s13). Then, the summary sentence creation device 10 executes an importance score prediction process for calculating the importance of each sentence of the sentence 3 analyzed in s13 (s15). Then, the summary sentence creation device 10 executes an annotation adding process of adding annotation information to each sentence of sentence 3 analyzed in s13 (s17). Then, based on the importance calculated in s15 and the annotation information added in s17, the summary creation device 10 executes a summary creation process for creating a summary of sentence 3 (s19). If the abstract creation device 10 receives a plurality of sentences 3, the processes of s13 to s19 are repeated by that number (s21).
Next, details of each process in the sentence group summarization process will be described. In the following description, the case where the abstract creation device 10 receives one sentence 3 will be described.

<Sentence analysis processing>
FIG. 6 is a flowchart for explaining the details of text analysis processing. First, the sentence dividing unit 111 of the summary sentence creation device 10 creates a plurality of sentences (character strings) by dividing the sentence 3 before and after the period (s131). Then, the summary sentence creation device 10 repeats the following processes of s135-s139 for each sentence (s133-s141).

That is, first, the morphological analysis unit 113 decomposes the character string of the target sentence into a plurality of morphemes (s135). Then, the dependency analysis unit 115 reconstructs each decomposed morpheme into a plurality of clauses, and specifies modification and modified relationships between the clauses (s137).

The summary sentence creation device 10 stores the character string of the sentence, the information on the morpheme specified in s135, and the information on the modification and modified relationship between clauses specified in s137 in the sentence analysis result 500 (s139). .

<Importance score prediction processing>
FIG. 7 is a flowchart for explaining the details of importance score prediction processing. First, the sentence importance prediction unit 133 of the abstract creation device 10 reads out the learned model 200 (s151).

Also, the sentence importance prediction unit 133 calculates the feature vector of each sentence of the sentence 3 based on the content of the sentence analysis result 500 created by the sentence analysis process (s153). Specifically, for example, the sentence importance predicting unit 133 identifies a feature amount that represents the feature of each sentence based on the morphemes, clauses, and modification and modified relationships between clauses of each sentence. Calculate the feature vector.

The sentence importance prediction unit 133 inputs the feature vector of each sentence calculated in s153 to the trained model 200 read out in s151, thereby calculating the importance of all the sentences of sentence 3, and calculating the importance of each sentence. An importance prediction value vector obtained by converting the degree into a vector format is output (s155). Then, the sentence importance prediction unit 133 stores this importance prediction value vector (s155).

<Annotation process>
FIG. 8 is a flowchart illustrating an example of annotation adding processing. First, the annotation adding unit 150 of the abstract creation device 10 reads the annotation database 300 (annotation dictionary 310 and annotation rules 330) (s171). Also, the annotation adding unit 150 reads the sentence analysis result 500 created in the sentence analysis process (s173).

Then, as described below, the annotation adding unit 150 adds annotations related to the four elements to each sentence of the sentence 3 registered in the sentence analysis result 500 (s175-s187). It should be noted that a sentence may be given multiple annotations, and a sentence may not be given any annotation at all.

First, the annotation adding unit 150 selects one sentence (hereinafter referred to as a selected sentence) in sentence 3 (s175). Also, the annotation adding unit 150 selects one of the annotation rules for the four elements and sub-elements defined by the annotation rule 330 (hereinafter referred to as a selection rule) (s177).

The annotation adding unit 150 checks the selected sentence against the selection rule to determine whether the selected sentence satisfies the selection rule (s179). Specifically, for example, the annotation adding unit 150 extracts the rule content 333 of the record related to the selection rule from the annotation rule 330 and extracts the dictionary data 315 of the record related to the selection rule from the annotation dictionary 310 . Then, the annotation adding unit 150 creates a specific rule by applying each word in the dictionary data 315 to the description part of the element in the rule content 333, and compares the specific rule with the selected sentence. , it is determined whether or not there is a word (or clause) that matches this specific rule in the selected sentence.

Here, a specific example of determination processing will be described.
(for location information ("where"))
For example, if the selection rule is an annotation rule for an element of location information, it is as follows. The selected sentence is "By Shinkansen, and the guidance at the time of delay was early." When certain record data is included, annotation information of "place" and "vehicle" is added to the selection sentence.

Also, for example, the selection sentence is "Inquiry about an accident that occurred at Shinjuku Station." If the record data "at the station" is included, annotation information of "location" and "station" is added to the selection sentence.

Also, for example, the selected sentence is "I was asked in English how to get on the subway at the south exit of Kanda Station." If the dictionary data 315 includes record data of "at the south exit of Kanda Station", annotation information of "place" and "other place" is added to the selected sentence.

(In the case of subject information (“who”))
If the selection rule is an annotation rule for a location subject element: For example, the selected sentence is "The green attendant responded politely." is included, the annotation information of "subject" is added to the selected sentence.

(In the case of time information ("when"))
When the selection rule is an annotation rule related to the element of time, it is as follows. For example, the selected sentence is "I'm worried about the delay in the announcement about the rush hour." When certain record data is included, the annotation information of "time" is added to the selection sentence.

(In the case of movement or behavioral information ("what's wrong"))
If the selection rule is an annotation rule for an action or action element, then: First, record data in which the selected sentence is "Because of an accident, I was late for work." is included, annotation information of "movement or action" is attached to the selected sentence.

When the annotation adding unit 150 determines that the selected sentence satisfies the selection rule as described above (s181: YES), the annotation adding unit 150 executes the process of s183 to be described later, and the selected sentence When it is determined that the selection rule is not satisfied (s181: NO), the annotation adding unit 150 executes the process of s185, which will be described later.

In s<b>183 , the annotation adding unit 150 stores the judgment result in the annotation adding result 151 . After that, the annotation adding unit 150 executes the process of s185.

In s185, if there is another annotation rule for which annotation addition has not been determined, the annotation adding unit 150 again executes the processes from s179 onward using that annotation rule as a new selection rule.

If there is no other annotation rule for which annotation addition has not been determined, and if there is another sentence for which annotation addition has not been determined, the annotation attachment unit 150 treats that sentence as a new selected sentence, and after s177 is executed again (s187). On the other hand, if there is no other sentence for which annotation addition has not been determined, the annotation addition unit 150 terminates the annotation addition processing (s189).

<Summary sentence creation process>
FIG. 9 is a flowchart for explaining an example of a summary sentence creation process. First, the abstract creating unit 170 of the abstract creating apparatus 10 reads the importance (predicted importance vector) of all the sentences of the sentence 3 created by the importance score prediction process (s191). Also, the abstract creation unit 170 reads the annotation information (annotation addition result 151) created in the annotation addition process (s193).

Based on the annotation information read in S193, the summary sentence creation unit 170 identifies all sentences associated with time information (“when”), and assigns a predetermined extraction priority to each identified sentence (when the sentence is A parameter representing the degree to which each element is indicated) is set (s195).

Specifically, for example, when the time information in the identified sentence is specific time or time information, the summary sentence creating unit 170 sets the extraction priority of the sentence to the maximum. Further, when the time information in the specified sentence is the time and the time information other than the time, the summary sentence creation unit 170 gives the sentence a higher extraction priority as the time indicated by the information is closer to the present time. set.

Based on the annotation information read in s193, the summary sentence creation unit 170 identifies all sentences associated with the location information (“where”), and sets an extraction priority for each identified sentence (s197). .

Specifically, for example, if the location information in the specified sentence is information about a train, station, or other location, the summary sentence creating unit 170 gives the sentence the highest extraction priority in this order. set.

The summary sentence creation unit 170 identifies all sentences associated with subject information (“who”) from the annotation information read in s193, and sets extraction priority to the identified sentences (s199).

Specifically, for example, the summary sentence creation unit 170 sets a higher extraction priority to a sentence closest to the beginning of the original sentence (sentence 3).

Based on the annotation information read in step s193, the summary sentence creation unit 170 identifies all sentences associated with the action or motion information (“What happened?”), and sets an extraction priority for each identified sentence. (s201).

Specifically, for example, the summary sentence creation unit 170 sets a higher extraction priority to a sentence closest to the beginning of the original sentence (sentence 3).

The summary sentence creation unit 170 identifies sentences with the highest extraction priority for each of the four elements determined in s195-s201 (s203). Note that when the highest extraction priority cannot be specified for each element (such as when there are multiple sentences with the same extraction priority), the summary sentence creation unit 170 places the element closest to the beginning of the original sentence (sentence 3). Maximize the extraction priority of a sentence.

Then, the summary sentence creation unit 170 sets a predetermined flag (extraction target flag) for each specified sentence, and stores the number of sentences for which this flag is set (the number of flagged sentences).

Next, based on the number of sentences in sentence 3, the summary sentence creating unit 170 sets a threshold (sentence extraction threshold) for the number of sentences in the abstract sentence (s205). For example, when the number of sentences in sentence 3 is n, the summary sentence creating unit 170 sets the maximum integer less than log ₂ (n+1) as the sentence extraction threshold.

The summary sentence creation unit 170 determines whether or not the number of flagged sentences set in s203 is greater than or equal to the sentence extraction threshold set in s205 (s207). If the number of flagged sentences is greater than or equal to the sentence extraction threshold (s207: YES), the summary sentence creation unit 170 executes the process of s209, and if the number of flagged sentences is less than or equal to the sentence extraction threshold (s207: NO). , the abstract creation unit 170 executes the process of s211.

In s209, the summary sentence creation unit 170 creates a summary sentence by combining sentences for which extraction target flags have been set in s203. Specifically, for example, the summary sentence creation unit 170 creates a summary sentence by combining each sentence in the order of appearance in the original sentence 3 .

In s211, the summary sentence creating unit 170 sets extraction target flags for the number of flagged sentences divided by the sentence extraction threshold value from the number of flagged sentences among the sentences for which the extraction target flag is not set. For example, the summary sentence creation unit 170 sets extraction target flags for sentences with relatively high importance calculated in the importance prediction process (sets extraction target flags in descending order of importance). After that, the process of s209 is performed.

Finally, the trained model creation unit 135 updates the learned model 200 by learning the relationship between the original sentence 3 and the summary sentence (summary sentence data 400) created in s209. Specifically, for example, the trained model creation unit 135 obtains the feature amount of each sentence in the original sentence 3 to obtain the probability that each sentence is included in the summary sentence, and calculates the calculated probability as the importance of each sentence. degrees. The learned model creation unit 135 performs machine learning by inputting the obtained importance and the summary sentence created in s209 to the learned model 200. FIG.

<Result display screen>
FIG. 10 is a diagram showing an example of a screen (result display screen) displayed by the abstract creating apparatus 10 after the abstract creating process. The result display screen 600 comprises an original sentence display column 601 displaying the sentence 3 that is the source of the abstract, and an abstract sentence display column 603 displaying the abstract. In the summary text of the summary text display field 603, portions corresponding to the time element 611, the location element 613, the subject element 615, and the action or motion element 617 are highlighted (for example, colored characters or underlined text). ) is provided.

As described above, the summary sentence creation system 1 of the present embodiment can, from the sentence 3 having a plurality of sentences related to the railway business, a sentence having information about the subject of the railway business, a sentence having information about the place in the railway business, After identifying sentences with information on time in railway business and sentences with information on actions or actions in railway business, sentences that clearly indicate subject, place, time, and information on actions or actions are identified. At least one piece of information is extracted, and a sentence 3 including the extracted sentence is created as a summary sentence corresponding to the sentence 3 . As a result, it is possible to create a summary that accurately describes important situations in the railway business. As a result, the abstract creating system 1 of the present embodiment can create an abstract that accurately grasps the gist of the sentence regarding the railway business.

In addition, the summary sentence creation system 1 of the present embodiment learns the relationship between the structure of the sentence about the railway business and the predetermined summary sentence associated with the sentence, and inputs the sentence about the railway business. As a value, a learned model creation process for creating a trained model that predicts the degree of importance indicating the importance of the sentence as a summary; and executing an importance prediction process for predicting the importance of each sentence in the sentences, and extracting sentences that well indicate each of the information based on the predicted importance in the summary sentence creation process.

In this way, by inputting the sentence 3 to the trained model 200, the degree of importance of each sentence in the sentence 3 (importance as a summary) is predicted, and based on this, a summary sentence is created to explain the situation. Able to create well-executed summaries.

Further, in the abstract creation system 1 of the present embodiment, when less than a predetermined number of sentences that well indicate each piece of information can be extracted in the abstract creation process, the degree of importance satisfies a predetermined condition. A predetermined number of additional sentences are extracted, and a summary sentence corresponding to the predetermined sentence is created including the additionally extracted sentences.

In this way, when less than a predetermined number of sentences that show the information of each element well can be extracted, a predetermined number of additional sentences whose importance satisfies a predetermined condition are additionally extracted to restore the original sentence. Regardless of the configuration of 3, it is possible to reliably create a summary including a situation explanation.

Further, in the abstract creation process, the abstract creation system 1 of the present embodiment extracts a sentence that well indicates the location information. In the case of location information, the sentences are preferentially extracted in this order.

By giving such priority to the location information in the railway business, it is possible to create a summary that accurately describes the location information.

Further, the abstract creation system 1 of this embodiment outputs the created abstract.

As a result, railway operators can check this summary and use it for business improvement.

Although the embodiments of the present invention have been described above, the embodiments of the present invention are not limited to those illustrated, and various modifications can be made without departing from the gist of the invention.

For example, in the present embodiment, a method using the annotation database 300 describing annotation rules has been described as a method of adding annotation information, but a method based on machine learning may be employed. For example, a learning model is created by converting past examples into numerical vectors, and new input sentences are converted into numerical vectors and applied to the learning model. Give.

Also, in the present embodiment, sub-elements are provided for the location element, but sub-elements may be provided for other elements.

Further, in the present embodiment, when the number of set extraction target flags is equal to or greater than the sentence number extraction threshold in the summary sentence creation process, the extraction target flags are set to new sentences in descending order of importance. , the extraction target flag may be set to the new sentence according to other conditions.

1 summary writing system, 3 sentences, 10 summary writing device, 20 sentence storage device

Claims (10)

A summary writing system having a processor and memory,
From a predetermined sentence having a plurality of sentences related to the railway business, a sentence having subject information in the railway business, a sentence having location information in the railway business, a sentence having time information in the railway business, and actions in the railway business or an annotation process that identifies a sentence having information on the action, respectively;
At least one or more sentences indicating each information of the subject information, the location information, the time information, and the action or action information are extracted from each of the specified sentences and extracted. a summary sentence creation process for creating a sentence containing the sentence as a summary sentence corresponding to the predetermined sentence;
How to write a summary. The summary sentence creation system
By learning the relationship between the structure of a sentence about the railway business and a predetermined summary sentence associated with the sentence, the sentence about the railway business is used as an input value to indicate the importance of the sentence as a summary. A learned model creation process for creating a learned model that predicts importance;
executing an importance prediction process for predicting the importance of each sentence in the predetermined sentence by inputting the predetermined sentence into the created trained model;
extracting a sentence indicating each of the information based on the predicted importance in the summary sentence creation process;
The method of creating a summary sentence according to claim 1. The summary sentence creation system includes:
In the summary sentence creation process, if less than a predetermined number of sentences indicating each information can be extracted, a predetermined number of sentences whose importance satisfies a predetermined condition are additionally extracted, and additionally extracted. Create a summary sentence corresponding to the predetermined sentence, including the sentence that
3. The method of creating a summary sentence according to claim 2. The summary sentence creation system includes:
When extracting the sentence indicating the location information in the summary sentence creation process, if the location information indicated by the sentence is train, station, or other location information, the sentences are extracted in this order. preferentially extract
The method of creating a summary sentence according to claim 1. 2. The abstract creating method according to claim 1, wherein said abstract creating system outputs said created abstract. having a processor and memory,
From a predetermined sentence having a plurality of sentences related to the railway business, a sentence having subject information in the railway business, a sentence having location information in the railway business, a sentence having time information in the railway business, and actions in the railway business or an annotation process that identifies a sentence having information on the action, respectively;
At least one or more sentences indicating each information of the subject information, the location information, the time information, and the action or action information are extracted from each of the specified sentences and extracted. a summary sentence creation process for creating a sentence containing the sentence as a summary sentence corresponding to the predetermined sentence;
Abstract writing system. By learning the relationship between the structure of a sentence about the railway business and a predetermined summary sentence associated with the sentence, the sentence about the railway business is used as an input value to indicate the importance of the sentence as a summary. A learned model creation process for creating a learned model that predicts importance;
executing an importance prediction process for predicting the importance of each sentence in the predetermined sentence by inputting the predetermined sentence into the created trained model;
extracting a sentence indicating each of the information based on the predicted importance in the summary sentence creation process;
The summary writing system according to claim 6. In the summary sentence creation process, if less than a predetermined number of sentences indicating each information can be extracted, a predetermined number of sentences whose importance satisfies a predetermined condition are additionally extracted, and additionally extracted. Create a summary sentence corresponding to the predetermined sentence, including the sentence that
The summary writing system according to claim 7. When extracting the sentence indicating the location information in the summary sentence creation process, if the location information indicated by the sentence is train, station, or other location information, the sentences are extracted in this order. preferentially extract
The summary writing system according to claim 6. 7. The abstract creation system according to claim 6, which outputs the created abstract.

Images