JP6905237B2 - Mail thread extractor, mail thread extraction method, and computer program - Google Patents

Description

The present invention relates to a mail thread extraction device for collectively extracting mail exchanges (mail threads).

Conventionally, various techniques for managing e-mail for each common topic have been proposed. For example, Patent Document 1 (Patent No. 5199449) discloses an e-mail data classification device for classifying e-mail data for e-mail thread display and the like. In this mail data classification device, a word contained in an e-mail and a response phrase which is a reply word to the word are stored in association with each other, a keyword related to the word contained in the e-mail to be classified is searched, and the response phrase is searched. Extract emails that contain, group them together and display them.

Further, a technique for collectively extracting e-mail exchanges by using "title", "destination", etc., which are information in the header part of the e-mail, has been widely used in the past.

Patent No. 5199449

However, thread extraction using keywords has a problem that it is difficult to obtain an appropriate extraction result unless the keywords are set properly. In addition, thread extraction using "title" and "destination" has a problem that even if the contents of emails are connected, if the "title" and "destination" are different, the threads cannot be properly organized. .. Furthermore, recently, the use of artificial intelligence (AI) is becoming more realistic, and extracting mail threads according to the content of mail using a trained model using a neural network is also an application field of AI. Is assumed as.

The present invention provides a mail thread extraction device, a mail thread extraction method, a computer program, and the like that can appropriately extract mail exchanges as mail threads without omission by using a trained model using a neural network. With the goal.

In order to achieve the above object, the mail thread extraction device of the present invention
A mail data storage unit that stores multiple e-mail data, and
A starting point mail data acquisition unit that acquires starting point mail data that is the starting point of a mail thread from the mail data storage unit, and a starting point mail data acquisition unit.
A response candidate data generation unit that generates response candidate data corresponding to the origin mail data using a learned model that has learned the correspondence between the mail data and the corresponding response mail data.
A response mail data acquisition unit that acquires response mail data similar to the response candidate data generated by the response candidate data generation unit from the mail data storage unit, and a response mail data acquisition unit.
It includes a mail thread generation unit that generates a mail thread by combining the origin mail data and the response mail data.

According to the present invention, a mail thread extraction device, a mail thread extraction method, a computer program, and the like capable of appropriately extracting e-mail exchanges as mail threads without omission by using a trained model using a neural network are provided. Can be provided.

It is a block diagram which shows the schematic structure of the mail thread extraction apparatus which concerns on 1st Embodiment of this invention. It is a figure which shows the example of vectorization of a document. It is a figure which shows an example of the mail data stored in the mail data storage part. It is a figure which shows an example of the user interface which specifies the origin mail data and the search period in 1st Embodiment of this invention. It is a figure which shows an example of the mail thread which concerns on 1st Embodiment of this invention. It is a figure which shows an example of the exchange of e-mail. It is a figure which shows the example of the set of the mail data and the response mail data used as teacher data. It is a figure which shows an example of the neural network for learning the correspondence relation of e-mail. It is a flowchart which shows an example of the processing flow of mail thread extraction which concerns on 1st Embodiment of this invention. It is a block diagram which shows the schematic structure of the mail thread extraction apparatus which concerns on 2nd Embodiment of this invention. It is a flowchart which shows an example of the processing flow of mail thread extraction in 2nd Embodiment of this invention. It is a figure which shows an example of the user interface which specifies the origin mail data and the search period in the 2nd Embodiment of this invention. It is a figure which shows an example of the mail thread which concerns on 2nd Embodiment of this invention. It is a figure which shows an example of the neural network for learning the similarity relation of the response mail data which concerns on 3rd Embodiment of this invention.

Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. The same or corresponding parts in the drawings are designated by the same reference numerals, and the description thereof will not be repeated.

[First Embodiment]
FIG. 1 is a block diagram showing a schematic configuration of a mail thread extraction system 100 according to a first embodiment of the present invention. The mail thread extraction system 100 extracts mail data as a starting point and mail data corresponding to the starting mail data according to a designated search condition, and collectively generates a mail thread by collecting the extracted mail data. However, the mail thread extraction system 100 simply extracts the response mail data depending on whether or not the text data of the subject or body includes a predetermined keyword to generate a mail thread, as in the conventional mail thread extraction system. The response mail data is extracted and a mail thread is generated by using the trained model generated based on a large amount of training data.

As shown in FIG. 1, the mail thread extraction system 100 includes an extraction device 1 and a learning device 2. The extraction device 1 and the learning device 2 do not need to be connected at all times. For example, the learner 2 can be configured as a cloud system.

The extraction device 1 includes a mail data analysis unit 11, a mail data storage unit 12, a starting mail data acquisition unit 13, a response candidate data generation unit 14, a response mail data acquisition unit 15, and a mail thread generation unit 16.

The mail data analysis unit 11 morphologically analyzes the mail data of the e-mail and then vectorizes the mail data. The vectorization of mail data is generally performed by a dictionary composed of morphemes of document data and their IDs. FIG. 2 shows the dictionary conversion of document data. Document data is converted into a dictionary by morphologically analyzing the underlying mail data and giving them unique IDs. Each ID represents the dimension of the analytic space composed of a dictionary. The vectorization of mail data is performed by morphologically analyzing the mail data to be vectorized, collating it with the morphemes of the dictionary, and expressing the corresponding number for each ID in a one-dimensional array (vector). Although the result of vector representation changes depending on the underlying document data and the method of dictionary conversion, the embodiment of the present invention is not limited.

The mail data storage unit 12 stores all the mail data vectorized by the mail data analysis unit 11. FIG. 3 shows an example of mail data stored in the mail data storage unit 12. The mail data stored in the mail data storage unit 12 includes, for example, a mail document, vectorized data, and a date and time.

The starting point mail data acquisition unit 13 acquires the mail data that becomes the starting point of the thread search within the designated search period from the mail data storage unit 12. Here, the search period and the designation of the mail data that is the starting point of the thread search are specified by the user through the user interface as shown in FIG. 4, for example.

The response candidate data generation unit 14 generates response candidate data corresponding to the starting mail data. The response candidate data generation unit 14 holds the trained model 14a, and the response candidate data is generated by inputting the starting mail data into the trained model 14a. The trained model 14a is generated by the learner 2 shown in FIG. The generation of the trained model 14a in the learner 2 will be described in detail later.

The response mail data acquisition unit 15 acquires mail data similar to the response candidate data generated by the response candidate data generation unit 14 from the mail data storage unit 12 as the response mail data that responds to the origin mail data. If there are multiple response mail data, acquire multiple response mail data. Here, a determination of similarity between the response candidate data and the mail data stored in the mail data storage unit 12 is obtained as an inner product of the vector of the response candidate data and the vector of the mail data stored in the mail data storage unit 12. .. In the method of determining the similarity using this vector, the method of obtaining the similarity as the inner product of the vectors does not limit the embodiment of the present invention. For example, the similarity between the vectors is determined by using the method based on the Euclidean distance. May be sought.

The mail thread generation unit 16 generates a mail thread by combining the starting mail data and at least one response mail data acquired from the mail data storage unit 12. For example, a mail thread in which mail exchanges are summarized is generated in the form shown in FIG. In the example of the mail thread shown in FIG. 5, an example is shown in which a starting mail entitled "Confirmation of the last day of the year" and four reply mails responding to the starting mail are displayed as threads.

Next, the configuration and function of the learner 2 will be described. As shown in FIG. 1, the learner 2 inputs mail learning data (teacher data), passes through the learning data analysis unit 21 and the data set generation unit 22, and causes the DNN 23 to learn, thereby responding to the extraction device 1. A trained model to be used in the candidate data generation unit 14 is created. The learning device 2 includes a learning data analysis unit 21, a data set generation unit 22, a DNN (deep neural network) 23, and a trained model storage unit 24. Further, as the learning data to be input to the learning device 2, a set of mail data and response mail data for the mail data is used as teacher data. FIG. 6 shows an example of e-mail exchanged in the form of a question and an answer. The Questions and Answers in FIG. 6 are associated and extracted as a set, and these sets are used as teacher data. FIG. 7 shows an example of a set of mail data used as teacher data and response mail data.

The learning data analysis unit 21 performs morphological analysis of the learning data and then vectorizes the learning data in the same manner as the mail data analysis unit 11 provided in the extraction device 1 described above. The learning data analysis unit 21 vectorizes a set of mail data and the corresponding response mail data as learning data.

The data set generation unit 22 generates a vectorized version of the input document of the teacher data and a label of the response mail data as a data set to be input to the DNN 23.

The DNN 23 is a deep neural network that associates a vectorized input mail data with a label of the response mail data. An image of learning performed by DNN23 is shown in FIG. As shown in FIG. 8, the response mail data with the correct answer label corresponding to the morphologically analyzed Question mail data is learned as Answer. For example, the response mail data corresponding to Question 1 is associated with Answer 1 for learning. As the neural network model used in DNN23, a neural network model such as a convolutional neural network (CNN) or a recurrent neural network (RNN, LSTM) can be applied, but the type of neural network model does not limit the present invention. do not have.

Next, the operation of the mail thread extraction device 100 described above will be described with reference to a flowchart. FIG. 9 shows the operation flow of the mail thread extraction device 100. In S101, it is determined whether or not the starting mail is specified. For the designation of the starting mail, for example, the starting mail and the search period are specified by the user through the user interface as shown in FIG. The mail thread extraction device 100 acquires the starting mail data within the designated search period from the mail data storage unit 12 (S102). Next, the mail thread extraction device 100 inputs the acquired origin mail data into the trained model 14a of the response candidate data generation unit 14 and generates response candidate data (S103). Then, the response mail data acquisition unit 15 acquires the response mail data similar to the response candidate data from the mail data storage unit 12 (S104). A mail thread is generated by combining the starting mail data acquired from the mail data storage unit 12 and at least one response mail data (S105). For example, the response mail data for the starting mail is collectively output (displayed) as a mail thread in the form shown in FIG. In the example of FIG. 5, a plurality of reply mails to the mail sent at 12:00 on December 10, 2018 with the subject "Confirmation of the last day of the year" are grouped in the mail thread. In the example of FIG. 5, the email displayed at the top of the email thread (reply from account aaa) is selected, and the text "I am going to work as usual" in the body of the email is in the right column. It is displayed in. As described above, according to the present embodiment, at least one response mail can be collectively displayed in a thread for one starting mail.

[Second Embodiment]
As described above, one specific embodiment of the present invention has been described, but the above-described embodiment is an example and does not limit the present invention. For example, in the first embodiment described above, at least one response mail data corresponding to one origin mail is acquired to generate a mail thread, but the mail data corresponding to the response mail data is acquired. It may be in the form of Hereinafter, a configuration different from that of the first embodiment will be described with respect to the mail thread extraction device according to the second embodiment of the present invention, which further acquires the response mail data corresponding to the response mail data and generates a mail thread. do.

FIG. 10 is a block diagram showing a schematic configuration of the mail thread extraction system 200 according to the second embodiment of the present invention. The end determination unit 17 determines whether or not the response mail data acquired from the mail data storage unit 12 is within the specified search period, and if the response mail data is within the search period, the start point mail data acquisition. The response mail data is sent to the unit 13. The starting mail data acquisition unit 13 inputs the received response mail data into the response candidate data generation unit 14, and the response candidate data generation unit 14 and the response mail data acquisition unit 15 perform a process of extracting the corresponding response mail data. .. This process is performed as long as the response mail data is a mail within the specified search period.

The origin mail data and the response mail data are stored in the mail thread storage unit 18. When the end determination unit 17 determines that the response mail data is out of the search period, the end determination unit 17 sends an end instruction to the mail thread generation unit 16. The mail thread generation unit 16 that has received the end instruction refers to the origin mail data and the response mail data stored in the mail thread storage unit 18 to generate a mail thread. The response mail data determined by the terminal determination unit 17 to be out of the search period is not reflected in the mail thread, and the mail thread is generated using the starting point mail data and the reply mail data already stored.

FIG. 11 shows a flowchart of the mail thread extraction method according to the second embodiment of the present invention. First, it waits for the search period and the starting mail to be specified (S201). For example, the starting mail and the search period are specified by the user using the user interface as shown in FIG. The starting mail data within the designated search period is acquired from the mail data storage unit 12 as input mail data (S202). The acquired input mail data is input to the trained model 14a of the response candidate data generation unit 14, and the response candidate data is generated (S203). Then, the response mail data acquisition unit 15 acquires response mail data similar to the response candidate data from the mail data storage unit 12 (S204). It is determined whether or not the response mail data acquired from the mail data storage unit 12 is the mail data within the designated search period (S205). When the response mail data is the mail data within the specified search period, the input mail data and the reply mail data are combined to generate a mail thread (S206) and store it in the mail thread storage unit 18 (S207). Then, the response mail data is sent to the starting mail data acquisition unit 13 as the next input mail data, and the process from S202 is repeated. On the other hand, in the determination of S205, when it is determined that the response email data acquired from the email data storage unit 12 is outside the specified search period, the mail thread already stored in the mail thread storage unit 18 is output (S208). ). For example, the response mail data for the starting mail is collectively output (displayed) as a mail thread in the form shown in FIG.

[Third Embodiment]
In the above-described embodiment, the degree of similarity between the response candidate data generated by the response candidate data generation unit 14 and the mail data stored in the mail data storage unit 12 is determined by determining the inner product of the response candidate data and the vector of the mail data. Although the form to be used has been illustrated, a form in which a plurality of response candidate data having a high degree of similarity is further extracted using a trained model trained using a neural network may be used. Hereinafter, the mail thread extraction system according to the third embodiment of the present invention will be described with a configuration different from that of the above-described embodiment.

FIG. 14 shows a similarity relationship learning model for extracting a plurality of response candidate data having a high degree of similarity. The similarity relationship learning model is a learning model for extracting highly similar data from Answer 1 to Answer N, which are response candidate data. For example, when Answer1 is extracted as response candidate data by the classification model, it is possible to extract response candidate data similar to Answer1 by using the similarity relationship learning model. By using the trained model learned by this similarity relationship learning model in the response candidate data generation unit 14 to generate a plurality of response candidate data, the response mail data acquisition unit 15 can acquire the response candidate data without omission. It will be possible. As a learning method of the similarity relationship learning model, supervised learning in which similar response candidate data is used as teacher data and unsupervised learning in which learning is performed only with response candidate data without using teacher data can be applied. The method of is not limited to the present invention.

As described above, one specific embodiment of the present invention has been described, but the above-described embodiment is an example and does not limit the present invention. For example, in the above embodiment, the generation of the trained model by supervised learning is illustrated, but the trained model may be generated by unsupervised learning. In that case, the labeling unit 24 is omitted.

Further, a part or all of the processing of each functional block of the above embodiment may be realized by a program. Then, a part or all of the processing of each functional block of each of the above embodiments is performed by a central processing unit (CPU), a microprocessor, a processor, or the like in a computer. Further, the program for performing each process is stored in a storage device such as a hard disk or a ROM, and is read and executed in the ROM or the RAM.

Further, each process of the above embodiment may be realized by hardware, or may be realized by software (including a case where it is realized together with an OS (operating system), middleware, or a predetermined library). Further, the mail classification system 100 may be realized by mixed processing of software and hardware.

Further, the execution order of the processing methods in the above-described embodiment is not necessarily limited to the description of the above-described embodiment, and the execution order can be changed without departing from the gist of the invention. Further, in the processing method in the above embodiment, some steps may be executed in parallel with other steps as long as the gist of the invention is not deviated.

A computer program that causes a computer to perform the above-mentioned method and a computer-readable recording medium on which the program is recorded are included in the scope of the present invention. Here, the type of computer-readable recording medium is arbitrary. Further, the computer program is not limited to the one recorded on the recording medium, and may be transmitted via a telecommunication line, a wireless or wired communication line, a network represented by the Internet, or the like.

The specific configuration of the present invention is not limited to the above-described embodiment, and various changes and modifications can be made without departing from the gist of the invention.

The present invention can also be described as follows.

The mail thread extraction device according to the first configuration of the present invention is
A mail data storage unit that stores multiple e-mail data, and
A starting point mail data acquisition unit that acquires starting point mail data that is the starting point of a mail thread from the mail data storage unit, and a starting point mail data acquisition unit.
A response candidate data generation unit that generates response candidate data corresponding to the origin mail data using a learned model that has learned the correspondence between the mail data and the corresponding response mail data.
A response mail data acquisition unit that acquires response mail data similar to the response candidate data generated by the response candidate data generation unit from the mail data storage unit, and a response mail data acquisition unit.
It is provided with a mail thread extraction unit that generates a mail thread by combining the starting mail data and the response mail data.

In this first configuration, a mail document data storage unit for storing mail document data is provided, and the starting mail document data serving as the starting point of the mail thread is acquired from the mail document data storage unit. Then, the response candidate data corresponding to the starting mail document data is generated by using the learned model in which the correspondence between the mail document data and the corresponding response mail document data is learned. Further, the response mail document data similar to the response candidate data is acquired from the mail document data storage unit, and a mail thread is generated based on the starting mail document data and the reply mail document data. As a result, compared to the conventional case where related emails are extracted by keyword search by a predetermined word and an email thread is generated, the corresponding emails are comprehensively handled based on the connection of the email body by using the learned model. It can be extracted to, and the thread extraction of mail can be performed appropriately.

In the mail thread extraction device according to the second configuration of the present invention, the response mail data acquired by the response mail data acquisition unit in the mail thread extraction device according to the first configuration is the target response of the mail thread. It is provided with a termination determination unit that determines whether or not the data is mail data.

According to this second configuration, the terminal determination unit for determining whether or not the response mail data is the target of the mail thread is provided, and when it is determined that the response mail data is the target mail of the mail thread, the above-mentioned By inputting the response mail data into the starting mail data acquisition unit and acquiring the mail data that responds to the response mail data, the desired mail thread can be obtained without omission.

The mail thread extraction method according to the present invention
The starting point mail data, which is the starting point of the mail thread, is acquired from the mail data storage unit in which a plurality of e-mail data are stored.
Using the trained model that learned the correspondence between the mail data and the corresponding response mail data, the response candidate data corresponding to the starting mail data is generated.
Response mail data similar to the response candidate data is acquired from the mail data storage unit, and is obtained.
A mail thread is generated by combining the origin mail data and the response mail data.

According to this mail thread extraction method, the starting mail data that is the starting point of the mail thread is acquired from the mail data storage unit in which a plurality of e-mail data are stored, and the correspondence between the mail data and the corresponding response mail data. Response candidate data corresponding to the origin mail data is generated by using the trained model trained in. Then, the response mail data similar to the response candidate data is acquired from the mail data storage unit, and the starting mail data and the response mail data are combined to generate a mail thread. As a result, compared to the conventional case where related emails are extracted by keyword search by a predetermined word and an email thread is generated, the corresponding emails are comprehensively handled based on the connection of the email body by using the learned model. It can be extracted to, and the thread extraction of mail can be performed appropriately.

The program according to the present invention is
The starting point mail data, which is the starting point of the mail thread, is acquired from the mail data storage unit in which a plurality of e-mail data are stored.
Using the trained model that learned the correspondence between the mail data and the corresponding response mail data, the response candidate data corresponding to the starting mail data is generated.
Response mail data similar to the response candidate data is acquired from the mail data storage unit, and is obtained.
This is a program for causing a computer to execute a process of generating a mail thread by combining the starting mail data and the response mail data.

The computer operated by this program acquires the starting point mail data, which is the starting point of the mail thread, from the mail data storage unit in which a plurality of e-mail data are stored, and determines the correspondence between the mail data and the corresponding response mail data. Using the trained trained model, response candidate data corresponding to the origin mail data is generated. Then, the response mail data similar to the response candidate data is acquired from the mail data storage unit, and the starting mail data and the response mail data are combined to generate a mail thread. As a result, compared to the conventional case where related emails are extracted by keyword search by a predetermined word and an email thread is generated, the corresponding emails are comprehensively handled based on the connection of the email body by using the learned model. It can be extracted to, and the thread extraction of mail can be performed appropriately.

A recording medium on which the above program is recorded is also one aspect of the present invention.

1 ... Extractor, 2 ... Learning device, 11 ... Mail data analysis unit 1, 12 ... Mail data storage unit, 13 ... Origin mail data acquisition unit, 14 ... Response candidate data generation unit, 15 ... Response mail data acquisition unit, 16 ... mail thread generation unit, 17 ... termination determination unit, 18 ... mail thread storage unit, 21 ... data analysis unit for learning, 22 ... data set generation unit, 23 ... DNN (deep neural network), 24 ... trained model, 100 ... mail thread extraction system, 200 ... mail thread extraction system

Claims (5)

A mail data storage unit that stores multiple e-mail data, and
A starting point mail data acquisition unit that acquires starting point mail data that is the starting point of a mail thread from the mail data storage unit, and a starting point mail data acquisition unit.
A response candidate data generation unit that generates response candidate data corresponding to the origin mail data using a learned model that has learned the correspondence between the mail data and the corresponding response mail data.
A response mail data acquisition unit that acquires response mail data similar to the response candidate data generated by the response candidate data generation unit from the mail data storage unit, and a response mail data acquisition unit.
A mail thread extraction device including a mail thread generation unit that generates a mail thread by combining the starting mail data and the response mail data. A terminal determination unit for determining whether or not the response mail data acquired by the response mail data acquisition unit is the response mail data targeted by the mail thread is provided.
The terminal determination unit according to claim 1, wherein when the response mail data is determined to be the response mail data to be the target of the mail thread, the response mail data is input to the starting mail data acquisition unit. Mail thread extractor. A method of extracting mail threads that is executed by a computer.
The starting point mail data, which is the starting point of the mail thread, is acquired from the mail data storage unit in which a plurality of e-mail data are stored.
Using the trained model that learned the correspondence between the mail data and the corresponding response mail data, the response candidate data corresponding to the starting mail data is generated.
Response mail data similar to the response candidate data is acquired from the mail data storage unit, and is obtained.
A mail thread extraction method for generating a mail thread by combining the starting mail data and the response mail data. The starting point mail data, which is the starting point of the mail thread, is acquired from the mail data storage unit in which a plurality of e-mail data are stored.
Using the trained model that learned the correspondence between the mail data and the corresponding response mail data, the response candidate data corresponding to the starting mail data is generated.
Response mail data similar to the response candidate data is acquired from the mail data storage unit, and is obtained.
A program for causing a computer to execute a process of generating a mail thread by combining the starting mail data and the response mail data. The starting point mail data, which is the starting point of the mail thread, is acquired from the mail data storage unit in which a plurality of e-mail data are stored.
Using the trained model that learned the correspondence between the mail data and the corresponding response mail data, the response candidate data corresponding to the starting mail data is generated.
Response mail data similar to the response candidate data is acquired from the mail data storage unit, and is obtained.
A recording medium in which a program for causing a computer to execute a process of generating a mail thread by combining the starting mail data and the response mail data is recorded.

Images