JP2020197851A - Mail analysis server, mail analysis method, and program - Google Patents

Abstract

To reduce a processing load on a user on a reception mail by improving the analysis accuracy of the reception mail.SOLUTION: A mail analysis server includes a determination unit for determining the presence/absence of replies to accumulated reception mails, a learning unit for analyzing at least one of characters included in a reception mail determined to have been replied and characters included in a reception mail determined not to have been replied, and constructing a machine learning model by machine learning using a neural network, and an estimation unit for estimating the necessity/unnecessity of a reply to the reception mail about the characters included in the received reception mail by using the machine learning model.SELECTED DRAWING: Figure 1

Description

One embodiment of the present invention relates to a mail analysis server, a mail analysis method, and a program.

In recent years, email has become one of the most frequently used and important tools in business and academia. The user receives a large amount of e-mails every day, and from the received e-mails, determines which one needs to be answered immediately and which one does not, and processes it. Users spend a great deal of time processing large amounts of email.

In Patent Document 1, the e-mail already received by the user is used as teacher data, and the degree of necessity of action (action priority) for each user is evaluated for the e-mail newly received by the user, and the result is evaluated. Describes the email system that can be presented.

In Patent Document 1, both an action e-mail such as a reply or a transfer and an e-mail in which no action is taken are machine-learned as teacher data. In addition, by weighting the keywords extracted from the actiond e-mail, the action required for the newly received e-mail is evaluated by the score.

International Publication No. 2016/117132

In Patent Document 1, the received mail stored in the receiving folder of the user's mail application is machine-learned as teacher data. Therefore, even if the account is the same, if the terminal device that receives the received mail is different, the analysis result may be different for each terminal device.

Further, in Patent Document 1, keywords included in both received mails that have been replied and received mails that have not been replied are extracted and evaluated by a score, but the dependency of keywords is not considered. Therefore, if the same keyword is included in both the received mail that has been replied and the received mail that has not been replied, the analysis accuracy of the received mail may decrease.

Therefore, one embodiment of the present invention aims to improve the analysis accuracy of the received mail and reduce the processing load on the received mail by the user.

The mail analysis server according to the embodiment of the present invention has a determination unit for determining whether or not there is a reply to the accumulated received mail, and it is determined that the characters included in the received mail determined to have been replied and not replied. A learning unit that analyzes at least one of the characters contained in the received mail and builds a machine learning model by machine learning using a neural network, and receives the characters contained in the received mail using the machine learning model. It has an estimation unit that estimates the necessity of replying to an email.

In the above configuration, the determination unit compares the accumulated received mail and the accumulated sent mail, determines whether or not the body of the sent mail includes a part of the body of the received mail, and the sent mail is the received mail. If a part of the body of the mail is included, it is determined that the received mail has been replied.

In the above configuration, the determination unit compares the accumulated received mail and the accumulated sent mail, determines whether or not the sent mail contains the incidental information of the received mail, and the incidental information of the sent mail is the received mail. If it contains incidental information, it is determined that the received mail has been replied.

In the above configuration, the determination unit further has a time calculation unit that calculates the time required for replying to the received mail determined to have been replied after determining whether or not there is a reply to the accumulated received mail. ..

In the above configuration, the learning unit analyzes the characters included in the received mail and the time required for replying, and constructs a machine learning model by machine learning using a neural network.

In the above configuration, the learning model is generated based on the received mail determined to have returned without distinguishing between a plurality of accounts.

In the above configuration, the machine learning model is updated based on the received mail determined to have been returned for each account.

In the above configuration, the neural network is a character-level convolutional neural network.

The mail analysis method according to the embodiment of the present invention determines whether or not there is a reply to the accumulated received mail, and the characters included in the received mail determined to be replied and the received mail determined not to be replied. Analyze at least one of the contained characters, build a machine learning model by machine learning using a neural network, and use the machine learning model to reply to the newly received received mail with the characters included in the received mail. An email analysis method that estimates the necessity.

In the above method, the presence or absence of a reply to the accumulated received mail is determined by comparing the accumulated received mail and the accumulated sent mail, and determining whether or not the sent mail contains a part of the body of the received mail. If the sent mail contains a part of the body of the received mail, it is determined that the received mail has been replied.

In the above method, the presence or absence of a reply to the accumulated received mail is determined by comparing the accumulated received mail and the accumulated sent mail, determining whether or not the sent mail contains incidental information of the received mail, and then sending the sent mail. If contains the incidental information of the received mail, it is determined that the received mail has been replied.

In the above method, after determining whether or not there is a reply to the accumulated received mail, the time required for replying to the received mail determined to have been replied is calculated.

In the above method, the characters included in the received mail and the time required for reply are analyzed, and a machine learning model is constructed by machine learning using a neural network.

In the above method, the neural network is a character-level convolutional neural network.

In the above method, the machine learning model is generated based on the received mail determined to have returned without distinguishing between a plurality of accounts.

In the above method, the machine learning model is updated based on the received mail determined to have been returned for each account.

The program according to the embodiment of the present invention is a program for causing a computer to execute the mail analysis method described above.

According to one embodiment of the present invention, it is possible to improve the analysis accuracy of the received mail and reduce the processing load of the user on the received mail.

It is a figure which shows the outline of the mail analysis server which concerns on one Embodiment of this invention. It is a figure which shows the display screen when the received mail is received together with the probability about the necessity of reply in the mail application. It is a figure which shows the display screen when the received mail is selected in the mail application shown in FIG. It is a block diagram explaining the hardware configuration of the mail analysis server which concerns on one Embodiment of this invention. It is a functional block diagram of the mail analysis server which concerns on one Embodiment of this invention. It is a figure explaining the process of generating a machine learning model in a mail analysis server. It is a figure explaining the process of analyzing the newly received received mail in the mail analysis server, and estimating the probability about the necessity of reply. It is a figure which shows the outline of the mail analysis server which concerns on one Embodiment of this invention. It is a functional block diagram of the mail analysis server which concerns on one Embodiment of this invention. It is a figure which shows the outline of the mail analysis server which concerns on one Embodiment of this invention. It is a functional block diagram of the mail analysis server which concerns on one Embodiment of this invention. It is a figure which shows the display screen when the received mail is received together with the probability about the necessity of reply in the mail application. It is a figure which shows the display screen when the received mail is selected in the mail application. It is a functional block diagram of the mail analysis server which concerns on one Embodiment of this invention. It is a functional block diagram of the mail analysis server which concerns on one Embodiment of this invention.

(First Embodiment)
In the present embodiment, a process in which the mail analysis server 10 according to the embodiment of the present invention outputs the necessity of replying to the newly received received mail will be described with reference to FIGS. 1 to 7.

In the present specification and the like, an e-mail has information including one or more characters in the body of the e-mail. The e-mail includes not only the body of the e-mail but also attachments such as ancillary information, document files, spreadsheet files, and presentation files. Here, the incidental information refers to information accompanying the document data of the mail, for example, information on the address, subject, and document data, operation information such as transmission and reply, operation execution date and time, creation date and time, and document. It may be information such as the creator of the data.

Further, in the present specification and the like, the sent mail means a mail created by a user who uses the mail analysis server 10 and sent to an account (email address) of an arbitrary user. Further, the received mail means a mail created by an arbitrary sender and received by the account (mail address) of the user who uses the mail analysis server 10. In addition, when it is simply described as e-mail, it does not distinguish between received e-mail, sent e-mail, forwarded e-mail, and deleted e-mail.

[Overview of mail analysis server]
FIG. 1 is a diagram illustrating a mail analysis server 10 according to an embodiment of the present invention. FIG. 1 shows a mail analysis server 10, mail servers 20_1 and 20_2, and terminal devices 30_1 and 30_2. The mail servers 20_1 and 20_2 send and receive e-mails via the network 40.

Further, the terminal device 30_1 is a terminal device of a user who uses the mail analysis server 10, and the terminal device 30_1 is a terminal device used by an arbitrary user. Further, the mail server 20_1 is a server that manages the accounts of users who use the mail analysis server 10, and the mail server 20_1 is a server that manages the accounts of arbitrary users. In the following description, when the mail servers 20_1 and 20_2 are not distinguished from each other, the mail server 20 is described. When the terminal devices 30_1 and 30_2 are not distinguished from each other, they are described as the terminal device 30.

The mail server 20 has a function of sending and receiving e-mails of registered accounts. Further, the mail server 20 has a function of holding mail data of e-mail for each account. Here, the mail data includes an address book, the text and incidental information of the e-mail, an attached file, and the like. Here, the mail data of the e-mail for each account is held in the recording medium of the mail server 20. The mail server 20_1 transmits the received mail received from the outside to the user's account from any user's account to the mail analysis server 10. Further, the mail server 20_1 transmits the outgoing mail sent from the user's account to the account of any user to the mail server 20_1.

The mail analysis server 10 analyzes the characters of the sent / received mail received from the mail server 20_1 and estimates the necessity of replying to the received mail. Here, the characters of the e-mail include at least the body of the e-mail, and may also include ancillary information and a quoted portion of other received e-mails. Ancillary information includes the IP address of the sender's mail server (Received field), sender information (From field), internal ID of the mail (Message-ID field), title of the mail (Subject field), and file format of the e-mail. Version information (MIME-Version field), email file format information (Content-Type field), destination information (To field), mail copy destination information (Cc field), mail creation date and time information (Date field) , Reply destination mail information (References field, In-Reply-To field) is included. The mail analysis server 10 estimates a probability of 0% to 100%, for example, assuming that the one with a high need to reply is 100%. The mail analysis server 10 has a function of transmitting the received received mail to the terminal device 30 together with the probability of whether or not a reply is necessary.

The terminal device 30 connects to the mail analysis server 10 and executes processing related to e-mail. The processing related to e-mail includes, for example, a function of inputting e-mail data for transmission sent from a user's account, a function of displaying e-mail data to be viewed, and the like. For example, when the terminal device 30 performs processing related to e-mail, it can be executed by using a mail application.

As shown in FIG. 1, the terminal device 30_2 transmits an e-mail to the source mail server 20_2, and the source mail server 20 transmits the e-mail to the destination mail server 20_1 via the network 40. The destination mail server 20_1 sends the received e-mail to the mail analysis server 10. The mail analysis server 10 analyzes the received e-mail and estimates the probability of the necessity of replying to the sender's account as 0% to 100%. The mail analysis server 10 transmits an e-mail to the terminal device 30 with a probability regarding the necessity of reply, and the terminal device 30 can receive the e-mail with a probability regarding the necessity of reply.

FIG. 2 is a diagram showing a display screen when a received mail is received together with a probability regarding the necessity of reply in the mail application 400 of the terminal device 30. The terminal device 30 displays the account 401, the action display area 402, and the received mail display area 404 by the mail application 400. The account 401 displays the account of the user who uses the mail application 400. Actions such as new mail, reply, reply to all, and forward are displayed in the action display area 402. The user selects one of the actions to compose the incoming email. In the received mail display area 404, the probabilities 406_1 to 406_5 regarding the necessity of reply are displayed in each of the received mails 405_1 to 405_5.

A predetermined threshold value may be preset by the user for the probability regarding the necessity of reply. The terminal device 30 may change the display of the received mail when the probability regarding the necessity of reply is equal to or higher than a predetermined threshold value. The terminal device 30 may, for example, highlight the probability of the necessity of the reply described in the received mail. For example, as shown in FIG. 2, the terminal device 30 replies depending on whether the probability regarding the necessity of reply is 80% or more, 60% to less than 80%, or 0% to less than 60%. Probability highlighting regarding the necessity of is different.

FIG. 3 is a diagram showing a display screen when received mail 405_1 is selected in the mail application 400 shown in FIG. As shown in FIG. 3, the received mail 405_1 displays the incidental information of the received mail 405_1 and the mail body together with the probability regarding the necessity of reply. In addition, the user may be prompted to reply by displaying an icon such as a reply along with the probability of the necessity of reply.

By confirming the probability regarding the necessity of reply in the terminal device 30, the user can easily determine whether the received mail has a high or low need for reply. Since the user only needs to process the received mail having a high probability of whether or not a reply is necessary, the processing load on the received received mail can be reduced.

[Hardware configuration of mail analysis server]
FIG. 4 is a block diagram illustrating a hardware configuration of the mail analysis server 10 according to the embodiment of the present invention. As shown in FIG. 4, the mail analysis server 10 includes a control unit 101, a ROM 102 (Read Only Memory), a RAM 103 (Random Access Memory), a recording medium 104, and a communication unit 105.

The control unit 101 is, for example, a CPU (Central Processing Unit) or the like, and controls the operation of a functional block described later. The control unit 101, for example, reads the operation program recorded in the ROM 102 and the recording medium 104, expands it in the RAM 103, and executes the operation program, thereby realizing the processing function of the mail analysis server 10 by each function block.

The recording medium 104 is, for example, a recording device such as an HDD (Hard Disk Drive) capable of permanently holding information and rewriting information. The recording medium 104 stores e-mail data of e-mails sent and received for each account acquired from the mail server 20. Further, the recording medium 104 stores teacher data based on the returned received mail, which will be described in detail later. Further, the recording medium 104 stores a machine learning model generated by machine learning, which will be described in detail later. The e-mail data, teacher data, and machine learning model of the e-mail for each account may be stored in different recording media.

The ROM 102 is a recording device capable of permanently holding information such as a non-volatile memory. The ROM 102 stores, for example, a program for executing a mail analysis process.

The RAM 103 is a rewritable storage device used for temporarily holding information such as a volatile memory. The RAM 103 is used as an expansion area for a program or the like of a functional block recorded in the ROM 102 and a storage area for storing intermediate data or the like output in the operation of the functional block.

The communication unit 105 is a communication interface. The mail analysis server 10 can connect to the terminal device 30 and the mail server 20 via the communication unit 105, and can send and receive mail to and from the terminal device 30 and the mail server 20.

In the present embodiment, for the sake of simplicity, the hardware configuration of the mail analysis server 10 has been described as having various blocks, but one embodiment of the present invention is not limited to this. For example, at least one of the various blocks may be configured as a device separate from the mail analysis server 10.

[Functional block diagram of mail analysis server]
FIG. 5 is a functional block diagram of the mail analysis server 10 according to the embodiment of the present invention. The mail analysis server 10 includes an e-mail acquisition unit 111, an e-mail storage unit 112, a determination unit 113, teacher data 114, a learning unit 115, a machine learning model 116, an estimation unit 117, and an output unit 118. The e-mail acquisition unit 111, the determination unit 113, the learning unit 115, the estimation unit 117, and the output unit 118 are processed by the control unit 101. Further, the e-mail storage unit 112, the machine learning model 116, and the teacher data 114 are stored in the recording medium 104.

First, a functional block used by the mail analysis server 10 for machine learning will be described.

The e-mail acquisition unit 111 acquires the sent / received e-mails stored in the mail server 20_1 at predetermined intervals. The acquired e-mail is used when the machine learning model 116 is constructed by the learning unit 115 described later. In addition, the e-mail acquisition unit 111 acquires the newly received e-mail received by the mail server 20_1. The newly received received mail that has been acquired is used by the estimation unit 117, which will be described later, to estimate the probability of the presence or absence of a reply.

The e-mail storage unit 112 stores the sent / received e-mail acquired by the e-mail acquisition unit 111. The acquired sent / received e-mail is stored in the e-mail storage unit 112 for each account.

The determination unit 113 reads the received mail and the sent mail stored in the e-mail storage unit 112 every predetermined period, and determines whether or not there is a reply for each received mail. The determination unit 113 compares the accumulated received mail and the accumulated sent mail, and determines whether or not the body of the sent mail includes a part of the body of the received mail. If the sent mail contains a part of the body of the received mail, it is determined that the received mail has been replied. Further, the determination unit 113 may compare the accumulated received mail and the accumulated sent mail, and if the incidental information of the transmitted mail includes the incidental information of the received mail, it may determine that the received mail has been returned. Here, the incidental information of the received mail is, for example, Message-ID, In-Reply-To, and References. In In-Reply-To, only one Message-ID to be returned is recorded. A plurality of Message-IDs of thread mails are recorded in References. For example, even if one of the sent mails is deleted, the reply status can be restored from the References of the subsequent mails. The received mail determined to have been replied and the received mail determined not to have been replied are stored in the teacher data 114.

The learning unit 115 analyzes the characters included in the received mail stored in the teacher data 114, and constructs a machine learning model by machine learning using a neural network. The constructed machine learning model is stored in the machine learning model 116. Here, as the neural network, a character-level convolutional neural network (Character-level Convolutional Neural Network) is used. Character-level convolutional networks perform machine learning on a character-by-character basis, not on a word-by-word basis. Therefore, it is resistant to unknown words such as proper nouns, industry terms, and industry-specific abbreviations, and errors.

In addition, the neural network relearns at predetermined intervals and updates the machine learning model 116. The neural network relearns the received mail stored in the teacher data 114 as new teacher data and updates the machine learning model 116.

Further, the construction of the machine learning model 116 may be executed without distinguishing a plurality of accounts, or may be executed for each account. For example, the machine learning model 116 is constructed for received emails that are determined to have replied, without distinguishing between a plurality of accounts. Further, the machine learning model 116 is updated based on the received mail determined to have been returned for each account. By updating the machine learning model 116 for each account, it is expected that the accuracy of the analysis result of the received mail for each account will be improved.

When the e-mail acquisition unit 111 acquires the newly received received e-mail, the estimation unit 117 analyzes the characters included in the newly received received e-mail by using the machine learning model 116, and needs to reply to the newly received e-mail. Estimate the probability of no.

The output unit 118 outputs the received mail to the terminal device 30 together with the probability regarding the necessity of the reply estimated by the estimation unit 117.

The newly received e-mail received by the e-mail acquisition unit 111 may be stored in the e-mail storage unit 112 and transmitted to the estimation unit 117. Alternatively, the received mail may be sent to the estimation unit 117 without being stored in the e-mail storage unit 112. Even if the e-mail acquisition unit 111 does not store the newly received received e-mail in the e-mail storage unit 112 at the timing when the newly received e-mail is acquired, the e-mail storage unit 112 mails the e-mail sent and received at predetermined intervals. It may be obtained from the server 20_1.

[Data processing method]
Next, the mail analysis method according to the embodiment of the present invention will be described with reference to FIGS. 6 and 7. The control unit 101 causes the mail analysis server 10 to execute each process.

First, the process of constructing the machine learning model on the mail analysis server 10 will be described with reference to FIG. The e-mail acquisition unit 111 acquires e-mails sent and received from the mail server 20_2 at predetermined intervals (step S301). The e-mail acquisition unit 111 stores the acquired e-mail in the e-mail storage unit 112 (step S302). Next, the determination unit 113 reads the sent mail and the received mail from the e-mail storage unit 112, and determines whether or not there is a reply to the received e-mail (step S303). Next, the determination unit 113 stores the received e-mail determined not to be replied in the teacher data (step S304). Next, the learning unit 115 analyzes the received mail determined to have returned as teacher data, and constructs a machine learning model by machine learning using a neural network (step S305). Next, the learning unit 115 stores the constructed machine learning model in the machine learning model 116 (step S306).

Next, the process of analyzing the newly received received mail on the mail analysis server 10 and estimating the probability regarding the necessity of reply will be described with reference to FIG. 7. The e-mail acquisition unit 111 acquires the newly received received e-mail from the mail server 20_2 (step S311). Next, the estimation unit 117 analyzes the characters included in the received mail by the neural network (step S312). Next, the estimation unit 117 estimates the probability regarding the necessity of replying to the received mail by using the machine learning model (step S313). Next, the output unit 118 outputs the received mail to the terminal device 30 together with the probability regarding the necessity of replying to the received mail (step S314).

When transmitting the transmitted mail from the terminal device 30_1 to the terminal device 30_2, it is not necessary to go through the mail analysis server 10. When the transmission mail is transmitted from the terminal device 30_1 to the terminal device 30_1, the transmission mail is transmitted from the terminal device 30_1 to the mail server 20_1. The mail server 20_1 transmits to the terminal device 30_2 via the mail server 20_2. When the mail server 20_1 acquires the sent mail from the terminal device 30_1, the mail server 20_1 stores the sent mail in the recording medium of the mail server 20_1. When the mail analysis server 10 acquires the sent mail, it may be obtained from the mail server 20_1 at predetermined intervals as described in step S301 of FIG.

The mail analysis server 10 according to the embodiment of the present invention uses a machine learning model to estimate the probability of the necessity of replying to the received mail from the characters included in the newly received received mail, thereby replying. The received mail can be sent to the terminal device 30 together with the probability regarding the necessity of the above. By confirming the probability regarding the necessity of replying in the terminal device 30, the user can easily determine whether the received mail has a high or low need for replying. Therefore, it is possible to reduce the processing load of the user for the received received mail.

In the past, since the number of occurrences of keywords included in received emails was only counted, it was not possible to consider the context and dependency of words. For example, "Please renew the SSL certificate. No reply is required." And "No need to renew the SSL certificate. Please reply." Cannot be distinguished. Since the mail server 20 according to the embodiment of the present invention uses a character-level convolutional neural network, it is possible to analyze an e-mail in consideration of the context and dependency of words.

In addition, in the past, analysis was performed at the word level, so if sentences containing words that are not in the dictionary, such as proper nouns, industry terms, and industry-specific abbreviations, could not be accurately divided into word sequences, the accuracy of analysis would be improved. There was a risk of deterioration. Since the mail server 20 according to the embodiment of the present invention analyzes at the character level, there is no possibility that the accuracy will be lowered even if a word that is not in the dictionary is included.

In addition, e-mails judged to be junk e-mails have a high degree of commonality in terms of characters and words used regardless of the account. Therefore, it makes sense to morphologically analyze emails that are determined to be annoying, regardless of the account.

In contrast, emails that require a reply have a very different tendency to require a reply depending on the user. Also, in emails that require a reply, there are many unknown words such as proper nouns, industry terms, and industry-specific abbreviations. Therefore, by analyzing the body of the e-mail at the character level rather than by morphological analysis, the accuracy of the probability regarding the necessity of reply can be further improved.

In addition, the mail analysis server 10 according to the embodiment of the present invention analyzes the characters included in the received e-mail. Therefore, if the accounts are the same, the probability regarding the necessity of reply output to the plurality of terminal devices 30 is the same for each of the plurality of terminal devices 30. Therefore, the user can receive the e-mail with a probability regarding the necessity of replying without depending on the receiving terminal device 30.

(Second Embodiment)
In the first embodiment, an example in which the mail analysis server 10 and the mail server 20 are configured by separate devices has been described, but in the present embodiment, the mail analysis server 10 and the mail server 20 are integrated into the same device. An example of the above will be described with reference to FIGS. 8 and 9.

[Overview of mail analysis server]
FIG. 8 is a diagram illustrating a mail analysis server 10A according to an embodiment of the present invention. FIG. 8 shows a mail analysis server 10A, a mail server 20_2, and terminal devices 30_1 and 30_2. In the present embodiment, the mail analysis server 10 has an e-mail transmission / reception function of the mail server 20_1 described in the first embodiment. Further, in the present embodiment, the mail analysis server 10 has a function of sending and receiving e-mails of the registered account. Further, the mail analysis server 10 has a function of holding an e-mail for each account.

[Functional block diagram of mail analysis server]
FIG. 9 is a functional block diagram of the mail analysis server 10A according to the embodiment of the present invention. The mail analysis server 10A shown in FIG. 9 is different from the mail analysis server 10 shown in FIG. 5 in that it further has a function of transmitting the sent mail transmitted from the terminal device 30_1 to the mail server 20_2. In this embodiment, a functional block having a function different from that of the first embodiment will be described in detail.

First, the process when the mail analysis server 10A newly receives the received mail from the mail server 20_2 will be described. When the e-mail acquisition unit 111A acquires the received e-mail received from the mail server 20_2, it stores it in the e-mail storage unit 112. Further, the e-mail acquisition unit 111A transmits the received e-mail to the estimation unit 117. Since the determination unit 113, the teacher data 114, the learning unit 115, the machine learning model 116, and the estimation unit 117 have the same functions as those described in the first embodiment, detailed description thereof will be omitted. Further, the output unit 118 outputs the received mail to the terminal device 30 together with the probability regarding the necessity of the reply estimated by the estimation unit 117.

Next, the process when the mail analysis server 10A sends the sent mail sent from the terminal device 30 to the mail server 20_2 will be described. When the e-mail acquisition unit 111A acquires the transmitted e-mail transmitted from the terminal device 30, it stores it in the e-mail storage unit 112. Further, the e-mail acquisition unit 111A transmits a transmission mail to the output unit 118A. The output unit 118A transmits an outgoing mail to the mail server 20_2.

(Third Embodiment)
In the present embodiment, FIGS. 10 and 10B show a mail analysis server 10B that enables e-mail exchange in the terminal device 30 by a user inputting or operating on a screen displayed on the browser of the terminal device 30. This will be described with reference to 11.

[Overview of mail analysis server]
FIG. 10 is a diagram illustrating a mail analysis server 10B according to an embodiment of the present invention. FIG. 10 shows a mail analysis server 10B, a mail server 20_2, and a terminal device 30.

[Functional block diagram of mail analysis server]
FIG. 11 is a functional block diagram of the mail analysis server 10B according to the embodiment of the present invention. The mail analysis server 10B shown in FIG. 11 does not output the e-mail to the terminal device 30, but displays the e-mail on the screen displayed on the browser of the terminal device 30. It is different from 10. In this embodiment, a functional block having a function different from that of the first embodiment will be described in detail.

When the e-mail processing unit 119 acquires the received mail received from the mail server 20_2, the e-mail processing unit 119 stores it in the e-mail storage unit 112. In addition, the e-mail processing unit 119 transmits the received mail to the estimation unit 117. The estimation unit 117 estimates the probability regarding the presence or absence of a reply, and stores the probability regarding the presence or absence of a reply together with the received mail stored in the e-mail storage unit 112. The e-mail processing unit 119 displays the received e-mail with the probability of presence / absence of a reply on the screen displayed on the browser of the terminal device 30 according to the received e-mail viewing request of the terminal device 30.

When the e-mail processing unit 119 acquires the generated outgoing mail on the screen displayed on the browser of the terminal device 30, it stores it in the e-mail storage unit 112. Further, the e-mail processing unit 119 transmits the transmitted e-mail to the mail server 20_2.

(Modification example 1)
In the first embodiment, an example of displaying all received received mails in the inbox of the mail application 400 has been described, but one embodiment of the present invention is not limited to this. Separately from the inbox, only received mails whose probability regarding the necessity of reply is equal to or more than a predetermined threshold may be displayed in the reply required tray. In addition, received mails below a predetermined threshold value may be displayed in the inbox.

FIG. 12 is a diagram showing a display screen when a received mail is received together with a probability regarding the necessity of reply in the mail application 400 of the terminal device 30. The difference from the display of the mail application 400 shown in FIG. 2 is that it has a reply-requiring tray in addition to the inbox. In the reply required tray, only received mails whose probability regarding the necessity of reply is equal to or higher than a predetermined threshold value can be displayed. In FIG. 12, received mails 405_1 to 405_4 having a probability of 80% or more regarding the necessity of reply are displayed. Incoming emails with a probability of less than 80% whether or not a reply is necessary are displayed in the inbox, so that the user can check them as appropriate. As a result, the user only has to first confirm and reply to the received mail having a high priority of the reply received in the reply required tray, so that the processing efficiency of the user for the received received mail is improved.

(Modification 2)
The probability of needing a reply to an incoming email may vary from account to account. This is because even for the same received mail, the tendency for a reply to be required differs greatly depending on the user.

FIG. 13 is a diagram showing a display screen when received mail 407_1 is selected in the mail application. As shown in FIG. 3, when "Suzuki of YYYY Co., Ltd." receives the received mail from "Yamada of ZZZZ Co., Ltd.", the probability of the necessity of reply is output as 80%. On the other hand, as shown in FIG. 13, when "Tanaka of YYYY Co., Ltd." receives the received mail from "Yamada of ZZZZ Co., Ltd.", the probability regarding the necessity of reply may be output as 50%. ..

The mail analysis server 10 according to the embodiment of the present invention updates the machine learning model for each account. Therefore, when a plurality of destinations are included in one received mail, the mail analysis server 10 can estimate the probability regarding the necessity of replying for each account.

(Modification 3)
In one embodiment of the present invention, an example in which characters included in a received email are targeted for machine learning has been described, but one embodiment of the present invention is not limited to this. The time taken to reply to the received mail may be the target of machine learning.

FIG. 14 is a functional block diagram of the mail analysis server 10C according to the embodiment of the present invention. The mail analysis server 10C shown in FIG. 14 further adds a time calculation unit 121 that calculates the time required for the determination unit 113 to reply to the received e-mail after determining whether or not there is a reply to the received e-mail. It is different from the mail analysis server 10 shown in FIG. 5 in that it has. Hereinafter, the function of the time calculation unit 121 will be described in detail.

The time calculation unit 121 calculates the time required to reply to the received mail. The target of the e-mail for which the time required for replying is calculated may be the reply mail replied to the received mail, or may be the target of all sent / received mails. When the reply mail replied to the received mail is targeted, the time required to reply is extracted from the reception date and time of the received mail and the transmission date and time of the sent mail replied to the received mail. .. Alternatively, the time required for replying may be extracted from the reception date and time of the received mail included in the incidental information of the reply mail and the transmission date and time of the reply mail. In addition, when targeting reply mails that have been replied to received mails, the time required to reply can be extracted even if the time required to reply to received mails that have not been replied is ∞. Good. The time required for the time calculation unit 121 to reply extracted is stored in the teacher data 114.

A threshold value may be set for the time required to reply to the received mail. For example, if the time required to reply is less than 24 hours, it is determined that the received mail is of high importance to be replied, and if it is 24 hours or more, it is determined that the received mail is of low importance. You may. The characters contained in the sent / received mail and the time required for the extracted reply are stored in the teacher data.

The learning unit 115 executes machine learning including not only the characters of the sent / received mail but also the time required for replying.

By including not only the characters of the received e-mail but also the time required for reply as the target of machine learning, the analysis accuracy regarding the necessity of replying the received e-mail can be further improved.

(Modification example 4)
In one embodiment of the present invention, the mail analysis servers 10, 10A to 10C have described an example including at least a determination unit 113, a learning unit 115, and an estimation unit 117, but the present invention is not limited thereto. The mail analysis server 10D may have at least an estimation unit 117, and may have a separate learning device for constructing a machine learning model.

FIG. 15 is a functional block diagram of the mail analysis server 10D and the learning device 50 according to the embodiment of the present invention. As shown in FIG. 15, a learning device 50 is provided separately from the mail analysis server 10D.

The learning device 50 includes an e-mail acquisition unit 111, an e-mail storage unit 112, a determination unit 113, teacher data 114, a learning unit 115, and a machine learning model 116. The determination unit 113 determines whether or not there is a reply to the accumulated received mail, and the learning unit 115 determines the characters included in the received mail determined to have been replied and the characters included in the received mail determined not to be replied. At least one of the above is analyzed, and a machine learning model is constructed by machine learning using a neural network. The detailed description of each configuration of the learning device 50 is as described in the first to third embodiments.

The mail analysis server 10D has an e-mail acquisition unit 122, an estimation unit 117, and an output unit 118. The e-mail acquisition unit 122 acquires the received e-mail from the mail server 20_1 that needs to estimate the necessity of reply. The estimation unit 117 estimates the necessity of replying to the characters of the received mail acquired by the e-mail acquisition unit 122 by using the machine learning model of the learning device 50. The output unit 118 outputs to the terminal device 30 whether or not a reply estimated by the estimation unit 117 is necessary.

As described above, the mail analysis server 10D and the learning device 50 may be different devices.

(Modification 5)
In one embodiment of the present invention, the learning unit 115 has described an example of constructing a machine learning model using a character-level convolutional neural network, but the present invention is not limited to this. In the learning unit 115, a machine learning model may be constructed by using a BERT (Biorectional Encoder Representations from Transfermers) as a neural network. BERT is a technique for pre-training linguistic expressions to obtain cutting-edge results in a variety of Natural Language Processing (NLP) tasks. BERT provides interactive pre-training without a teacher. Unsupervised means that BERT was trained using only the plain text corpus.

The learning unit 115 learns the received mail stored in the teacher data 114 as teacher data by using the pre-learned model of BERT, and constructs a machine learning model. When the estimation unit 117 acquires the newly received received mail by the e-mail acquisition unit 111, the machine learning model 116 by BERT is used to analyze the characters included in the newly received received mail and reply to the newly received mail. Estimate the probability of needing. In BERT, it is possible to consider both the context before and after the word of interest, so that the accuracy of the probability regarding the necessity of replying the received mail can be further improved.

10, 10A, 10B: Mail analysis server, 20: Mail server, 30: Terminal device, 40: Network, 101: Control unit, 102: ROM, 103: RAM, 104: Recording medium, 105: Communication unit, 111, 111A : E-mail acquisition unit, 112: E-mail storage unit, 113: Judgment unit, 114: Teacher data, 115: Learning unit, 116: Machine learning model, 117: Estimate unit, 118, 118A: Output unit, 119: E-mail Processing unit, 400: mail application, 401: account, 402: action display area, 404: received mail display area, 405: received mail, 406: probability regarding necessity of reply, 407: received mail

Claims (17)

A judgment unit that determines whether or not there is a reply to the accumulated received mail, and
A learning unit that analyzes at least one of the characters contained in the received mail determined to be replied and the characters contained in the received mail determined not to be replied, and builds a machine learning model by machine learning using a neural network. When,
A mail analysis server having an estimation unit that estimates the necessity of replying to the received mail by using the machine learning model for characters included in the received mail. The determination unit
The accumulated received mail and the accumulated sent mail are compared, and it is determined whether or not the body of the sent mail includes a part of the body of the received mail.
The mail analysis server according to claim 1, wherein when the sent mail includes a part of the body of the received mail, it is determined that the received mail has been replied. The determination unit
The accumulated received mail and the accumulated sent mail are compared, and it is determined whether or not the sent mail contains incidental information of the received mail.
The mail analysis server according to claim 1, wherein when the incidental information of the sent mail includes the incidental information of the received mail, it is determined that the received mail has been returned. 1. The determination unit further includes a time calculation unit that calculates the time required for replying to the accumulated received mail after determining whether or not there is a reply to the accumulated received mail. The mail analysis server described in. The mail analysis server according to claim 4, wherein the learning unit analyzes characters included in the received mail and the time required for the reply, and constructs the machine learning model by machine learning using a neural network. The mail analysis server according to claim 1, wherein the machine learning model is constructed based on the received mail determined to have returned without distinguishing a plurality of accounts. The mail analysis server according to claim 1, wherein the machine learning model is updated based on the received mail determined to have returned for each account. The mail analysis server according to claim 1, wherein the neural network is a character-level convolutional neural network. Determine if there is a reply to the accumulated received mail,
Analyze at least one of the characters contained in the received mail determined to be replied and the characters contained in the received mail determined not to be replied, and build a machine learning model by machine learning using a neural network.
A mail analysis method for estimating the necessity of replying to a newly received received mail by using the machine learning model. Whether or not there is a reply to the accumulated received mail is determined.
The accumulated received mail and the accumulated sent mail are compared, and it is determined whether or not the sent mail contains a part of the body of the received mail.
The mail analysis method according to claim 9, wherein when the sent mail includes a part of the body of the received mail, it is determined that the received mail has been replied. Whether or not there is a reply to the accumulated received mail is determined.
The accumulated received mail and the accumulated sent mail are compared, and it is determined whether or not the sent mail contains incidental information of the received mail.
The mail analysis method according to claim 9, wherein when the sent mail contains incidental information of the received mail, it is determined that the received mail has been returned. The mail analysis method according to claim 9, wherein after determining whether or not there is a reply to the stored received mail, the time required for replying to the stored received mail is calculated. The email analysis method according to claim 12, wherein the character included in the received email and the time required for the reply are analyzed, and the machine learning model is constructed by machine learning using a neural network. The mail analysis method according to claim 9, wherein the neural network is a character-level convolutional neural network. The email analysis method according to claim 9, wherein the machine learning model is generated based on the received email determined to have returned without distinguishing a plurality of accounts. The email analysis method according to claim 9, wherein the machine learning model is updated based on the received email determined to have been replied to for each account. The mail analysis method according to any one of claims 9 to 16.
A program that lets a computer run.

Images