FR3095876A1 - Method of filtering a message to be sent to a recipient, and associated filtering device - Google Patents

Abstract

Process for filtering a message to be sent to a recipient, and associated filtering device The invention relates to a method for filtering a message (MS) to be sent to a recipient, said method comprising the following steps implemented by a filtering device: - obtaining (E320) of at least one context element (EC) for sending said message (MS), by semantic analysis of at least part of said message (MS), - obtaining (E330 ) an identifier (IDD) corresponding to the recipient of the message (MS), - filtering (E350) of the message (MS) as a function of said at least one context element (EC) and of said identifier (IDD). Figure for the abstract: Fig. 3.

Description

Method for filtering a message to be sent to a recipient, and associated filtering device

The present invention relates to the general field of telecommunications, and more particularly relates to a method of filtering an electronic message to be sent to a recipient.

In order to send an e-mail message from a terminal, an identifier of the message recipient, such as a telephone number or an e-mail address or a name, must be entered in the message recipient field.

Errors in filling in the recipient field are however common, often resulting in the message being sent to the wrong recipient.

A filling error can result from a careless error when filling in the recipient field, consisting for example of using the “reply to all” function instead of the “reply” function to reply to an email received.

The careless error can also be made in a situation where a user's mobile phone receives an SMS message from a first interlocutor, while the user is chatting by SMS messages with a second interlocutor. It is common, in such a situation, that the user makes a mistake when replying to the SMS message of the first interlocutor in the discussion thread of the conversation with the second interlocutor.

There is therefore a need to reduce the risk of recipient error before sending an electronic message.

The present invention relates to a method for filtering a message to be sent to a recipient, said method comprising the following steps implemented by a filtering device:
- obtaining at least one context element for sending said message, by semantic analysis of at least part of said message,
- obtaining an identifier corresponding to the recipient of the message,
- filtering the message according to said at least one context element and said identifier.

Filtering the message based on at least one context element and the identifier effectively reduces the risk of recipient error.

In a particular embodiment, the method further comprises a step of calculating a recipient error risk coefficient as a function of said at least one context element and of said identifier, implemented by the filtering device or a server, the filtering of the message being carried out as a function of said recipient error risk coefficient.

Taking into account the content of the message and the recipient makes it possible to precisely calculate a risk of recipient error. By filtering the message according to the risk coefficient, the invention makes it possible to effectively reduce the risk of recipient error.

In a particular embodiment, the method is implemented upon receipt of a command to send said message.

In a particular embodiment, the filtering step comprises:
- the blocking of the sending of the message and the emission of an alert intended for a user of the filtering device when the risk coefficient is greater than a risk threshold value,
- the sending of the message when the risk coefficient is lower than the risk threshold value, or following the reception of a validation command after the transmission of the alert.

In a particular embodiment, the step of calculating the risk coefficient comprises the following sub-steps:
- obtaining a set of conversation profile elements, associated with the identifier corresponding to the recipient and comprising at least one profile element,
- calculation of the risk coefficient according to the result of a search for said at least one context element in the set of conversation profile elements.

In a particular embodiment, the risk coefficient is also calculated according to the result of a search for the identifier corresponding to the recipient in a set of identifiers of recipients associated with said at least one context element.

In a particular embodiment, the risk coefficient is calculated according to the result of a comparison between said at least one context element and the identifier corresponding to the recipient. Said at least one compared context element is typically the subject of a sentence of said at least part of the message.

In a particular embodiment, said message is associated with a marker indicating that the message is at risk, the risk coefficient being calculated as a function of said marker.

In a particular embodiment, said at least one context element is associated with at least one marker, each marker indicating:
- the syntactic function, in a sentence of said at least part of the message, of the associated context element, or
- a position of the associated context element in said at least part of the message, or
- that the associated context element is at risk, said marker then being determined from at least one other context element,
the risk coefficient being calculated as a function of said at least one marker.

In a particular embodiment, when said at least one marker indicates that the syntactic function of said context element is a subject or an object complement, the set of conversation profile elements is obtained from the identifier corresponding to the recipient and said at least one context element associated with said marker.

In a particular embodiment, said at least one context element and the identifier corresponding to the recipient are compared when said at least one marker indicates that the syntactic function of said at least one context element is a subject or an object complement .

The invention further relates to a filtering device capable of filtering a message to be sent to a recipient, said filtering device comprising:
- a module for obtaining at least one context element for sending said message, by semantic analysis of at least part of said message,
- a module for obtaining an identifier corresponding to the recipient of the message,
- a module for filtering the message according to said at least one context element and said identifier.

In a particular embodiment, the filtering device comprises a module for calculating a recipient error risk coefficient as a function of said at least one context element and of said identifier, the message filtering module filtering the message by function of the recipient error risk coefficient.

In a particular embodiment, the different steps of the filtering method are determined by computer program instructions.

Consequently, the invention also relates to a computer program, on an information medium, this program comprising instructions adapted to the implementation of the steps of a method according to the invention.

This program may use any programming language, and be in the form of source code, object code, or intermediate code between source code and object code, such as in partially compiled form, or in any other desirable form.

The invention also relates to an information medium readable by a computer, and comprising instructions of a computer program as mentioned above.

The information carrier can be any entity or device capable of storing the program. For example, the medium may comprise a storage means, such as a ROM, for example a CD ROM or a microelectronic circuit ROM, or even a magnetic recording means, for example a hard disk.

On the other hand, the information medium can be a transmissible medium such as an electrical or optical signal, which can be conveyed via an electrical or optical cable, by radio or by other means. The program according to the invention can in particular be downloaded from an Internet-type network.

Alternatively, the information carrier may be an integrated circuit in which the program is incorporated, the circuit being adapted to execute or to be used in the execution of the method in question.

Other characteristics and advantages of the present invention will become apparent from the description given below, with reference to the appended drawings which illustrate an example of embodiment devoid of any limiting character. In the figures:

FIG. 1 schematically represents a system comprising a filtering device according to an exemplary embodiment of the invention;

FIG. 2 schematically represents the filtering device of the system of FIG. 1;

FIG. 3 represents, in the form of a flowchart, the main steps of a method for filtering a message to be sent to a recipient, according to an exemplary embodiment of the invention;

Figure 4 schematically represents a data structure that can be used in the filtering method of Figure 3.

FIG. 1 schematically represents a system 100 comprising a filtering device 110, capable of implementing a method of filtering a message to be sent to a recipient, according to an exemplary embodiment of the invention.

The filtering device 110 is typically integrated into a terminal 120. For example, the filtering device 110 can be integrated into a messaging tool 130 of the terminal 120, the messaging tool 130 being software installed on the terminal 120, allowing a user of the terminal 120 to send and receive electronic messages via a telecommunications network 140.

No limitation is attached to the nature of the terminal 120, nor to the telecommunications network 140. Thus, the terminal 120 can be a fixed terminal or a mobile terminal, such as a portable telephone, for example of the “smartphone” type, a digital tablet, or a personal computer. In addition, the telecommunications network 140 can be a wired or wireless, mobile or fixed telecommunications network, and therefore can be an Internet network, a Wifi network, a Bluetooth network, an NFC network, or a fixed telephone network or mobile (3G, 4G etc.).

In addition, each electronic message that can be sent or received via the messaging tool can be an "SMS" type message (for "Short Message Service", in English terminology), an "MMS" type message (for "Multimedia Messaging Service", in English terminology), an email, an instant messaging message, etc.

The filtering device 110 may comprise a reception module 111, a module 112 for obtaining at least one context element, a module for obtaining 114 an identifier, a calculation module 116 and/or a module for filtering 118, the roles of which will be described below, with reference to FIG. 3.

The system 100 comprises a server 150, the filtering device 110 and the server 150 being able to communicate, via the terminal 120, by means of the telecommunications network 140 or another wired or wireless telecommunications network (not shown) .

As shown in Figure 2 , the terminal 120 has the conventional architecture of a computer. The terminal 120 comprises in particular a processor 200, a read only memory 202 (of the “ROM” type), a rewritable non-volatile memory 204 (of the “EEPROM” or “Flash NAND” type for example), a rewritable volatile memory 206 (of the "RAM"), and a communication interface 208.

The ROM 202 of the terminal 120 constitutes a recording medium in accordance with an exemplary embodiment of the invention, readable by the processor 200 and on which is recorded a computer program P1 in accordance with an exemplary embodiment of the invention. Alternatively, the computer program P1 is stored in the rewritable non-volatile memory 204.

The computer program P1 enables the filtering device 110 to implement the filtering method according to an exemplary embodiment of the invention, or at least a part of this method.

This computer program P1 can thus define functional and software modules of the filtering device 110, configured to implement the steps of a filtering method in accordance with an exemplary embodiment of the invention, or at least one part of these steps. These functional modules are based on or control the hardware elements 200, 202, 204, 206 and 208 of the terminal 120 mentioned above. They can include in particular here the reception module 111, the module 112 for obtaining at least one context element, the module for obtaining 114 an identifier, the calculation module 116 and the filtering module 118.

The rewritable non-volatile memory 204 can store an SD data structure whose role is described below with reference to FIG. 3.

In addition, the server 150 can also have the conventional architecture of a computer, and thus in particular comprise a processor, a read only memory (of the "ROM" type), a rewritable non-volatile memory (of the "EEPROM" or "Flash NAND” for example), a rewritable volatile memory (“RAM” type), and a communication interface.

The ROM can constitute a recording medium in accordance with an exemplary embodiment of the invention, readable by the associated processor and on which is recorded a computer program in accordance with an exemplary embodiment of the invention. . Alternatively, the computer program is stored in the associated rewritable non-volatile memory. The computer program may implement at least part of the filtering method according to an exemplary embodiment of the invention.

FIG. 3 represents a method of filtering an electronic message MS to be sent to a recipient, according to an exemplary embodiment of the invention.

The method is implemented by the filtering device 110 described with reference to Figure 1. Alternatively, the method is implemented by the system 100, also described with reference to Figure 1.

As mentioned above, MS message is an electronic message such as email, SMS type message, MMS type message, instant messaging message, etc.

The MS message typically includes a header and a body. The body of the message typically includes text and/or an image etc., forming information to be transmitted to the recipient(s). The header includes a recipient field, in which one or more identifiers of the recipient(s) can be entered, and can include a subject field, a sender field, a date field, etc. The header thus includes information allowing the transmission of the information contained in the body of the message.

The method is implemented before sending the electronic message MS.

For example, the method is implemented upon reception, that is to say following the reception, by the reception module 111, of a CMD command to send the message MS (step E310), the CMD command being typically sent following an action by the user aimed at sending the message MS, for example a click on a send button.

As a variant, the method is implemented upon receipt of a command to verify the recipients of the message MS, this command being typically issued following an action by the user aimed at verifying the risk of recipient error, for example a click on a verification button.

As a variant, the method is implemented automatically, as soon as at least part of the text of the body of the message MS and of the recipients have been written. Thus, the method is typically implemented after the recipient field and the body of the message MS have been filled in at least partially.

The obtaining module 112 of the filtering device then obtains, in a step E320, at least one context element EC for sending the message MS, by semantic analysis of at least part of the message MS, for example by analysis semantics of at least part of the body text of the message MS. The context element EC corresponds to a keyword of the part of the text analyzed.

Each EC context element obtained by semantic analysis is typically a lexical unit such as a word or a group of words, for example a named entity. The context element thus takes the form of a character string.

By "named entity", we mean a lexical unit such as a word or a group of words (a linguistic expression) referring to an entity of the concrete world, and thus being able to designate a proper name, such as a place name. , a name and/or a first name of a person or a name of an organization. This lexical unit constitutes the name of this “entity”, which thus becomes a named entity. In practice, a named entity is most often unique, but can have several names (London/Londres, Paris/La capital de la France/La ville lumière). Named entities are used in particular in the fields of automatic language processing and text corpus analysis.

Semantic analysis is typically performed on all of the MS message body text, or on a portion of the MS message body text, such as the first sentence of the MS message body text. The context element(s) EC obtained form a summary of the message.

Semantic analysis is based in particular on two functionalities which are lexical analysis and syntactic analysis. Lexical analysis makes it possible to split the text into lexemes (words, expressions, named entities), also called lexical units. Syntactic analysis relies on grammatical rules to define what functions words have within a text, and the relationships between them (e.g. relationship between subject and object).

The semantic analysis thus makes it possible to obtain the context element(s) EC, by segmenting the text in order to obtain one or more lexical unit(s), then by labeling this or these lexical units, the label indicating for example that the associated lexical unit is a personal name, a company name, etc.

Semantic analysis can be based on an algorithm using formal grammars and possibly a priori knowledge, for example taking the form of a list of first names, cities or countries.

Semantic analysis is typically based on machine learning algorithms, such as a random conditional field algorithm, a support vector machine algorithm, or a finite state transducer algorithm.

Semantic analysis can also be based on algorithms for measuring distances between words, typically a distance between a word in the sentence and a named entity).

Neural networks can be used.

In addition, in a sub-step E325 of step E320 for obtaining, at least part of the context element(s) EC obtained in step E320, for example each context element, can be associated with one or more markers QM. Each MQ marker can indicate:
- the syntactic function, in a sentence of said at least part of the message MS, of the associated context element, and therefore for example indicate that the associated EC context element is a subject or an object complement of a sentence of said at least part of the message MS, or
- a position of the associated EC context element in said at least part of the message MS, or
- that the associated EC context element is at risk, the marker MQ then being able to be determined from at least one other EC context element.

The semantic analysis of step E320 can thus be carried out in order to find one or more subjects and/or sentence object complements of the message MS.

In a step E330, the obtaining module 114 obtains at least one identifier IDD corresponding to the recipient of the message MS. More specifically, the obtaining module 114 searches in the recipient field of the message MS for one or more IDD identifiers, each IDD identifier possibly corresponding to a different recipient.

Each identifier can be a telephone number, an e-mail address, a surname and/or a first name, or even a nickname of the recipient.

Step E330 of obtaining an identifier can be performed in parallel with step E320 of obtaining at least one context element, or before or after step E320 of obtaining at least one element of context.

The filtering module 118 then filters the message MS according to the context element(s) EC obtained in step S320 and the identifier(s) IDD obtained in step E330.

Typically, in a step E340, the calculation module 116 calculates a recipient error risk coefficient CR, as a function of at least one context element EC obtained in step E320 and at least one identifier IDD obtained at step E330.

The risk coefficient CR is for example calculated by searching for each context element EC or part of the context elements EC obtained in step E320 in a set EEP of conversation profile elements associated with the recipient.

The EEP set of conversation profile elements comprises one or more EP profile elements, each EP profile element having typically been obtained by semantic analysis of a message having previously been sent to the recipient, typically during a learning phase . Each EP profile element can therefore be a lexical unit such as a word or a group of words, for example a named entity.

More precisely, each EP profile element obtained by semantic analysis of a previous message sent to the recipient can be added to the EEP set of profile elements after several occurrences of this EP profile element in the previous message or messages sent to the recipient, or on its first occurrence. In addition, each EP profile element can be added to the EEP set of profile elements after user validation or without user validation. User validation is especially useful when the profile item is confidential data.

The set EEP of conversation profile elements thus forms a conversation history grouping together key words or standard expressions generally used when the user of the terminal 110 sends an electronic message to this recipient.

The EEP set of profile elements is stored in association with an IDE identifier of the sender and the IDD identifier of the recipient, in a memory of the terminal 120 (for example in the rewritable non-volatile memory 204) or in the server 150. Alternatively, the EEP set of profile items is stored in association with the recipient's IDD identifier.

Alternatively, the EEP set of profile items is stored in association with the sender's IDE identifier, the recipient's IDD identifier, and an EPS profile item marked as the subject or subject complement. a phrase, typically corresponding to a person's name. An EP profile element obtained by semantic analysis of a previous message is then added to the EEP set only if the previous message also includes the EPS profile element marked as being a subject or a subject complement. Thus, it is possible to find the EEP set of profile elements associated with a sender and a recipient communicating about a particular EP profile element.

A sender/receiver pair can therefore be associated with several EEP sets of profile elements, each EEP set of profile elements also being associated with a different EPS profile element.

Storing the EEP set in the terminal 120 is advantageous when the EP profile elements are sensitive or confidential data, while storing the EEP set in the server 150 frees up the memory space of the terminal 110.

The EEP set of profile elements is typically stored in a data structure SD, such as a database. The data structure SD can thus group together several sets EEP1-EEPn of profile elements, each set EEP1-EEPn being associated with a different recipient, a different sender/recipient couple, or a tripled sender/recipient/marked profile element, and therefore each set being stored in association with the identifier IDD of the recipient, the pair of sender/recipient identifiers IDE/IDD, or the triplet sender/recipient/profile element marked IDE/IDD/EPS (see FIG. 4 ).

One or more EP profile items from the EEP set may additionally be stored in association with a marker indicating that the associated EP profile item is at risk. For example, the name of a company executive or an EP profile element associated with a confidential notion can be considered to be a risky EP profile element.

The calculation step E340 includes a sub-step E342 of obtaining the set EEP of conversation profile elements associated with the identifier IDD corresponding to the recipient. The EEP set of conversation profile elements can be obtained from the memory of the terminal 120 using the IDD identifier corresponding to the recipient and/or the EPS profile element marked as being a subject or a subject complement. Alternatively, a get request, comprising the IDD identifier corresponding to the recipient, the IDE identifier corresponding to the sender and/or the EPS profile element marked as being a subject or a subject complement can be transmitted intended for the server 150, the EEP set then being sent by the server 150 and then received by the calculation module 116 of the filtering device 110, via the terminal 120 and the network 140.

Then, in a sub-step E344 of the calculation step E340, the risk coefficient CR is calculated according to the result of a search for at least one context element EC obtained in the step E320 of obtaining in the set EEP of conversation profile elements obtained in sub-step E342.

Each context element EC obtained in step E320 can be searched for in sub-step E344. As a variant, only part of these context elements EC is searched. For example, only the context element(s) EC marked as being the subject or an object complement of a sentence at sub-step E325 are searched for.

The search for an EC context element can be performed by searching the EEP set for an EP profile element identical to or corresponding to the EC context element.

When the EP profile element and the EC context element are named entities, they can match when the character string of the EC context element is identical to the character string of the EP profile element, when these character strings are close, for example when they differ by one or a few characters (the match is then sought via so-called elastic distance measurements, which tolerate small differences between the two strings), and/or when the EP profile element is associated with a marker indicating that this EP profile element is an equivalent of the EC context element (for example in the case of a name, a first name and/or a nickname ), which makes it possible to compare the context element EC with all its variants. Moreover, the presence of capital letters does not impact the equivalence between the two character strings.

A first correlation coefficient CC1 can then be calculated as a function of the search, the first correlation coefficient CC1 being equal to a value included in an interval of correlation values, comprising a minimum correlation value and a maximum correlation value, by example an interval between 0 and 1.

In the case where no context element EC sought is found in the set EEP, the first correlation coefficient CC1 is equal to the minimum correlation value, therefore for example equal to 0.

In the case where all the context elements EC are found in the set EEP, the first correlation coefficient CC1 is equal to the maximum correlation value, therefore for example equal to 1.

The first correlation coefficient CC1 is for example calculated according to the following formula:
CC1 = n2/n1,
where CC1 is the first correlation coefficient, n1 is the number of context elements EC obtained at step E320, and n2 is the number of context elements EC found in the set EEP.

In the case where a context element EC marked as being the subject or the object complement of a sentence is found in the set EEP, the first correlation coefficient CC1 is high, that is to say above above the median value of the interval, for example equal to 0.8 when the median value of the interval is 0.5 (for an interval between 0 and 1). Indeed, if a context element EC marked as subject or object complement of a sentence of the message MS is found in the set EEP, this means that the user of the terminal 120 (i.e. say the sender or the sender) and the recipient of the message are in the habit of discussing this element of context EC. It is therefore likely that there is no recipient error.

For example, the first correlation coefficient CC1 can then be calculated according to the following formula:
CC1 = m + m (n2/n1),
where CC1 is the first correlation coefficient CC1, m is the median value of the interval (for example equal to 0.5), n1 is the number of context elements EC obtained in step E320, and n2 is the number of EC context elements found in the EEP set.

A high value of the first correlation coefficient CC1 calculated in sub-step E344 signifies that the context element(s) EC conform to the profile element(s) EP of the set EEP, and therefore that the content of the text of the message MS corresponds to the conversation history between the user and the recipient of the MS message. In other words, this means that the user and the recipient of the message MS have already exchanged about the context element(s) EC, and therefore there is probably no recipient error.

On the contrary, a low value of the first correlation coefficient CC1 calculated in sub-step E344 can mean that the context element(s) EC does not correspond to the profile element(s) EP of the set EEP, and therefore that the content of the MS message text does not match conversation history between user and MS message recipient. In other words, this means that the user and the recipient of the message MS are not used to exchanging about the context element(s) EC, and therefore there could be a recipient error. .

The risk coefficient CR can then be calculated as a function of the first correlation coefficient CC1, so that if the first correlation coefficient is lower than the median value of the interval, the risk coefficient CR is higher than a threshold value of risk SR, described in more detail below, and such that if the first correlation coefficient is greater than the median value of the interval, the risk coefficient CR is less than the risk threshold value SR.

For example, the risk coefficient CR can be calculated from the following formula:
CR = (1-CC1)/(1-m)*SR,
where CR is the risk coefficient, CC1 is the first correlation coefficient, m is the median value of the interval and SR is the risk threshold value.

The risk coefficient can also be calculated according to the result of a search for each IDD identifier corresponding to a recipient obtained in step E330 in a set of recipient EID identifiers (sub-step E345), so as to obtain a second correlation coefficient CC2, the second correlation coefficient being typically equal to a value included in an interval comprising a minimum correlation value and a maximum correlation value, for example an interval between 0 and 1. The set of identifiers EID can be associated with an EC context element obtained in step E320, typically an EC context element marked as being a subject or an object complement. This search makes it possible to determine whether the user is in the habit of exchanging about the context element EC with the recipient.

More precisely, in the case where all the identifiers IDD are found in the set of identifiers EID, the second correlation coefficient CC2 can take on a high value, that is to say greater than the median value of the interval (which is for example equal to 0.5). In the case where an identifier IDD obtained in step E330 is not found, the second correlation coefficient CC2 can take a lower value, that is to say lower than the median value of the interval.

The second correlation coefficient CC2 can be calculated from the following formula:
CC2 = ABS((n4/n3) – m/2),
where CC2 is the second correlation coefficient, ABS means absolute value, n3 is the number of IDD identifiers obtained in step E330, and n4 is the number of IDD identifiers found in the set of EID identifiers and m is the median value of the interval of correlation values.

The risk coefficient can then be calculated based on the second correlation coefficient CC2, such that if the second correlation coefficient CC2 is less than the median value of the interval, the risk coefficient CR is greater than the threshold value of risk SR, and so that if the second correlation coefficient CC2 is higher than the median value of the interval, the risk coefficient CR is lower than the risk threshold value SR.

The risk coefficient CR can be calculated from the following formula:
CR = (1-CC2)/(1-m)*SR,
where CR is the risk coefficient, CC2 is the second correlation coefficient, m is the median value of the interval (for example equal to 0.5 when the interval is between 0 and 1) and SR is the threshold value of risk.

Of course, the risk coefficient CR can be calculated by combining steps E342, E344 and E345. For example, the sub-steps E342 and E344 are implemented in order to calculate a first correlation coefficient CC1, the sub-step E345 is implemented in order to calculate a second coefficient CC2, a third coefficient CC3 is obtained, the third coefficient CC3 being equal to the smallest correlation coefficient between the first correlation coefficient CC1 and the second correlation coefficient CC2:
CC3 = min(CC1; CC2).

The risk coefficient CR can then be calculated from the following formula:
CR = (1-CC3)/(1-m)*SR,
where CR is the risk coefficient, CC3 is the third correlation coefficient, m is the median value of the interval of correlation values, and SR is the risk threshold value.

The risk coefficient CR can also be calculated according to the result of a comparison between each context element EC or part of the context elements EC obtained in step E320, and each identifier IDD corresponding to a recipient of the message MS ( under step E346).

More specifically, each context element EC marked as subject or object complement of a sentence of the message MS can be searched among the identifier(s) IDD obtained in step E330 for obtaining an identifier. If at least one context element EC corresponds to an identifier IDD obtained in step E330 for obtaining an identifier, the risk coefficient CR can take on a high value, that is to say greater than the threshold value of SR risk. Indeed, putting the subject or an object of a sentence as the recipient (i.e. the person we are talking about) is a classic recipient error.

We consider, for example, that the user writes in the body of a message "Benjamin has started proofreading the document", while the recipient field includes an identifier corresponding to a Benjamin. In this situation, the subject of the sentence, Benjamin, should not be the recipient of the message, because Benjamin already knows that he has started proofreading the document.

This type of careless error can be even more embarrassing when the body text of the message includes a value judgment, such as "Severine found X ineffective", while the recipient field includes an identifier corresponding to Severine. The message cannot be sent to Séverine, because it relates to one of these judgments. Also, Séverine might not want her judgment to be propagated.

If no context element EC corresponds to the identifiers IDD, the error risk coefficient CR can take a low value, that is to say lower than the risk threshold value SR.

“Matches” means that the IDD identifier and the EC context element are identical or equivalent. The identifier IDD and the context element EC are identical when the character string of the context element EC is identical to the character string of the identifier IDD (the correspondence is then sought via so-called elastic distance measurements , which tolerate small differences between the two strings). The IDD identifier and the EC context element can be equivalent when the character string of the IDD identifier differs by one or a few characters, or when the EC context element is marked as being an equivalent of the IDD identifier. Moreover, the presence of capital letters does not impact the equivalence between the two character strings.

In this sub-step E346, the position of the compared EC context element can be used in order to assess even more precisely the risk of recipient error. Indeed, the user can name the recipient in the text of the body of the MS message, in particular at the beginning to greet him, or at the end. The marker indicating the position of the context element EC can therefore be used, so that if at least one context element EC, associated with a marker indicating that the context element EC is positioned in the middle of the message MS or the part of the message MS (hereafter middle marker) corresponds to an identifier IDD obtained in step E330, the risk coefficient CR can take a value greater than the risk threshold value SR. In the other cases, that is to say if no EC context element associated with a medium marker corresponds to an identifier IDD obtained in step E330, if an EC context element associated with a marker indicating that the 'EC context element is positioned at the start or at the end corresponds to an IDD identifier obtained in step E330, and/or if no EC context element associated with a marker indicating that the EC context element is positioned at the start or at the end does not correspond to an identifier IDD obtained in step E330, the risk coefficient CR can take a value lower than the risk threshold value SR.

Of course, the risk coefficient CR can be calculated by combining steps E342-E344 and/or E345, and E346. For example, the sub-steps E342-E344 and/or E345 can be implemented in order to calculate the correlation coefficient CC1, CC2 and/or CC3. Then, the sub-step E346 is implemented, so that the correlation coefficient CC1, CC2 and/or CC3 calculated previously is modified in the event of a match.

More precisely, the correlation coefficient CC1, CC2 and/or CC3 can be reduced if a context element EC marked as subject or object complement of a sentence of the message MS corresponds to an identifier IDD corresponding to a recipient.

The correlation coefficient CC1, CC2 and/or CC3 is for example divided by a constant K1, for example between 1 and 20 and typically equal to 10, in order to obtain a fourth correlation coefficient CC4. The fourth correlation coefficient CC4 is equal to a value included in an interval of correlation values, comprising a minimum correlation value and a maximum correlation value.

If no context element EC corresponds to an identifier IDD corresponding to a recipient, the correlation coefficient CC1, CC2 and/or CC3 calculated in sub-step E344 and/or E345 is not modified.

The constant K1 is typically configurable according to the degree of sensitivity to the risk sought.

The risk coefficient CR is then calculated from the modified correlation coefficient CC4 or not CC1, CC2 and/or CC3 following the implementation of sub-step E346. The risk coefficient CR can for example be calculated from the following formula:

CR = (1-CCX)/(1-m)*SR,

where CR is the risk coefficient, CCX is the correlation coefficient CC1, CC2, CC3 and/or CC4 obtained at the end of the implementation of step S346, m is the median value of the interval of values of correlation (for example m=0.5), and SR is the risk threshold value.

Consider, for example, two Human Resources managers A and B who, in the context of their professional activity, are used to discussing the various employees of the company.

Managers A and B regularly discuss an employee P, citing her in their discussions. Before each sending of a message from manager A to manager B (or vice versa), the method according to the invention is implemented in order to check the consistency of the text that manager A (or manager B) wants to send to manager B ( manager A), in order to determine a risk of recipient error.

A semantic analysis is thus carried out for each message. It is then determined that employee P is the subject of certain sentences of the message, such as "P finds Mrs. Y very competent". This learning phase therefore makes it possible to learn a conversation profile between manager A and manager B, when manager A and manager B talk about employee P as the subject of their sentences.

A set of conversation profile elements is thus created, this set being associated with manager A, manager B and employee P. Other sets of conversation profile elements can be determined during this phase. learning for other people/subjects Q, A, S T etc. Each set is stored in a data structure.

In a use phase, manager A wants to write a message to manager B, such as an email about employee P in which he writes "P is doing too many project meetings" or "P is finding his incompetent Z project and wrote it”. These sentences correspond to value judgments on P, they are therefore sensitive.

A new semantic analysis is then performed (step E320) in order to obtain context elements or named entities of the message (including employee P, subject of the sentence), corresponding to the exchange context of the message.

The set of profile elements associated with manager A, manager B and employee P can then be obtained (sub-step E342) then the context elements obtained in sub-step E320 can be compared with the profile elements of the whole, in order to calculate the correlation coefficient (sub-step E344). The context element “employee P” being the subject of a sentence and being found in the set, the correlation coefficient takes on a high value, for example 0.8. The formula CC1 = m + m (n2/n1) or CC1 = n2/n1, described above, can for example be used to obtain this value.

The “employee P” context element is then searched for (under step E346) among the identifiers corresponding to the recipients entered in the recipient field of the message.

If manager A makes a mistake and puts employee P in the recipient field of the message, the “employee P” context element is found among the identifiers. The correlation coefficient is then divided by the constant K1=10, and then becomes low (0.08). The risk coefficient deduced from the correlation coefficient is then high.

The risk coefficient can also be calculated as a function of MQR risk markers, such as an MQR marker associated with the message MS, indicating that the message MS is at risk, and/or an MQR marker indicating that an element of EP profile is at risk.

More specifically, step E340 may include a sub-step E348 in which it is determined whether the message MS is associated with a risk marker MQR, and/or whether an EC context element corresponds to an EP profile element stored in association with an MQR risk marker indicating that the EP profile element is at risk.

The risk marker MQR associated with the message MS indicates for example that the message MS is confidential. The MQR risk marker can be automatically associated with the MS message when the sender or the recipient exercises a profession in which confidential data is handled.

If this is the case, the error risk coefficient can take on a high value, that is to say greater than the risk threshold value SR. Indeed, the message MS then has a great chance of containing confidential data, and a recipient error can therefore be particularly troublesome.

Of course, the risk coefficient can be calculated by combining the steps E342-E344 and/or E345, and E348, or even the steps E342-E344 and/or E345, E346 and E348.

For example, the sub-steps E342-E344, E345 and/or E346 can be implemented in order to calculate a correlation coefficient CC1, CC2, CC3 and/or CC4. Then, the sub-step E348 is implemented, so that the correlation coefficient CC1, CC2, CC3 and/or CC4 calculated previously is modified when the message MS is associated with the risk marker MQR or if a context element EC corresponds to an EP profile item stored in association with a risk marker MQR indicating that the EP profile item is at risk. More precisely, the correlation coefficient CC1, CC2, CC3 and/or CC4 can then be reduced, for example by being divided by a constant K2, for example between 1 and 20 and typically equal to 10, in order to obtain a fifth correlation coefficient CC5. The fifth correlation coefficient CC5 is equal to a value included in an interval of correlation values, comprising a minimum correlation value and a maximum correlation value.

When the message MS is not associated with the risk marker MQR, the correlation coefficient CC1, CC2, CC3 and/or CC4 calculated in step E344, E345 and/or E346 is not modified. The constant K2 is typically configurable according to the degree of sensitivity to the risk sought.

The risk coefficient CR can then be calculated based on the correlation coefficient CC1, CC2, CC3, CC4 and/or CC5, so that if the correlation coefficient is lower than the median value of the interval of correlation values, the risk coefficient CR is higher than the risk threshold value SR, and so that if the correlation coefficient is higher than the median value of the interval, the risk coefficient CR is lower than the risk threshold value SR.

As before, the risk coefficient CR can for example be calculated from the following formula:
CR = (1-CCX)/(1-m)*SR,
where CR is the risk coefficient, CCX is the correlation coefficient CC1, CC2, CC3, CC4 and/or CC5 obtained at the end of the implementation of step S348, m is the median value of the interval and SR is the risk threshold value.

As a variant, the sub-step E345 can be implemented only if a risk marker MQR is associated with the message MS.

We consider, for example, that managers A and B discuss the recruitment strategy for the next 5 years that the CEO of a large company wishes to pursue. It is essential to carefully check the recipients of this type of message so as not to send them to a person not concerned by this strategic information, such as an employee Z, for example.

Similar to the previous example for managers A and B, in a learning phase, a set of conversation profile items associated with the CEO is created, where the CEO is a profile item that can be associated with a marker indicating that this profile item is at risk.

After the learning phase, during a usage phase, manager A writes a message to manager B explaining the CEO's strategy to reduce R&D staff, then inadvertently puts employee Z in copy of this message .

Since the sender of this message is a human resources manager, the message is associated with an MQR risk marker. Steps and sub-steps E320, E330, E348 and E345 can then be implemented. Indeed, in this case, it is the message as a whole that is sensitive to all recipients.

More precisely, in the present case, it is determined in step E348 that the message is associated with a risk marker MQR. The step E345 of searching for the identifiers corresponding to the recipients in a set of identifiers corresponding to the “CEO” context element is then performed.

Since the identifier corresponding to employee Z is not found in the set of identifiers, the value of the correlation coefficient resulting from the implementation of step E345 is 0.2. In addition, the context element “PDG” corresponding to a profile element associated with a marker indicating that this profile element is at risk, and, the message being associated with a risk marker, the correlation coefficient is divided by the constant K2, so as to obtain a new correlation coefficient of 0.02. The risk coefficient CR is therefore high, i.e. greater than the risk threshold value SR.

The constant K2 thus makes it possible to increase the effectiveness of detecting an inconsistency before sending the message.

When several recipient identifiers are obtained in step E330, the sub-step(s) E342, E344, E345, E346 and/or E348 can be implemented for each recipient identifier, so as to calculate a risk coefficient per recipient . The risk coefficient with the maximum value is then selected for the implementation of step E350 below.

As a variant, the filtering device 110 sends the context element(s) EC and the identifier(s) IDD to the server 150 via the terminal 120, the server 150 then calculating the recipient error risk coefficient CR by implementing the under steps E342, E344, E345, E346 and/or E348, then sending this risk coefficient to the filtering device 110.

In a step E350, the filtering module 118 filters the message MS according to the risk coefficient CR obtained in step E340.

The filtering step E350 typically includes a sub-step E352 of comparing the risk coefficient to the risk threshold value SR.

The risk threshold value SR may depend on one or more pieces of information concerning the sender and/or the recipient of the message, such as for example the profession exercised by the sender and/or the recipient. In addition, the SR risk threshold value can be set by the user. Alternatively, the risk threshold value SR can be automatically configured, depending on user actions, for example in response to a previous alert issued during a previous implementation of the sub-step E356 described below .

When the risk coefficient CR is lower than the risk threshold value SR, the filtering module 118 sends the message, via the terminal 120 (substep E354).

When the risk coefficient CR is greater than the risk threshold value SR, the filtering module 118 blocks the sending of the message and sends an alert AL to the user (sub-step E356). The AL alert can be visual and/or audible, and typically indicates that the risk of recipient error is high.

The alert AL may further comprise a request for validation, by the user, of the recipient(s) of the message MS. The user can then validate or invalidate the recipient(s) of the message MS, by performing an action on the terminal 120. If the user validates the recipient(s) of the message MS, a validation command CV, also called a send command of the message MS, is received by the filtering module 118 (sub-step E357), which then sends the message MS, via the terminal 120 (sub-step E358).

If the user invalidates the recipient(s) of the MS message, the sending of the MS message remains blocked by the filter module 118 (sub-step E359). The filtering module 118 can then transmit, to the user, a request for correction of the recipient or recipients. Steps E310, E320, E330, E340 and E350 can then be repeated in order to again verify a risk of recipient error.

Steps E310, E320, E330, E340 and E350 are typically implemented during a so-called process use phase.

Claims (14)

Method for filtering a message (MS) to be sent to a recipient, said method comprising the following steps implemented by a filtering device (110):
- obtaining (E320) at least one context element (EC) for sending said message (MS), by semantic analysis of at least part of said message (MS),
- obtaining (E330) an identifier (IDD) corresponding to the recipient of the message (MS),
- filtering (E350) of the message (MS) as a function of said at least one context element (EC) and of said identifier (IDD). Method according to claim 1, further comprising a step of calculating (E340) a recipient error risk coefficient (CR) as a function of said at least one context element (EC) and of said identifier (IDD), implemented by the filtering device (110) or a server (150),
the filtering (E350) of the message (MS) being carried out as a function of said recipient error risk coefficient (CR). Method according to claim 1 or 2, implemented upon receipt (E310) of a command (CMD) to send said message (MS). Method according to claim 2 or 3, in which the filtering step (E350) comprises:
- the blocking of the sending of the message and the transmission (E356) of an alert intended for a user of the filtering device (110) when the risk coefficient (CR) is greater than a risk threshold value ( SR),
- the sending (E354, E358) of the message when the risk coefficient (CR) is lower than the risk threshold value (SR), or following the reception (E357) of a validation command after the transmission ( E356) of the alert. Method according to any one of Claims 2 to 4, in which the step of calculating (E340) the risk coefficient comprises the following sub-steps:
- obtaining (E342) a set (EEP) of conversation profile elements, associated with the identifier (IDD) corresponding to the recipient and comprising at least one profile element (EP),
- calculation of the risk coefficient (CR) according to the result of a search (E344) of said at least one context element (EC) in the set (EEP) of conversation profile elements. Method according to Claims 2 to 5, in which the risk coefficient (CR) is calculated according to the result of a search (E345) of the identifier (IDD) corresponding to the recipient in a set of identifiers (EID) of recipients associated with said at least one context element (EC). Method according to any one of Claims 2 to 6, in which the risk coefficient (CR) is calculated according to the result of a comparison (E346) between the said at least one context element (EC) and the identifier ( IDD) corresponding to the recipient. Method according to any one of Claims 2 to 7, in which the said message (MS) is associated with a marker (MQR) indicating that the message (MS) is at risk, the risk coefficient (CR) being calculated (E348) as a function of said marker (MQR). Method according to any one of Claims 2 to 8, in which the said at least one context element (EC) is associated with at least one marker, each marker indicating:
- the syntactic function, in a sentence of said at least part of the message (MS), of the associated context element (EC), or
- a position of the associated context element (EC) in said at least part of the message (MS), or
- that the associated context element (EC) is at risk, said marker then being determined from at least one other context element (EC),
the risk coefficient (CR) being calculated as a function of said at least one marker. Method according to Claim 5 and Claim 9, in which, when the said at least one marker indicates that the syntactic function of the said context element (EC) is a subject or an object complement, the set (EEP) of elements conversation profile is obtained from the identifier (IDD) corresponding to the recipient and said at least one context element (EC) associated with said marker. Method according to Claim 7 and Claim 9, in which the said at least one context element (EC) and the identifier (IDD) corresponding to the recipient are compared when the said at least one marker indicates that the syntactic function of the said at least one element of context (EC) is a subject or an object complement. Filtering device (110) capable of filtering a message (MS) to be sent to a recipient, said filtering device (110) comprising:
- a module for obtaining (112) at least one context element (EC) for sending said message (MS), by semantic analysis of at least part of said message (MS),
- a module for obtaining (114) an identifier (IDD) corresponding to the recipient of the message,
- a filtering module (118) of the message (MS) according to said at least one context element (EC) and said identifier (IDD). Computer program (P1) comprising instructions for the execution of the steps of the method according to any one of Claims 1 to 11 when the said program is executed by a computer. Computer-readable recording medium on which is recorded a computer program (P1) comprising instructions for carrying out the steps of the method according to any one of Claims 1 to 11.