Method for automatically generating a response to a request contained in an electronic message

Abstract

The invention relates to a computer-implemented method (100) for automatically generating at least one first response (R1) to at least one request (Q1) contained in an electronic message (M1), the method comprising: ■ receiving (REC) the electronic message (M1); ■ extracting (EXT1) a first dataset (ENS1) and a second dataset (ENS2) from the message (M1), so as to generate an input vector (VE) from the second dataset (ENS2); ■ automatically detecting (DET) the at least one first request (Q1) by implementing an algorithm; ■ extracting (EXT2) a list of at least one digital document of interest ({Di}i) by implementing a first learned function (FA1), the first learned function (FA1 receiving the input vector (VE) as input; ■ implementing (IMP) a second learned function (FA2) so as to generate the first response (R1) to the at least one request (Q1).

Description

[0001] METHOD FOR AUTOMATICALLY GENERATING A RESPONSE TO A REQUEST CONTAINED IN AN ELECTRONIC MESSAGE

[0002] Field of invention

[0003] The field of the invention is that of methods and systems for automatically generating responses to requests contained in electronic messages.

[0004] State of the art

[0005] There are a variety of electronic document management (EDM) solutions available for managing information in the form of electronic documents within organizations.

[0006] Since the amount of information managed by EDM software can be large, it can sometimes be difficult and time-consuming to search for specific information among all the stored electronic documents. A typical case is, for example, when a user is looking for specific information related to an organization of interest that has an EDM.

[0007] There are also solutions for analyzing digital documents based on machine learning models. State-of-the-art solutions generally allow the extraction of either data of interest relating to the type of document in question in order to classify it, or data of interest to exploit the document in a certain way by means of operations such as anonymization, verifications on named entities, etc.

[0008] One of the aims of the invention is to propose a solution facilitating the search by a user for precise information within information stored in an EDM, which is simple and quick to implement.

[0009] Summary of the invention

[0010] According to one aspect, the invention relates to a computer-implemented method for automatically generating at least one first response, encoded in a plurality of discrete symbols in a natural language, to at least one first request contained in an electronic message, the method comprising:

[0011] - Receiving said electronic message within a messaging system, said message comprising a set of data and being associated with a first user; - Extracting a first set of data and a second set of data from said message, so as to generate an input vector from the second set of data;

[0012] - Automatic detection of said at least one first request by implementing, on the first set of data, a detection algorithm;

[0013] - Extraction of a list of at least one digital document of interest by implementing an extraction algorithm receiving input data as input;

[0014] - Implementation of a second learned function generated from a pre-trained large language model, the second learned function receiving as input a set of data from said electronic message, relating to said at least one first request and called effective set, and said digital documents of interest from said list, so as to generate said first response to the at least one first request.

[0015] By request is meant a query representative of a search for particular information, which can be formulated in the form of a question, but also of a declarative sentence formulating the query, or of a set of words representative of the query and which is encoded in a plurality of discrete symbols in a natural language.

[0016] In this patent application, an electronic message is understood to mean a communication transmitted electronically from a sender to one or more recipients via an electronic device such as a computer, a smartphone, a tablet and via messaging software that allows at least one individual to be contacted. A sender is understood to mean the entity that sends the electronically transmitted communication. The sender is typically an individual. A recipient is understood to mean an entity associated with an electronic storage space to which the communication is transmitted. The recipient may be an individual or a robot, so that in the latter case, the method according to the invention is fully automated. A learned function is understood to mean, for example, a function generated from a machine learning model trained from a training domain.In another example, a learned function is a predefined function for extracting data from a data set to generate a prediction, said data set being able to evolve over time by adding or selecting data or enriching data.

[0017] Thus, the method according to the invention makes it possible to obtain a response to a request based on the analysis of selected documents of interest, relevant to developing the response. The method makes it possible to obtain such a response with a high degree of reliability and robustness. Furthermore, an advantage is to be able to extract information from documents that do not need to be classified.

[0018] According to a first aspect, the invention relates to a computer-implemented method for automatically generating at least one first response, encoded in a plurality of discrete symbols in a natural language, to at least one first request contained in an electronic message, the method comprising:

[0019] - Receiving said electronic message within a messaging system, said message comprising a set of data and being associated with a first user;

[0020] - Extracting a first data set and a second data set from said message, so as to generate an input vector from the second data set;

[0021] - Automatic detection of said at least one first request by implementing, on the first set of data, a detection algorithm;

[0022] - Extracting a list of at least one digital document of interest by implementing a first learned function generated from a machine learning model trained from a first training domain, said first learned function receiving said input vector as input;

[0023] - Implementation of a second learned function generated from a pre-trained large language model, the second learned function receiving as input a set of data from said electronic message, relating to said at least one first request and called effective set, and said digital documents of interest from said list, so as to generate said first response to the at least one first request.

[0024] Thus, thanks to the method according to the invention, a user having a request vis-à-vis an organization of interest can automatically receive a response to his request, on the basis of a set of documents of interest automatically selected by an artificial intelligence system.

[0025] According to one embodiment, the response generated and transmitted automatically to the user includes a request, for example a question contained in an email sent to the user, asking him about missing information to respond to the first request transmitted by the user.

[0026] According to one example, the artificial intelligence system configured to process the request transmitted by the user detects that information is missing from the database to provide a satisfactory answer to the user, for example on the basis of a score compared with a predefined threshold. According to this example, the system is configured to identify the missing information to provide the answer and to generate a new request, for example a question, which will be transmitted to the user so that the latter can provide the missing information to answer his request.

[0027] Missing information may include information missing from the database, such as information missing from a document, such as a date of birth, or information missing from a document, such as an ID.

[0028] According to the preceding example, the request transmitted by the user may be a new request, or a response from the user to a first request for which the system has already provided a response to the user. This is for example a case in which the user sends a new request and / or new documents as attachments in response to the response automatically transmitted by the system to his first request. According to one embodiment, the system is configured to extract information contained in a request from a user, whether it is a first request or a subsequent request in response to a response already transmitted by the system, and to compare said information with information contained in the database, and to generate a response based on said comparison.

[0029] In an illustrative example, the system receives an email from a user including an identity card on which a date of birth appears, and the system identifies in the database that the date of birth does not match that appearing on the identity card submitted by the user. In this example, the system is configured to automatically generate a response informing the user of this inconsistency, and requesting the user to respond in turn, for example by providing justification or a new correct document.

[0030] Thus, according to the preceding embodiments and examples, the artificial intelligence system is configured to generate at least one task following the processing of the request transmitted by the user. The task(s) include, for example, sending a new request to the user, or generating a suggestion for an email to be transmitted by an operator to the user, who can therefore enrich or correct the suggested email, or a task indicating to an operator an action to be carried out, for example a modification of documents in the database.

[0031] According to one embodiment, the response transmitted by the system of the invention to the user in response to the first request, or to any other subsequent request, is transmitted by email to the user.

[0032] According to one embodiment, the response transmitted by the system of the invention in response to the first request, or to any other subsequent request, is transmitted to a third-party system comprising a user interface with which the user can interact, such as an application, a web platform, or any other system allowing the user to receive and read the response. The transmitted response may comprise a notification or an alert informing the user of an action to be taken, for example to provide missing information or documents, or to indicate to the user that his request has been processed or is being processed. According to this example, the user responds to the notification or alert either via the third-party system or by email.

[0033] According to one embodiment, the first request and / or any subsequent request, and / or any alert or notification, is automatically transmitted to a third-party user other than the first user who originated the first request, for example an insurance broker. In this case, the system of the invention operates as an interface between the first user who originated the first request and the other user.

[0034] According to one embodiment, the system is configured to automatically generate an automatically generated response justification and / or response suggestion and / or alert and / or notification in response to a user's first request.

[0035] In one embodiment, the system is configured to automatically generate and display a chat window on a third-party system including a display, such as a web platform or an application. In this example, a user, such as the user who initiated the first request, or a third-party user, such as an insurance broker, may interact to process any request from the user who initiated the request. It may also be a chatbot configured to automatically provide information to the user in real time about their requests, or to ask questions or suggest actions to the user to respond to their request.

[0036] According to one embodiment, the first data set and the second data set may comprise common data. The first data set also comprises at least one first data item that does not belong to the second data set and the second data set comprises at least one second data item that does not belong to the first data set.

[0037] In another embodiment, the first data set and the second data set are disjoint.

[0038] Thus, extracting the first data set and the second data set makes it possible to provide a more condensed volume of data as input to the algorithm for detecting the at least one first request and also to use a more condensed volume of data to generate the input vector. The method can then be implemented in a faster, more efficient and more robust manner.

[0039] Advantageously, the method further comprises generating a context from the first data set, and wherein, when implementing said second learned function, said second learned function also receives said context as input.

[0040] Thus, the second learned function receiving a complementary context relating to the user's request, the generated response is all the more precise and reliable.

[0041] Advantageously, the generation of the context can be implemented from the first set of data and at least one digital document of interest.

[0042] Advantageously, the method further comprises a step of extracting a third data set configured to determine rules for extracting the first data set and / or the second data set.

[0043] Advantageously, the third data set is configured to determine rules for extracting the first data set.

[0044] Thus, extracting the third data set makes the step of detecting at least one first request more efficient.

[0045] Advantageously, the third data set is configured to determine rules for extracting the second data set.

[0046] Thus, extracting the third dataset makes the input vector generation step more efficient.

[0047] Advantageously, the third data set is configured to determine rules for extracting the first data set and the second data set.

[0048] Thus, extracting the third data set makes the step of detecting the at least one first request and the step of generating the input vector more efficient.

[0049] According to another aspect, the invention relates to a computer-implemented method for automatically generating at least a first response, encoded in a plurality of discrete symbols in a natural language, to at least a first request contained in an electronic message, the method comprising: - Receiving said electronic message within a messaging service, said message comprising a set of data and being associated with a first user;

[0050] - Extracting a third data set configured to determine rules for extracting a first data set and a second data set

[0051] - Extracting said first data set and said second data set from said message, so as to generate an input vector from the second data set;

[0052] - Automatic detection of said at least one first request by implementing, on the first set of data, a detection algorithm;

[0053] - Extraction of a list of at least one digital document of interest by implementing an extraction algorithm receiving input data as input;

[0054] - Implementation of a second learned function generated from a pre-trained large language model, the second learned function receiving as input a set of data from said electronic message, relating to said at least one first request and called effective set, and said digital documents of interest from said list, so as to generate said first response to the at least one first request.

[0055] According to another aspect, the invention relates to a computer-implemented method for automatically generating at least one first response, encoded in a plurality of discrete symbols in a natural language, to at least one first request contained in an electronic message, the method comprising:

[0056] - Receiving said electronic message within a messaging system, said message comprising a set of data and being associated with a first user;

[0057] - Extracting a third data set configured to determine rules for extracting a first data set and a second data set; - Extracting said first data set and said second data set from said message, so as to generate an input vector from the second data set;

[0058] - Automatic detection of said at least one first request by implementing, on the first set of data, a detection algorithm;

[0059] - Extracting a list of at least one digital document of interest by implementing a first learned function generated from a machine learning model trained from a first training domain, said first learned function receiving said input vector as input;

[0060] - Implementation of a second learned function generated from a pre-trained large language model, the second learned function receiving as input a set of data from said electronic message, relating to said at least one first request and called effective set, and said digital documents of interest from said list, so as to generate said first response to the at least one first request.

[0061] Advantageously, the at least one digital document of interest is contained in a corpus of digital documents stored in a collaborative workspace.

[0062] Thus, with the method according to the invention, the user receives a response resulting from a search on a corpus of documents which can represent a significant mass of resources, and this in a very short time. Advantageously, the at least one digital document of interest comprises at least one document attached to the electronic message.

[0063] Advantageously, at least one document among the at least one digital document of interest is a document having a specific structure and formalism, with predefined distinct fields and intended to collect a defined type of information.

[0064] Thus, the method according to the invention makes it possible in particular to process business documents, which may contain a specific and local lexical field. Advantageously, the effective set is the set of data included in said electronic message.

[0065] Thus, in practice, the method according to the invention is capable of automatically processing an electronic message containing a request and generating a corresponding response in return.

[0066] Advantageously, the method further comprises a step of encoding the at least one first request into a prompt formed from another plurality of discrete symbols in a natural language, and in which said effective set comprises the prompt.

[0067] Thus, the method according to the invention makes it possible to generate in an intermediate manner a prompt which can represent an augmented version, i.e. contextualized and more precise, thus increasing the chances of generating an automatic response which is as appropriate as possible for the user who sent his request in the electronic message.

[0068] According to one embodiment, the step of encoding the at least one first request into the prompt is carried out from said at least one first request, a history of errors and correct responses obtained during previous implementations of the method according to the invention.

[0069] Thus, advantageously, the performance of the method for generating a correct response to at least one first request is improved as the method according to the invention is implemented through the use of the history.

[0070] Advantageously, the detection algorithm is the second function learned.

[0071] Advantageously, the detection algorithm comprises a third learned function generated from a natural language processing model trained from a second training domain.

[0072] Advantageously, the step of implementing the second learned function also makes it possible to generate at least one reference to a passage of at least one digital document of interest from said list of at least one digital document of interest, said passage containing information relating to said first response. Thus, the method according to the invention makes it possible to provide the automatically generated response with evidence or supporting documents taken from one or more digital documents of interest from the extracted list.

[0073] Advantageously, the method further comprises:

[0074] - Acquisition of at least one additional data item defined by a second user by means of a user interface, and the second learned function also receives as input the at least one additional data item.

[0075] Thus, the method according to the invention may be guided by additional information. This additional information may be new documents, or information comprising a clarification or correction of information. Advantageously, the second user is the first user. In another example, the second user is another user different from the first user.

[0076] Advantageously, the method further comprises:

[0077] - Transmission of said at least one first response and, where applicable, of said at least one reference, so that said at least one first response and, where applicable, said at least one reference are made accessible to said first user.

[0078] Advantageously:

[0079] - the transmission is a transmission of a second electronic message to a messaging system of said first user,

[0080] - at least one reference is, where appropriate, transmitted in the form of a hyperlink linked to an address referring to a memory location containing said passage from at least one digital document of interest from said list.

[0081] In this way, the user receives the automatic response to his request directly, through the same communication channel he used to send his request.

[0082] Advantageously:

[0083] - said at least one first request comprises a plurality of requests,

[0084] - said at least one first response comprises a plurality of responses corresponding to each of the requests among the plurality of requests. Advantageously, the extraction step is implemented by implementing a natural language automatic processing model on the data set.

[0085] Advantageously, the method further comprises a step of acquiring at least one additional data item defined by a second user by means of a user interface, and the second learned function also receives as input the at least one additional data item.

[0086] For example, the at least one additional data can be an additional document.

[0087] Advantageously, the method comprises, prior to receiving the electronic message:

[0088] - Segmentation of at least one original electronic document so as to obtain a plurality of individual electronic documents, each individual electronic document corresponding to an individual electronic message, said electronic message being one of the individual electronic messages.

[0089] According to a second aspect, the invention also relates to a system comprising an electronic terminal of a user comprising at least one computer, a memory, a display and a communication interface for transmitting messages over a data network to at least one first server comprising hardware resources for executing the first learned function, the second learned function, the third learned function and a memory for recording the models, which generated these learned functions, in order to execute the steps of the method of the invention.

[0090] By "in order to execute the steps of the method of the invention" is meant that the system is configured to automatically implement the steps of the method of the invention.

[0091] Brief description of the figures

[0092] Other characteristics and advantages of the invention will emerge on reading the detailed description which follows, with reference to the appended figures, which illustrate: Fig. 1: an example of steps which can be implemented to generate at least one response to at least one request contained in an electronic message according to the method of the invention;

[0093] Fig. 2: a system of the invention comprising different resources for implementing the method of the invention;

[0094] Fig. 3: An example of a structured document model that can be used to train a machine learning model generating a learned function used in the method according to the invention;

[0095] Fig. 4: An example of a flowchart illustrating an implementation of a large language model used in a step of the method according to the invention.

[0096] The invention relates to a method 100 for automatically generating at least one response Ri to at least one request Qi contained in an electronic message Mi and a system 200 for implementing the method 100.

[0097] Figure 1 shows an embodiment of a set of steps that can be performed to implement the method 100 and will be described in detail below.

[0098] Figure 2 represents an example of a set of elements of the system 200 of the invention making it possible to implement the method 100 of the invention. Figure 2 represents more particularly a data network NETi which can be the internet network. A first user terminal Ti allows access to a first remote server SERVi which makes it possible to host in particular a collaborative workspace. This space includes in particular memory resources for storing received digital documents and processing them automatically. A second server SERV2 includes the means making it possible to generate models and to execute certain functions and algorithms of the method 100 of the invention.

[0099] According to the embodiments, the models and configurations of the learned functions can be stored directly on the server hosting the collaborative workspace on which the method of the invention is executed. However, in a preferred mode, a server dedicated to their storage and execution is implemented.

[0100] The first user terminal T1 can send and receive information to and from the first remote server SERVi and to and from the second server SERV2 by means of electronic communication systems. For example, the first user terminal Ti can communicate with the first remote server SERVi and the second server SERV2 via electronic messaging allowing sending and receiving electronic mails. In another example, the first user terminal T1 can communicate with the first remote server SERVi and the second server SERV2 via instant messaging, where for example a user of the first user terminal T1 can interact with a chatbot integrated into the collaborative workspace of the first remote server SERVi or integrated on a platform hosted by the second server SERV2.

[0101] Advantageously, the collaborative workspace hosted by the first remote server SERVi stores in its memory resources digital documents received from organizations. For example, the organizations can be insurance companies, real estate agencies, social landlords, using the memory resources of the collaborative workspace for the storage of their digital documents.

[0102] The digital documents stored in the collaborative workspace are, for example, a technical specification, a user manual, an assembly manual, an accounting document such as a balance sheet, a quote, an invoice, a letter, a plan, a certificate, an official document, a medical certificate, a form, a digitized document or any other document produced by an organization, an automated process or an individual. The documents processed within the framework of the invention are preferably documents that relate to a type of document or a family of documents. The type of document can be defined by a generic name or a label allowing it to be associated with a type of document. It is understood that an invoice or a pay slip can be, in this respect, types of document in an organization as well as a technical specification of a product.

[0103] The digital documents may include various formats such as a .txt format, a .pdf format, a .png format, a Jpg format, a Json format, an .xml format, a .doc or docx format, or any other document format that can encode a plurality of discrete symbols in a natural language. In some embodiments, the digital documents may be messages from an email or instant messaging service. In one embodiment, the digital documents are image files, video files, or audio files. In one embodiment, the documents are data sets received in real time, such as a video stream or an audio stream.

[0104] One of the advantages provided by the method 100 according to the invention, as will be described below, is to automatically provide responses to user requests based on the digital documents stored in the memory resources of the collaborative workspace hosted by the first remote server SERVi and / or documents attached to an electronic message containing the request(s).

[0105] It is assumed that a first user of the first user terminal Ti presents a request Qi to an organization as exemplified previously and called the organization of interest. The organization of interest has a corpus of digital documents stored in the memory resources of the collaborative workspace hosted on the first remote server SERVi. The first user sends the request Qi within an electronic message Mi such as an email sent from an email address of the first user, to an email address of the organization of interest or a representative of the organization of interest, or an instant message sent to a member of the organization of interest. The electronic message Mi can thus be sent from a communication interface such as a network card allowing data to be sent to a remote entity such as a data server.A response Ri to the request Qi will be generated automatically according to the method 100 according to the invention and as illustrated in FIG. 1.

[0106] In one embodiment, in a reception step REC, the electronic message Mi is received within a recipient messaging service linked to the organization of interest. The recipient messaging service may be, for example, an electronic mailbox linked to the organization of interest, or an interface of a client software of an instant messaging service linked to the organization of interest. The electronic message Mi comprises a set of data ENS including the request Qi. For example, when the electronic message Mi is an electronic mail, the set of data ENS comprises metadata (or headers) and a message body. The metadata comprises, among other things, the sender, the recipient, the subject, the sending date. The message body is encoded in the form of a text in a natural language or multiple parts (for example, a text and attachments).

[0107] Once the electronic message Mi has been received, in a first extraction step EXTi, a first data set ENSi and a second data set ENS2 are extracted from the electronic message M1. The first data set ENS1 and the second data set ENS2 may comprise common data. Advantageously, the second filtering data set ENS2 is intended for the generation of an input vector VE consisting of components. It will be explained later how the input vector VE is used as input to a first learned function FA1 in order to extract a list of at least one digital document of interest {Di}i.

[0108] In some embodiments, prior to the first extraction step EXT1, a so-called preliminary extraction step makes it possible to extract a third set of data from the electronic message Mi. The third set of data is configured to define rules for extracting the first set of data ENS1 and / or the second set of data ENS2 which will be applied during the first extraction step EXT1. The third set of data may be specific to the organization of interest. For example, the third set of data may consist of so-called configuration data specific to the organization of interest.

[0109] In other embodiments, the first extraction step EXT1 is implemented by implementing a natural language processing model on the ENS dataset.

[0110] The first extraction step EXT1 is advantageously implemented by components of the first remote server SERV1 or of the second remote server SERV2. For example, the extraction step EXT1 is advantageously implemented by a processing unit of the server SERV1 or of the second remote server SERV2. By processing unit is meant an electronic component or a plurality of electronic components comprising, for example, a computer, comprising a set of at least one processor, and possibly a memory operatively coupled to the computer.

[0111] In an automatic detection step DET, the request Qi contained in the electronic message Mi is automatically detected by implementing a detection algorithm on the first data set ENSi.

[0112] According to one example, the detection algorithm detects a question mark symbol “?” in the ENS dataset. In another example, the detection algorithm may use regular expressions (or regex). In another example, the detection algorithm corresponds to a second learned function FA2 generated from a pre-trained large language model LLM and which is used in another step IMP of the method 100 according to the invention. In yet another example, the detection algorithm comprises a third learned function FA3 different from the second learned function FA2, and generated from a natural language processing model NLP trained from a second training domain.

[0113] The automatic detection step DET is advantageously implemented by components of the first remote server SERV1 or the second remote server SERV2. For example, the detection step DET is advantageously implemented by a processing unit of the first remote server SERV1 or the second remote server SERV2.

[0114] In a second extraction step EXT2, a list of at least one digital document of interest {Di}i is extracted by implementing a first learned function FA1 receiving as input the input vector VE generated from the second data set ENS2. Advantageously, the at least one digital document of interest {Di}i is contained in the corpus of digital documents available to the organization of interest and stored in the memory resources of the collaborative workspace hosted by the first remote server SERV1. Alternatively or in addition, the at least one digital document of interest {Di}i comprises at least one document attached to the electronic message M1.

[0115] The at least one digital document of interest {Di}i represents one or more documents relevant for the purpose of generating the response Ri to the request Qi. For this purpose, the components of the input vector VE comprise a plurality of input data making it possible to select the documents of interest {Di}i by a correspondence link with characteristic elements of the request Qi resulting from the second data set ENS2 which was extracted during the first extraction step EXT1. The at least one digital document of interest {Di}i is for this purpose filtered, or in other words extracted, on the basis of specific characteristics or criteria which constitute the components of the input vector VE. These criteria may be based on keywords, categories, metadata.

[0116] For example, the input vector VE may include as components a person's name, an address, an organization's name, a due date, a delivery date, or a term corresponding to an action or a descriptive characteristic.

[0117] Advantageously, the first learned function FA1 is generated from a machine learning model MOD trained from a first training domain. Examples of architecture for the MOD model are given later. The first learned function FA1 is configured to extract documents having a correspondence link with the components of the input vector VE. In certain embodiments, the first learned function FA1 is configured to find in a set of documents the documents comprising data corresponding to all or part of the components of the input vector VE.

[0118] For example, the first learned function FA1 makes it possible to automatically extract from a document, by a first prediction, a characteristic date DATE1 , such as a due date, by a second prediction, an entity named N0M1 such as an organization name, and by a third prediction, a document type TYPE1 , such as an invoice or a quote. Advantageously, the first learned function FA1 can be a combination of learned functions. For example, the first learned function FA1 can be a combination of a matching algorithm, a date extractor and a classifier. The extraction of the three characteristic data {DATE1 , N0M1 , TYPE1} makes it possible to generate a prediction based on all of the first, second and third predictions. This prediction allows a classification of the document which can lead to the extraction of this document during the extraction step EXT2.

[0119] However, according to another example, the extraction of a digital document of interest does not necessarily implement an extraction of data from said document. For example, the first learned function FA1 can extract said document based on an analysis of its name or its type.

[0120] The MOD machine learning model is for example a pre-trained language model, such as those based on neural networks, which can be used to identify any word in documents. In other embodiments, the MOD machine learning model is a classifier that has been trained on one or more annotated datasets to learn to recognize words corresponding to the annotations in new documents. In other embodiments, the MOD machine learning model is a natural language processing (NLP) model, measuring the semantic similarity between words in a document and the list of elements constituting the input vector VE. Algorithms such as Word Embeddings or BERT can be used.

[0121] For example, in the case of documents with a defined structure, such as a pay slip, a form, a typical quote from an organization, a typical energy performance diagnostic (EPD) report from an organization that include predefined distinct fields intended to collect a defined type of information, the MOD model can be a machine learning model that has been trained to recognize the fields and the value of the data present in these fields. The VE input vector advantageously includes certain fields among its components in this case. Such documents can follow a predefined model.

[0122] Figure 3 represents an example of a predefined model, in which a certain number of objects OB1, OB2, OB3 representing headings, illustrations or paragraphs can be labeled to indicate characteristic data fields. In the model of Figure 3, a first date field noted DAT1 and a second date field DAT2 representing for example respectively a mail date and a signature date. In this example, the NMi field is that of a named entity for example a name and an address of an organization receiving the document. Another field NM2 makes it possible to label another named entity such as another organization name, for example an organization publishing the document. In this example, a signature zone is indicated SIG1 in Figure 3 and represents an area where a signature is expected.

[0123] According to an exemplary embodiment, a document model such as that of Figure 3 can be used initially to generate a first training of a machine learning model such as the MOD model. According to another example, the MOD model is constructed from real documents and allows a consistency check to be carried out with predictions calculated by learning functions.

[0124] Advantageously, the first learned function FA1 is implemented by one or more components of the first server SERV1 or the second server SERV2. For example, the first learned function FA1 is advantageously implemented by a processing unit of the second server SERV2.

[0125] Once the list of at least one digital document of interest {Di}i has been extracted, a second learned function FA2 is implemented, receiving as input a set of data from the electronic message M1 and relating to the request Q1, and called the effective set ENSe, and the at least one digital document of interest {Di}i, so as to automatically generate as output a response Ri, encoded in a plurality of discrete symbols in a natural language, to the request Q1. Advantageously, in order to generate the response Ri, the second learned function FA2 also receives as input a context CONT generated from the first set of data ENS1 extracted during the first extraction step EXT1. By context, we mean a set of information (or in other words, data) available, for example past, which influences the meaning of a word, a group of words or a sentence included in the request Q1. The generation of such a context CONT will be described later.

[0126] In one embodiment, the effective set ENSe consists of all the ENS data of the electronic message M1. In another embodiment, the effective set ENSe is a subset of data included in the ENS data set of the electronic message Mi and the automatic detection step DET previously described advantageously comprises a step of encoding the detected request Qi into a prompt P formed from another plurality of discrete symbols in a natural language. By prompt, is meant a computer command in a natural language and intended to be received as input to a computer program. Advantageously, the prompt P resulting from the encoding represents an augmented version of the request Qi, that is to say, capable of including contextual information, or representing a more precise formulation of the request Qi. Advantageously, the prompt P is generated by a query engineering technique or in English “prompt engineering”.For example, the encoding may be preceded by a step of receiving details on the context CONT or on the form of information sought included in the request Qi, or else a history HIST of feedbacks FB of errors and good responses obtained during previous implementations of the method 100. In this case, the effective set ENSe advantageously includes the prompt resulting from the encoding step.

[0127] The HIST history can be in the form of a database or a file. The HIST history can include new data that is aggregated over time. Updating the HIST history with new data can be done automatically or by user action.

[0128] Figure 4 illustrates an example implementation of the second learned function FA2, receiving as input a prompt P and the digital documents of interest {Di}i. Figure 4 illustrates that the prompt P can be generated from the question Q1 and the history HIST. The second learned function FA2 generates as output an answer Ri to the question Q1. An annotation of the answer Ri, indicating for example whether it is correct or incorrect, is stored in the form of a feedback FB (correct answer or error) and returned in the history HIST.

[0129] Advantageously, the second learned function FA2 was generated from a pre-trained large language model LLM. A large language model is a natural language processing model with parameters and the ability to understand and generate natural language text. Examples of large language models that can be used include Mistral IA™, the LLaMA™ model, or the Falcon 180B™ model.

[0130] Advantageously, the second learned function FA2 is configured to generate the response Ri on the basis of the effective set ENSe and the at least one digital document of interest {Di}i, and where appropriate the context CONT. For example, the second learned function FA2 performs an analysis of the at least one digital document of interest {Di}i, so as to extract and exploit values ​​of interest from data and / or information close to or corresponding to the components of the input vector VE found in the digital documents of interest {Di}i to generate the response Ri. Among these data, it may be a text field, a date field or a number. These data may include indicators of the presence of iconography, a signature, a header or even initials, etc.This data may also include characteristic values ​​of a geometric area of ​​a page of the document, of one or more characteristic positions, or even a page number of a document, etc.

[0131] As mentioned previously, a step of generating a CONT context from the set of ENS data contained in the electronic message M1 can be carried out, so that the second learned function FA2 also receives the CONT context as input.

[0132] For example, the context CONT may comprise metadata of the electronic message M1 (such as the title of the message, a particular term in the title such as the indication “Urgent”, the recipients in copy, the date of the electronic message M1) and / or other documents such as a set of past electronic messages originating from the same sender, a set of documents relating to the same tenant.... Advantageously, the context CONT is generated by applying learned functions or combinations of learned functions similar to the first learned function FA1 (applied during the extraction step EXT2) on the first data set ENS1.

[0133] An example of a learned function that can be used to generate the context is a learned function configured to process textual data, for example to detect specific expressions or dates (start date of a service, end date of validity, etc.). The first learned function FAi can, in a different way, be configured to process vector data.

[0134] In another example, the CONT context is generated by a large language model receiving as input a prompt called a context prompt formulating a request to obtain specific data such as the name of the company concerned by the Mi electronic message, or a type of document required in the Mi electronic message. The context prompt can for example formulate a request to create a .json file with a plurality of information.

[0135] Advantageously, metadata can be assigned manually, by users, to the electronic message Mi, or more generally to any other document used or processed in the method 100 according to the invention. The users can be users of the collaborative space of the organization of interest hosted on the first remote server SERVi.

[0136] Advantageously, a request to add metadata that may include a link is automatically sent by email when information is missing within the CONT context.

[0137] Advantageously, the implementation of the second learned function FA2 also makes it possible to generate at least one reference REF at the output to a passage of at least one digital document of interest from the list of at least one digital document of interest {Di}i. The passage includes information relating to the response Ri and can serve as proof accompanying the response Ri.

[0138] In one example, the reference comprises the digital document of interest in question where the passage has been marked, for example highlighted or boxed. In another example, the reference is a hyperlink linked to an address referring to a memory location comprising the passage of the digital document of interest. In another example, the reference comprises the name of the digital document of interest as well as a page, paragraph or line number or an indication of a section of the digital document of interest. In yet another example, the reference comprises a set of data extracted from the digital document of interest, for example in the form of a table comprising data values ​​of the digital document of interest. In yet another example, the second learned function FA2 extracts from a plurality of documents of the same type or of different types several portions of documents to produce a single edited document from these portions.

[0139] Advantageously, the implementation of the second learned function FA2 also makes it possible to generate a confidence index of the response Ri generated by the second learned function FA2. The confidence index is a measure that evaluates the reliability associated with the prediction represented by the response Ri. For example, the confidence index is expressed in the form of a probability or a numerical value.

[0140] Advantageously, the second learned function FA2 is implemented by one or more components of the first server SERV1 or the second server SERV2. For example, the second learned function FA2 is advantageously implemented by a processing unit of the second server SERV2.

[0141] In order to make the generated response Ri (and where appropriate the at least one reference REF) available to the first user of the first user terminal T1, the method 100 according to the invention advantageously comprises a step of transmitting the response Ri and, where appropriate, the at least one reference REF. The transmission step is advantageously implemented by means of electronic communication systems allowing communications between the first server SERV1 (or the second server SERV2) and the first user terminal T1. Alternatively, these are electronic communication systems allowing communications between the first server SERV1 (or the second server SERV2) and a second user terminal used by a second user different from the first user who will transmit the response Ri to the first user.

[0142] In some embodiments, the transmission is a transmission of a second electronic message M2 to the user. According to one example, the second electronic message M2 comprises the response Ri and, if applicable, the at least one reference REF as generated following the implementation IMP of the second learned function FA2. In another example, the second electronic message M2 comprises a version of the response Ri modified by a second user, such as a member of the organization of interest. According to one example, the second electronic message M2 may be an email sent from the email address to which the electronic message M1 was sent to the email address from which the electronic message M1 was sent.In another example, when the electronic message M1 is an instant message, the second electronic message M2 may be an instant message sent by a member of the organization of interest or by a chatbot and responding to the electronic message M1.

[0143] When the second learned function FA2 has generated at least one reference REF to a passage of at least one digital document of interest from the list of at least one digital document of interest {Di}i, the at least one reference REF can be transmitted for example in the form of a document attached to the second electronic message M2 or a hyperlink referring to a resource of a remote entity comprising the digital document of interest or an original document created to include extracted data or an area of ​​interest of the digital document of interest and allowing opening.

[0144] In some embodiments, the method 100 according to the invention may comprise one or more additional steps.

[0145] Thus, in certain embodiments, the method 100 according to the invention advantageously comprises a step of acquisition, by a second user, by means of a user interface of a second user terminal, of at least one additional data item, before the step of implementation IMP of the second learned function FA2. In this case, during the implementation IMP of the second learned function FA2, the second learned function FA2 also receives as input the at least one additional data item.

[0146] According to one example, the at least one additional data item comprises information relating to at least one additional digital document of interest Dcomp, i.e., a digital document containing information likely to be relevant for the generation of the response Ri to the request Q1. For example, the at least one additional data item may comprise the name of the additional digital document of interest Dcomp and / or an indication relating to a passage of this additional digital document of interest Dcomp and / or the additional digital document of interest Dcomp as a whole. For example, the additional digital document of interest Dcomp is a document that is not on the list extracted during the second extraction step EXT2.In another example, the digital document of additional interest Dcomp is included in the list extracted during the second extraction step EXT2 and the at least one additional data item includes a passage or a reference to a specific passage from this digital document of additional interest Dcomp.

[0147] In certain embodiments, the method 100 according to the invention comprises, before the step of implementation IMP of the second learned function FA1 generating the response Ri, a step of preliminary implementation of the second learned function FA2. This preliminary implementation step can take place for example when there is uncertainty about the response Ri to be generated on the basis of the digital documents of interest in the list of at least one digital document of interest {Di}i. An example corresponds to the case where several digital documents of interest in the list correspond to different versions of the same initial digital document having slight differences, and for which it is necessary to know which one must be taken into account for the step of implementation IMP of the second learned function FA2.This is the case, for example, where a plurality of quotes for the same task are extracted during the EXT2 extraction step, and the countersigned quote, or the most recent quote in time, must be taken into account.

[0148] Subsequent to the preliminary implementation step, the second learned function FA2 generates as output at least one complementary question encoded in a second plurality of discrete symbols of a natural language, intended for a third user. The third user may advantageously be a member of the organization of interest who can provide complementary information Rcomp to assist in the implementation step IMP of the second learned function FA2. In another example, the third user may be the first user who originated the request Q1.

[0149] Advantageously, the at least one complementary question is transmitted to the third user by means of an electronic communication system. Then, the method 100 further comprises a step of acquiring complementary information Rcomp in response to said complementary question by the third user by means of a user interface of a third user terminal. Advantageously, during the step of implementing IMP the second learned function, the second learned function also receives as input the complementary information Rcomp.

[0150] In some embodiments, the at least one request Qi comprises a plurality of requests. In this case, the automatic detection step DET comprises the detection, by the detection algorithm, of the plurality of requests.

[0151] For example, requests are processed one by one, without using the response to other requests for a particular question.

[0152] In another example, requests are processed simultaneously, in order to produce a response consistent with their combination (e.g., using a large language model, which will respond more accurately to two concurrent requests that are consistent with each other than to two requests taken separately).

[0153] In another example, the response to some of the requests may specify the context for the other requests, particularly when the type of information produced matches one of the information types defined in the context.

[0154] In yet another example, the requests each have their own context, particularly if they are presented distinctly (in punctuation, such as dashes or numbering) or spaced out in the body of the Mi email by particular formatting such as separate paragraphs.

[0155] In some embodiments, the electronic message Mi is contained in an original digital document comprising a plurality of individual messages, the electronic message Mi being one of the individual messages. For example, the original digital document is part of the corpus of digital documents stored in the memory resources of the collaborative workspace hosted on the first remote server SERVi. The method 100 then further comprises, prior to the step of receiving REC the electronic message Mi, a step of segmenting the original digital document so as to obtain a plurality of individual electronic documents, each individual electronic document corresponding to one of the plurality of individual messages.

[0156] For example, the segmentation may correspond to a set of pages of the original digital document. The segmentation may be performed automatically by one or more components of the first remote server SERVi that extracts the information from the original digital document. This extraction may be corrected, replaced, or performed by a user, such as a member of the organization of interest. This manual correction may improve the next segmentation performed automatically. Advantageously, a learned function generated from a model trained on a message typology is applied to detect the beginning (first page, header, etc.) of documents or messages of this typology. In another example, clustering techniques, particularly hierarchical clustering, may be used.

[0157] Learned functions FAi, FA2, MOD models, LLM and training data

[0158] The method 100 of the invention comprises executing at least one learned function generated by a machine learning model trained from a training data set called training domain. The first learned function FA1, the second learned function FA2 and the third learned function FA3 are part of the at least one learned function.

[0159] According to one embodiment, the first learned function FA1, the second learned function FA2 and the third learned function are generated by at least one machine learning model or by a plurality of machine learning models.

[0160] According to an alternative, the method of the invention comprises the execution of at least one learning function which is configured to query and / or extract data within a history. By history, we mean a set of data accumulated over time, which can evolve and gradually aggregate new data. The history can take the form of a database or a file. The updating of the history with new data can be carried out automatically, for example from a user action. The learned functions can be generated by machine learning models having specific architectures. According to one example, a network architecture can be a machine learning model of the RNN type, meaning in the English literature "Recurrent Neural Network" or a model of the LSTM type, meaning in the English literature "Long short-term memory" can be used.In another example, the machine learning model can be a Transformer, such as GPT-3, which stands for "generative Pre-Training Transformer" and is a model based on the Transformer architecture, i.e., some layers of the model have the structure of a Transformer.

[0161] An advantage of using a pre-trained network, for example of the Transformer type, is to use their good capabilities to process data defining documents comprising discrete symbols encoded in a natural language. According to one embodiment, the invention is compatible with the use of already pre-trained networks already existing, for example on platforms accessible from the internet.

[0162] A pre-trained BERT-type network, designated in the literature as “Bidirectional Encoder representations from Transformers”, whose model also includes certain layers having the structure of a Transformer, can also be implemented within the framework of the present invention.

[0163] In another example, a BART (Bidirectional and Auto-Regressive Transformers) type language model can be used to generate representations of digital documents.

[0164] For example, when few classes are addressed by a classifier, for example between 5 and 10 classes, a machine learning model architecture may include the implementation of binary trees such as "Random Forest" or "XGBoost" or a convolutional neural network called CNN.

[0165] According to another example, when a large number of classes are addressed by a classifier, a machine learning model architecture may include the implementation of a Transformer model, for example of the BERT type, or a convolutional neural network called CNN. The machine learning models may be configured within the scope of the invention to extract data from digital documents such as dates, addresses, proper names, tables, signatures, structured information, sentiments and / or target sentences. A “regex” type model architecture or a library such as “spacy” comprising different types of models, or a CNN or a “transition-based models” type model may be used.

[0166] The models used to extract objects, named entities can implement a "transformer" type model or a BERT type model.

[0167] For example, CLIP or RoBERTa models can be used. The CLIP model analyzes the visual content of a document as an image and produces a vector that maximizes the dot product with text "similar to the image content". Another example of a model that can be used is the Mixtral / Mistral model trained on a large amount of annotated data.

[0168] According to one embodiment, an architecture for recognizing patterns or motifs, such as a signature, in a digital document may implement a "YOLO" type or "autoencoders" type model.

[0169] As mentioned above, the large language model LLM that generated the second learned function FA2 can be the Mistral IA™ company model, the LLaMA™ model, or the Falcon 180B™ model.

[0170] Advantageously, when a model used for the implementation of the method 100 corresponds to a pre-trained model, the data from previous implementations of the method 100 (for example, from a classification, extraction or matching step) are used to re-train this model.

[0171] In some embodiments, the first learned function FA1, the second learned function FA2, and / or the third learned function may comprise regular expressions (regex).

[0172] Examples

[0173] Example 1:

[0174] In this example, the organization of interest is an insurance company and the first user is a client of the insurance company or a broker of the insurance company. The insurance company has a collaborative workspace with memory resources where it can store various documents, such as member contracts, invoices, scanned versions of annual premium calls, information statements, scanned versions of payment checks, among others.

[0175] Advantageously, members of the insurance company annotated all or part of these documents in order to associate metadata with each document. Alternatively, the annotations or metadata could be extracted automatically. Examples of metadata are the document type, a document owner, or any other type of information specific to the insurance sector.

[0176] The broker or client writes one or more requests, the response to which is included in one of the documents.

[0177] This request can be sent in a Mi electronic message by email or any electronic means of communication, or via an API and its interfaces and the use of a specific field dedicated to requests.

[0178] Examples for the Qi request are:

[0179] - What is the amount of guarantees for electronic activity for my RCP?

[0180] - What is the next due date?

[0181] - What is my insurance premium?

[0182] - Who is my insurer?

[0183] - Is bicycle theft in front of my premises covered by my insurance?

[0184] According to the method 100 according to the invention, the request(s) are extracted from the electronic message Mi.

[0185] According to the method 100 according to the invention, a data set called second data set ENS2 is extracted from the message M1, making it possible to search for the relevant documents concerned, or to directly carry out a free search among a document indexing base and a set of documents of interest are extracted.

[0186] Then the second learned function FA2 is implemented so as to generate responses to each request, and propose one or more responses, optionally accompanied by extracts of documents from the extraction and / or a hyperlink allowing to point to or display the extracts to provide a justification of the response.

[0187] Example 2:

[0188] In this example, the organization of interest is a real estate agency.

[0189] In the first case, the first user is a client of the agency who is searching for real estate.

[0190] In a second case, the first user is an agency agent who transcribes a client's property search.

[0191] According to an example, the first user writes the description of a real estate search in the form of a Qi request including several characteristics of the property sought, such as surface area, location, number of rooms, for example, and integrates the description into an Mi electronic message.

[0192] The characteristics are extracted from the message. Based on the extracted characteristics, according to the method 100 according to the invention, a list of documents describing goods having a correspondence link with the characteristics, for example, HTML pages. For example, the response Ri generated by the second learned function FA2 can be a list of descriptions of each corresponding good in the list of extracted documents. The second learned function FA2 can advantageously, when generating the response Ri, highlight certain characteristics of goods particularly in line with the requested search.

[0193] For example, the second learned function FA2 can advantageously attach to the response Ri a link to the contact details of the person who owns a particular property, or is in charge of this property.

[0194] In another example, query Q1 may be a statistical query, such as a question about the average surface area of ​​real estate sold over a given period. When implementing the first learned function FA1, the list of digital documents of interest {Di}i is dependent on the given period.

[0195] Example 3:

[0196] In this example, the organization of interest is a social landlord and the first user is a tenant of a housing unit managed by the social landlord. For example, the social landlord scans incoming mail en masse and stores the digital documents resulting from the scanning in the memory resources of the SERVi server. Then, a digital document segmentation step can be implemented in order to split the digital documents resulting from the scanning into individual documents, each containing an individual message.

[0197] Advantageously, one or more members of the social landlord manually generate metadata to enrich each individual message. For example, the metadata could be a type of mail, the department to which it is addressed, the property it concerns, and the current tenant. Alternatively, the metadata can be extracted automatically.

[0198] The first user writes one or more Qi requests, the response to which is included within one or more of the individual documents.

[0199] The Qi request can be sent in a Mi electronic message, for example by email or any electronic means of communication. The social landlord's EDM constitutes a collaborative workspace which can be hosted within the SERVi server.

[0200] Examples of Qi requests are:

[0201] - What is the amount of work voted on at the last AGM for this property?

[0202] - How many tenants gave notice of departure in September?

[0203] - What is the outcome of the structural diagnosis carried out?

[0204] - Has planning permission been granted?

[0205] - When is the invoice due for carpentry work?

[0206] - What is the amount of the charge adjustment?

[0207] According to the method 100 according to the invention, during an extraction step EXTi, a data set named second data set ENS2 is extracted from the electronic message M1. The second data set ENS2 will make it possible to search for relevant documents to respond to the request Q1, such as the letters concerned. Alternatively, a free search among a mail indexing database can be carried out. In certain cases, during a step of adding additional information, the first user can cite a name or information to identify a document, or even attach an additional document to the electronic message Mi, so that the latter is taken into account during the step of implementing the second learned function FA2.

[0208] A list of stored mails is retrieved, representing the list of documents of interest, which have been determined to contain a response to the Q1 request(s).

[0209] The second learned function FA2 is implemented to generate responses to each request, and then proposes one or more responses, optionally accompanied by extracts of the mail from the extracted list where the response or a response element was found, and possibly a hyperlink generated to provide a justification for the response, for example by referring to the social landlord's EDM and in particular the mail in question or an extract from it, so as to be able to display it for example.

[0210] Responses are returned to the first user by any means of communication.

[0211] Example 4:

[0212] In this example, the first user is a user of a database of technical documents such as clinical trial results or scientific publications.

[0213] The query Q1 may consist of searching for an efficacy rate or an optimal dose of a drug to treat a pathology. The second learned function FA2 can be implemented to extract the most relevant documents from the database with respect to the query Q1.

[0214] The aspects of the invention are not abstract but closely related to the technological implementation involving specific hardware and software integrations. In one implementation, the system uses interconnected data servers for receiving, identifying or detecting, selecting, extracting, processing and storing data, combined with computer models operating in separate environments. The use of predefined configurations to generate lists of documents of interest, extract information of interest, evaluate the results and classify or prioritize them according to predefined criteria demonstrates a specific and non-generic application of artificial intelligence technology.

[0215] In one or more embodiments, the system includes dedicated computing units, such as GPUs or TPUs, optimized for running machine learning models, such as the MOD machine learning model or the LLM large machine learning model. These units may be hosted in a server.

[0216] Each compute unit can apply trained model weights on specific training datasets, generate outputs through a sequence of matrix computations and predictions, and optimize response generation by applying predefined constraints and error control algorithms or cost functions during execution.

[0217] The machine learning model execution device may include software modules designed to load pre-trained language models into memory, dynamically configure model prompts, allowing results to be customized to linguistic, cultural, or domain-specific requirements.

[0218] In one or more embodiments, the device for executing the machine learning models may utilize a scalable cloud-based infrastructure, such that, for example, servers hosted in the cloud dynamically allocate computing resources to execute instances of the computer programs implementing the MOD and LLM machine learning models, clusters of virtual machines provide redundancy and scalability to process large volumes of test data, and containerized environments ensure reproducibility and isolation of different models during execution.

[0219] Expressions such as "comprise," "include," "incorporate," "contain," "is," and "have" should be interpreted in a non-exclusive manner when construing the description and associated claims, i.e., interpreted to allow for the presence of other elements or components that are not explicitly defined. Reference to the singular should also be interpreted as a reference to the plural and vice versa.

[0220] The articles "a" and "an" may be used in connection with various elements and components of the compositions, methods or structures described herein. This is merely for convenience and general meaning of the compositions, methods or structures. Such description includes "one or at least one" of the elements or components. Furthermore, in this document, articles in the singular also include a description of a plurality of elements or components, unless it appears from a specific context that the plural is excluded.

[0221] As used in the specification and in the claims, the expression "at least one", for example in reference to a first request, is to be understood as meaning at least one element selected from one or more of a plurality of elements and constituting said first request, such as a question contained in an email, but not necessarily including at least one of each specifically listed element in the plurality of elements and not excluding any combination of elements in the plurality of elements. This definition also allows for the optional presence of elements other than the specifically identified elements in the plurality of elements to which the expression "at least one" refers, whether or not related to the specifically identified elements.

[0222] The expression "and / or", as used in the specification and in the claims, is to be understood to mean "either or both" of the elements so joined, i.e., elements which are present conjunctively in some cases and disjunctively in other cases. Multiple elements listed with "and / or" are to be interpreted in the same way, i.e., "one or more" of the elements so joined. Other elements may optionally be present in addition to the elements specifically identified by the "and / or" clause, whether or not they are related to the specifically identified elements.

[0223] A person skilled in the art will readily appreciate that various elements, features, and parameters disclosed in the description may be modified and that various disclosed embodiments may be combined without departing from the scope of the invention. For example, various aspects of the present disclosure may be used alone, in combination, or in a variety of arrangements not specifically described in the embodiments described in the foregoing and is therefore not limited in its application to the details and arrangement of components set forth in the foregoing description or illustrated in the drawings. For example, aspects described in one embodiment may be combined in any manner with aspects described in other embodiments.

[0224] Having described above several aspects of at least one embodiment, it is appropriate to appreciate the various alterations, modifications and improvements that persons skilled in the art can readily make to this embodiment. These alterations, modifications and improvements are intended to constitute aspects of the present disclosure. Accordingly, the foregoing description and drawings are given by way of example only.

[0225] According to another aspect, the invention relates to a computer program product comprising instructions which, when the program is executed by a computer, such as the system of the invention, for example by means of the computer interacting with the other components of said system, for example the memories, the communication interface, and / or the display, and / or interacting with components external to the system, such as a data server, for example the first server SERVi, lead the latter to implement the steps of the method of the invention. The computer program product is for example configured to automatically implement the steps of the method of the invention.

[0226] According to another aspect, the invention relates to a non-transitory medium, on which the computer program product is recorded. The term "non-transitory medium" means that the medium is a tangible medium. It is not a transient signal.

Claims

CLAIMS 1. A computer-implemented method (100) for automatically generating at least one first response (Ri), encoded in a plurality of discrete symbols in a natural language, to at least one first request (Qi) contained in an electronic message (Mi), the method comprising: ■ Reception (REC) of said electronic message (Mi) within a messaging system, said message comprising a set of data (ENS) and being associated with a first user; ■ Extraction (EXTi) of a first data set (ENSi) and a second data set (ENS2) from said message (M1), so as to generate an input vector (VE) from the second data set (ENS2); ■ Automatic detection (DET) of said at least one first request (Q1) by executing, on the first data set (ENS1), a detection algorithm; ■ Extraction (EXT2) of a list of at least one digital document of interest ({Di}i) by executing a first learned function (FA1) generated from a machine learning model (MOD) trained from a first training domain, said first learned function (FA1) receiving as input said input vector (VE); ■ Execution (IMP) of a second learned function (FA2) generated from a pre-trained large language model (LLM), said second learned function (FA2) receiving as input a set of data from said electronic message (M1), relating to said at least one first request (Q1) and called effective set (ENSe), and said digital documents of interest ({Di}i) from said list, so as to generate said first response (Ri) to the at least one first request (Q1).

2. Method (100) according to claim 1, further comprising generating a context (CONT) from said first data set (ENSi), and wherein, during the execution (IMP) of said second learned function (FA2), said second learned function (FA2) also receives as input said context (CONT).

3. Method (100) according to one of the preceding claims, in which the at least one digital document of interest ({Di}i) is contained in a corpus of digital documents stored in a collaborative workspace.

4. Method (100) according to one of the preceding claims, in which the at least one digital document of interest ({Di}i) comprises at least one document attached to the electronic message (M1).

5. Method (100) according to one of the preceding claims, in which said effective set (ENSe) is the data set (ENS) included in said electronic message (M1).

6. Method (100) according to one of claims 1 to 5, further comprising a step of encoding said at least one first request (Q1) into a prompt formed from another plurality of discrete symbols in a natural language, and in which said effective set (ENSe) comprises said prompt.

7. Method (100) according to the preceding claim, in which the step of encoding into said prompt is carried out from said at least one first request (Q1) and a history (HIST) of errors and correct responses obtained during previous implementations of the method (100).

8. Method (100) according to one of the preceding claims wherein said detection algorithm is said second learned function (FA2).

9. Method (100) according to one of the preceding claims, further comprising: ■ Acquisition of at least one additional data item defined by a second user by means of a user interface, and in which the second learned function (FA2) also receives as input the at least one additional data item.

10. Method (100) according to one of the preceding claims, further comprising: ■ Transmission of said at least one first response (Ri) so that said at least one first response (Ri) and, where applicable, said at least one reference (REF) are made accessible to said first user.

11. Method (100) according to the preceding claim, in which: ■ the transmission is a transmission of a second electronic message (M2) to a messaging system of said first user.

12. Method (100) according to one of the preceding claims, in which: ■ said at least one first request comprises a plurality of requests, ■ said at least one first response (Ri) comprises a plurality of responses each corresponding to one request among the plurality of requests.

13. Method (100) according to one of the preceding claims, in which the extraction step (EXT1) is implemented by executing a natural language automatic processing model on the data set (ENS).

14. Method (100) according to one of the preceding claims, comprising, prior to reception, of said electronic message (M1): ■ Segmentation of at least one original electronic document so as to obtain a plurality of electronic documents individual, each individual electronic document corresponding to an individual electronic message, said electronic message (Mi) being one of the individual electronic messages.

15. System (200) comprising an electronic terminal of a user (Ti) comprising at least one computer, a memory, a display and a communication interface for transmitting messages over a data network to at least one first server (SERVi) comprising hardware resources for executing the first learned function (FAi), the second learned function (FA2), the third learned function (FA3) and a memory for recording the models (MOD, LLM, NPL) in order to execute the steps of the method (100) according to any one of claims 1 to 14.

16. Computer program product comprising instructions which cause the system according to claim 15 to execute the steps of the method according to any one of claims 1 to 14.

17. A non-transitory computer-readable medium on which the computer program product according to claim 16 is recorded.

Images