Method and device for personalising an audiovisual stream

Abstract

The invention relates to a method for personalising a first audiovisual digital stream (FAVo), comprising steps of: - determining (E90), using a knowledge base of at least one word from a user, at least one word (GMMC1) which is obtained from the first audiovisual digital stream (FAVo) and is poorly known by a user to which the first digital stream (FAVo) is to be rendered; and - personalising the first audiovisual digital stream (FAVo) into a personalised audiovisual stream (FAVp), the personalisation comprising sending, to a generative language model, a request (RQ1) asking it to explain the at least one word (GMMC1).

Description

[0001] Method and device for personalizing an audiovisual stream

[0002] Background

[0003] The present disclosure is in the field of personalization of a digital audiovisual stream, i.e. a digital stream comprising visual data and / or audio data, for the purpose of presenting it to a user. It applies to any type of digital audiovisual stream, regardless of its content, for example to purely textual documents, to images, to sounds, to mixed documents comprising for example images inserted into text pages, in particular web pages, ...

[0004] The present disclosure finds a privileged but non-limiting application in the more particular context of the interaction between a user and a dialoguing software agent (or conversational software agent) (in English chatbot).

[0005] Such agents are now widely deployed on the Internet, for example for online sales or help applications.

[0006] These agents are called conversational in that they allow a dialogue or conversation to be established between a user and a machine hosting the dialoguing agent, a conversation in which a user can ask questions in natural language to the agent (in prompt English), the agent returning to the user an audiovisual stream in response to these questions, for example a statement, a new question, a list of choices, etc.

[0007] In the current state of the art, the responses provided by the agent offer a certain level of personalization based, for example, on information contained in a user profile.

[0008] For some applications at least this personalization of responses is not satisfactory.

[0009] Summary

[0010] The present disclosure provides a solution for improving the personalization of responses sent to a user by a dialog agent, and generally for improving the personalization of any audiovisual stream intended for a user.

[0011] Thus and according to a first aspect, the disclosure relates to a method for personalizing a first audiovisual digital stream, said method comprising sending, to a generative language model, a request asking it to explain at least a first word obtained from the first audiovisual digital stream.

[0012] In at least one embodiment, the method comprises determining said at least one first word using at least one knowledge base of at least one word by a user.

[0013] In at least one embodiment, said at least one first word is a word that is little or poorly known by a user to whom the first digital stream is to be rendered.

[0014] For example, the disclosure relates to a method of personalizing a first audiovisual digital stream, said method comprising steps of:

[0015] - determination, using at least one knowledge base of at least one word by a user, of at least one word obtained from the first audiovisual digital stream and little or poorly known by a user to whom the first digital stream must be rendered;

[0016] - personalization of the first audiovisual digital stream into a personalized audiovisual stream, said personalization comprising sending, to a generative language model, a request asking it to explain said at least one word.

[0017] Thus, and in general, the present disclosure proposes a solution aimed at improving the intelligibility of an audiovisual stream, for example of a document, before rendering it (i.e. restoring it, presenting it, providing it) to a user.

[0018] More specifically, when it is detected, for example, that an audiovisual stream intended to be rendered to a user contains one or more words that the user is assumed not to know or to know poorly, the present disclosure proposes to render to the user an audiovisual stream likely to be more easily understood by the user.

[0019] When the first audiovisual stream is a text document, the words can be extracted directly from the document. If the first audiovisual stream is of another type, obtaining the words may require preprocessing of the stream, for example, character recognition to obtain the words from an image or speech recognition to obtain the words from sounds.

[0020] In accordance with the present disclosure, the expression "at least one word" may refer to "at least one group of at least one word", a "group of at least one word" designating either one word or several consecutive words, optionally separated by separators (such as spaces or punctuation marks).

[0021] Different types of knowledge bases may be used in the context of this disclosure. For example, said at least one knowledge base may be chosen from:

[0022] - a knowledge base containing words or groups of words known to a user; or

[0023] - a knowledge base containing words or groups of words that are poorly known or little known to a user; or

[0024] - a knowledge base comprising words or groups of words associated with clues to a user's knowledge of these words or groups of words, or

[0025] - a combination of at least two of these knowledge bases.

[0026] A knowledge index can, for example, be a discrete variable "KNOWN, UNKNOWN, LITTLE KNOWN" or a probability of knowledge. Many techniques can be used to determine the probability that a user knows little or little about a word or group of words. Words or groups of words can, for example, be classified according to their occurrence in documents published on the Internet, depending on whether they belong to a more or less specialized field. A knowledge base such as that published by patent document FR3111728A1 can also be used. Any other mechanism associating a probability of knowledge with a word or group of words can be used.In this disclosure, the expression "A has at least one word unknown or little known to user B" is a simplified expression meaning "a word or group of words of A is determined from the knowledge base to be unknown or little known to user B".

[0027] In at least one embodiment of the present disclosure, the personalization method comprises at least one sending, to said generative language model, of an additional request asking it to explain at least one word or group of words included in the response to said request previously sent to the generative language model.

[0028] In at least one embodiment of the present disclosure, the personalization method comprises rendering, to the user to whom the first digital stream was to be rendered, a personalized audiovisual digital stream obtained from a response to at least one of the requests mentioned above (first request or additional requests).

[0029] In at least one embodiment, additional queries are sent to the generative language model as long as the responses received to these queries include at least one word or group of words that is little or poorly known to the user.

[0030] This embodiment aims to ensure that the personalized audiovisual digital stream delivered to the user does not include any word or group of words determined to be little or poorly known to the user.

[0031] In at least one embodiment, a maximum predetermined number of additional requests is sent to the generative language model. This embodiment makes it possible to obtain the personalized audiovisual digital stream in a reasonable time (for example to render the personalized audiovisual digital stream to the user in a reasonable time in a live implementation (“real time”) and possibly to avoid the method entering an endless loop).

[0032] Queries sent to the natural language model can explicitly include the word or group of words to be explained.

[0033] For example, a request might be something like, "Please clarify what the word X means."

[0034] Queries sent to the natural language model can include a natural language expression allowing the generative language model to determine the words or groups of words to be explained.

[0035] For example, a request might be something like "Please clarify the meaning of the first sentence of your last answer." Upon receiving such a request, the generative language model sends a response explaining the words in the first sentence of its last answer without these words having been specified in the request.

[0036] In one embodiment of this disclosure, at least one query further requests the generative language model to further explain at least one word or group of words taking into account at least one characteristic of the user. This characteristic can be obtained from a profile of the user.

[0037] This characteristic can be determined from a type of the first audiovisual digital stream.

[0038] For example, if the first audiovisual digital stream is a web page intended for children, it can be determined that the user is a child.

[0039] For example, if the user is a child, a request might be something like "Please clarify what the word X means by adapting your answer for a child" or "Please clarify the meaning of the first sentence of your last answer for a child."

[0040] At least one characteristic can be obtained from a profile of a group of individuals to which the user belongs, for example a profile of a company employing the user. For example, a knowledge index of certain words can be assigned to a user employed by this company taking into account the field of activity (health, telecommunications, etc.) of the employing company.

[0041] In at least one embodiment of the method of the present application, said personalized audiovisual digital stream is obtained at least by:

[0042] - replacing at least one word of the first audiovisual digital stream with at least one word included in at least one said response; and / or

[0043] - by inserting near at least one word of the first audiovisual digital stream, at least one word included in at least one said response.

[0044] In at least one embodiment of the method of the present application, said at least one replaced and / or inserted word is at least one word that is little or poorly known to a user to whom the first digital stream is to be rendered.

[0045] Thus, in at least one embodiment of the present disclosure, the personalized audiovisual digital stream is obtained at least by:

[0046] - replacing the words or groups of words in the first audiovisual digital stream that are little or poorly known to the user with words or groups of words included in at least one response from the generative language model; and / or

[0047] - by inserting near the words or groups of words of the first audiovisual digital stream that are little or poorly known to the user, words or groups of words included in at least one of these responses.

[0048] Correlatively, the invention relates to a device for personalizing a first audiovisual digital stream, this device comprising:

[0049] - a module for personalizing the first audiovisual digital stream configured to send, to a generative language model, a request asking it to explain at least a first word obtained from the first audiovisual digital stream.

[0050] The device of the present application may be configured to implement the method of the present application, in any of its embodiments.

[0051] Thus, in some embodiments, the device comprises

[0052] - a determination module configured to determine, using a knowledge base of at least one word by a user, at least one word obtained from the first audiovisual digital stream and little or poorly known by a user to whom the first digital stream must be rendered;

[0053] - a personalization module for the first audiovisual digital stream configured to send, to a generative language model, a request asking it to explain said at least one word.

[0054] The personalized audiovisual digital stream obtained from a response to said request is intended to be rendered to the user by a rendering module.

[0055] It should be noted that this restitution module is not necessarily a module of the personalization device.

[0056] The personalization device may be adapted to implement the method presented in this disclosure, in any of its embodiments.

[0057] In one embodiment of this disclosure, the personalization device comprises a rendering module. This rendering module makes it possible to render the personalized audiovisual stream and also, in at least one embodiment, the first audiovisual stream. This embodiment allows the user to compare the audiovisual stream and the personalized audiovisual stream.

[0058] This restitution module is, for example, a screen, a printer, a magnetic support, a speaker, a monitor, a camera, glasses or a virtual reality headset.

[0059] In one embodiment, the first audiovisual stream is obtained according to a first type, the method comprising a conversion of the first audiovisual stream or the personalized audiovisual stream either before or after the personalization for the purpose of its rendering.

[0060] The first audiovisual stream and the personalized audiovisual stream are, for example, each a digital text document, or a sound, or an image.

[0061] The method and device for personalizing this disclosure can in particular be used to promote understanding of an audiovisual digital stream obtained on a website.

[0062] In a particular embodiment, the different steps of the personalization method are determined by computer program instructions or are implemented by a silicon chip which comprises transistors adapted to constitute logic gates of non-programmable hard-wired logic.

[0063] Consequently, the invention also relates to a computer program on an information medium, this program being capable of being implemented in a controller computer, this program comprising instructions adapted to the implementation of the steps of a personalization method as described above, in its different embodiments.

[0064] This program may use any programming language, and may be in the form of source code, object code, or code intermediate between source code and object code, such as in a partially compiled form, or in any other desirable form.

[0065] The invention also relates to a computer-readable information medium, and comprising instructions of a computer program as mentioned above, in its various embodiments. The information medium can be any entity or device capable of storing the program. For example, the medium can comprise a storage means, such as a ROM, a non-volatile memory of the flash type or even a magnetic recording means, for example a hard disk. Furthermore, 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 a network of the Internet type.Alternatively, the information carrier may be an integrated circuit in which the program is incorporated, the circuit being adapted to perform or to be used in the performance of the method in question.

[0066] Brief description of the drawings:

[0067] Other characteristics and advantages of the present invention will emerge from the description given below, with reference to the appended drawings which illustrate exemplary embodiments thereof which are not in any limiting nature. In the figures:

[0068] [Fig. 1] Figure 1 shows a personalization device according to certain embodiments of this disclosure;

[0069] [Fig. 2] Figure 2 represents, in the form of a flowchart, steps of a personalization method in accordance with this disclosure, implemented for example by the device of Figure 1;

[0070] [Fig. 3] Figure 3 shows a personalization device according to certain embodiments of this disclosure;

[0071] [Fig. 4] Figure 4 represents, in the form of a flowchart, steps of a personalization method in accordance with this disclosure, implemented for example by the device of Figure 3;

[0072] [Fig. 5] Figure 5 depicts a hardware architecture of a personalization device in accordance with this disclosure.

[0073] Description of the embodiments

[0074] Figure 1 represents a device DIS1 for personalizing an audiovisual digital stream (first stream within the meaning of the disclosure) in accordance with at least one embodiment of this disclosure.

[0075] Figure 2 represents, in the form of a flowchart, steps of a personalization method in accordance with this disclosure, implemented for example by the device of Figure 1.

[0076] In this embodiment, the DIS1 device only receives audiovisual streams corresponding to text documents and it systematically analyzes them to determine whether they include words or groups of words that are little or poorly known to a user to whom these streams are likely to be rendered (rendered, or produced). If it is determined that they include one or more words or groups of words that are little or poorly known to a user, these streams are personalized using a generative language model, before their (optional) rendering to the user.

[0077] The DIS1 device includes a MODJDBT module for obtaining or receiving a first FAVo audiovisual digital stream.

[0078] In the detailed example, the device DIS1 is configured to access a knowledge base BC, which includes for at least one word or group of words and at least one user an index on knowledge of this word or group of words by this user.

[0079] The DIS1 device includes a determination module MOD_DET configured to determine, using the knowledge base BC, whether a textual audiovisual digital stream, for example the textual audiovisual digital stream FAVo, includes at least one word or group of words that is little or poorly known by a user to whom this stream is intended to be rendered.

[0080] The DISI device includes a generative language model MLG. This model is, for example, a model of the GPT 4, GPT 3.5, PaLM 2, Lima 2, Falcon, Alpaga or Bloom type (registered trademarks) known to those skilled in the art.

[0081] The DIS1 device includes a MOD_PERS module for customizing the first FAVo audiovisual digital stream configured to send an RQi request to the MLG generative language model asking it to explain the words or groups of words GMMCi of the first FAVo audiovisual digital stream determined to be unknown or poorly known.

[0082] The DIS1 device is configured to produce as output either the first audiovisual digital stream FAVo if it did not contain any words considered little or poorly known by the user to whom this stream is to be rendered, or a personalized stream FAVP obtained from the response to its request by the generative language model MLG.

[0083] Figure 2 represents in the form of a flowchart a personalization process implemented by the personalization device of Figure 1.

[0084] It includes a step E10 for obtaining the first FAVo audiovisual digital stream.

[0085] During a step E90, it determines, using the knowledge base BC, whether the flow contains at least one word or group of words GMMCi that is little or poorly known by a user to whom the first digital flow FAVo must be returned.

[0086] If this is not the case, the result of test E90 is negative and the first stream is rendered (or alternatively stored or transmitted for the purpose of such rendering) during a step E120.

[0087] If at least one word is little or poorly known, the result of the E90 test is positive and the method includes a step E100 of personalizing the first flow by sending, to the generative language model MLG, a request asking to explain this or these GMMCi words.

[0088] In the embodiment described here, if the response from the generative language model MLG includes at least one word or group of words that is unknown or poorly known to the user, an additional request is sent to the MLG model asking it to explain this or these words. In the embodiment described here, as long as the responses include words that are little or poorly known, an additional request is sent to the MLG model to ask it to explain these words.

[0089] When the response to the query does not include an unknown or poorly understood word or group of words, a personalized audiovisual digital stream obtained from this response is produced so that it can be rendered to the user (or alternatively stored or transmitted for such rendering, possibly delayed in time).

[0090] For example, in this personalized audiovisual digital stream, all the little or poorly known words or groups of words from the first FAVo audiovisual digital stream are replaced by words received in at least one response from the MLG model.

[0091] Figure 3 shows a DIS2 audiovisual stream personalization device according to certain embodiments of this disclosure.

[0092] Figure 4 represents in flowchart form the main steps of a method for personalizing a digital document according to certain embodiments of this disclosure.

[0093] The DIS2 device comprises an obtaining module MODJDBT configured to obtain, during a step E10, a first audiovisual digital stream FAVo intended to be returned to a user.

[0094] The first audiovisual digital stream FAVo can be of different types. For example, it represents a digital text document, an image or a digital voice document.

[0095] The rendering of the first digital audiovisual stream FAVo by a rendering module MOD-RES, included or not in the DIS2 device, can be of different natures. The digital audiovisual stream FAVo can for example be displayed on a screen, printed by a printer, recorded on a medium (for example a hard disk), rendered in the form of a voice message by a loudspeaker, projected in the form of an image by a camera, for example on the screen of glasses or a virtual or augmented reality headset, rendered in the form of sounds by an earpiece, reproduced by the gestures of an avatar.

[0096] The above-mentioned screen, printer, stand, speaker, monitor, camera, glasses, headset, avatar and virtual reality headset are examples of MOD_RES rendering modules within the meaning of this disclosure.

[0097] The first audiovisual digital stream FAVo can be obtained in different ways. For example, it can be produced by and / or received from different SRC sources. For example, a source can correspond to a communication module of the device, this communication module being configured to receive the first audiovisual stream via a network, an email client, an Internet browser, a microphone, a camera, a storage medium, a file

[0098] The first digital audiovisual stream FAVo may optionally be converted into a digital audiovisual stream FAVc before being rendered or restored. For example, a digital textual (or text-like) document may be synthesized to produce a digital voice document that can be restored by a loudspeaker or converted into an image to be projected onto a screen. For this purpose, in a particular embodiment of this disclosure, the personalization device DIS2 comprises a conversion module MOD_CONV configured to convert an audiovisual stream of a certain format (or type) into an audiovisual stream of another format (or type) adapted to the rendering module used (for example by voice synthesis).

[0099] In some embodiments of the invention, the method may be implemented for all FAVo audiovisual streams obtained or received by the device. In other embodiments, this implementation may be conditional, for example it may be performed on demand.

[0100] The user can, for example, configure a (binary) parameter of the personalization device to request or not an analysis of the received streams. According to another example, the method can be implemented in the form of a service that can be invoked by a computer application (for example, by the computer application that generated or produced the audiovisual stream that will be personalized (if applicable).

[0101] In one embodiment, a module of the audiovisual stream cut-off rendering device (e.g., a web browser) may or may not request that an audiovisual stream be personalized.

[0102] Thus, in certain embodiments of the invention, all FAVo audiovisual streams obtained or received by the device can be analyzed with a view to possibly being personalized.

[0103] In other embodiments, the personalization method may comprise a step E20 for determining whether a first audiovisual stream FAVo must be analyzed with a view to possibly being personalized.

[0104] For example, in some embodiments, a first voice command (predetermined for example) may be provided so that all audiovisual streams obtained after this first command are analyzed.

[0105] For example, in some embodiments, a second voice command (predetermined for example) may be provided so that all audiovisual streams obtained after this second command are not analyzed.

[0106] For example, in some embodiments, audiovisual streams may or may not be analyzed based on the SRC source that produced or sent that stream.

[0107] For example, in some embodiments, audiovisual streams are analyzed or not based on metadata associated with this digital document.

[0108] In some embodiments, as in the illustrated example, if an audiovisual stream is not to be analyzed, the result of step E20 is negative.

[0109] In certain embodiments, as in the illustrated example, the personalization method comprises a step E30 for determining whether the first audiovisual stream FAVo must be converted before rendering, i.e. before being rendered to the user.

[0110] For example, if this is not the case, the result of test E30 is negative and the audiovisual digital stream FAVo is rendered (without conversion) during a step E40. If the first digital stream must be converted, it is converted during a step E50 and the converted audiovisual stream FAVoc is rendered during a step E60 (or alternatively stored or transmitted for such possible subsequent rendering).

[0111] In certain embodiments, as in the illustrated example, if it is determined that the first audiovisual stream FAVo obtained in step E10 must be analyzed with a view to possibly being personalized, the result of step E20 is positive.

[0112] During a step E70, it is determined whether the first audiovisual stream FAVo is of text type. If this is the case, the result of step E70 is positive. Otherwise, the first audiovisual stream FAVo is converted into a text-type audiovisual stream (step E80). The device DIS1 can use the conversion module MOD_CONV for this purpose.

[0113] Steps E20, E30, E40, E50, E70 and / or E80 may be optional in certain embodiments.

[0114] Subsequently, we denote by FAVA the digital audiovisual stream to be analyzed. This is either the first digital audiovisual stream obtained in step E10 if this stream is of text type, or an audiovisual stream FAVoc digital document of text type obtained by conversion (step E80) of the first stream FAVo.

[0115] In certain embodiments, as in the example illustrated, the device DIS2 comprises a determination module MOD_DET configured to determine, during a step E90, using the knowledge base BC, whether the digital stream to be analyzed FAVA comprises at least one word or group of words GMMG which are little or poorly known to the user of the rendering device.

[0116] This knowledge base is for example a BD knowledge base as described in patent document FR3111728A1.

[0117] This patent document discloses a method for analyzing text / image / sound type content received, sent or searched by a user, analyzing this content, extracting information from it and associating in a knowledge base with at least some of this information, an index of knowledge of the user of this information.

[0118] This knowledge index can be a function of a delay since the user's last encounter with the given information and a forgetting curve.

[0119] In a particular embodiment of patent document FR3111728A1, it is considered that a word or group of words is little or poorly known by a user, if the knowledge index associated with the user for this word or group of words is lower than a predetermined threshold.

[0120] In certain embodiments of the present application, it is considered that a group of words is little or poorly known by a user or a member of a group of individuals based on at least one knowledge index associated with this group of words, in at least one knowledge base of this user or this group of individuals. For example, it can be considered that the word or group of words is little or poorly known if the knowledge index (or the average of the knowledge indices, the maximum or minimum value of the knowledge indices) is lower (respectively higher) than a certain value (acting as a threshold), for example a configurable (predetermined) value. In the example illustrated, GMMCi denotes a group of words of the FAVA audio-video stream to be analyzed determined to be little or poorly known by the user of the rendering module.

[0121] In certain embodiments of this disclosure, the determination module MOD_DET calls (step E100) on a module MOD_PERS for personalizing the device DIS2 configured to personalize the analyzed FACA audiovisual stream. The call to this module may comprise, as an input parameter, an assignment of an objective of explaining at least one group of little or poorly known GMMCi words. This request for personalizing the first audiovisual digital stream may be carried out by sending, to the generative language model MLG, a request REQ asking it to explain the word or group of words.

[0122] In some embodiments of this disclosure, the determination module MOD_DET may also request the generative language model to explain at least one determined GMPj word by taking into account at least one characteristic of the user or of a group to which this user belongs. These words may be determined from a profile of the user or from a profile of a group to which the user belongs from a reference database BDREF.

[0123] The MOD_PERS personalization module is, for example, in certain embodiments, a natural language processing module. It may, for example, comprise the architecture of a Pre-trained Generator Transformer (in English "Generative Pre-trained Transformer"). Such a generator transformer may, for example, have been trained on a very large number of audiovisual documents in at least one certain language (such as English) to answer questions (or prompts according to English terminology) in natural language (for example, in the same language as the training documents (such as English).

[0124] For example, the determination module MOD_DET provides as input to the personalization module MOD_PERS a prompt of the type "Please reformulate "XXXXX" (i.e. the content of the digital document (FAVA) by specifying the notions of (GMMG) and (GMPj)" where DOCA and GMMG and GMPj represent respectively the audiovisual flow (digital document) FAVA to be analyzed, the groups of little or poorly known words determined from the knowledge base and words determined from a user profile in the reference base BDREF.

[0125] The query may specify that the words must be reformulated or explained taking into account at least one characteristic of the user (age, sex, nationality) or of a group (company, profession, etc.) to which this user belongs.

[0126] In the embodiment described here, as long as the responses to these requests include words or groups of words determined to be unknown or little known to the user, the analysis module sends additional requests to the personalization module to ask it to explain these words or groups of words.

[0127] The personalization module MOD_PERS produces a second audiovisual stream FAVp from the first audiovisual stream FAVo, in which the unknown or little-known words are replaced or explained by words or groups of words included in the responses of the MLG model. In certain embodiments, as in the illustrated example, the personalization method may comprise a step E1 10 for determining whether the second stream (i.e. the personalized audiovisual stream) FAVP must be converted before being rendered (rendered) by the rendering module MOD_RES. This decision may be made on the basis of a user parameter of the personalization device DISP2.

[0128] If this is not the case, the result of test E110 is negative and the personalized audiovisual stream FAVP is rendered during a step E120 (or alternatively stored or transmitted, with a view to such possible rendering).

[0129] If the personalized audiovisual stream FAVp must be converted, it is converted during a step E130 and the converted personalized audiovisual stream FAVPC is rendered during a step E140.

[0130] In the embodiment of figure 2, the first audiovisual stream FAVo may possibly be rendered to the user during a step E150 (or alternatively stored or transmitted, with a view to such possible rendering).

[0131] Steps E110 and E130 may be optional in certain embodiments.

[0132] Figure 5 shows the hardware architecture of a personalization device according to certain embodiments of this disclosure.

[0133] In this example, the personalization device has the hardware architecture of a computer. It comprises a random access memory 2 (for example a RAM memory), a read-only memory 3 (for example a ROM memory or a hard disk) and a processing unit 1 (equipped for example with a processor, and controlled by a computer program PG stored in the read-only memory 3, this computer program comprising instructions for implementing the steps of a personalization method as described above, for example with reference to figures 2 or 4. At initialization, the code instructions of the computer program PG are for example loaded into the random access memory 2 before being executed by the processor of the processing unit 1.

Claims

CLAIMS

1. Method for personalizing a first audiovisual digital stream (FAVo), said personalization comprising sending, to a generative language model, a request (RQi) asking it to explain at least one first word (GMMCi) obtained from the first audiovisual digital stream (FAVo).

2. Device (DIS1, DIS2) for personalizing a first audiovisual digital stream (FAVo), this device comprising: - a module (MOD_PERS) for customizing the first audiovisual digital stream (FAVo) configured to send, to a generative language model, a request (RQi) asking it to explain at least one first word (GMMCi) obtained from the first audiovisual digital stream (FAVo).

3. Personalization method according to claim 1 where said method comprises, or personalization device (DIS1, DIS2) according to claim 2 comprising a determination module (MOD_DET) configured for: a determination (E90) of said at least one first word using at least one knowledge base of at least one word by a user.

4. Personalization method according to claim 3, or personalization device (DIS1, DIS2) according to claim 3, where said at least one first word is a word little or poorly known by a user to whom the first digital stream (FAVo) is to be rendered.

5. Personalization method according to any one of claims 1 or 3 to 4 comprising, or personalization device (DIS1, DIS2) according to any one of claims 2 to 4 configured for, at least one sending, to said generative language model, of a complementary request (RQk) asking it to explain at least one word (GMMCk) included in the response (RPk-i) to said request (RQk-i) previously sent.

6. Personalization method according to any one of claims 1 or 3 to 5, or personalization device (DIS1, DIS2) according to any one of claims 2 to 5, where said personalized audiovisual digital stream (FAVP) is obtained at least from a response (RPk) to the last request sent (RQi, RQk).

7. Method for personalizing any one of claims 1 or 3 to 6, or device (DIS1, DIS2) for personalizing according to any one of claims 2 to 6, where a said request (RQi, RQk) explicitly comprises said at least one word (GMMCi, GMMCk) to be explained or an expression in natural language allowing the generative language model to determine it.

8. Personalization method according to any one of claims 1 or 3 to 7, or personalization device (DIS1, DIS2) according to any one of claims 2 to 7 where at least one said query (RQi, RQk) further requests the generative language model to explain at least one word (GMPj) determined by taking into account at least one characteristic of a user to whom the first digital stream (FAVo) must be returned, or of a group to which a user belongs to whom the first digital stream (FAVo) must be returned.

9. Personalization method according to any one of claims 1 or 3 to 8, or personalization device (DIS1, DIS2) according to any one of claims 2 to 8, wherein said personalized digital audiovisual stream (FAVp) is obtained at least by: - replacing at least one word (GMMCi) of the first audiovisual digital stream (FAVo) with at least one word included in at least one said response (RP1, RPk); and / or - by inserting near at least one word (GMMCi) of the first audiovisual digital stream (FAVo), at least one word included in at least one said response (RP1, RPk).

10. Personalization method according to claim 9, or personalization device (DIS1, DIS2) according to claim 9, where said at least one replaced and / or inserted word is at least one word little or poorly known to a user to whom the first digital stream (FAVo) is to be rendered.

11. Personalization method according to any one of claims 1 or 3 to 10, or personalization device (DIS1, DIS2) according to any one of claims 2 to 10, wherein said at least one knowledge base of at least one word by a user is chosen from: - a knowledge base containing words known to the user; or - a knowledge base containing words that are poorly known or little known to the user; or - a knowledge base comprising words associated with clues to a user's knowledge of these words; or -a combination of at least two of these knowledge bases.

12. Personalization method according to any one of claims 1 or 3 to 11, comprising, or personalization device (DIS1, DIS2) according to any one of claims 2 to 11 configured for, a conversion (E30, El 10) of the first audiovisual stream (DOCo) and / or said personalized audiovisual stream.

13. Personalization method according to any one of claims 1 or 3 to 12, comprising, or personalization device (DIS1, DIS2) according to any one of claims 2 to 12 comprising a rendering module (MOD_RES) configured for, a rendering (E120, E140) of said first audiovisual digital stream (FAVo) and / or of a personalized audiovisual digital stream (FAVP).

14. Personalization method according to claim 13, or personalization device (DIS1, DIS2) according to claim 13, wherein said personalized audiovisual digital stream (FAVP) is obtained from a response (RPi) to at least one said request (RQi).

15. Personalization method according to any one of claims 1 or 3 to 14, or personalization device (DIS1, DIS2) according to any one of claims 2 to 14, where said audiovisual digital stream (FAVo) is obtained (E10) on a website.

Images