Method for generating a response by a language model

Abstract

A method is proposed for generating a response by a response generator comprising a language model, said method comprising, in response to a user request (USR REQ): - generating a response adapted to relevant information determined from a user context using the response generator (RESP GEN).

Description

Method for generating a response using a language model

[0001] The present invention relates to the general field of language models, in particular large natural language models, also called LLMs (for "Large Language Models"), used to automatically generate responses from user inputs.

[0002] It relates in particular to a process for improving the quality and relevance of the responses produced by a model of this type.

[0003] Large language models (LLMs) are now widely used to perform a variety of tasks involving natural language understanding and generation. These tasks include conversational dialogue, intelligent voice assistance, automated text generation, language translation, and human-computer interaction.

[0004] However, despite their remarkable performance, these models have several known technical limitations. In particular, they tend to occasionally produce so-called "hallucinatory" responses—that is, erroneous answers formulated with apparent certainty. Furthermore, their effectiveness can be limited by their inherently static knowledge and the date of their last training, rendering them incapable of integrating dynamic contextual information specific to the immediate environment or the user.

[0005] To overcome these limitations, there is a technique called "Retrieval-Augmented Generation" (RAG), which consists of supplementing the inputs of language models with external contextual information taken from one or more reference documents.

[0006] The RAG process generally relies on three steps: prior conversion of one or more external documents into vector representations suitable for a similarity calculation; automatic selection of the text passages most semantically close to the user input; construction of an enriched instruction combining the user input and the selected passages, before submitting it to the LLM model to generate a response.

[0007] This technique has the advantage of enriching the model's knowledge base and reducing hallucinations by forcing the model to rely on real and verifiable information.

[0008] However, traditional RAG techniques themselves have a fundamental limitation: they only select information with a direct semantic similarity to the explicit user input, thus ignoring potentially highly relevant contextual information whose relevance is implicit or indirect, such as the user's usage habits, their immediate environmental context, or the precise state of the device being used. Consequently, the model's response may lack precision or relevance in certain specific usage contexts. Summary

[0009] This disclosure improves the situation.

[0010] To this end, a response generation method is proposed using a response generator comprising a language model, said method comprising, in response to a user request (USR REQ): generating a response adapted to relevant information determined from a user context using the response generator (RESP GEN).

[0011] Advantageously, the process may include: - determining relevant contextual information by means of an adaptive predictive model applied to basic contextual information (PRED) from the user context including environmental data, data relating to a platform used and historical interaction data of the user (BAS CONT INF);

[0012] Advantageously, the process can include, during a user request:

[0013] - construct a modified instruction by combining at least the said user request and the said relevant contextual information (MOD PRPT),

[0014] the response generation is obtained by applying the modified instruction as input to the response generator.

[0015] Advantageously, this basic contextual information can be transformed into a feature tensor according to a predefined feature model, with the relevant contextual information being determined by the adaptive predictive model using the feature tensor.

[0016] Advantageously, the adaptive predictive model includes a continuous updating of conditional probabilities associated with relevant contextual information based on said contextual information.

[0017] Advantageously, the adaptive predictive model can include an adaptive reinforcement learning model, updating its parameters based on positive or negative feedback received from the user environment.

[0018] Advantageously, the step of constructing the modified instruction may further include a selection of information from at least one reference document carried out by augmented generation by retrieval, said selection being based on a semantic similarity with the user query.

[0019] Advantageously, the step of constructing the modified instruction can thus include an automatic selection of information from at least one reference document, carried out by a retrieval-augmented generation technique (known as the "RAG" technique, for "Retrieval-Augmented Generation").

[0020] Advantageously, platform data may include information describing the type or condition of an electronic device used by the user.

[0021] Advantageously, the step of generating a response may include a command for an action on said platform.

[0022] Advantageously, the feature tensor resulting from the transformation step can be stored in a contextual feature database, said database being continuously updated by new user interactions.

[0023] This document also concerns a response generator using a language model capable of generating a response adapted to relevant information determined from a user context using the language model (RESP GEN).

[0024] This document also relates to a computer device comprising a processing circuit for implementing the process according to one of the preceding claims.

[0025] This document also relates to a computer program containing instructions for implementing the aforementioned process, when this program is executed by a processor of a processing circuit.

[0026] Any suitable hardware and / or software may be used for the practical implementation of the proposed technique. Generally, although aspects of the proposed technique may be described in this document as a process, device, system, procedure, or method, it should be noted that the proposed technique may also cover computer memory that can be connected to a processor, possibly connected to a communication interface. This memory stores instructions that, when executed by such a processor, enable the implementation of the processes, devices, systems, procedures, or methods described in this document.

[0027] The proposed technique can be implemented within a cloud computing environment, where the described processes and methods can be distributed across multiple servers. The use of cloud services can, for example, facilitate the use of artificial intelligence for element extraction and representation generation, remote access to the virtual environment and editing tools, and collaboration among multiple connected users.

[0028] Other features, details, and advantages will become apparent upon reading the detailed description below and analyzing the attached drawings, on which: Fig. 1

[0029] schematically illustrates one form of implementation of the process according to this document. Fig. 2

[0030] schematically illustrates one form of implementation of the process according to this document.

[0031]

[0032] schematically illustrates one embodiment of the generator according to this document.

[0033] The drawings and description can not only help to better understand the proposed technique, but also contribute to its definition, where appropriate.

[0034] The term "language model" refers to a computer model, generally based on machine learning techniques, particularly deep learning, trained to automatically generate text or other forms of coherent linguistic content from textual input. Typically, this model is trained on large amounts of textual data to acquire a statistical representation of language. Common examples of language models include large language models (LLMs). These models are used in numerous applications such as conversational dialogue, virtual assistants, machine translation, and automatic text generation.

[0035] The term "user request" refers to an input provided by the user, which can take various forms: an explicit instruction or command formulated in writing or by voice command converted into text, or an implicit command resulting from a non-textual action by the user, such as clicking on a graphical interface or triggering a sensor automatically converted into a textual representation usable by a language model.

[0036] More specifically, a "user request" refers to a command formulated by a user with the explicit aim of acting on a technical device or electronic equipment. This request is not simply a descriptive or narrative text entry, but rather an operational instruction intended to produce a concrete technical effect on a given object or system. For example, the request might aim to turn a home automation device on or off ("turn on the television"), select a specific channel or digital content ("display channel 1"), or configure a connected device ("set the thermostat to 21°C"). Thus, a user request is clearly distinguished from free or purely informative prose text, in that it expresses an intention to act on or control a specific technical object.

[0037] The term "user context" refers to all the information relating to the specific conditions under which the user makes their request.

[0038] Laillustre une forme d'animation du processus de génération réponse, selon une forme d'animation de l'invention.

[0039] This includes, in response to a user request (USR REQ): generating a response adapted to relevant information determined from a user context using the response generator (RESP GEN).

[0040] In particular, the response generation process includes: determining relevant contextual information by means of an adaptive predictive model applied to basic contextual information (PRED) from the user context including environmental data, data relating to a platform used and historical interaction data of the user (BAS CONT INF).

[0041] In particular, the response generation process includes one or more of the following successive steps: obtaining a user request (USR REQ); capturing and extracting basic contextual information from a user context including environmental data, data relating to a platform used and historical interaction data of the user (BAS CONT INFO); applying an adaptive predictive model, implementing incremental learning allowing continuous updating of its parameters during an inference phase, in order to determine relevant contextual information based on said basic contextual information (PRED); constructing a modified instruction by combining at least said user request and said relevant contextual information (MOD PRPT);apply said modified instruction as input to said language model in order to generate a response adapted to relevant information determined from a user context using the language model (RESP GEN).;

[0042] The user context can be defined by basic contextual information which may include environmental data, such as the current time, date, geographical location of the user, surrounding physical or ambient conditions (temperature, brightness, presence detected by sensor, noise level, detection of human presence, etc.).

[0043] This basic contextual information may also include data relating to a platform in use, that is, technical information about the electronic equipment used by the user (e.g., smartphone, computer, smart TV, smart speaker, home automation device, etc.), as well as its current state (battery level, connectivity status, cellular network availability, etc.). It may also include information about the applications or services currently active or open on the equipment, as well as information about the software or hardware version of said platform.

[0044] Finally, basic contextual information can include historical user interaction data. This data corresponds, for example, to a complete or partial history of previous actions performed by the user, such as previously formulated requests or commands ("Turn on the TV to channel 1," "Set my alarm for 7 a.m. tomorrow"), physical interactions with a graphical interface (clicks on specific buttons, frequent navigation through certain menus), as well as stored data reflecting the user's explicit or implicit preferences (for example, favorite TV channels, usual settings for brightness, volume, or desired room temperature). This information can also include the user's recurring habits, such as regularly checking a specific application at set times each day.

[0045] The "appropriate response to relevant information determined from a user context," generated by the language model, can take several forms depending on the request and the specific user context. This response can be a direct textual response, such as explanatory or conversational information ("The current temperature in the living room is 22°C"), or a recommendation ("You should activate silent mode for your meeting"). It can also consist of a structured technical command, expressed as computer instructions directly interpretable by a device (for example, "turn on the television, channel 1, volume at 15"), or an explicit call to an application programming interface (API), allowing the automatic execution of a concrete action on a platform (for example, "call the thermostat.set_temperature API with the parameter 21°C").Thus, the appropriate response is characterized by its immediate contextual relevance, which can be used either to directly inform the user or to automatically control a device or associated technical platform.

[0046] This process allows for the automatic generation of personalized responses based on the user's context, thereby improving relevance and significantly reducing the errors or hallucinations frequently observed in traditional language models. This approach ensures a better fit with the user's specific expectations and increases the overall effectiveness of user interactions.

[0047] "Relevant contextual information" can correspond, for example, to a prediction of the television channel the user is likely to want in a specific context (e.g., Channel 1 at 8:00 p.m. on Saturday night), to a specific application that the user is likely to open at a specific time or place (e.g., automatically opening the "Maps" application every morning at 8:00 a.m. when the user leaves home), or to the optimal configuration of home automation equipment in response to a situation detected by the context (e.g., automatically setting the heating temperature to 22°C when the user returns home after 6:00 p.m.).

[0048] The term "adaptive predictive model" refers to a computer model capable of automatically generating predictions based on collected contextual data, while continuously adapting its own parameters during the inference phase.

[0049] This process allows for the integration of diverse contextual information, such as environmental, platform, and historical interaction data, going beyond direct semantic similarity to better understand the user's implicit or implicit intentions, thereby increasing the relevance of the generated responses. Through continuous updating of the adaptive predictive model's parameters, the process progressively improves its predictions, constantly adapting to user habits, preferences, and evolving behaviors. Furthermore, the proactive integration of this contextual information significantly reduces the risk of hallucinations inherent in traditional models, ensuring consistent and technically reliable responses.Finally, this process is particularly suited to a wide variety of applications such as intelligent control of home equipment, simplified navigation in mobile applications or improved voice interaction with embedded systems, overall enhancing the user experience.

[0050] The adaptive predictive model can implement incremental learning allowing continuous updating of its parameters during an inference phase.

[0051] Incremental learning refers to a machine learning method in which the parameters of a predictive model are updated progressively and continuously as new data becomes available during the inference phase—that is, during the model's active use. Unlike traditional methods, which require a separate initial training phase, this approach allows the model to adapt directly and immediately to new contextual information or changes observed in the user environment.

[0052] Using incremental learning within the predictive model offers the advantage of continuous, dynamic adaptation to changes in user behavior or variations in the context of use. Through this ongoing updating of its parameters, the model can progressively improve the accuracy and relevance of its predictions, ensuring a consistent fit with the user's actual expectations, habits, and preferences. This not only allows for a more effective response to the user's immediate needs but also guarantees a high and stable quality of responses produced by the language model over time.

[0053] The process may include, during a user request: - obtaining basic contextual information from the user context including environmental data, data relating to a platform used and historical interaction data of the user (BAS CONT INF).

[0054] The process may include, during a user request:

[0055] - construct a modified instruction by combining at least the said user request and the said relevant contextual information (MOD PRPT),

[0056] the response generation is obtained by applying the modified instruction as input to the response generator.

[0057] The term "modified instruction" refers to an enriched input provided to the language model, incorporating both the user's initial request and relevant contextual information determined by the adaptive model. This modified instruction allows the language model to produce a more targeted response, taking into account not only the user's explicit request but also implicit contextual information from the immediate usage context.

[0058] This basic contextual information can be transformed into a feature tensor according to a predefined feature model, with the relevant contextual information being determined by the adaptive predictive model using the feature tensor.

[0059] The term "feature tensor" refers to a multidimensional numerical data structure used to represent, in a structured way, the various contextual information collected. Each dimension of the tensor can correspond to a particular category or type of contextual data (e.g., temporal, spatial, historical, or platform-related).

[0060] The "predefined feature model" refers to a predefined specification determining what contextual information is relevant to capture and how to represent it numerically in the tensor.

[0061] As an example, temporal (current time), spatial (GPS location) and historical (last TV channel watched) data can be numerically encoded in different dimensions of the tensor. For example, the current time could be encoded in a first dimension ("8pm" = value 20), the GPS location in a second and a third dimension (latitude and longitude respectively), and the user history in another dimension ("last channel watched" = channel 1, coded for example with the value 1).

[0062] This tensor thus formed allows the adaptive predictive model to immediately have a compact, accurate, and structured representation of all the contextual information necessary for prediction.

[0063] Using a feature tensor enables an efficient and actionable representation of contextual data, thus facilitating its algorithmic processing by the adaptive predictive model. This approach, in particular, accelerates predictive calculations, improves the accuracy of the relevant contextual information determined, and ensures greater flexibility in integrating new features. By structuring contextual data as a predefined tensor, the process significantly improves the operational efficiency of the adaptive model and the relevance of the resulting predictions.

[0064] The adaptive predictive model includes a continuous updating of conditional probabilities associated with relevant contextual information based on said contextual information.

[0065] In particular, the adaptive predictive model may include a Bayesian model allowing the continuous updating of conditional probabilities associated with relevant contextual information based on said contextual information.

[0066] A Bayesian model is a probabilistic model that explicitly represents the relationships of interdependence between various contextual information in the form of conditional probabilities. In this context, conditional probabilities are quantitative measures that reflect the probability that an event or contextual characteristic is relevant given the presence of other contextual characteristics.

[0067] The incremental operation of the Bayesian model consists of a continuous and automatic update of these conditional probabilities each time new basic contextual information is captured and extracted. This continuous update allows the model to constantly reflect the current state of the usage context as well as the evolution of the user's habits or preferences.

[0068] More precisely, a Bayesian model is a probabilistic model based on the application of Bayes' theorem to automatically calculate predictions from contextual input information. In particular, this model relies on the determination and continuous updating of conditional probabilities.

[0069] Formally, for an output variable, here a relevant contextual piece of information, representing the desired prediction, and a set of explanatory variables (here the basic contextual information) denoted The Bayesian model calculates the following conditional probability:

[0070]

[0071]

[0072] Or :

[0073] - is the probability being sought, that is, the probability that the relevant event or contextual information occur given the basic contextual information .

[0074] - is the likelihood (probability of basic contextual information) knowing the occurrence of ).

[0075] - is the prior probability of , regardless of basic contextual information .

[0076] - is the total probability of the observed basic contextual information.

[0077] The model is called "adaptive" because the conditional probabilities are continuously updated through an incremental learning process. Thus, each time new contextual observations appear (for example, new user actions or choices), the conditional probabilities are updated according to Bayes' rule:

[0078]

[0079] Or :

[0080] - represents the new observations,

[0081] - the previously known observations.

[0082] For example, consider a concrete case of predicting a user's favorite television channel at a specific time. Let's assume the following explanatory variables (basic contextual information):

[0083] - : current time (for example 8:00 PM),

[0084] - : day of the week (for example, Saturday),

[0085] - : identified user (for example Sophie).

[0086] Let the variable to be predicted representing Sophie's favorite channel.

[0087] The Bayesian model uses learned conditional probabilities, for example:

[0088]

[0089]

[0090]

[0091] These probabilities are initially estimated based on historical data and updated as the user interacts with the system. When the model receives the current inputs (Saturday, 8 PM, Sophie), it directly calculates the string with the highest probability, in this case string 1 with an 80% probability.

[0092] If, for example, the user starts to consistently prefer channel 2 over the course of several Saturdays, the conditional probability will be gradually adjusted towards:

[0093]

[0094] as new observations are made. This mechanism allows the model to remain relevant and responsive to the actual evolution of user preferences, thus ensuring high accuracy and adaptability of predictions over time.

[0095] The adaptive predictive model may include an adaptive reinforcement learning model, updating its parameters based on positive or negative feedback received from the user environment.

[0096] The adaptive reinforcement learning model is a model that continually learns to select optimal actions or predictions based on positive or negative feedback received through interaction with its user environment.

[0097] Reinforcement learning is an iterative process, where the model selects one action from several possible ones based on the observed state (here, basic contextual information), then receives feedback (reward or penalty) indicating the relevance or irrelevance of that action, and then adjusts its internal parameters to maximize future rewards.

[0098] More specifically, the functioning of such a reinforcement learning model can be mathematically represented by a Markov decision process (MDP).

[0099] Such a process primarily includes the following elements:

[0100] - a set of observable states corresponding here to basic contextual information,

[0101] - a set of possible actions representing the relevant contextual information that can be provided,

[0102] - a reward function quantifying the immediate usefulness of carrying out the action in the state ,

[0103] - a policy , defined as a function that associates each state to an action ,

[0104] - and a state transition function indicating the probability of evolution from one state to another.

[0105] At each instant, the model observes a current state , corresponding to the current basic contextual information, and chooses an action according to a policy Following the execution of this action (for example, automatically recommending a particular application on a smartphone), the system receives an immediate reward. This reward can be determined by explicit or implicit user feedback, such as validating or rejecting the proposed recommendation.

[0106] As a concrete example, let's consider an adaptive recommendation system on a smartphone using adaptive reinforcement learning. When the user wakes up at 7:30 am (contextual state ) and picks up their smartphone, the model must choose the most relevant application to display immediately on the home screen (action Initially, the model offers a weather application, but the user prefers to open their calendar directly. The initial incorrect choice results in negative feedback ( ), which leads the model to gradually reduce the probability of displaying this weather application in this specific context. Conversely, when the model correctly recommends the calendar in a similar situation later, it receives positive feedback ( ), thus progressively increasing the future probability of displaying the agenda under these conditions.

[0107] Mathematically, updating the model parameters is generally done using a method called "incremental reinforcement learning," typically using a method like Q-learning. In this case, each state-action pair is assigned a value called the "Q value" ( ), which represents the cumulative expected value of future rewards obtained by performing the action in the state .

[0108] With each new interaction with the environment, these Q values ​​are updated according to the following rule:

[0109]

[0110] Or :

[0111] - is a learning rate controlling the model's adaptation speed,

[0112] - is a discounting factor determining the importance given to future rewards,

[0113] - is the value maximum possible in the following state .

[0114] The major advantage of this adaptive approach is the model's ability to continuously adjust its predictions based on the user's actual and evolving preferences. Indeed, unlike traditional static approaches, this model improves directly through interaction with the user environment, dynamically adapting its choices to observed real needs while remaining capable of responding to sudden or gradual changes in behavior or habits. This ensures a high level of relevance and personalization in the predictions provided to the language model, contributing to significantly more tailored and accurate responses, which are therefore better accepted by end users.

[0115] The step of constructing the modified instruction may further include a selection of information from at least one reference document carried out by augmented generation by retrieval, said selection being based on a semantic similarity with the user query.

[0116] The step of constructing the modified instruction can thus include an automatic selection of information from at least one reference document, carried out by a retrieval-augmented generation technique (known as the "RAG" technique, for "Retrieval-Augmented Generation").

[0117] In practice, this technique relies on a prior representation of the reference document in a vector format suitable for semantic similarity calculations with the user query. Typically, the reference document is divided into text fragments ("chunks"), each converted into a digital representation vector ("embedding") by a specialized model, such as a Transformer model (e.g., Sentence-BERT, GPT, etc.). Similarly, the user query is converted into a representation vector using the same procedure.

[0118] Once these vector representations are obtained, the selection of relevant information from the document is performed by calculating a vector similarity, typically using a metric such as cosine similarity. Formally, if the vector representing the user query is noted and each vector representing a fragment of the document is noted Semantic similarity can be calculated according to the following relationship:

[0119] ​

[0120] Document fragments exhibiting a high degree of vector similarity to the user query (e.g., exceeding a predefined threshold) are automatically selected. These selected fragments represent information semantically related to the user query and are automatically added to the modified instruction passed to the language model, in addition to the relevant contextual information determined elsewhere.

[0121] For example, if a user asks "How do I configure my smart thermostat?", the RAG technique will perform an automated semantic search in a reference document (e.g., the thermostat's digital user manual). From all the fragments available in the manual, those with a high semantic similarity to the query (e.g., "initial setup of the smart thermostat" or "advanced settings of the smart thermostat") will be automatically identified, selected, and incorporated into the instruction provided to the language model.

[0122] The technical advantages of this RAG technique are numerous. First, by enriching the instruction with semantically relevant information from one or more reference documents, the language model has additional up-to-date and accurate knowledge, considerably reducing the risks of hallucination or generation of erroneous responses.

[0123] Furthermore, this approach allows the system to provide accurate and contextualized answers, even if the model itself has not been directly trained on the specific information contained in the reference documents under consideration.

[0124] Finally, the use of automated selection based on semantic similarity allows for rapid, dynamic and targeted extraction of relevant information, thus significantly improving the relevance, consistency and overall quality of the responses generated by the model.

[0125] The modified instruction construction step can combine information selected by augmented generation by retrieval, with relevant contextual information determined by the adaptive predictive model.

[0126] For example, for a user request "Turn on the television", the process uses the RAG technique to select, from a technical document, the instructions to turn on the device, while simultaneously integrating a contextual prediction such as the current favorite television channel (e.g. "Channel 1 at 8pm"), determined by the adaptive model.

[0127] This combination improves the quality and relevance of the generated responses, offering both technical accuracy and dynamic personalization tailored to the user's immediate expectations.

[0128] Environmental data can include temporal and / or spatial data.

[0129] By "temporal data" we mean information describing or indicating a position or reference in time, such as the current time, the precise date, the day of the week, or a specific period of the day (for example: "Monday 8:30 a.m.", "evening", "weekend", "school holiday period").

[0130] "Spatial data" refers to information relating to a physical location or position, such as precise geographic coordinates provided by a GPS system, the identification of a specific place ("user's home", "office", "living room"), or the detected presence of the user in a particular room.

[0131] For example, a temporal environmental data point could be "Friday at 6:00 PM," while a spatial data point could be "user present in the living room." This information allows the process to accurately take into account the spatiotemporal context of use in order to generate appropriate responses.

[0132] Platform data may include information describing the type or condition of an electronic device used by the user.

[0133] The term "type of electronic device" refers to a technical identification of the device used, such as smartphone, tablet, connected TV, smart speaker, home automation device or smartwatch.

[0134] The term "state of an electronic device" refers to information concerning the current conditions of use or operation of the device, such as the current battery charge level ("battery at 20%"), the state of the network connection ("Wi-Fi connection active", "weak network signal"), the applications currently open or active ("Netflix application playing"), or the hardware or software version used ("smartphone running Android version 13").

[0135] This data allows the process to take into account precisely the technical and functional characteristics of the device in order to finely adapt the responses generated to the user request.

[0136] The step of generating a response may include a command for an action on said platform.

[0137] Such a command can, for example, be carried out in the form of a call to an application programming interface, or API, with parameters determined by said generated response.

[0138] The term "application programming interface" or API refers to a predefined software interface that allows the system to automatically send operational commands to an external device or service.

[0139] For example, when a generated response is "set the living room thermostat to 21°C", the process can automatically trigger a command, such as a call to the connected thermostat's API, with the parameter "temperature = 21". Similarly, a response such as "display channel 1 on the television" can lead to a command, for example, through a call to the television's API with specific parameters (e.g., "channel = 1").

[0140] This ensures the direct, reliable and precise execution of concrete technical actions on the platform used by the user.

[0141] The feature tensor resulting from the transformation step can be stored in a contextual feature database, said database being continuously updated by new user interactions.

[0142] After each interaction with the user, the captured contextual data can thus be structured and saved as a tensor in a dedicated database.

[0143] This database then progressively accumulates tensors representing the complete history of usage contexts encountered by the system, allowing the adaptive predictive model to have at all times a complete and up-to-date representation of user behavior and the associated context.

[0144] The advantage of this continuous storage in a dedicated database is twofold. Firstly, it allows for the efficient and immediate consideration of evolving user behavior, ensuring rapid adaptation of the model to observed changes. Secondly, it significantly improves the model's predictive accuracy through the progressive enrichment of its knowledge base.

[0145] Thus, with each new user interaction, the adaptive model can leverage all previously encountered contexts to refine its predictions. This dynamic mechanism ensures continuous improvement in the relevance and effectiveness of the responses provided by the language model, thereby optimizing the overall user experience.

[0146] Laillustrate in more detail a form of realization of the process of automatic generation of adapted responses by a large natural language model (LLM), enriched by an adaptive predictive model.

[0147] The process begins with the reception of a user request (UTT), for example, in the form of an explicit command such as "Turn on the TV." Simultaneously, a complete user context is captured, including environmental data (ENV), data relating to the technical platform used (PLTF), data characterizing the current task (TSK), and user-specific data (USR), such as interaction history, preferences, or specific usage conditions like geographic location or current time. This contextual data is captured and then processed (DTC) according to a predefined feature model (FTR MOD) to form a structured tensor (FTV).

[0148] The adaptive predictive model (PRED), which implements incremental learning, uses this feature tensor to perform inference (INF) in order to select relevant contextual information (RELC). For example, in the context of a query such as "Turn on the TV," this model could determine with 90% probability the user's preference for the TF1 television channel at that precise moment.

[0149] This incremental learning (ILN), for example Bayesian or reinforcement learning, allows the adaptive model to continuously update its parameters as it receives new contextual data from successive interactions with the user. Thus, contextual predictions become increasingly accurate over time, constantly adapting to the user's evolving habits and preferences.

[0150] This relevant contextual information is then combined with a dynamically selected ontology step using a context-retrieval augmented generation (OSEL) technique. This step involves extracting information from a structured knowledge base (ONT), which precisely describes the available technical actions (e.g., the technical command to select string 1). All of this information constitutes a modified or enriched instruction (PRPT), which is then passed to the natural language model (LLM) for the generation (GEN) of an automatic response (OUT).

[0151] The modified instruction also includes additional data (ADDDTA) which may include specific instructions to guide the language model's behavior, few-shot learning, or explicit supplementary instructions to enhance the relevance and technical accuracy of the generated responses. Through this integration of implicit and explicit contexts, the model produces a tailored response, often in the form of a structured technical command that can be directly executed by the device (for example, activating the command "set-tv-on(channel=1)").

[0152] This embodiment of the process according to the invention thus allows for better consideration of the user's implicit intentions, considerably improves the contextual accuracy of the responses generated by the language model, and significantly reduces the hallucination phenomena encountered with classical models.

[0153] Laillustre one embodiment of the response generator, according to one embodiment of the invention.

[0154] The response generator (1) by a language model is capable of generating a response adapted to relevant information determined from a user context using the language model.

[0155] In particular, the response generator (1) may include one or more of the following devices:

[0156] - a user interface (10) capable of receiving a user request,

[0157] - a receiver (11) of at least one captured contextual information,

[0158] - an extractor (12) of at least one contextual piece of information from a captured or received user context

[0159] - an analyzer (13) capable of determining contextual information by means of an adaptive predictive model,

[0160] - an instruction generator (14) based on the user request and contextual information,

[0161] - a computer (15) capable of implementing the language model according to the generated modified instruction, the computer providing said response.

[0162] A particular embodiment of the generator is a computer device. The computer device includes a processing circuit for implementing the method according to the invention.

[0163] The term "computer device" refers to any electronic device capable of automatically executing a set of software instructions, such as a personal computer, a computer server, an embedded device or a smart mobile device (e.g., smartphone, tablet or connected home automation device).

[0164] The term "processing circuit" refers to a hardware and / or software unit, such as a processor (CPU), a graphics processing unit (GPU), a microcontroller, or any other programmable electronic component, configured to perform the operations necessary to carry out the claimed process.

[0165] The computer system may include a user interface configured to receive a user request.

[0166] The computer system may include one or more sensors configured to capture environmental data.

[0167] The computer system may include an acquisition module configured to collect data relating to an electronic platform used by the user.

[0168] The computer system may include memory configured to store a history of user interactions.

[0169] The computer system may include a control module configured to transmit the generated response in the form of executable instructions.

[0170] The computer system may include a user terminal configured to transmit the user request to a remote server, said server being configured to run the response generator and return the appropriate response to said user terminal.

[0171] The user terminal can be a personal computer or a mobile device, such as a smartphone, for example.

[0172] The computer system may include a server configured to centralize the acquisition of contextual information captured by one or more clients, and to dynamically update the adaptive predictive model based on this contextual information.

[0173] The IT system may include one or more clients configured to collect local environment and platform data, transmit this data to a server, and receive responses tailored to their context of use.

Claims

Method of generating a response by means of a response generator comprising a language model, said method comprising, in response to a user request (USR REQ): generating a response adapted to relevant information determined from a user context using the response generator (RESP GEN). Method according to the preceding claim, wherein the method comprises:- determining relevant contextual information by means of an adaptive predictive model applied to basic contextual information (PRED) from the user context including environmental data, data relating to a platform used and historical interaction data of the user (BAS CONT INF); Method according to the preceding claim, wherein the adaptive predictive model implements incremental learning allowing continuous updating of its parameters during an inference phase. A method according to any one of claims 2 or 3, wherein said basic contextual information is transformed into a feature tensor according to a predefined feature model, the relevant contextual information being determined by the adaptive predictive model using the feature tensor. A method according to the preceding claim, wherein the adaptive predictive model includes a continuous updating of conditional probabilities associated with relevant contextual information based on said contextual information. A method according to any one of claims 2 to 5, wherein the adaptive predictive model includes an adaptive reinforcement learning model, updating its parameters based on positive or negative feedback received from the user environment. A method according to any one of claims 2 to 6, wherein the method comprises, upon a user request: - constructing a modified instruction by combining at least said user request and said relevant contextual information (MOD PRPT), the generation of the response being obtained by applying the modified instruction as input to the response generator. Method according to claim 4, characterized in that the step of constructing the modified instruction further comprises a selection of information from at least one reference document carried out by an augmented generation by retrieval, said selection being based on a semantic similarity with the user query. Method according to the preceding claim, wherein the modified instruction construction step combines the information selected by the augmented generation by retrieval, with the relevant contextual information determined by the adaptive predictive model. A method according to any one of the preceding claims, wherein the data relating to the platform includes information describing the type or state of an electronic device used by the user. A method according to any one of the preceding claims, wherein the step of generating a response includes a command for an action on said platform. A method according to claim 5, wherein the feature tensor resulting from the transformation step is stored in a contextual feature database, said database being continuously updated by new user interactions. Response generator including a language model, the generator being able to generate a response (RESP GEN) adapted to relevant information determined from a user context in response to a user request. Computer device comprising a processing circuit for implementing the process according to any one of claims 1 to 12. Computer program comprising instructions for implementing the method according to any one of claims 1 to 12, when this program is executed by a processor of a processing circuit.

Images