Method for controlling the rendering of text generated from textual data originating from at least one embedded system of a vehicle.

Abstract

The present invention relates to a method and device for controlling the rendering of text generated from textual data originating from at least one vehicle's embedded system, said method comprising the following steps: receiving (51) textual data from said at least one embedded system; generating (52) the output text of a language model trained during a learning phase and fed as input the received textual data; obtaining (53) a confidence score output from a confidence score determination model trained during the learning phase and fed as input the textual data and the text generated as output from the trained language model; and rendering (54) the text generated as output from the trained language model from a vehicle's embedded system as a function of the confidence score obtained. Figure 5

Description

Title of the invention: Method for controlling the rendering of a text generated from textual data from at least one embedded system of a vehicle. technical field

[0001] The present invention relates to generative artificial intelligence, that is to say, to artificial intelligence systems capable of generating text, images or other media in response to prompts.

[0002] In particular, the present invention relates to generative artificial intelligence used to generate text in various fields such as vehicle driver assistance systems. Technological background

[0003] Modern vehicles incorporate an increasing number of on-board systems that contribute to vehicle control by assisting the driver and / or to the comfort of the driver and vehicle passengers. Some of these systems are known as ADAS (Advanced Driver-Assistance System).

[0004] Some AD AS systems can use artificial intelligence to implement generators and other natural language processing according to particular road situations.

[0005] The emergence of so-called Large Language Models (LLMs) has recently revolutionized the field of natural language processing and generative artificial intelligence because these models are capable of generating texts close to those that a human can generate, answering questions, translating texts into different languages, and even writing poetry. They have been used in various applications ranging from chatbots to writing assistants (Metz, C. (2021). AI language model GPT-3 reproduces human-like biases. MIT Technology Review).

[0006] However, one problem with LLM models is their tendency to hallucinate (fabricate), that is, these LLM models respond with a text (a sequence of words) that is clearly false but is presented as a certain fact. This problem can be significant, particularly in applications where the accuracy and reliability of the generated texts are essential (McCoy, T., Pavlick, E., & Linzen, T. (2021). Right for the Wrong Reasons: Diagnosing Syntactic Heuristics in Natural Language Inference. arXiv preprint arXiv:2104.08696).

[0007] The fabrication of LLM models is related to the more general problem of neural networks known as out-of-distribution (OOD) generalization. Indeed, like any trained neural network, LLM models perform poorly during the interference phase when the textual data presented to them deviates from the (statistical) probability distributions of the textual data used for their training.

[0008] However, LLM models can also respond to texts that may not correspond to the textual data presented as input. This is an identified problem with LLM models, which is similar to quantifying the uncertainty in the response of these LLM models.

[0009] There are several ways to quantify the uncertainty of neural networks, but they generally require training these networks multiple times with different parameter configurations, such as the Monte Carlo method described in the article by Gai, Y, & Ghahramani Z (“Dropout as a Bayesian Approximation: Representing Model Uncertainty in Deep Learning”. 2016. Proceedings of the 33rd International Conference on Machine Learning. https: / / arxiv.org / abs / 1506.02142. Other methods use Bayesian neural networks based on parameter distributions, such as the method described in the article by Tishby, Levin, and Solia (“Consistent inference of probabilities in layered networks: predictions and generalizations.” In: International 1989 Joint Conference on Neural Networks, 403-409 vol.2, 1989).

[0010] Prior art methods succeed in quantifying uncertainty, some even offering theoretical guarantees, but they all require computing resources that are proportional to the number of parameters. Consequently, they become more expensive as the number of parameters in the LLM model increases, making them highly unsuitable for large LLM models.

[0011] It is therefore necessary to define a method which quantifies the response uncertainty of an LLM model which does not require computing resources which are proportional to the number of parameters of these LLM models, i.e. to the size of these LLM models. Summary of the present invention

[0012] One object of the present invention is to solve at least one of the problems of the technological background described above.

[0013] Another object of the present invention is to improve the performance of a deep neural network used for the implementation of an LLM model.

[0014] Another object of the present invention is to provide a response uncertainty of an LLM model.

[0015] According to a first aspect, the present invention relates to a method for controlling the rendering of text generated from textual data originating from at least one embedded system of a vehicle, said method comprising the following steps: - receiving textual data from said at least one embedded system; - generation of output text from a language model trained during a learning phase and fed input by the received textual data; - obtaining a confidence score output from a confidence score determination model trained during the learning phase and fed input by the textual data and the text generated output from the trained language model; and - rendering the text generated output from the trained language model from an on-board system of the vehicle according to the confidence score obtained.

[0016] The method is advantageous because it allows the uncertainty of the text generated by a language model trained during a learning phase to be quantified, based on a confidence score calculated from a confidence score determination model trained during the learning phase. The parameters of the language model are thus adjusted to optimize the confidence scores associated with the texts generated by the language model during training, thereby limiting the uncertainties in the responses provided by this language model. The method then makes it possible to generate texts that can be used in applications implemented by a vehicle such as an ADAS system.

[0017] Furthermore, by rendering generated text based on a confidence score, a user can quickly be alerted to the veracity of the generated text. Indeed, users can quickly identify potentially unreliable generated texts based on the confidence score values. This advantage is all the more valuable in applications where accuracy is essential, such as, for example, a cockpit assistant implementing the process.

[0018] The method is also advantageous because the confidence scores are determined from internal states of the language model and does not require the generation of several texts for the same textual data in order to determine a probability of confabulation as is the case in the prior art.

[0019] According to a particular and non-limiting embodiment of the present invention, during the training phase, texts are generated as output from the language model to be trained when a set of training text data is presented as input to the language model; internal parameters of the language model to be trained are adjusted to minimize a first loss function quantifying the differences between the texts generated as output from the language model to be trained and the expected texts defined for the training text data set; confidence scores are generated as output from the model of confidence score determination to be trained when the training text dataset and the generated output texts of the language model to be trained are presented as input to the confidence score determination model to be trained; and internal parameters of the confidence score determination model to be trained are adjusted to minimize a second loss function quantifying differences between the generated output confidence scores of the confidence score determination model to be trained and the expected confidence scores that are defined from the generated output texts of the language model to be trained and the expected texts defined for the training text dataset.

[0020] According to a particular and non-limiting embodiment of the present invention, the method further comprises a step of obtaining a first vector representation, a second vector representation and a third vector representation defined in the same vector space from a model of projection of textual data in the vector space; - the first vector representation being a representation of the set of textual data as input to the language model; - the second vector representation being a representation defined in the vector space of the text generated as output from the language model; - the third vector representation being a vector representation of the expected and defined texts for the set of textual training data; - the first and second vector representations being presented as input to the confidence score determination model to be trained and the expected confidence scores being defined from the second and third vector representations.

[0021] According to a particular and non-limiting embodiment of the present invention, the language model is a large language model comprising a deep neural network whose input is fed with tokens obtained by converting textual data present as input to the language model and adapted to provide as output tokens that correspond to the text generated once these tokens have been converted into textual data.

[0022] According to a particular and non-limiting embodiment of the present invention, the language model is an autoregressive model based on a transformer-type architecture.

[0023] According to a particular and non-limiting embodiment of the present invention, the confidence score determination model is a deep neural network.

[0024] According to a second aspect, the present invention relates to a device for controlling the rendering of text generated from textual data originating from at least one embedded system of a vehicle, the device comprising a memory associated with a processor configured for the implementation of the steps of the process according to the first aspect of the present invention.

[0025] According to a third aspect, the present invention relates to a vehicle, for example of the automobile type, comprising a device as described above according to the second aspect of the present invention.

[0026] According to a fourth aspect, the present invention relates to a computer program which includes instructions adapted for carrying out the steps of the process according to the first aspect of the present invention, in particular when the computer program is executed by at least one processor.

[0027] Such a computer program may use any programming language, and be in the form of source code, object code, or an intermediate code between source code and object code, such as in a partially compiled form, or in any other desirable form.

[0028] According to a fifth aspect, the present invention relates to a computer-readable recording medium on which is recorded a computer program comprising instructions for carrying out the steps of the process according to the first aspect of the present invention.

[0029] On the one hand, the recording medium can be any entity or device capable of storing the program. For example, the medium can include a storage means, such as a ROM, a CD-ROM or a microelectronic circuit-type ROM, or a magnetic recording means or a hard disk drive.

[0030] On the other hand, this recording medium can also be a transmissible medium such as an electrical or optical signal, such a signal being able to be transmitted via an electrical or optical cable, by conventional or radio frequency, by self-directing laser beam, or by other means. The computer program according to the present invention can, in particular, be downloaded from an Internet-type network.

[0031] Alternatively, the recording medium may be an integrated circuit in which the computer program is incorporated, the integrated circuit being adapted to execute or to be used in the execution of the process in question. Brief description of the figures

[0032] Other features and advantages of the present invention will become apparent from the description of the particular and non-limiting embodiments of the present invention below, with reference to the attached Figures 1 to 5, in which:

[0033] [Fig-1] schematically illustrates an example of a rendering control process of a text generated from textual data originating from at least one embedded system of a vehicle according to a particular and non-limiting embodiment of the present invention.

[0034] [Fig.2] schematically illustrates a language model learning phase and a confidence score determination model according to a particular and non-limiting embodiment example of the present invention.

[0035] [Fig.3] schematically illustrates a deep neural network based on a transformer-type architecture according to a particular and non-limiting embodiment of the present invention.

[0036] [Fig.4] schematically illustrates a device configured to control the rendering of text generated from textual data from at least one embedded system of a vehicle, according to a particular and non-limiting embodiment of the present invention.

[0037] [Fig.5] illustrates a flowchart of the different stages of a process for controlling the rendering of a text generated from textual data from at least one embedded system of a vehicle, according to a particular and non-limiting embodiment of the present invention. Description of examples of achievements

[0038] A method and a rendering control device for text generated from textual data from at least one vehicle's embedded system will now be described in the following with joint reference to Figures 1 to 5. The same elements are identified with the same reference signs throughout the following description.

[0039] A vehicle, according to the present invention, corresponds, for example, to a vehicle with an internal combustion engine, with electric motor(s), or even a hybrid vehicle with an internal combustion engine and one or more electric motors. The vehicle thus corresponds, for example, to a land vehicle, such as a car, a truck, a bus, or a motorcycle.

[0040] The vehicle includes a device 2 of the [Fig.4].

[0041] The vehicle includes at least one embedded system which is configured to provide textual data, i.e. either a word or a sequence of words from natural language.

[0042] According to a particular and non-limiting embodiment of the present invention, the textual data may be representative of a state of at least one component of the vehicle and / or of a road situation in which the vehicle is located.

[0043] According to a particular and non-limiting embodiment of the present invention, one of these embedded systems may include a touchscreen and memory configured for input and recording natural language words. Another An example might be a voice recording system comprising a microphone and memory.

[0044] According to a particular and non-limiting embodiment of the present invention, an on-board vehicle system may be an ADAS system. For example, the system may be a road sign detection system that provides word sequences as soon as it detects a road sign. For example, a word sequence 'speed limit 80 km / h' may be provided by this ADAS system as soon as an 80 km / h speed limit sign is detected. Another ADAS system may provide word sequences as soon as weather conditions change.

[0045] These examples of AD AS systems are given only by way of example and the present invention extends to any type of AD AS system which would provide sequences of words according to a state of at least one organ of the vehicle and / or a road situation in which the vehicle is located.

[0046] According to a particular and non-limiting embodiment of the present invention, an embedded vehicle system may be a display system comprising interface means with a vehicle user that enable the retrieval of textual data and the audio rendering of text. These interface means may, for example, include a touchscreen, a microphone and / or a speaker controlled by a computer, or a voice control system comprising a microphone and / or a speaker. The embedded system may include a memory configured to store natural language words.

[0047] Fig. 1 schematically illustrates an example of a process for controlling the rendering of text generated from textual data from at least one embedded system of a vehicle according to a particular and non-limiting embodiment of the present invention.

[0048] This process corresponds to an inference phase of a language model 100 trained during a learning phase and a confidence score determination model 400 also trained during said learning phase.

[0049] In a first operation, a text data 101 is received from at least one on-board system of the vehicle. The received text data 101 can be a word or a sequence of words in natural language or any other representation of a word or a sequence of words in natural language such as tokens.

[0050] In a second operation, a text 102, that is to say a sequence of at least one natural language word or at least one token, is generated as output of the language model 100 trained during a learning phase when the language model 100 is fed as input by the received textual data 101.

[0051] In a third operation, a confidence score 402 is obtained as output from a confidence score determination model 400 trained during the phase learning when the trained confidence score determination model 400 is fed input by text data 101 and text 102 generated as output from the trained language model 100.

[0052] In a fourth operation, the generated text 102 output from the trained language model 100 is rendered from an embedded system 200 of the vehicle according to the confidence score 402 obtained.

[0053] For example, the generated text 102 can be rendered on an in-vehicle display with a brightness or color that depends on the confidence score 402 obtained. An icon can be added to the text displayed on a screen when the confidence score 402 exceeds a threshold value, for example.

[0054] According to one variant, the process may further include an operation of comparing the confidence score 402 obtained to a threshold value, and the generated text 102 may be rendered according to a result of the comparison of the confidence score 402 obtained and the threshold value.

[0055] According to a particular and non-limiting embodiment of the present invention, the threshold value is determined during a validation phase of the language model by an operator.

[0056] Fig. 2 schematically illustrates a learning phase of the language model 100 and the confidence score determination model 400 according to a particular and non-limiting embodiment of the present invention.

[0057] In a first operation of the learning phase, texts 104 can be generated as output of the language model 100 to be trained when a set of textual training data 103 is presented as input to the language model 100 to be trained.

[0058] In a second operation of the learning phase, internal parameters of the language model 100 to be trained can be adjusted to minimize a first loss function 502 quantifying (500) the differences between the texts 104 generated as output of the language model 100 to be trained and the texts 501 expected and defined for the training textual dataset 103.

[0059] In a third operation of the learning phase, confidence scores 403 can be generated as output of the confidence score determination model 400 to be trained when the training textual dataset 103 and the texts 104 generated as output of the language model 100 to be trained can be presented as input to the confidence score determination model 400 to be trained.

[0060] In a fourth operation of the learning phase, internal parameters of the confidence score determination model 400 to be trained can be adjusted to minimize a second loss function 702 quantifying (700) the differences between the confidence scores 403 generated as output from the confidence score determination model. confidence score 400 to be trained and expected confidence scores 602 which are defined (600) from the texts 104 generated as output from the language model 100 to be trained and the texts 501 expected and defined for the training textual dataset 103.

[0061] According to a particular and non-limiting embodiment of the present invention, in a fifth operation of the learning phase, a first vector representation (301 of [Fig.1], 303 of [Fig.3]), a second vector representation (302 of [Fig.1], 304 of [Fig.2]) and a third vector representation (305 of [Fig.2]) defined in the same vector space are obtained from a model 300 of projection of textual data in the vector space.

[0062] The first vector representation (301, 303) is a representation, defined in the vector space, of the set of textual data (101, 103) as input to the language model 100.

[0063] The second vector representation (302, 304) is a representation, defined in vector space, of the text 104 generated as output from the language model 100.

[0064] The third vector representation 305 is a vector representation, defined in the vector space, of the expected and defined texts 501 for the training textual data set 103.

[0065] The first (301, 303) and second (302, 304) vector representations are presented as input to the confidence score determination model 400 to be trained and the expected confidence scores 602 are defined (600) from the second (304) and third (305) vector representations.

[0066] This embodiment is advantageous because it allows the set of textual data 101, the generated texts 102 and the expected texts 501 to be expressed in the same vector space, thus allowing their differences to be quantified.

[0067] For example, model 300 is a sentence transformation model such as the trained neural network proposed by Reimers in 2019 commonly called "all-mpnet-base-v2". This neural network makes it possible to match sentences or paragraphs of text with a vector defined in a 768-dimensional vector space.

[0068] According to a particular and non-limiting embodiment of the present invention, the first loss function 502 is of the maximum likelihood type.

[0069] According to a particular and non-limiting embodiment of the present invention, the second loss function 702 is of the mean squared error type.

[0070] According to a particular and non-limiting embodiment of the present invention, the language model is an LLM (Large Language Model), that is to say, a language model that includes a deep neural network comprising a large number of internal parameters typically on the order of a billion or more and a converter of textual data into tokens and vice versa.

[0071] LLM models are mathematical functions whose input and output are lists of numbers (tokens). To enable them to accept textual data as input and provide textual output, a conversion is applied to this textual data to obtain tokens (numbers), and a conversion is applied to the tokens to obtain textual data from which the output texts of the LLM model are generated. This conversion is implemented by a bijective function (called a tokenizer) based on a predefined mapping (vocabulary). This bijective function is generally first fitted to a complete training textual dataset, then frozen, before the deep neural network is trained.Another characteristic of the bijective function is to compress the textual data presented as input to the LLM model to save computation time during the training and inference of the deep neural network.

[0072] According to a particular and non-limiting embodiment of the present invention, illustrated in [Fig. 3], the LLM model is an autoregressive LLM model based on a transformer-type architecture as described, for example, in the article by Vaswani et al. (Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A., Kaiser, U., & Polosukhin, I. (2017). Attention is All You Need. In Advances in Neural Information Processing Systems. Curran Associates, Inc.). The transformer architecture is a neural network model that relies on self-attention mechanisms to capture the relationships between tokens in a sequence, enabling parallel processing and efficient context learning.

[0073] Basically, an autoregressive LLM based on a transformer-type architecture comprises an encoder (or 'tokenizer') 110, a decoder (also called a 'tokenizer') 130, and a deep neural network 120. The encoder 110 encodes textual data x, such as words or sentences in natural language, which are present at its input to the LLM model, into a sequence of tokens y, i.e., numbers ('token'), which are present at the input to the deep neural network, based on a predefined mapping (also called a vocabulary) between words and / or sentences in natural language and tokens. The decoder 130 performs the inverse operation by decoding a sequence of tokens z (output of the deep neural network) to obtain textual data x, on the basis of which a text is generated.The deep neural network 120 operates on a sequence of tokens y as input (output of the encoder) and generates a new sequence of tokens z as output (input of the decoder).

[0074]

[0075]

[0076]

[0077]

[0078]

[0079]

[0080]

[0081]

[0082]

[0083]

[0084]

[0085] The 120 deep neural network comprises a set of artificial neuron layers. For example, each artificial neuron is a perceptron, that is, a linear classifier generally comprising several inputs and a single output, and characterized by an activation function, weights (or synaptic coefficients) and a bias (or threshold). For example, a perceptron with n inputs (x? ..., xn ) and a single output o is defined by nweights (Wj, ...,wn) and a bias (or threshold) 0; > 0 0 otherwise The output o then results from applying the Heaviside function to the postsynaptic potential z given by: Z = -6 with a non-linear activation function H(x) given for example by: „ _ _ „, .[0 if x < 0 VxeP,€ € H(x) = , . , It if x > 1 The internal parameters of the neural network 120 are then these weights and biases for the set of perceptrons. The present invention is not limited to this definition of perceptron nor to the use of other basic elements forming a layer of the neural network. It is also not limited by the number of perceptrons (or other basic elements) used per layer, nor by the number of layers. The internal parameters of the neural network are usually weights and biases, regardless of the basic elements of the neural network layers. During the training phase of a deep neural network, the network provides a set of output tokens when a set of tokens is presented as input. The input and output token sets correspond to words or sequences of words in natural language. The deep neural network is trained when it provides output token sets that correspond to expected words or sequences of words. More specifically, during the training phase, the internal parameters of the langal20 model's deep neural network are optimized using loss first estimation, which can be of the maximum likelihood type. The langal20 model calculates the probability distribution over a given vocabulary for each token in a training sequence present as input to the deep neural network, and the goal is to maximize the probability that the next 'true' token in the training sequence is predicted. This is achieved by minimizing a function of cross-entropy loss defined between the predicted probabilities and the next real token in the learning sequence.

[0086] According to a particular and non-limiting embodiment of the present invention, during the learning phase, the internal parameters of the deep neural network can be optimized by backpropagation and gradient descent based, for example, on the method described by Robbins and Monro (Robbins, H. and S. Monro (1951). “A Stochastic Approximation Method.” In The Annals of Mathematical Statistics 22.3, pp. 400-407).

[0087] The learning phase continues until the deep neural network of the language model 120 reaches a level of performance satisfactory for predicting the next token in a token sequence.

[0088] According to a particular and non-limiting embodiment of the present invention, the language model can undergo several learning phases: a pre-training phase on large public datasets such as the 'Common Crawl' database and a refinement phase on datasets dedicated to a specific application.

[0089] According to a particular and non-limiting embodiment of the present invention, the confidence score determination model is a deep neural network.

[0090] For example, the confidence score determination model is a neural network-type sentence transformation model which, during the learning phase, was trained using an optimizer commonly called Adam (Kingma DP and Ba, J. “Adam: A method for stochastic optimization,” in Proceedings of the 3rd International Conference on Learning Representations (ICLR) (2015).

[0091] When the textual dataset 103 and the generated texts 104 are comparable with each other then the confidence score determination model 400 can determine confidence scores 403 during the learning phase when the textual dataset 103 and the generated texts 104 are presented as input to the confidence score determination model.

[0092] If the textual dataset 103 and the generated texts 104 are not comparable to each other, then they are represented by the first 303 and second 304 vector representations, as explained previously, and the confidence scores 403 are determined during the learning phase when the first 303 and second 304 vector representations are presented as input to the neural network (confidence score determination model).

[0093] When the generated texts 104 and the expected texts 501 are comparable, then the expected scores 602 can be determined to quantify the differences between the generated texts 104 and the expected texts 501. When the generated texts 104 and the expected texts 501 are not comparable, then the generated texts 104 and the expected texts are represented by the second 304 and third 305 vector representations and the expected confidence scores 602 are determined from the second 304 and third 305 vector representations.

[0094] The neural network (confidence score determination model) is then trained to minimize the second loss function 702 between confidence scores 403 and confidence scores 602.

[0095] According to a particular and non-limiting embodiment of the present invention, the second loss function is of the mean squared error type defined by:

[0096]

[0097] Where S denotes the second loss function 702 between a generated confidence score 403 and b denotes an expected confidence score 602.

[0098] It is advantageous to define the second loss function 702 as the mean squared error function S above rather than a cosine function because the function S returns a scalar value between -1 and 1, where 1 means that a and b have identical semantics, 0 means that a and b are unrelated, and -1 means that a and b have opposite semantics. However, we are only interested in a similarity measure that ranges from "unrelated" to "identical," so it makes more sense to take the L2 (Euclidean norm) distance between a and b. Dividing this number by the sum of the norms and subtracting it from 1, the function S defines a normalized similarity score between 0 and 1, thanks to the triangle inequality.

[0099] [Fig.4] schematically illustrates a device 2 configured to control the rendering of text generated from textual data from at least one embedded system of a vehicle, according to a particular and non-limiting embodiment of the present invention.

[0100] Device 2 corresponds for example to a device embedded in the vehicle, for example a computer.

[0101] Device 2 is, for example, configured to implement at least one step of the process described opposite Figures 1 and 2. Examples of such a device 2 include, but are not limited to, embedded electronic equipment such as a vehicle's on-board computer, an electronic control unit such as an ECU (Electronic Control Unit), a smartphone, a tablet, or a laptop computer. The elements of device 2, individually or in combination, may be integrated into a single integrated circuit, into several integrated circuits, and / or into discrete components. Device 2 may be implemented in the form of electronic circuits or software (or computer) modules, or a combination of electronic circuits and software modules.

[0102] The device 2 comprises one (or more) processor(s) 20 configured to execute instructions for carrying out at least one step of the process and / or for executing instructions from the software embedded in the device 2. The processor 800 may include integrated memory, an input / output interface, and various circuits known to those skilled in the art. The device 2 further comprises at least one memory 801, for example, volatile and / or non-volatile memory, and / or includes a memory storage device that may include volatile and / or non-volatile memory, such as EEPROM, ROM, PROM, RAM, DRAM, SRAM, flash, magnetic disk, or optical disk.

[0103] The computer code of the embedded software(s) including the instructions to be loaded and executed by the processor is for example stored on memory 801.

[0104] According to various particular and non-limiting embodiments, the device 2 is coupled in communication with other similar devices or systems and / or with communication devices, for example a TCU (Telematic Control Unit), for example via a communication bus or through dedicated input / output ports.

[0105] The deep neural network (of the language model and / or the confidence score determination model) can be trained by a remote device, such as a server or a series of servers, communicating with device 2 via a communication network. The internal parameters of the deep neural network (of the language model 100 and / or the confidence score determination model 400) can then be transmitted by this remote device to device 2, which receives them. Device 2 is then able to implement the rendering control process when text data 101 from an embedded system of the vehicle is presented as input to the trained deep neural network (of the language model 100).

[0106] According to a particular and non-limiting embodiment of the present invention, the device 2 comprises an 802 block of interface elements for communicating with external devices, for example, a remote server or the cloud, other nodes of the ad hoc network. The interface elements of the 802 block comprise one or more of the following interfaces: - Radio frequency (RF) interface, for example, Wi-Fi® type (according to IEEE 802.11), for example in the 2.4 or 5 GHz frequency bands, or Bluetooth® type (according to IEEE 802.15.1), in the 2.4 GHz frequency band, or Sigfox type using UBN (Ultra Narrow Band) radio technology, or LoRa in the 868 MHz frequency band, LTE (Long- Long-term Evolution” or in French “Long-term Evolution”), LTE-Advanced (or in French LTE-advanced); - USB interface (from the English "Universal Serial Bus" or "Universal Serial Bus" in French); - HDMI interface (from the English "High Definition Multimedia Interface", or "High Definition Multimedia Interface" in French); - LIN interface (from the English "Local Interconnect Network", or in French "Réseau interconnecté local").

[0107] Data, for example internal parameters of the trained deep neural network (of the language model 100 and / or the confidence score determination model 400), are for example loaded into the device 2 via the interface of the 802 block using a Wi-Fi® network such as according to IEEE 802.11, an ITS G5 network based on IEEE 802.1 Ip or a mobile network such as a 4G (or 5G) network based on the LTE (Long Term Evolution) standard defined by the 3GPP consortium, in particular an LTE-V2X network.

[0108] According to another particular and non-limiting embodiment, the device 2 includes a communication interface 803 which enables communication with other devices (such as other computers in the embedded system) via a communication channel 804. The communication interface 803 corresponds, for example, to a transmitter configured to transmit and receive information and / or data via the communication channel 804. The communication interface 803 corresponds, for example, to a wired network of the CAN (Controller Area Network), CAN FD (Controller Area Network Flexible Data-Rate), FlexRay (standardized by ISO 17458) or Ethernet (standardized by ISO / IEC 802-3) type.

[0109] Internal parameters of the trained deep neural network (of the language model 100 and / or the confidence score determination model 400) can be received via the communication interface 803.

[0110] According to a particular and non-limiting embodiment of the present invention, the device 2 can provide output signals to one or more external devices, such as a display screen 805, touch or non-touch, one or more loudspeakers 806 and / or other peripherals 807 via output interfaces 808, 809 and 810 respectively. According to a variant, one or more of the external devices is integrated into the device 2.

[0111] Figure 5 illustrates a flowchart of the different stages of a process for controlling the rendering of text generated from textual data from at least one embedded system of a vehicle, according to a particular and non-limiting embodiment of the present invention.

[0112] In a step 51, text data is received from said at least one embedded system.

[0113] In a step 52, text is generated as output from a language model trained during a learning phase and fed as input by the received textual data.

[0114] In a step 53, a confidence score is obtained as output from a confidence score determination model trained during the learning phase and fed as input by the textual data and the text generated as output from the trained language model.

[0115] In a step 54, the text generated as output from the language model trained from an embedded system of the vehicle as a function of the confidence score obtained.

[0116] According to one variant, the variants and examples of the operations described in relation to Figures 1 to 3 apply to the steps of the process in [Fig. 5].

[0117] Of course, the present invention is not limited to the embodiments described above but extends to a method for controlling the rendering of text generated from textual data originating from at least one embedded system of a vehicle, which would include secondary steps without departing from the scope of the present invention. The same would apply to a device configured for implementing such a method.

[0118] The present invention also relates to a vehicle, for example an automobile or more generally an autonomous land-powered vehicle, comprising device 2 of [Fig.5].

Claims

Demands

1. A method for controlling the rendering of text generated from textual data from at least one vehicle-mounted system, said method comprising the following steps: - receiving (51) textual data from said at least one vehicle-mounted system; - generating (52) the output text from a language model trained during a learning phase and fed as input the received textual data; - obtaining (53) a confidence score from the output of a confidence score determination model trained during the learning phase and fed as input the textual data and the text generated as output from the trained language model; and - rendering (54) the text generated as output from the trained language model from a vehicle-mounted system as a function of the confidence score obtained.

2. A method according to claim 1, wherein during the training phase, - texts are generated as output from the language model to be trained when a set of training text data is presented as input to the language model to be trained; - internal parameters of the language model to be trained are adjusted to minimize a first loss function quantifying the differences between the texts generated as output from the language model to be trained and the expected texts defined for the set of training text data; - confidence scores are generated as output from the confidence score determination model to be trained when the set of training text data and the texts generated as output from the language model to be trained are presented as input to the confidence score determination model to be trained;and - internal parameters of the confidence score determination model to be trained are adjusted to minimize a second loss function quantifying differences between the confidence scores generated as output of the confidence score determination model to be trained and the expected confidence scores which are defined from the texts generated as output of the language model to be trained; train and the expected and defined texts for the set of training textual data.

3. A method according to claim 2, further comprising a step of obtaining (55) a first vector representation, a second vector representation and a third vector representation defined in the same vector space from a model of projection of textual data into the vector space; - the first vector representation being a representation of the set of textual data as input to the language model; - the second vector representation being a representation defined in the vector space of the text generated as output from the language model; - the third vector representation being a vector representation of the expected texts defined for the set of textual training data;- the first and second vector representations being presented as input to the confidence score determination model to be trained, and the expected confidence scores being defined from the second and third vector representations.

4. A method according to any one of the preceding claims, wherein the language model is a large-scale language model comprising a deep neural network whose input is fed with tokens obtained by converting textual data present as input to the language model and adapted to provide as output tokens that correspond to the text generated once these tokens have been converted into textual data.

5. A method according to claim 4, wherein the language model is an autoregressive model based on a transformer-type architecture.

6. A method according to any one of the preceding claims, wherein the confidence score determination model is a deep neural network.

7. A computer program comprising instructions for carrying out the method according to any one of the preceding claims, when such instructions are executed by a processor.

8. Computer-readable recording medium on which a computer program comprising instructions is recorded 19 for carrying out the steps of the process according to any one of claims 1 to 6.

9. Device (2) for controlling the rendering of text generated from textual data from at least one vehicle-mounted system, said device (2) comprising a memory associated with at least one processor configured for carrying out the steps of the method according to any one of claims 1 to 6.

10. Vehicle comprising device (2) according to claim 9.

Images