EP4272108A1

EP4272108A1 - Method for automatically understanding multiple instructions in natural language expressions

Info

Publication number: EP4272108A1
Application number: EP21848298.2A
Authority: EP
Inventors: Thierry Roger; Alexy CARLIER
Original assignee: Individual
Current assignee: Individual
Priority date: 2020-12-31
Filing date: 2021-12-31
Publication date: 2023-11-08
Also published as: WO2022144847A1

Abstract

A method is for automatically understanding expressions which represent a multiple instruction comprising at least two single instructions, each one representing an order. This method comprises, in the case of an expression being received, a first step (100) in which the multiple instruction represented by the expression received is automatically segmented into single instructions, and a second step (110) in which an application field is automatically determined from a group of application fields for each of the single instructions of the segmentation.

Description

Automatic understanding of multi-statements in natural language expressions

The present invention relates to methods (or methods) for the automatic understanding of textual expressions in natural language (resulting from spoken or written expressions), and more precisely to methods which are based on the automatic segmentation of multiple instructions contained in expressions (or sentences) in natural language.
As known to those skilled in the art, automatic multi-statement segmentation is a recent subfield of the general field of Spoken Language Understanding (SLU).
The automatic multi-statement segmentation method currently offering the "best" results is the AI-based method proposed by Rohan Shet in the article "Segmenting multi-intent queries for spoken language understanding" (published in 2019). It uses both word embedding (or "word embedding ElMo") which exploits the structure of words, recurrent neural networks (or RNNs ("Recurrent Neural Networks")) based on LSTM type architectures (" Long Short-Term Memory”), and a Named-Entity Recognition (NER) method to identify relevant textual objects (separator word, start of segment, end of segment). The NER method consists of extracting information from a text by searching for textual objects (of the word or group of words type) that can be classified into categories such as proper names, companies, places, quantities.
This method of automatic segmentation of multiple instructions does not provide complete satisfaction because of its fairly high error rate, and therefore cannot be used on its own to generate automatic scripts intended to control or feed automated functions. . This high error rate results from the fact that artificial intelligence does not know the context of each simple instruction and therefore requires a very large amount of quality learning data to configure its neural networks, which is very difficult to produce since they must relate to very many often very different applications and very many sets of instructions.
There are also two solutions marketed under the registered trademarks OkGoogle® and Siri® and managing very diverse fields of application. But OkGoogle® handles some multi-statements when they include single statements separated by the word "and", and Siri® apparently doesn't handle multi-statements at all.
It may thus be desirable to provide a process for the automatic understanding of textual expressions in natural language (resulting from spoken or written expressions) which makes it possible to overcome at least some of the aforementioned problems and constraints.

une première étape dans laquelle on segmente de façon automatisée l’instruction multiple représentée par l’expression reçue en instructions simples, et
une seconde étape dans laquelle on détermine de façon automatisée un champ d’application parmi un ensemble de champs d’application pour chacune des instructions simples de la segmentation.

a first step in which the multiple instruction represented by the expression received is segmented in an automated manner into simple instructions, and
a second step in which a scope of application is determined automatically from among a set of scopes for each of the simple instructions of the segmentation.

Thus, it is now possible to determine for each multiple instruction, represented by a received expression, simple instructions, then it is possible to determine the processing context targeted by each of these determined simple instructions, which makes it possible to significantly improve the understanding of the multiple instructions. possibly relating to different fields of application.
The comprehension method according to the invention may comprise other characteristics which may be taken separately or in combination, and in particular:
- in a first embodiment, in its first step, each elementary sub-part of said received expression can be compared with separators contained in a set of separators (predefined or configurable), in order to determine each separator contained in the received expression , then we can consider that the elementary sub-parts located before or after a determined separator are representative of a simple instruction;
- in this first embodiment, in the first step, a set of separators can be used comprising punctuation marks chosen from a group comprising a comma, a semicolon, a period and ellipsis and chosen separation words in another group comprising the words and, then, then, finally, after and also;
- in a second embodiment, in its first step, an artificial intelligence can be used to classify each elementary sub-part of the expression received in a category which is chosen from among a first category of separators and a second category of words other than separators, then to consider that the elementary sub-parts located before or after an elementary sub-part classified in the first category of separators are representative of a simple instruction;
- in a third embodiment, in its first step, an artificial intelligence can be used to determine in the expression received each elementary sub-part representing a beginning of a simple instruction and each elementary sub-part representing an end of a simple instruction , then to consider that the elementary sub-parts comprised between an elementary sub-part included and representing a beginning of simple instruction and another elementary sub-part included and representing an end of simple instruction following this beginning of simple instruction are representative of the same simple instruction;
- in the second or third embodiment, in the first step, artificial intelligence and a named entity recognition method can be used to classify each elementary sub-part of the expression received;
- in the presence of the very last option, in the first stage one can use artificial intelligence and the method of recognition of named entities in association with a so-called conditional random field layer to carry out the classification of each elementary sub-part of the expression received;
- in its second step, an artificial intelligence can be used to determine a context for each simple instruction, then to determine the scope of application of each simple instruction according to the context determined for the latter.

à segmenter de façon automatisée l’instruction multiple représentée par l’expression reçue en instructions simples, et
à déterminer de façon automatisée un champ d’application parmi un ensemble de champs d’application pour chacune des instructions simples de la segmentation.

automatically segmenting the multiple statement represented by the received expression into single statements, and
to determine in an automated way a field of application among a set of fields of application for each of the simple instructions of the segmentation.

The invention will be better understood using the following description, given solely by way of example and made with reference to the accompanying drawings in which:
the schematically and functionally illustrates an example of a communicating electronic object, comprising an understanding device according to the invention, and coupled to a communication network to which communication equipment is also connected,
the schematically and functionally illustrates an embodiment of an understanding device according to the invention, and
the schematically illustrates an example of an algorithm implementing a method of understanding according to the invention.
The aim of the invention is in particular to propose a method of comprehension, and an associated comprehension device 1, intended to allow the automatic comprehension of textual expressions in natural language (resulting from written or spoken expressions) which are each representative of a multiple instruction comprising at least two simple instructions each representing an order (or a command).
We have schematically and functionally represented on the an example of a communicating electronic object 2, comprising an understanding device 1 according to the invention, and coupled to a communication network 3 to which a communication device 4 is also connected.
The comprehension method, according to the invention, is implemented when an electronic object 2 receives, via at least one interface that it understands, a textual expression (or sentence) in natural language (resulting from a written expression or spoken) and which must be understood to control or drive or launch at least one function or program that this electronic object 2 provides or includes or which is provided or included by at least one other electronic object to which this electronic object 2 is coupled (directly or indirectly, for example via a communication network (possibly wireless)). It will be noted that the aforementioned interface can be an input interface (possibly communication) or a man/machine interface (such as for example a microphone or a keyboard (possibly digital and possibly forming part of a touch screen)).
It is considered in what follows, by way of illustrative example, that the electronic object 2 is a communication device used by a user. For example, this communication equipment 2 can be a mobile intelligent telephone (or "smartphone"), an electronic tablet, a game console, a connected watch, or a computer (fixed or portable). But the electronic object 2 can be of any type provided that it can receive textual expressions in natural language (resulting from written or spoken expressions) representative of multiple instructions. Thus, it could be, for example, a communication server or a device (possibly a household appliance) or a communicating robot or a connected (or communicating) sensor or control equipment or control (for example of an alarm system, an opening locking system, a surveillance system, a lighting system, a signaling system, a heating/air conditioning system, a vehicle, a machine, an installation, or a building).
Furthermore, it is considered here, by way of illustrative example, that the electronic object 2 (here a communication device) receives textual expressions (resulting from written or spoken expressions) provided by its user and that it must understand to drive at least one function ensured or controlled by a communication device 4 (for example a device for controlling a home automation installation equipping a building).
As illustrated without limitation by the example algorithm of the , an understanding method according to the invention comprises first 100 and second 110 steps which can be implemented at least partially by an understanding device 1 according to the invention or else by a dedicated computer program executed on at least one equipment with processor(s) or at least one computer. This dedicated computer program includes instructions for executing the steps of the comprehension method and can be downloaded from a communication network and/or recorded on a computer-readable medium and/or executable by at least one processor (or one processing) of a device (here at least the communication device 2).
In the first step 100, of the comprehension method, one (the comprehension device 1) automatically segments the multiple instruction which is represented by an expression (in natural language) received into simple instructions (at least two).
In the second step 110, of the comprehension method, one (the comprehension device 1) automatically determines a field of application from among a set of fields of application for each of the simple instructions of the segmentation which has been determined in the first step 100. Here, "scope" means a contextual domain grouping together a set of at least one electronic device or at least one function or at least one program (local or remote) concerned by a command of a simple instruction and which can be accessed via an interface opened by a computer program.
Thanks to the invention, it is now possible to determine for each multiple instruction, represented by a received expression, simple instructions, then it is possible to determine the processing context targeted by each of its determined simple instructions. This makes it possible to significantly improve the understanding of expressions (or multiple instructions), possibly relating to different fields of application, and therefore to significantly reduce the comprehension error rate, thus allowing the use of multiple instructions to command or power automated functions after translation into an automation script.
For example, in the second step 110, an artificial intelligence can be used to determine a context for each simple instruction, then to determine the scope of application of each simple instruction according to the context determined for the latter. The context of a simple statement can be a class of scopes chosen from several predefined ones. If no context is recognized for a simple statement, then this simple statement can be associated with an inactive class. An inactive class means that we consider a classification of type "OTHER" or "UNDEFINED" meaning that we are not able to determine a context corresponding to the initially possible contexts. For example, possible contexts could be "interact with a printer" or "interact with a coffee maker", and a natural language request could be "what's the weather tomorrow?" ". In this case, the request cannot be classified as a printer or a coffee maker and therefore we have an inactive class.
To perform the context determination one can use a context classifier based on a BERT type model, proposed by Jacob Devlin in the article "BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding") (published in 2019), based on "Transformers" architecture (proposed by Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin in "attention is all you need", 2017) , with preferably a linear output layer whose output size corresponds to the number of detectable contexts plus one unit (the inactive class). It is this linear layer that will allow the final detection of the context. Such a model requires less training (or learning) data than that required for a recurrent neural network (or RNN) based on an LSTM-type architecture, and moreover converges more quickly than an RNN.
The context classifier uses to operate a dataset containing a few hundred sentences (or expressions) in which classes of application fields are represented. Preferably, the sentences (or expressions) are distributed in a balanced way between the different classes. It should be noted that an addition of context within the artificial intelligence ensuring the function of context classifier requires a new training of the latter, but this new training is inexpensive in time because the number of data necessary is reasonable.
It will also be noted that it is also possible to carry out a determination of the context between simple instructions of the same complex instruction in order to obtain expressions (or scripts resulting from the understanding of these expressions) that are even more natural and more reliable input of electronic objects implementing the functions concerned.
By way of example, we can have two fields of application, one being a power strip coupled to lamps of four different colors (and therefore each color of which can be chosen), the other being a decoder box (or " set-top box") to choose a TV or radio channel. If a user provides the expression "put the green then put RMC Découverte afterwards if the dog is in the room then you will wash it", this represents a multiple instruction that the comprehension device 1 will segment into three segments: a first segment "put on the green" (first simple instruction), a second segment "put RMC Découverte" (second simple instruction), and a third segment "if the dog is in the room then you will wash it" (third simple instruction appearing as the form of a ruler). Then, the comprehension device 1 will associate the three segments respectively with three fields of application: "put on the green" is associated with the "power strip" (this first association constitutes a first intention), "put RMC Découverte" is associated with the set-top box (this second association constitutes a second intention), and “if the dog is in the room then you will wash it”. In the latter case, no field of application makes it possible to detect the presence of a dog or even to wash it, and therefore the rule cannot be associated with a known field of application, which prohibits the determination of a third association and therefore of a third intention. In this example only the first and second intents can then be transformed into low-level commands and parameters: from the first intent is extracted the command "turn on" and the parameter "green", and from the second intent is extracted the command “put the television channel” and the parameter “RMC Découverte”.
At least three embodiments can be considered for implementing the first step 100.
In a first embodiment, in the first step 100 one (the comprehension device 1) can compare each elementary sub-part of the expression received with separators which are contained in a set of separators (predefined or configurable), in order to to determine each separator contained in the received expression. Here, "elementary subpart" means a word or a punctuation mark. Then, we can consider that the elementary sub-parts which are located before or after a determined separator are representative of a simple instruction.
In other words, from a set of separators, the comprehension device 1 will cut (or segment) a multiple instruction into a list of simple instructions separated two by two by a separator recognized within this set. Each recognized separator is then removed (and therefore not integrated into a simple instruction).
For example, in the first step 100 one can use a set of separators comprising, on the one hand, punctuation marks which are chosen from a group comprising a comma, a semicolon, a period and ellipsis, for example, and, on the other hand, separating words which are chosen from another group comprising the words "and", "then", "then", "finally", "after" and "also", for example.
As an example, a user can provide the expression “Turn on the green light and turn off the red light finally put ARTE on”. In this case, the comprehension device 1 will transform this expression into a series of simple instructions [“Turn on the green lamp”, “turn off the red lamp”, “turn on ARTE”].
This first embodiment is based on a classic elementary subpart to elementary subpart (or "word by word") lexical analysis based on rules and knowledge of the set of separators that are usable in a given language. It results from the fact that a person usually formulates a multiple statement by separating his single statements only by separators. Its implementation does not require the use of artificial intelligence, and does not require training (or learning) data. For example, this implementation may be based on a script that performs a comparison filter in which each elementary subpart of a received expression is compared against the set of separators. In practice, the script can be represented by an algorithm of adding spaces around the special and punctuation characters (',', ';', '-', '', ...) of the multi-statement, and then comparing each elementary subpart of the expression to the separators in the set of separators, then breaking the expression at each separator that was appended with space and then matched within the set of separators.
In a second embodiment, in the first step 100 one (the comprehension device 1) can use an artificial intelligence to classify each elementary sub-part of the expression received in a category which is chosen from among a first category of separators and a second category of words other than separators. Then, one (the understanding device 1) can use the classification results of this artificial intelligence to consider that the elementary sub-parts which are located before or after an elementary sub-part classified in the first category of separators are representative of a simple statement.
In other words, the artificial intelligence classifies each elementary subpart of the expression into one of two categories: either it is a separator or it is not a separator. This makes it possible to determine the separators in an expression without requiring an exhaustive list of separators. As soon as a separator is found, the multi-statement is segmented without this separator being integrated into one of the simple statements.
As an example, a user can provide the expression “Turn on the green lamp while putting ARTE on”. In this case, the comprehension device 1 will transform this expression into a series of simple instructions ["Turn on the green lamp", "turn on ARTE"].
For example, in the first step 100, for the second embodiment, artificial intelligence and a method of recognition of named entities (or NER ("Named-Entity Recognition") can be used to perform the classification of each sub- elementary part of the expression received. In this case, a BERT type architecture can be combined with the NER method. This makes it possible to further improve the understanding of the expressions and therefore to further reduce the failure rate.
It will be noted that in the first step 100 one can use artificial intelligence and the method of recognition of named entities (or NER) in association with a so-called conditional random field layer (or CRF ("Conditional Random Field")) to carry out the classification of each elementary sub-part of the expression received. This further improves the comprehension of expressions and therefore further reduces the failure rate.
In a third embodiment, in the first step 100 one (the comprehension device 1) can use an artificial intelligence to determine in the expression received each elementary sub-part which represents a beginning of a simple instruction and each sub-part elementary which represents a simple end of instruction. Then, one (the comprehension device 1) can use this artificial intelligence to consider that the elementary sub-parts which are included between an elementary sub-part included and representing a beginning of simple instruction and another elementary sub-part included and representing a simple instruction end following this simple instruction start are representative of the same simple instruction.
In other words, the artificial intelligence considers that all the elementary subparts that are included between what it considers to be a simple instruction start and a simple instruction end (both inclusive) constitute this simple instruction .
As an example, a user can provide the expression “Turn on the green lamp while putting ARTE on”. In this case, Comprehension Device 1 will determine that the word "ignite" is the start of a first simple instruction, the word "green" is the end of the first simple instruction, the word "putting" is the start of the second simple statement and the word "ARTE" is the end of the second simple statement, then will transform this expression into a sequence of simple statements ["Turn on the green lamp", "put ARTE"].
For example, in this third embodiment, it is possible to use an artificial intelligence with BERT type architecture and the method of recognition of named entities (or NER), possibly in association with a conditional random field layer (or CRF ("Conditional Random Field")) to perform simple statement start and end determination of a received expression. This further improves the comprehension of expressions and therefore further reduces the failure rate.
It is recalled here that the BERT architecture, which can be used for NER in the second and third embodiments described above, uses its own technique for cutting expression into tokens (or "tokens" - complete word or piece of word) and its own technique of word embedding (or "word-embedding" - the tokens are represented as vectors which are directly used by the architecture to make a prediction). The determined tokens are sent to the BERT architecture, and downstream of the latter a linear layer is placed in order to carry out a classification of each of the tokens in various predefined classes (a "softmax" function (classic in the field of intelligence artificial) is applied at the output of the linear layer). Without the CRF layer, for each token, a label is determined according to the maximum emission score, so that the labels of the different tokens are set independently of each other, but aberrations may appear. This is why it is advantageous to additionally use a CRF layer which reduces these aberrations. Indeed, the addition of a CRF layer makes it possible to take into account not only the emission scores of the labels but also the scores of possible transitions from one label to another. The scores that constitute the transition matrix are the parameters of the CRF layer that are optimized during the training (or learning) of the artificial intelligence. With the CRF layer, the labels of an expression are no longer chosen independently, because for each expression a sequence of labels is decided in such a way as to minimize the scores of emissions and transitions of the sequence of labels.
It should be noted that adding a CRF layer at the output of the BERT architecture can be done via PyTorch (open source machine learning Python software library) by instantiating an optimizer for the parameters of the CRF model (using Adam), then sending the output of the NER part into the CRF layer (changing the forwarding method), then summing the cost of the processing performed by the NER part and the cost of the processing performed by the CRF layer ( during backpropagation, PyTorch automatically calculates the gradients relating to the parameters of the CRF layer at the same time as those of the BERT architecture when calling the backward method on the sum of the costs ), then by updating the parameters of the CRF layer (call to the “step” method of the optimizer).
In the case of the second embodiment, it is possible, in addition to what is described above, to use three classes: "B-SEP" (beginning of separator) and "I-SEP" (continuation of separator after a B -MS), and “O” (other cases). For each token, the label retained is the one associated with the maximum score. The received expression is split at the location where a separator is detected ("B-SEP" and optionally at least one "I-SEP"), this separator is ignored and a list of simple statements is obtained.
As an example, a user can provide the expression “Turn on the red lamp while putting ARTE on”. In this case, the comprehension device 1 determines the sequence of words Light the red lamp while putting ARTE on, and associates respectively with these words the sequence of labels O O O O B-SEP I-SEP O O, which allows it to arrive at the segmentation "Turn on the red lamp" "put ARTE" (because "everything" is labeled "B-SEP" and "en" is labeled "I-SEP").
In the case of the second embodiment with a CRF layer as a complement at the output, the labeling and the segmentation method remain identical. However, the choice of labels is no longer obtained solely from the output scores of the BERT architecture. It is indeed the CRF layer which decides on the most probable sequence of labels.
In the case of the third embodiment, it is possible, in addition to what is described above, to use three classes: "BOS" ("Begin Of Sentence" - beginning of expression) and "EOS" ("End Of Sentence - end of phrase), and "O" ("other" - other case). For each label, the one whose score is maximum is retained, as in the case of the second embodiment (with NER part but without CRF layer). Each simple instruction at the end of the segmentation starts at a label BOS and ends at a label EOS, the unframed words are ignored.
As an example, a user can provide the expression “Turn on the red lamp while putting ARTE on”. In this case, the comprehension device 1 determines the sequence of words Turn on the red lamp while turning on ARTE, and respectively associates these words with the labels BOS O O EOS O O BOS EOS, which allows it to end up with the segmentation "Turn on the red lamp” “putting ARTE”.
In the case of the third embodiment with a CRF layer as a complement at the output, the labeling and the segmentation method remain identical.
It will be noted that the learning of the artificial intelligences, used by the comprehension device 1, is done by means of sets of learning data which can be characteristic instructions in a given language and/or lines of data produced from sentence structures (or "masterspins") in which certain words cycle through lists (this production is called "content spinning"). An example of four expressions generated from the masterspin {Allume | Turn off} the {green | red} gives the four possibilities "Turn on the green lamp", "Turn on the red lamp", "Turn off the green lamp" and "Turn off the red lamp". Instead of using data lines produced from masterspins, one can use training datasets generated from the integrated speech recognition platform SNIPS, for example on the necessary instruction domains.
As illustrated without limitation on the , the comprehension device 1 comprises at least one processing unit 5 and at least one memory 6 capable of performing the operations consisting, in the event of reception of an expression, of performing the operations consisting of segmenting the multiple instruction in an automated manner which is represented by this expression received in simple instructions, and in automatically determining a scope of application among a set of scopes for each of the simple instructions of the segmentation.
The processing unit 5 can, for example, be a processor, possibly a digital signal processor (or DSP (“Digital Signal Processor”)). Memory 6 stores data files and computer program(s) whose instructions are intended to be executed by processing unit 5 (or processor). This memory 6 can, for example, be of the RAM ("Random Access Memory") type. This processing unit 5 and this memory 6 are arranged to perform together the operations contributing to the automatic understanding of expressions (written or spoken).
In the example illustrated without limitation on the , the comprehension device 1 is arranged in the form of a computer equipping the communication equipment 2 of the user. But the understanding device 1 could be part of a computer providing at least one other function within the communication equipment 2. As illustrated, the computer 1 can also include an input interface 7 for receiving at least one other function. minus the expressions to be understood in order to use them in calculations or processing, possibly after having formatted and/or demodulated and/or amplified them, in a manner known per se, by means of a digital signal processor 8. In addition, this computer 1 can also comprise an output interface 9, in particular for delivering the associations of scope and simple instructions (or intentions) as well as any associated low-level commands and parameters, and/or a mass memory 10.
The computer 1 can be made in the form of a combination of electrical or electronic circuits or components (or “hardware”), possibly micro-programmed or micro-wired, and software modules (or “software”).
It will also be noted that the invention is not limited to the embodiments described above. It will indeed appear to those skilled in the art that various modifications can be made to the embodiments described above, in the light of the teaching which has just been disclosed to them. In the detailed presentation of the invention which is made above, the terms used must not be interpreted as limiting the invention to the embodiments set out in the present description, but must be interpreted to include therein all the equivalents whose provision is within the reach of those skilled in the art by applying their general knowledge to the implementation of the teaching which has just been disclosed to them.
It is clear that an understanding process/device such as the one described above makes it possible to make sentences or expressions in natural language more intuitive by accepting several simple instructions at the same time, to make the selection of the automation field transparent for the user or the machine, to significantly limit misunderstandings of multiple instructions and to reduce manual processing during the integration phases in new component configurations. In addition, they simplify the writing and execution of scripts by adding intelligence in the creation and execution of these.

Claims

Method for automatic understanding of expressions representative of a multiple instruction comprising at least two simple instructions each representative of an order,
characterized in that it comprises, upon receipt of an expression:

a first step (100) in which the multiple instruction represented by said expression received is segmented in an automated manner into simple instructions, and

a second step (110) in which a scope of application is determined automatically from among a set of scopes for each of said simple instructions of said segmentation.
Method according to Claim 1, characterized in that in the said first step (100), each elementary sub-part of the said expression received is compared with separators contained in a set of separators, in order to determine each separator contained in the said expression received, then it is considered that the elementary sub-parts situated before or after a determined separator are representative of a simple instruction.
Method according to Claim 2, characterized in that in the said first step (100), a set of separators is used comprising punctuation marks chosen from a group comprising a comma, a semicolon, a period and suspension points, and separation words chosen from another group comprising the words and, then, then, finally, after and also.
Method according to Claim 1, characterized in that in the said first step (100), an artificial intelligence is used to classify each elementary sub-part of the said expression received in a category chosen from among a first category of separators and a second category of words other as separators, then to consider that the elementary sub-parts located before or after an elementary sub-part classified in said first category of separators are representative of a simple instruction.
Method according to Claim 1, characterized in that in the said first step (100), an artificial intelligence is used to determine in the said expression received each elementary sub-part representing a beginning of a simple instruction and each elementary sub-part representing an end of 'simple instruction, then to consider that the elementary sub-parts comprised between an elementary sub-part included and representing a start of a simple instruction and another elementary sub-part included and representing an end of a simple instruction following this start of simple instruction are representative of the same simple instruction.
Method according to claim 4 or 5, characterized in that in said first step (100) said artificial intelligence and a named entity recognition method are used to perform said classification of each elementary sub-part of said received expression.
Method according to claim 6, characterized in that in said first step (100) said artificial intelligence and said named entity recognition method are used in association with a so-called conditional random field layer to perform said classification of each sub-part element of said received expression.
Method according to one of Claims 1 to 7, characterized in that in the said second step (110), artificial intelligence is used to determine a context of each simple instruction, then to determine the field of application of each simple instruction according of the context determined for the latter.
Computer program downloadable from a communication network and/or recorded on a computer-readable medium and/or executable by at least one processor, characterized in that it comprises instructions for the execution of the steps of a understanding according to any one of Claims 1 to 8, when the said program is executed on at least one piece of equipment with processor(s) (2) or at least one computer.
Comprehension device (1) for the automatic comprehension of expressions representative of a multiple instruction comprising at least two simple instructions each representative of an order,
characterized in that it comprises at least one processing unit (5) and at least one memory (6) capable, in the event of reception of an expression, of performing the operations consisting of:

automatically segmenting the multiple statement represented by said received expression into single statements, and

automatically determining a scope from among a set of scopes for each of said simple instructions of said segmentation.