FR3135802A1 - Method for supervised generation of a virtual semantic graph of specialized knowledge - Google Patents

Abstract

The invention is a method (100) for generating a knowledge graph, implemented by computer, comprising the following steps: - providing (103) several triples corresponding respectively to knowledge assertions; - sequence between triplets (104) aimed at organizing them into a topological graph; - definition (106) of types of semantic concepts, of semantic concepts (107) to describe and qualify the types of entities and attributes recognized among the elements of the topological graph, - instantiations (108) of defined semantic concepts, instances being generated by classification of the elements composing the triplets and associated with each other so as to also inherit the relationships of the topological graph, so as to generate a semantic knowledge graph. Figure for abstract: figure 3

Description

Method for supervised generation of a virtual semantic graph of specialized knowledge

The invention concerns knowledge management, in particular the supervised generation and feeding of virtual semantic graphs of specialized knowledge, in other words dedicated to a certain application or analysis.

Knowledge graphs aim to structure the knowledge acquired in a predetermined domain according to their reciprocal relationships. When they are constructed virtually via computer means, they allow the rapid or even automatic analysis by these means of the knowledge of the graph, in particular the determination of hidden links between entities which nevertheless seemed distant from each other, the detection of semantic patterns recurring, the generalization of facts to areas larger than the use case concerned. These graphs thus form tools for modeling complex processes of socio-technical systems, for example for the analysis of the management of certain accidents, so as to improve the management of system processes and even to prevent these accidents.

Several methods of organizing knowledge are known in the state of the art, but each is adapted to a particular field and has its own drawbacks. For example, the FRAM method (for “ Functional Resonance Analysis Method ”), used in the case of accident analysis, requires a large quantity of data to be provided in a precise and complex framework, resulting in a knowledge graph obeying a syntax to be acquired by its users and difficult to generalize to other domains. The AHP method (for “ Analytic Hierarchy Process ”) generates a graph on the basis of initial hypotheses which may subsequently turn out to be false and result in overly complex links between the entities in the graph. We also know methods aimed at creating ontologies, such as the “ M ethondology ” method, the “ Gospl ” method or the “ Gellish ” method. However, they do not allow the use of a common framework for all areas of knowledge organization.

Generally speaking, it is difficult to identify a method for supervised generation of a specialized knowledge graph that can be generalized to all domains.

It is also difficult to identify a method allowing the aggregation of knowledge according to the same framework for numerous sources of knowledge with varied profiles, and in particular a method allowing the contribution of knowledge by a neophyte in the targeted field.

Finally, it is difficult to obtain a virtual graph including relationships that are easy to analyze, the links between the entities in the graph often becoming more and more complex as the graph is enriched.

The invention aims in particular to enable the supervised generation of a specialized knowledge graph using the same method regardless of the targeted domain.

It also aims to enable the structuring of a large quantity of knowledge, coming from varied and heterogeneous sources, into a graph with relationships that are simple to understand and analyze.

Finally, it aims to enable the provision of knowledge in a simple manner by any neophyte in the field concerned by the graph.

To this end, the subject of the invention is a method for generating a virtual semantic graph of specialized knowledge, implemented by computer, in which the following steps are implemented:

- provision of several triplets corresponding respectively to knowledge assertions, each triplet comprising two nouns linked by a predicate;

- aggregation of triples based on names so as to link the triples together and organize them into a topological graph;

- detection of semantic patterns in the topological graph;

- depending on the detected semantic patterns, definition of types of generic semantic concepts, definition of semantic concepts, each semantic concept belonging to a type of generic semantic concept, and definition of possible associations between the semantic concepts so as to form a semantic model comprising the concepts associated with each other;

- instantiations of semantic concepts, each instance being associated with at least one of the defined concepts, the instances being generated by classification of the content of the topological graph, so as to generate a knowledge graph.

Thus, the step of providing triplets corresponds to the provision of knowledge. In the context of the invention, a triplet, or “ triplet ”, is a group of three verbal or numerical elements associated with each other to represent an assertion of knowledge. The expression “ knowledge assertion ” concerns the formulation of testimony, a fact, or a hypothesis to be verified. This contribution is reduced to its simplest expression since it only involves indicating two names, corresponding to two entities or to an entity and an attribute of this entity, and a predicate, corresponding to the link between these two names: an interaction, a qualification, for example. Thus, any person, not necessarily an expert in the field concerned by the knowledge graph, can contribute their knowledge. This supply step is in particular carried out via a software or web application, allowing a user to fill in fields predestined for the triplet collection.

The aggregation step makes it possible to obtain a topological graph of the knowledge provided. Thus, after this step, an expert in the domain targeted by the knowledge has a topological graph that he can analyze. This graph is, however, devoid of semantic contributions since it only includes the triples provided by the different sources, without generalizations having been made or categories having been created to group the knowledge. This triplet aggregation step is carried out as the triplets are entered, manually or automatically. It uses computer means recognizing identical or similar nouns (by orthographic or grammatical criteria) already used in other triplets, so as to gradually connect the triplets around the similar nouns that they include. The means make the triplets entered appear on a screen with the portion of the topological graph which concerns them.

The step of detecting semantic patterns in the topological graph aims to identify recurring links between the different elements of the topological graph, such as similar predicates between semantically close entities. It therefore aims to identify generic types of interaction which can be applied to numerous elements constituting the nodes of the topological graph, and therefore to suggest more general concepts than the case study of the topological graph. In this sense, it allows us to prepare for the following steps. This step can be carried out manually or using machine learning algorithms that have been trained to detect and recognize semantic patterns and/or links. In this case, machine learning is carried out beforehand on pre-existing ontologies forming learning databases.

The next step corresponds to the definitions of the types of generic semantic concepts. By " generic concept type ", we can in particular designate two general types: generic " entities ", such as " Resource ", " Agent ", on the one hand, or " Process ", " Event ", on the one hand, on the other hand, and the types of concepts forming generic “ attributes ”, to “name” an identity, a context, or a qualification. The concepts forming “ attributes ” are thus intended to qualify the concepts forming “ entities ”. Concept types represent the primary components of the semantic data model, i.e. the semantic structure. The choice of these types of concepts and the choice of the semantic model, that is to say the concepts themselves, are therefore interdependent.

The concept definition step corresponds to the selection of semantic classes, called “concepts” or “semantic concepts”, belonging to the types of generic concepts defined above. and the definition of their semantic characteristics. By “ concept ”, we designate terms useful for classifying elements of the topological graph as entities or as entity attributes, in a manner adapted to the objective of the case study. These concepts can be defined from scratch, but they can alternatively be selected by the user from existing generic ontologies (or upper ontologies ), provided in a database. At this stage we also define the characteristics and properties of these concepts. Finally, we define the possible relationships and associations between these concepts, in particular between the concepts belonging to the “entity” concept types on the one hand, and between the concepts of “entity” and “attribute” on the other hand.

All of these concepts form the reference ontology of the knowledge graph of interest. This ontology is called “ hybrid ” because the concepts are extracted on the one hand from general ontologies (or “ upper ontologies ”), applicable to numerous domains, and on the other hand from the ontologies of specific domains relevant to the domain of interest. These external concepts are selected, aggregated and adapted in their definitions as concepts of the semantic data model, considering the scope and application of the knowledge graph of interest. Any knowledge provided, through the triples, is part of this reference ontology.

These definitions allow us to fit into a very generic framework, which can be used in many different use cases. They also make it possible to simplify as much as possible the knowledge graph which will then be generated, since each entity in the graph belongs to one concept among a small number of concepts.

This step is carried out manually by an expert in the field of application, or semi-automatically using computer resources. The concept definition step is carried out concurrently with the concept type definition step, but the concepts can then be supplemented as knowledge claims are discovered or provided. from triples to the topological graph.

The concept instantiation step is the step allowing the transformation of the topological graph into a semantic knowledge graph strictly speaking. It is these instances of concepts that appear in the final knowledge graph.

This instantiation and association step to generate the semantic knowledge graph can be carried out manually by the expert via computer interaction means, such as a software or web application, or automatically by Machine Learning means having were trained to produce instances conforming to an ontology and a semantic model of reference data, from a topological graph.

We can predict that the concepts of this ontology themselves include semantic subclasses. For example, the “ hospital ” subclass could belong to the “infrastructure” concept, itself belonging to the “ Resource ” concept type. We can then also create instantiations of these subclasses.

By classifying in this way, as instances of semantic concepts, all the triplets provided in the topological graph while respecting the logic of the semantic modeling of the data defined in the previous step, we produce a knowledge graph in which it is easy to navigate and deduce hidden links between the elements of the graph, generalize the content of the topological graph, in particular by evolving between different levels of granularity and by adjusting the resolution of information extracted from the graph.

The method allows the traceable processing of knowledge which can be very numerous and heterogeneous and provided by very different sources, not experts in the field, in the form of simple triplets. Finally, by conforming to a very generic semantic model, the method can be used in all relevant domains, without anyone having to learn and apply complicated syntax or rules to bring the knowledge to the graph.

The invention may also include one or more of the following optional features, taken alone or in combination.

Preferably, each knowledge assertion corresponds to:

- either to a qualification, in which case the predicate of this triplet includes a connecting verb to qualify one of the two nouns of this triplet by the other noun of this triplet;

- either to an action, in which case the predicate of this triplet includes an action verb to describe an interaction between the two nouns of the triplet,

and in which, preferably, at least one of the names includes a temporal or spatial reference so as to chronologically or geographically position the statement to which the triplet corresponds.

The temporal or spatial reference can also make it possible to position chronologically or spatially statements to which correspond at least some of the other triples comprising the same name.

Advantageously, the provision of the triplets and the organization of the triplets in a topological graph are carried out via a computer module for editing and structuring text.

Preferably, the generic semantic concept types are generic entities or generic attributes, the attributes corresponding to identification, context or qualification characteristics associated with the entities.

Advantageously, the definition of semantic concepts is carried out via a computer module for detecting recurring relational patterns in the topological graph.

Preferably, the instantiations of the concepts and the associations of these instances between them are carried out via a computer module for transforming the topological graph into a semantic knowledge graph.

Advantageously, the method further implements a step of manipulating the knowledge graph through computer interaction means, so as to navigate through the entire knowledge graph and to edit the knowledge graph.

According to the invention, there is also provided a method of providing feedback, implemented by computer, in which the following steps are implemented:

- choice of an existing knowledge graph, the knowledge graph having been generated in accordance with the method described above;

- provision of a triplet corresponding to feedback, the triplet comprising two nouns linked by a predicate;

- transformation of the triple into instantiation of one or more concepts and integration of the instance(s) into the knowledge graph.

Advantageously, the transformation is carried out automatically by a computer module for instantiation and integration of one or more concepts in the knowledge graph according to the triplet.

According to the invention, there is also provided a computer program comprising instructions which, when the program is executed by a computer, lead it to implement the steps of the method described above.

According to the invention, there is also provided a computer-readable recording medium comprising instructions which, when executed by a computer, lead it to implement the steps of the method described above.

Brief description of the figures

The invention will be better understood on reading the description which follows, given solely by way of example and made with reference to the appended drawings in which:

there is a diagram of means of implementing the invention;

there is a diagram of a computer program of the invention;

there is a diagram of a process according to the invention;

there is a diagram of a step in the process of ;

there is a diagram of a step in the process of ;

there is a diagram of a step in the process of ;

there is a diagram of a step in the process of ;

there is a diagram of a step in the process of ;

there is a diagram of another method according to the invention.

detailed description

There illustrates a user 1 and computer means 2 allowing the implementation of methods 100 and 200 which will be described below.

User 1 is a natural person wishing to generate a knowledge graph from a plurality of pieces of knowledge, as part of the method 100, or add knowledge to a tree already generated as part of the method 200.

The computer means 2 include interaction means 21 such as a screen and a keyboard, or other man-machine interfaces, calculation means 22 such as a processor, and a computer memory 23, which is physical but could be virtual. Within the memory 23 is recorded a computer program 24, called “Core”, allowing the implementation, when executed by the user 1, of the steps of the processes 100 and 200.

The architecture of program 24 is schematized on the . Any user interface 25 is thus connected to the program 24 via a proxy module 26, which is connected directly or indirectly to all of the other modules forming the program 24. The module 27 uses the open source software TiddlyWiki and its TiddlyMap extension to allow the editing of text data to be transformed into a topological graph. It is therefore a computer module for editing and organizing text. Thus, user 1 uses, through program 24, the functions of this module to process the raw data provided in the form of triplets (see below). Module 28 allows the transformation of raw data, triplets, into semantic data, instances of concepts. It uses an Apache Jena Fuseki SPARQL server, which supports the “ OWL ” language (or “ Web Ontology Language ”). Module 29 includes the PostgreSQL library and allows the recording of SPARQL queries performed on the knowledge graph. It is connected to module 30 called “ Core Back - end ”. Module 31 is the “ Core Front -end ” module, it allows the injection of Javascript code into module 27, queries in module 28 and is also linked to module 30. The output of program 24 is carried out at the level of module 32, connected to module 28. Module 32 uses the OntotextGraphDB tool, which allows the user to visualize and manipulate the knowledge graph generated at the end of process 100.

Alternatively, it is entirely possible to provide tools other than those mentioned for each of these modules. For example, the module 27 can exploit human-machine interfaces providing different ways of interaction 21 for entering and editing data, and the module 32 can include another tool for visualizing and manipulating the semantic knowledge graph for extraction and use of analysis results.

In reference to the , we will now describe a method 100 for generating a knowledge graph. It is implemented by the user 1 thanks to the computer means 2 which execute the program 24. It is this program 24 which allows the implementation of the steps of the method 100 described below. This process 100 is inspired by the theory of categories, toposes (or topoi) and “sketches, graph theory, but also grounded theory and “ multiper spe ctivity ”. In the example described below, it is applied to the management of a natural disaster. The knowledge graph that it generates aims to determine causality and hidden relationships between the entities involved in disaster management, so as to generate alternative management scenarios to improve the management of future crises of the same type, or even crises of other types. Although the examples here come from crisis management, the method 100 can be applied to all fields, in particular because it includes the method for defining ontologies dedicated to different study cases, allowing its implementation in very varied fields and frameworks of application.

Beforehand, in a step 101, called “ Scope graph ”, the user 1 defines the scope of application of the process 100 which will follow. It defines the sources of knowledge which will be used to provide knowledge assertions, selects the names of targets, subjects or objects, which could be proposed, selects action verbs and attribution verbs which can also be used during step 103 of providing triplets, for the description of the desired or expected relationships between these elements. Module 27 assists the user in defining this perimeter, which results in a graph, which can be described as a graph of initial terms, an example of which is provided in .

Following the definition of the scope of application of the method 100, in a step 102, the user 1 collects knowledge assertions. This involves having documents, testimonies, explanations, and all other types of information which will be provided to the computer means 2 to form the corpus of data available to generate and supply the semantic knowledge graph . These knowledge assertions, collected in particular from the sources defined in the previous step, can be collected by feedback processing solutions, ranging from basic approaches allowing text editing, to machine learning tools , such as MAXQDA , which help to highlight the essential elements of the knowledge collected. Atlas.ti software can also allow the processing of large textual, graphic, audio and video data, and enable collaborative analysis. We can also use NLP (for “ Natural Language Processing ”) technologies. However, these technologies do not produce original data, they simply allow part of the meaning to be extracted from certain data provided. Alternatively, the data may be provided by other computing means, such as video recording means and/or augmented reality software.

Once the knowledge assertions have been collected, we begin step 103. This involves providing triplets, represented schematically in . A triplet consists of two nouns, or noun phrases, connected by a predicate usually including a verb. These nouns and verbs can be chosen from those defined in step 101 of defining the scope of the application of the method. These triplets correspond either to a qualification, in which case the predicate of this triplet includes a connecting verb to qualify one of the two nouns of this triplet by the other noun of this triplet or to an action, in which case the predicate of this triplet includes an action verb V _a to describe an interaction between the two nouns of the triplet. In the context of a qualification, we can consider that the second name is therefore in fact an attribute of the first name. There thus illustrates three triplets: a triplet of interaction between the name N ₁ and the name N ₂ , with the predicate V _l1 , a triplet of attribution, or qualification, attributing the name a ₁ to the name N ₂ by means of the predicate Va , and another triplet of interaction, this time between the name N ₂ and a name N ₃ by means of a predicate V _l2 . In the context of natural disaster management taken as an example here, a triplet can for example include the nouns “Hospital” and “helicopter runway”, linked by the predicate “possesses”. This involves providing the knowledge that, in the context of this crisis, a hospital entity is equipped with a helicopter landing strip. In addition, a name can also represent a temporal or geographical reference, if relevant for the case study, so as to chronologically or spatially position the information to which the triple corresponds.

All the knowledge available to user 1, and which may have been collected from numerous and heterogeneous sources, is thus divided into triplets allowing the simplest expression of this knowledge. User 1 is therefore not necessarily an expert in the target domain. The knowledge can also be provided, in the form of triplets, by each or some of the sources having the knowledge, rather than by the single user 1.

This step 103 is carried out via interface 25, organized to intuitively allow the user to simply provide the two names and the predicate of each of the triples. For example, it can present fields to fill in to place the three elements of the triplet. It can propose names and attributes already provided previously, to avoid duplication of names concerning the same element.

In step 104, the user is helped by computer means to aggregate the triples so as to form a so-called “ topological ” graph. This is made possible by the functions of module 27 of program 24. Thus, by considering all of the triples, the identical names or which refer to the same element are merged so as to connect the triples together. For example, all the triples comprising the name “Madame X” are connected around the name “Madame An example of a topological graph obtained is that of the , about triplets linked to the management of a natural disaster.

It should be noted that this topological graph is in itself already a useful result of the method 100, since it presents in it the set of knowledge assertions, reorganized so as to avoid redundancies and to group similar elements together, in allowing the coherent aggregation of several heterogeneous sources of information

The next step 105 is the step of detecting semantic patterns in the topological graph. It aims to identify recurring links between the different elements of the topological graph, such as triples including similar predicates, for names possibly belonging to similar semantic concepts. It therefore aims to allow the user to identify generic types of interaction which can be applied to numerous elements constituting the nodes of the topological graph, and therefore to suggest concepts more general than the case study of the graph. topological. This step is carried out manually by the user, assisted by module 27. Alternatively, it could be carried out automatically via a machine learning module having been trained to detect and recognize semantic patterns in pre-existing ontologies forming learning bases.

In step 106, the user determines a semantic data structure. This structure is for example that of the . It reveals two types of concepts, corresponding to “ generic entities ”: “ agents ” on the one hand, which can also be called “ resources ”, and “ processes ” or “ events ” on the other hand. There also reveals eight other types of concepts, which are not “ generic entities ” but “ generic attributes ”. In this structure, the eight types of concepts forming attributes accompany the two types of concepts forming entities, to semantically allow their identification, characterization and qualification. These eight types of generic attributes form geospatial references, qualifications, conditions, contexts, key parameters, to be associated with entities. The semantic structure also indicates what types of links are allowed between entities on the one hand and between entities and attributes on the other hand. As described below, any triplet of knowledge is intended to be transformed into instances of concepts belonging to these types of entities or attributes.

Step 107 is carried out concomitantly with step 106 by means of the functions of module 28. This involves creating the semantic data model, by means of the structure, that is to say the types of concepts, from step 106. This involves defining, or selecting from pre-existing ontologies, concepts, and possibly their subclasses, belonging to the types of concepts of the semantic structure chosen in step 106. defines for example the concept “ Infrastructure ”, belonging to the concept type “ Resource ”, a generic entity. We can also define subclasses belonging to concepts. For example, we can define the subclass “ hospital ” of the concept “ infrastructure ”.

The definition of a concept includes properties or characteristics, the definition of the compatibility of the concept with concepts belonging to other types of concepts. With regard to properties for example, we can in particular define attribute concepts which are specific to a particular type of entity concept, but also generic attributes, in particular " qualification " concepts, which can be applied to all types of concept. Similarly, regarding compatibilities between concepts, one can define that a specific attribute concept can be compatible with one or more entity concepts, but not with all entity concepts in the model. On the other hand, a qualification attribute concept could, for example, be associated with concepts of all types. All of these choices lead to what is called the “ semantic data model ”. Thus, in the case of study concerning the management of a natural disaster, we define the concept of “ People ” belonging to the type of concept “ Agent/Resource ”. An instance of this “ People ” concept could then be linked to:

- another instance of the same concept, identified by two different “ social entity ” attribute concepts (for example an “ Individual ” belonging to a “ Community ”),

- an instance of another concept of the same type “ Agent / Resource ” (for example an “ Individual ” requiring “ Food Resources ”),

- a different type of concept, such as the “ Process/ Event ” type (for example an “ Individual ” suffering from an “ Impact ”), and/or

- an instance of a concept which expresses an attribute, such as the “ condition ” (for example an “ Individual ” which would be a “ Disaster Person ”).

This step is carried out via the user interface 25. Alternatively, this step can be carried out using a machine learning module having been trained to propose semantic concepts from a set of ontologies of possibly relevant domains , based on predefined structures of semantic data models, based on the theories of toposes and sketches.

All of these concepts form a reference ontology of the use case. The ontology is called hybrid because it is both based on specifically selected generic semantic concepts, coming from external ontologies, and adapted as concepts specific to the use case.

These entity and attribute classification structures are intrinsically linked to mathematical concepts from category theory such as morphisms, functors, and monoids. Thus, by describing knowledge using these structures, we obtain recurring and repetitive patterns, which we call universal morphisms. The possible concepts are limited. Although they can be defined from scratch by the user, it is advantageous that they are selected by user 1 from " upper ontologies " and from ontologies of domain, grouped into concepts. Optionally, a list of choices of concepts or higher ontologies is provided to the user on their screen. These concepts (and, if relevant, their subclasses) will be used as described below to define the instances of the semantic knowledge graph.

In step 108, user 1 controls the generation of the knowledge graph. This involves producing, by means of module 28, the instances of the different concepts chosen in the previous step, the entities and their attributes, their relationships, classified on the basis of the semantic model of the data, so as to transform (or semantically enrich) the content of the topological graph. The generation of semantic instances requires the definition of “ concordance rules ” associating the concepts with the data types of the topological graph. The user, or a semantic classification algorithm, therefore browses the topological graph and classifies the instances of concepts in a manner consistent with these rules, and with the concepts defined in step 107. In other words, it is the semanticification of the triples provided, the generation of a knowledge graph organized according to the data which constitutes its corpus in the form of a topological graph. The semanticification of triples is based on category theory. It allows the use of this ontology in other cases in the same domain, as well as simplified functional analyzes in the knowledge graphs subsequently generated.

In the case of studying the management of a natural disaster, it is for example a question of describing semantically, by classification with several concepts, the entity “ Saint Nazaire Hospital ”. This entity is an instance of the “ Infrastructure ” concept, a “ Resource ” type entity, classified by the specific attribute concept of “ Contextual Sector : Health ”. The properties provided in the concepts are defined in a manner consistent with the reference ontology, and in such a way as to account for the triplets provided in step 103. Thus, starting on the one hand from a triplet including the name “ Hospital », the name “ Saint Nazaire ” and a predicate comprising the verb “ to be called ”, and on the other hand another triplet comprising the nominal group “ Saint Nazaire Hospital ”, the nominal group “ 1500 people ” and the predicate “ employs ”, the user creates the “ Saint Nazaire Hospital ” instance of the “ hospital ” subclass (an “ Infrastructure ” of type “ Resource ” and including the Contextual Sector attribute “Health”), he associates the number of “1500” to the “ key parameter: Number of employees ” attribute, and the name “Saint-Nazaire” to the “ identity: name ” attribute.

These instances are associated with each other in accordance with the triples provided in step 103 and then form the knowledge graph. Thus, if we have a triplet including the nominal group “ Saint Nazaire Hospital ”, the name “ Madame X ” and the predicate “ directs ”, user 1 creates the instance “ Madame : Individual " itself an attribute of the concept " People ", assigns it the subclass " Director " of the concept " condition: role ", in the contextual sector " Health ", and links the instance " Saint Nazaire Hospital " to the “Madame X ” instance, via a connection between the two entities in accordance with the chosen semantic model.

Alternatively, this step 108 can be carried out automatically by an instantiation module, or even a “ computer module for transforming the topological graph into a knowledge graph ”. The module takes the elements of the topological graph and creates instances of an entity concept, characterized by attribute concepts. It may in particular be a machine learning module trained for this purpose.

The knowledge graph is then generated. It allows you to visualize links that were not directly identifiable before, and to navigate through all the knowledge in a simple and organized way.

As we have seen, user 1 is at least assisted by program 24 at each step of the process, but some of these steps can be carried out automatically by Machine Learning algorithms. In general, the extraction of semantic classes from pre-existing ontologies for the vocabulary of concepts as well as the application of the set of rules necessary to transform a topological instance into a semantic instance, can therefore be carried out by interactive database processing. or by graphical interface. The design of system 24 involves the integration of semi-automatic incremental construction of ontologies implementing processes known as “ontological learning”. The algorithm can be trained to choose semantic classes for the definition of concepts, on the basis of a semantic structure and a determined topological graph. Alternatively, it may have been trained to propose classes based on certain terms from a topological graph and a given semantic model.

In an optional step not illustrated, user 1 performs a query on the knowledge graph so as to extract the knowledge he wishes, or so as to identify knowledge not directly visible, for example by applying one or more filters. For this purpose, it uses module 29. An example of uses of the types of concepts, concepts and subclasses on a portion of the knowledge graph produced for the analysis of a crisis management case, is shown in section .

In defining a query, user 1 can also define new ontological classes inferred by rules of combinations or relationships between concepts defined in the semantic model of the data.

The process thus allows the organization of all kinds of structured, unstructured or semi-structured data in a knowledge graph with significant inference capacity. It allows the discovery of knowledge and its extraction by means of its conceptual representation at different levels of abstraction. It ultimately allows the analysis of complex socio-technical systems of any kind, in order to improve or even aid decision-making.

We will now describe the process 200 with reference to the . This process aims to provide feedback or a body of information in an already existing knowledge graph. It is implemented by user 1 using the same modules as those which implemented method 100.

In step 201, user 1 chooses, via interface 25, a knowledge graph from several possible ones, all the graphs having been generated in accordance with method 100. The chosen graph corresponds to the domain or use case to which he wants to contribute.

In step 202, user 1 provides a triplet corresponding to elements concerning his feedback, the triplet comprising two names linked by a predicate in the same way as the triplets of step 101 of method 100.

In step 203, user 1 orders the transformation of the triple into instantiation of one or more concepts and the integration of the instance(s) into the knowledge graph. Once ordered, this transformation and this integration are carried out automatically by a non-illustrated computer module for instantiation and integration of one or more concepts in the knowledge graph according to the triplet. In particular, this is a machine learning module that has been trained to modify a knowledge graph based on a provided triplet.

The method of the invention is applicable to any field and in this sense aims to contribute to the standardization of knowledge management solutions. Among the targeted areas are, in addition to crisis management, all industrial areas and in particular all industrial processes, such as mining, but also resource management in general, including the analysis of material flows. primary and secondary, and engineering the resilience of sociotechnical systems.

The graph can be used as a basis for other processes. It can make it possible to collect and structure in a more efficient manner the data necessary for the implementation of methods of analysis of sociotechnical systems already known, requiring a large quantity of input data and the management of complex relationships between them, like the “FRAM”.

If the knowledge claims provided to it include quantitative data, it can be used to evaluate the performance of organizations or carry out analyzes suggesting the best options to follow to improve a process, whether in a factual manner, it is that is to say at the operational level, and conceptually, that is to say at a strategic level or in terms of generic rules. Thus, it can make it possible to accelerate the updating of learning data in accordance with the knowledge learned, its use can be oriented to the understanding of cascade effects, to the anticipation of cause and effect relationships, to the definition of alternative scenarios.

The invention is not limited to the embodiments presented and other embodiments will be clear to those skilled in the art.

Claims (11)

Method (100) for generating a virtual semantic graph of specialized knowledge, implemented by computer (2), in which the following steps are implemented:
- supply (103) of several triplets corresponding respectively to knowledge assertions, each triplet comprising two names (N ₁ , N ₂ , a ₁ , N ₃ ) connected by a predicate (V _l1 , V _a , V _l2 );
- aggregation (104) of the triples according to the names so as to connect the triples together and to organize them into a topological graph;
- detection (105) of semantic patterns in the topological graph;
- depending on the detected semantic patterns, definition (106) of types of generic semantic concepts, definition (107) of semantic concepts, each semantic concept belonging to a type of generic semantic concept, and definition of possible associations between the semantic concepts in a manner to form a semantic model comprising the concepts associated with each other;
- instantiations (108) of semantic concepts, each instance being associated with at least one of the defined concepts, the instances being generated by classification of the content of the topological graph, so as to generate a knowledge graph. Method (100) according to the preceding claim, in which each knowledge assertion corresponds:
- either to a qualification, in which case the predicate (V _a ) of this triplet includes a connecting verb to qualify one of the two nouns (N ₂ , a ₁ ) of this triplet by the other noun of this triplet;
- either to an action, in which case the predicate (V _l1 , V _l2 ) of this triplet includes an action verb to describe an interaction between the two nouns (N ₁ , N ₂ , N ₃ ) of the triplet,
and in which, preferably, at least one of the names includes a temporal or spatial reference so as to chronologically or geographically position the statement to which the triplet corresponds. Method (100) according to any one of the preceding claims, in which the supply (103) of the triplets and the organization of the triplets into a topological graph (104) are carried out via a computer module (27) for editing and structuring of text. Method (100) according to any one of the preceding claims, wherein the generic semantic concept types are generic entities or generic attributes, the attributes corresponding to identification, context or qualification characteristics associated with the entities. Method (100) according to any one of the preceding claims, in which the definition of semantic concepts is carried out via a computer module for detecting recurring relational patterns in the topological graph. Method (100) according to any one of the preceding claims, in which the instantiations (108) of the concepts and the associations of these instances between them are carried out via a computer module for transforming the topological graph into a semantic knowledge graph. Method (100) according to any one of the preceding claims, further implementing a step of manipulating the knowledge graph through computer interaction means (2), so as to navigate through the entire knowledge graph and to edit the knowledge graph. Method (200) for providing feedback, implemented by computer (2), in which the following steps are implemented:
- choice (201) of an existing knowledge graph, the knowledge graph having been generated in accordance with any one of the preceding claims;
- provision (202) of a triplet corresponding to feedback, the triplet comprising two nouns linked by a predicate;
- transformation (203) of the triple into instantiation of one or more concepts and integration of the instance(s) into the knowledge graph. Method (200) according to the preceding claim, in which the transformation (203) is carried out automatically by a computer module for instantiation and integration of one or more concepts in the knowledge graph according to the triplet. Computer program (24) comprising instructions which, when the program is executed by a computer (2) lead it to implement the steps of the method (100) for generating a virtual semantic graph of specialized knowledge according to any one of claims 1 to 7 or the steps of the method (200) for providing feedback according to claim 8 or 9. A computer-readable recording medium (23) comprising instructions which, when executed by a computer, cause the computer to implement the steps of the method (100) for generating a virtual semantic graph of specialized knowledge according to any one of claims 1 to 7 or the steps of the method (200) for providing feedback according to claim 8 or 9.