INTERFACE AND CONVERSATIONAL METHOD WITH A KNOWLEDGE SYSTEM OF THE EXPERT SYSTEM TYPE

Abstract

The invention relates to a conversational interface (1) with a knowledge-based system (5), comprising: - an extraction module (11) configured to extract from a user's question (3) relating to a search for a decision, corresponding decision search inputs which are adapted to be used by said knowledge-based system (5) in order to generate a decision to said decision inputs, - an introspection module (13) configured in case said decision search inputs are insufficient, to determine additional decision search inputs which are missing for the knowledge-based system to generate the decision, and - a reconstitution module (15) configured to reconstitute said additional decision search inputs from the introspection module (13) or the decision from the knowledge-based system (5) into a natural language text. Figure for abstract: Fig.1.

Description

Title of the invention: INTERFACE AND CONVERSATIONAL METHOD WITH A KNOWLEDGE SYSTEM EXPERT SYSTEM TYPE Technical field

[0001] The present invention relates to the field of expert systems and more particularly their interfaces with users. STATE OF THE PRIOR ART

[0002] An expert system is a software program capable of answering questions by reasoning based on facts and logical rules. The expert system includes knowledge bases and logical rules as well as an inference engine. The latter uses the facts and rules to answer an expert question posed by a user.

[0003] In general, the expert system is equipped with a query interface based on entering inputs and then generating a response based on models of sentence gaps. These gaps are filled by elements resulting from a calculation carried out by the expert system. In addition, some expert systems implement algorithms to find the rule to be executed. Other expert systems allow the next interaction with the user to be chosen.

[0004] However, all these expert systems do not allow for a real dialogue with a user and in particular, do not allow for explanations to be provided. Furthermore, these systems do not allow for enriching the rule bases or for having a choice of selection of the interaction with the user.

[0005] Thus, the object of the present invention is to remedy the aforementioned drawbacks by proposing a conversational interface with a knowledge-based system (expert system, planning system, constraint solver) making it possible to dialogue in natural language with the latter in a precise manner and by providing explanations. Statement of the invention

[0006] The subject of the invention is a conversational interface with a knowledge-based system, comprising:

[0007] -an extraction module configured to extract from a question or description relating to decision-making, the value of a subset of the corresponding system inputs which are suitable for use by said knowledge-based system in order to generate a decision based on said inputs,

[0008] -an introspection module configured in case said inputs are insufficient, determine additional inputs that the knowledge-based system is missing to generate the decision, and

[0009] - a reconstruction module configured to reconstruct said inputs and the knowledge used during the decision-making process into a natural language text.

[0010] This makes it possible to provide a text interface for conversing with a knowledge-based system. Thus, a user can converse in natural language with the knowledge-based system which is adapted to generate a decision in response to information provided by the user while requesting additional information from the user if the information already provided is not sufficient for decision-making.

[0011] The knowledge system comprises, for example, fuzzy logic rule bases and an inference engine configured to deduce decisions from the inputs using said rules. Advantageously, the introspection module is configured to determine said additional decision search inputs by searching among all the fuzzy logic rules those which could not be triggered and the reason for this non-triggering.

[0012] Thus, the introspection module relies on fuzzy logic rules to activate several rules at the same time rather than one rule at a time in a sequence of rules.

[0013] Advantageously, the extraction module is further configured to extract from an explanation request the corresponding stimuli (justification request inputs) necessary to enable said knowledge-based system to provide an explanation, and in that the reconstruction module is configured to produce a natural language explanation text from the explanation generated by the knowledge-based system.

[0014] This allows the user to interact with the knowledge-based system to justify the decision provided.

[0015] Advantageously, the extraction module is further configured to extract from a rule editing request, the corresponding stimuli to enable said knowledge-based system to update the rule bases.

[0016] Advantageously, the editing of the rule bases includes a deletion, a modification, or an addition of at least one rule in the rule bases.

[0017] This makes it possible to supply the rule bases by editing them linguistically.

[0018] Advantageously, the extraction module is initially configured to extract from conversational information provided by the user context entries allowing the knowledge system to determine the context of the conversation and to load the logical rules adapted to said context desired by the user. This may concern the search for a decision, the request for an ex- application to a decision, or editing of rule bases.

[0019] Advantageously, the extraction module comprises advanced techniques for automatic processing of natural language NLP, and in that the reconstruction module comprises advanced techniques for automatic generation of NLG text.

[0020] This allows for a better human-machine interface.

[0021] Alternatively, the extraction module includes basic techniques including parser or regular expressions, and the reconstruction module includes regenerated text models.

[0022] Advantageously, the conversational interface comprises a conversion module for converting a voice message into a text message intended for the extraction module, and a voice synthesis module for converting a text response from the knowledge-based system into a voice response.

[0023] Advantageously, the knowledge-based system is an expert system, a planning system, or a constraint solver.

[0024] The invention also relates to a system for controlling a treatment plant comprising a tank equipped with sensors and adapted to contain liquids. The control system comprises a knowledge-based system and a conversational interface according to any one of the preceding characteristics, the conversational interface being configured to interact with the knowledge-based system according to measurements made by the sensors and / or operators relating to parameters representative of the level of contamination of the tank.

[0025] The invention also relates to a conversational method with a knowledge-based system, comprising the following steps:

[0026] - extract from a question relating to a search for a decision, entries corresponding decision research which are adapted to be used by said knowledge-based system to generate a decision to said decision inputs,

[0027] - determining in case said decision research inputs are insufficient, additional decision research inputs that the knowledge-based system lacks to generate the decision, and

[0028] - reconstruct said additional decision-making search inputs or said decision in a natural language text.

[0029] Advantageously, the determination of said additional decision-making search inputs is carried out by searching among fuzzy logic rules recorded in rule bases of the knowledge-based system, those which could not be triggered and the reason for this non-triggering.

[0030] Advantageously, the method comprises:

[0031] -an initialization phase for extracting from conversational information (user-provided) context inputs allowing the knowledge-based system to determine the context of the conversation and to load fuzzy logic rules according to said context, and

[0032] - a dialogue phase including the search for a decision, the request for an explanation of a decision, or the editing of the rule bases.

[0033] According to a variant, the conversational method comprises a single phase in which the context is re-determined each time.

[0034] Other advantages and characteristics of the invention will appear in the detailed non-limiting description below. Brief description of the drawings

[0035] Embodiments of the invention will now be described, by way of non-limiting examples, with reference to the accompanying drawings, in which:

[0036] [Fig.l] very schematically illustrates a conversational interface with a knowledge-based system, according to one embodiment of the invention;

[0037] [Fig.2] is a block diagram illustrating a method of conversational interfacing with a knowledge-based system in relation to [Fig.l], according to a preferred embodiment of the invention;

[0038] [Fig. 3] very schematically illustrates a control and monitoring system for a treatment plant comprising a conversational interface, according to a preferred embodiment of the invention; and

[0039] Figs. 4A-4H illustrate curves representing the parameters to be monitored in the treatment plant using fuzzy logic, according to a preferred embodiment of the invention. DETAILED DESCRIPTION OF THE INVENTION

[0040] The principle of the invention is to provide a textual interface for conversing with a knowledge-based system, in particular a fuzzy one, either to obtain answers or to feed its rule base.

[0041] According to the present invention, knowledge-based system means any system of the expert system type, planning system, constraint solver, or others.

[0042] [Fig.l] very schematically illustrates a conversational interface with a knowledge-based system, according to one embodiment of the invention.

[0043] The conversational interface 1 is configured to connect a human operator or user 3 with the knowledge-based system 5. Three main types of interaction with the knowledge-based system 5 can be distinguished.

[0044] A first interaction II is a consultation allowing the user 3 to interact with the knowledge-based system 5 to provide it with information from which it will automatically extract the inputs it needs to make a decision. It is a dialogue because it may be that user 3 has not provided all the necessary information, and in this case, the conversational interface 1 allows the knowledge-based system 5 to ask him for the values ​​he is missing.

[0045] A second interaction 12 is a dialogue between the user 3 and the knowledge-based system 5 in order to justify a decision. The knowledge-based system 5 provides a first justification and the user 3 can request additional information or ask more specific questions.

[0046] A third interaction 13 is a provisioning of rules allowing the rule base 7 included in the knowledge-based system 5 to be edited linguistically.

[0047] The knowledge-based system 5 comprises knowledge and logic rule bases 7 and an inference engine 9. Advantageously, the logic rules comprise fuzzy logic rules. The inference engine 9 is configured to process different forms of logic rules using the knowledge and rule bases 7 in order to deduce facts from the information received from the users and / or data from the sensors (not shown).

[0048] The conversational interface 1 comprises an extraction module 11, an introspection module 13, and a reconstitution module 15. The relationships between these elements are described in relation to [Fig.2]

[0049] Indeed, [Fig.2] is a block diagram illustrating a method of conversational interfacing with a knowledge-based system in relation to [Fig.l], according to a preferred embodiment of the invention.

[0050] In a first phase, called the initialization phase PI, the user 3 connects to the knowledge-based system 5 to start a conversation.

[0051] The extraction module 11 is initially configured to analyze the conversational information provided by the user 3 in order to extract context entries allowing the knowledge-based system 5 to determine 12 the context of the conversation. The knowledge-based system 5 determines 12 the rule(s) to be loaded from the knowledge and rule bases 7 according to the context desired by the user. This context may concern the search for a decision, the request for an explanation for a decision, or the editing of the rule or knowledge bases.

[0052] Once the rules and / or knowledge are loaded, a second phase, called dialogue phase P2 begins in which there can be a dialogue among three possible interactions or requests.

[0053] The first interaction II is a consultation during which an expert question is asked by the user 3 concerning a search for a decision. The question contains inputs and their values. The extraction module 11 is configured to extracts from the expert question asked, corresponding decision search inputs according to a formal representation and provides them to the reasoning or inference engine 9. These decision search inputs are adapted to be used by the knowledge-based system and more particularly, by the inference engine 9 in order to generate a decision or a response to the decision search inputs.

[0054] If the decision-making search inputs are sufficient, then the knowledge-based system 5 and more particularly, the inference engine 9 relies on the logical rules and knowledge to deduce or generate a decision whose content is formally represented to the question of the user 3.

[0055] The reconstruction module 15 is then configured to reconstruct the formal representation of the content of the decision and its uncertainty from the inference engine 9 of the knowledge-based system 5 into a textual decision in natural language intended for the user.

[0056] However, in the event that the decision search inputs are insufficient, the introspection module 13 is configured to determine by introspection (i.e. by analysis of the trace) additional decision search inputs that are missing from the inference engine 9 to generate the decision or the response.

[0057] More particularly, the introspection module 13 is configured to determine the additional decision search inputs by searching among all the fuzzy logic rules those which could not be triggered and the reason for this non-triggering. The introspection uses the logic rules. For example, in classical logic, if we have A or B or C and A is true, then the values ​​of B and C are not useful. On the other hand, if A is false, a value is required for B and C. Similarly for the conjunction, if we have A and B and C and A is false, then the values ​​of B and C are not useful, etc. (see the description in relation to Figs. 4A-4H).

[0058] Thus, the introspection module 13 allows the knowledge-based system 5 to intelligently indicate which inputs it is missing (i.e. the minimum necessary subset) in order to be able to produce a relevant decision.

[0059] The reconstitution module 15 is then configured to reconstitute the additional decision-making search inputs from the introspection module 13 into a textual request for additional information in natural language intended for the user 3. Thus, the reconstitution module 15 returns a question to the user 3 and awaits their response, which must contain details regarding the situation.

[0060] Advantageously, the inference engine 9 included in the knowledge-based system 5 is configured to deduce the decisions or responses from the decision-making search inputs using logic rules including fuzzy logic rules stored in rule bases 7.

[0061] Fuzzy logic extends classical Boolean logic with partial truth values. It replaces the truth value of a proposition between true or false by a degree of truth, to be chosen for example in the domain of definition of a function. In fuzzy logic, there are therefore different levels of truth in the satisfaction of a condition. Thus, the inference engine 9 can embody more or less strict logics. The same application can thus call upon different implementations judiciously chosen according to the context.

[0062] The second interaction 12 is the justification of a decision. Indeed, once the decision is provided, the user 3 can request an explanation or a justification of this decision. In this case, the extraction module 11 is configured to extract from the user's request for explanation corresponding justification request entries necessary to enable the knowledge-based system 5 to provide an explanation. Then, an explanation generation module 17 included in the knowledge-based system 5 is used to generate a first explanation. The reconstruction module 15 is configured to transform the first explanation generated by the explanation generation module 17 into an explanation text in natural language intended for the user 3. If the user is not convinced, he can ask for clarifications and we find ourselves in a dialogue loop.Different approaches can be used to gather the important elements of the decision.

[0063] The third interaction 13 is a request for editing rules or knowledge. In this case, the extraction module 11 is configured to extract from the editing request, corresponding editing entries necessary to allow an update module 18 of the knowledge-based system 5 to update the rule or knowledge bases 7. The request and the editing of the rule or knowledge bases 7 may include a deletion, a modification, or an addition of at least one rule in the rule bases and / or information in the knowledge bases. The rule and knowledge bases are updated and the user is notified of the success or failure of his request.

[0064] It will be noted that as a variant, instead of having two initialization and dialogue phases, we can have only one phase in which the context is redetermined each time.

[0065] Advantageously, the extraction module 11 includes advanced automatic natural language processing (NLP) techniques for analyzing the user's text using stochastic, probabilistic or statistical methods.

[0066] Alternatively, the extraction module 11 comprises basic regular expression techniques or parser syntactic analysis techniques based on a formal language defined by a set of syntax rules.

[0067] Advantageously, the reconstitution module 15 includes advanced techniques for automatic generation of NLG “Natural Language Generation” text by transforming a formal representation of content into a textual form intended for the user.

[0068] Alternatively, the reconstruction module 15 comprises regenerated text models “templates”.

[0069] Furthermore, the conversational interface may comprise a voice conversion module 19 and a voice synthesis module 21. The conversion module 19 is configured to convert a voice message produced by the user 3 into a text message intended for the extraction module 11. The synthesis module 21 is configured to convert a text response from the knowledge-based system 5 into a voice response.

[0070] [Fig.3] very schematically illustrates a control and monitoring system for a treatment plant comprising a conversational interface, according to a preferred embodiment of the invention.

[0071] This example is an application of the conversational interface in a control system 31 and monitoring of a treatment plant 33. The treatment plant 33 comprises a tank 35 or basin filled with liquids. The control system 31 comprises a set of sensors 37 equipping the tank 33, a knowledge-based system 5, and a conversational interface 1. The conversational interface 1 is configured to interact with the knowledge-based system 5 according to the description in relation to Figs. 1 and 2.

[0072] The set of sensors 37 is intended to measure the values ​​of several parameters representative of the level of contamination of the tank 35. By way of example, this set of sensors 37 comprises sensors for measuring concentrations of calcium carbonate, 1,4-dioxane, NO2, particles, as well as sensors for measuring the flow rate of the liquid and the opening of the valves 39 of the tank.

[0073] The sensors 37 may be configured to provide the measurements to the knowledge-based system 5 periodically, for example every five minutes. Depending on the values ​​or levels measured by the sensors 37, the knowledge-based system 5 is configured to determine the level of contamination risk of the tank 35.

[0074] However, in certain cases it is useful for the knowledge-based system 5 to take into account other parameters or events which require deferred analyses and which can be requested from the user 3 via the conversa interface. tional 1. One of these parameters is the concentration of bacteria in tank 35 which allows the risk of contamination to be better determined. Another parameter is the turbidity (i.e. cloudy nature of the liquid) of tank 35 or the basin.

[0075] It should be noted that certain analyses may require a relatively long duration (for example several hours) and in this case, they are only requested to be carried out if they are necessary to obtain the decision.

[0076] Advantageously, the risk of contamination of the tank 35 is determined by applying fuzzy logic rules to the different parameters to be monitored in the purification station 33.

[0077] Fuzzy logic allows each parameter to be characterized by allowing different degrees of veracity. For example, the value of a parameter can be described as high, medium, low or other. This allows for rules that are closer to natural language, thus allowing for the formulation of a simpler, more nuanced explanation that is better suited to the application.

[0078] Figs. 4A-4H illustrate curves representing the parameters to be monitored in the treatment plant using fuzzy logic, according to a preferred embodiment of the invention.

[0079] These curves represent, using fuzzy logic, the degree of veracity or confidence rate (between 0 and 1) as a function of the values ​​of a physical, chemical or biological parameter.

[0080] More particularly, the three curves 41, 42, 43 in [Fig.4A] represent in a Cartesian coordinate system three levels of bacterial concentration according to different confidence levels. The abscissa represents the concentration expressed in number per gram (nb / gram) of bacteria while the ordinate represents the confidence level between 0 and 1. The first 41, second 42, and third 43 curves describe the bacterial concentrations according to low, medium, and high levels respectively. For example, for the first curve 41, a concentration below 25 nb / gram is considered low with a confidence level of 0.5 or a concentration below 10 nb / gram is considered low with a confidence level of 0.9.

[0081] For example, if a confidence level of 0.5 is chosen, then these three curves imply that a concentration below approximately 25 nb / gram presents a low level, that a concentration between approximately 25 nb / gram and 70 nb / gram presents a medium level, and that a concentration greater than approximately 70 nb / gram presents a high level.

[0082] The four curves 45, 46, 47, and 48 in [Fig.4B] represent four levels of calcium carbonate concentration. The abscissa represents the concentration expressed in milligrams per liter (mg / l) of calcium carbonate, and the ordinate represents the confidence level between 0 and 1. The first 45, second 46, third 47, and fourth 48 curves describe the calcium carbonate concentrations according to low, medium, high, and very high levels respectively.

[0083] For example, if a confidence level of 0.5 is chosen, then the four curves indicate that a concentration below about 10 mg / l is a low level, that a concentration between about 10 mg / l and 55 mg / l is a medium level, that a concentration between about 55 mg / l and 100 mg / l is a high level, and that a concentration greater than about 100 mg / l is a high level.

[0084] Curve 51 in [Fig.4C] represents the confidence level as a function of the concentration expressed in milligrams per kilogram (mg / Kg) of 1,4-dioxane. For example, a concentration greater than about 5500 mg / Kg is considered high.

[0085] Curve 53 in [Fig.4D] represents the confidence level as a function of the concentration expressed in micrograms per cubic meter (microg / m3) of NO2 level. According to this example, a concentration greater than approximately microg / m3 is considered high.

[0086] The two curves 55, 56 in [Fig.4E] respectively represent low and high levels of particle concentration expressed as number per gram (nb / g). For example, if a confidence level of 0.5 is chosen, then the two curves 55, 56 show that a concentration below about 30 nb / g presents a low level and that a concentration above about 30 nb / g presents a high level.

[0087] The three curves 61, 62, 63 in [Fig.4F] represent respectively low, medium, and high levels of a risk rate. For example, if a confidence rate of 0.5 is chosen, then these three curves 61, 62, 63 imply that a risk rate below approximately 35% presents a low level, that a risk rate between approximately 35% and 65% presents a medium level and that a risk rate above approximately 65% ​​presents a high level.

[0088] Curve 65 in [Fig.4G] represents the confidence rate as a function of the liquid flow rate expressed in liters per second (1 / s). For a confidence rate of 0.5, the flow rate is considered low below 25 1 / s.

[0089] Curve 67 in [Fig.4H] ​​represents the confidence rate as a function of the valve opening expressed as a percentage. For a confidence rate of 0.5, the valve opening is considered low below 75%.

[0090] All these curves can be stored in the knowledge bases. In addition, fuzzy logic rules based on the data of these curves are stored in the rule bases 7 of the knowledge-based system 5.

[0091] Fuzzy logic rules include formulas that connect propositions by connectives. The connectives of conjunction (and), disjunction (or), and implication (then) are represented in the following in bold italic characters.

[0092] The output of each rule designates the level (low, medium, high) of the “risk” of contamination. By default, the risk level is initially equal to zero “null”. The logic rules are activated to determine the current level of risk based on the different parameters.

[0093] By way of example, the rule bases may include, without limitation, the following rules:

[0094] 1. If the calcium carbonate level is high and (the NO2 level is high and the 1,4-dioxane level is high), then the risk is high.

[0095] 2. If the calcium carbonate level is high and (((the NO2 level is not high) or (the NO2 level is high and (the 1,4-dioxane level is high))) and the bacteria concentration is high), then the risk is high.

[0096] 3. If the calcium carbonate level is low and (((the NO2 level is not high) or (the NO2 level is high and (the 1,4-dioxane level is not high))) and ((the bacteria concentration is not high) and the particle concentration is high)), then the risk is medium.

[0097] 4. If the calcium carbonate level is low and (((the NO2 level is not high) or (NO2 level is high and (1,4-dioxane level is not high))) and ((bacteria concentration is not high) and ((particle concentration is not high) and ((upstream rate is not low) and downstream rate is low)))), then the risk is medium.

[0098] 5. If the calcium carbonate level is low and (((the NO2 level is not high) or (the NO2 level is high and (the 1,4-dioxane level is not high))) and ((the bacteria concentration is not high) and ((the particle concentration is not high) and ((the upstream rate is not low) and (the downstream rate is not low))))) then the risk is low.

[0099] 6. If the calcium carbonate level is low and (((the NO2 level is not high) or (NO2 level is high and (1,4-dioxane level is not high))) and ((bacteria concentration is not high) and ((particle concentration is not high) and (upstream flow rate is low and valve opening is high)))), then the risk is medium.

[0100] 7. If the calcium carbonate level is low and (((the NO2 level is not high) or (NO2 level is high and (1,4-dioxane level is not high))) and ((bacteria concentration is not high) and ((particle concentration is not high) and (upstream flow rate is low and (valve opening is not high))))), then the risk is medium.

[0101] Some scenarios relating to the activation of the above rules can be summarized as follows:

[0102] a) No rule is activated, in this case the risk is at its default value: Zero.

[0103] b) Rule 1 is activated, i.e. the calcium carbonate levels, nitrogen dioxide and 1,4-dioxane are high. The response calculated by the knowledge-based system 5 from the outputs of the sensors 37 can therefore be sent directly as is to the operator 3 via the conversational interface 1. The risk here is high.

[0104] c) The calcium carbonate level is high but not the nitrogen dioxide level. Rule 2 could be activated, but the information on the bacteria concentration is missing. The same goes for the other rules. A message is therefore sent via conversational interface 1 to operator 3 and the knowledge-based system 5 waits for his response concerning the bacteria concentration. For example, one hour later, operator 3 enters the results of the analysis and the inference can continue. If the bacteria concentration is high, rule 2 is activated and conversational interface 1 indicates to the operator that the risk is high.

[0105] d) The operator asks why the risk is high. By introspection, the conversational interface 1 selects the only activated rule, which is the second one and transmits the following information, verified in the premise: "because the calcium carbonate level is high and the NO2 level is not high and the bacteria concentration is high".

[0106] Of course, various modifications can be made by those skilled in the art to the invention which has just been described, solely by way of non-limiting examples.

Claims

Claims

1. Conversational interface with a knowledge-based system, characterized in that it comprises: - an extraction module (11) configured to extract from a question relating to a search for a decision, corresponding decision search inputs which are adapted to be used by said knowledge-based system (5) in order to generate a decision to said decision inputs, - an introspection module (13) configured in case said decision search inputs are insufficient, to determine additional decision search inputs which are missing from the knowledge-based system to generate the decision, and - a reconstitution module (15) configured to reconstitute said additional decision search inputs from the introspection module (13) or the decision from the knowledge-based system (5) into a natural language text.

2. Conversational interface according to claim 1, characterized in that the knowledge-based system comprises bases (7) of fuzzy logic rules, and in that the introspection module (13) is configured to determine said additional decision-making search inputs by searching among all the fuzzy logic rules those which could not be triggered and the reason for this non-triggering.

3. Conversational interface according to claim 1 or 2, characterized in that the extraction module (11) is further configured to extract from an explanation request corresponding justification request inputs necessary to enable said knowledge-based system (5) to provide an explanation, and in that the reconstruction module (15) is configured to produce a natural language explanation text from the explanation generated by the knowledge-based system.

4. Conversational interface according to any one of the preceding claims, characterized in that the extraction module (11) is further configured to extract from a rule editing request, corresponding editing entries to enable said knowledge-based system (5) to update the rule bases.

5. Conversational interface according to claim 4, characterized in that that the edition of the rule bases (7) includes a deletion, a modification, or an addition of at least one rule in the rule bases.

6. Conversational interface according to any one of the preceding claims, characterized in that the extraction module (11) is initially configured to extract from conversational information provided by the user context entries allowing the knowledge-based system to determine the user's desired context of the conversation and to load the logical rules adapted to said context.

7. Conversational interface according to any one of the preceding claims, characterized in that the extraction module (11) comprises NLP natural language automatic processing techniques, and in that the reconstruction module (15) comprises NLG text automatic generation techniques.

8. Conversational interface according to any one of claims 1 to 6, characterized in that the extraction module (11) comprises syntactic analysis or regular expression techniques, and in that the reconstruction module (15) comprises regenerated text models.

9. Conversational interface according to any one of the preceding claims, characterized in that it comprises a conversion module (19) for converting a voice message into a text message intended for the extraction module (11), and a synthesis module (21) for converting a text response from the knowledge-based system (5) into a voice response.

10. Conversational interface according to any one of the preceding claims, characterized in that the knowledge-based system (5) is an expert system, a planning system, or a constraint solver.

11. 1 Control system (31) of a purification station (33) comprising a tank (35) equipped with sensors (37) and adapted to contain liquids, characterized in that it comprises a knowledge-based system (5) and a conversational interface (1) according to any one of the preceding claims, the conversational interface (1) being configured to interact with the knowledge-based system (5) according to measurements carried out by the sensors (37) and / or operators (3) relating to parameters representative of the level of contamination of the tank (35).

12. 12. Conversational method with a knowledge-based system, characterized in that it comprises the following steps: - extracting from a question relating to a search for a decision, corresponding decision search inputs which are suitable for use by said knowledge-based system (5) in order to generate a decision to said decision inputs, - determining in case said decision search inputs are insufficient, additional decision search inputs which are missing from the knowledge-based system to generate the decision, and - reconstructing said additional decision search inputs or said decision into a natural language text.

13. 3Conversational method according to claim 12, characterized in that the determination of said additional decisional search inputs is carried out by searching among fuzzy logic rules recorded in rule bases of the knowledge-based system, those which could not be triggered and the reason for this non-triggering.

14. 4Conversational method according to claim 13, characterized in that it comprises: - an initialization phase (PI) for extracting context entries from conversational information allowing the knowledge-based system to determine the context of the conversation and to load the fuzzy logic rules according to said context, and - a dialogue phase (P2) comprising the search for a decision, the request for an explanation of a decision, or the editing of the rule bases.

15. 5Conversational method according to claim 12 or 13, characterized in that it comprises a single phase in which a context is redetermined each time.

Images