User assistance through semantic graph analysis

Abstract

The present description relates to assistance implemented by a computing device, comprising: - establishing a current semantic graph of a text obtained from a current assistance request message (S34) received by the computing device, - searching for at least one graph similar to the current semantic graph (S36), in a business knowledge database (BDD) listing process steps associated with semantic graphs, in order to identify a process step (S37) corresponding to the current message, and - generating a signal (SIG) comprising recommendation data (S39) relating at least to the identified step.

Description

Description Title: User assistance through semantic graph analysis Technical field

[0001] This description relates to data processing occurring in computer applications, in particular but not limited to office applications. Prior art

[0002] In the context of the "digital enterprise", the processes on which users work, or to which these users contribute, require several IT applications and numerous people involved to move forward over time: customer, operator, support team, partner, voice or chat systems.

[0003] Experience is the sum of trials and errors or trials and successes to achieve a goal. For a given user, experience allows a "process" to move forward and therefore takes into account their sensitivity and history of intervention in such a process.

[0004] When a user is asked to intervene on a specific part of the process and is required to interact with a third party or system (via voice exchange, chat or email), the user is not aware of what has been done in an equivalent way during the execution of other instances of the same process and what the useful business knowledge of people with a role similar to that of the user has been and the impact, positive or negative, of the exploitation of this knowledge.

[0005] The user generally only has access to the history of the current instance of the process and to his own experiences.

[0006] Feedback and its use are often difficult and take some time. However, it is important to use this feedback, especially that of the different stakeholders. To collect feedback, it is possible to naturally approach other business stakeholders to ask them questions about a common situation / problem in order to determine if they have experienced a similar situation or if they recommend other business stakeholders to consult. However, this means of searching for information can be costly in terms of time and can be ineffective if a business stakeholder has a limited network of contacts. Summary

[0007] This disclosure improves the situation.

[0008] An assistance method implemented by a computer device is proposed, comprising: - establish a current semantic graph of a text obtained from a current message, a request for assistance received by said computer device, - search, in a business knowledge database listing process steps corresponding to semantic graphs, for at least one graph similar to the current semantic graph, to identify a process step corresponding to the current message, - generate a signal comprising recommendation data relating at least to the identified step.

[0009] A "process" is a set of steps that must be carried out, for example by a user in a professional setting. The steps in a process can, for example, be linked together in a series of steps.

[0010] It is then proposed to provide relevant recommendations to a user on a process in progress, and in particular on a step to be carried out in this process and possibly on one or more steps which follow the step to be carried out (to guide the user in the process). These recommendations can be relevant thanks in particular to: - the richness (in terms of number and / or diversity of elements for example), or even the exhaustiveness, of the database, on the one hand, and - looking for similarity(ies) between the current graph and the graphs in the database, on the other hand.

[0011] This application relates in particular to an assistance method implemented by a computer device, comprising: - establish a current semantic graph of a text obtained from a current message, requesting assistance, received by said computer device, - search, in a business knowledge database listing process steps in correspondence with semantic graphs, at least one graph of which at least one part is similar to the current semantic graph, to identify at least one step of at least one process corresponding to the current message, - generate a signal comprising recommendation data (S39) relating at least to the identified step.

[0012] In at least one embodiment, said database lists steps of several business processes.

[0013] In at least one embodiment, similarity scores with the current semantic graph can be estimated for at least part of the graphs in the database, to search for at least one graph having the best similarity score (i.e. the similarity score representative of the highest similarity with said current semantic graph). The generated signal can then comprise recommendation data relating at least to the step identified as corresponding to the graph having the aforementioned best similarity score, for example.

[0014] For example, the similarity measurement can be performed with respect to nodes and / or edges and / or a graph structure.

[0015] In at least one embodiment, the database is searched for a plurality of graphs having better similarity scores (i.e., similarity scores representative of the highest similarities with said current semantic graph), and the method further comprises an estimation of an efficiency score for respecting business rules, this efficiency score being able to be assigned to at least part of the graphs of the aforementioned plurality of graphs.

[0016] In at least one embodiment where the database contains recommendation data associated with each graph in the database, such an embodiment may, for example, make it possible to order a list of recommendations respectively associated with the graphs, according to their effectiveness scores.

[0017] Thus, in such an embodiment, the method may comprise an ordering of recommendation data associated with the graphs of the aforementioned plurality of graphs, according to the respective effectiveness scores assigned to the graphs of the plurality of graphs. Thus, the generated signal may comprise the recommendation data ordered according to these effectiveness scores.

[0018] In at least one embodiment, said similarity scores are weighted by said effectiveness scores.

[0019] In at least one embodiment, said method comprises storing the data of the current semantic graph in the database (BDD), said storage being carried out conditionally, as a function of a similarity between said current semantic graph and the graphs of the database corresponding to the at least one identified step, so as to store in the database only the graphs least similar to each other.

[0020] For the construction of the corpus of graphs of the database, it is possible to provide, in one embodiment, a storage of data of the current graph in the database if the best similarity score is associated with a similarity lower than a similarity associated with a first similarity score value.

[0021] Thus, in at least one embodiment, said storage of the data of the current graph in the database (BDD) is triggered if the similarity scores of the graphs of the database corresponding to the at least one identified step are representative of a similarity lower than a similarity associated with a first similarity score value.

[0022] In at least one embodiment, where the similarity scores are increasing as the similarity increases, said storing of the data of the current graph in the database (BDD) is triggered if the highest similarity score of the graphs in the database corresponding to the at least one identified step is lower than said first similarity score value.

[0023] For example, if the similarity scores are increasing with the similarity, then the storage of the data of the current graph occurs if its similarity score is lower than the first value mentioned above. For example, this first value can be a threshold value, which can be fixed or variable depending for example on a number of graphs already stored for the same process in the database.

[0024] Thus, in such an implementation, it is chosen to save the current graph by attaching it to a current step (the identified step, mentioned above) if this graph is sufficiently different from the others graphs in the database, also attached to the same step. If no graph in the database is similar to the current graph beyond the first value mentioned above (with scores increasing according to similarity), then the information contained in the current graph is new and should be saved in the database. Otherwise (similarity scores beyond the first value), it is possible to simply create a link that persists the information provided by the current graph in the database. This link can be an association between an instance identifier of the current process, a step identifier and a graph identifier that is most similar in the database. Generally speaking, the data in a graph in the database can, for example, be referenced by correspondence of an instance identifier and a step identifier, for example. A "process instance" typically means an actual realization of the process.

[0025] The aforementioned implementation, aimed at storing graphs that "differ" from those stored in the database, can help to list all the ideas exchanged in the messages, therefore all the experiences, in order to participate in offering a wide variety of possible solutions to the user. If the corresponding graph is different from the graphs in the database, it is preferred to save this graph in the database with its specificities. If the graph is similar to the other graphs, a link is recorded between the instance of the current process and the similar graph in the database. Thus, unnecessary duplication of business information that may be relevant can be limited.

[0026] In at least one embodiment, said similarity score is estimated in the background by said computing device.

[0027] In at least one embodiment, said generated signal is rendered on an output user interface of said computing device.

[0028] In at least one embodiment, the generated signal may include recommendation data relating to at least one step following the identified step in the process of the identified step. Such an embodiment may assist the user in visualizing the sequence of steps in the process and thus organizing their own completion of the current step.

[0029] In at least one embodiment, generation of the signal comprising the recommendation data may be triggered at least based on a mood detected in the current message.

[0030] In at least one embodiment, the generation of the signal comprising the recommendation data can be triggered at least upon detection of an interrogative form, specific to a question, in the current message.

[0031] For example, if a question form is detected and / or if the mood detected in the message (beyond a threshold for example) is fear and / or anger and / or sadness, then a work computer of the user, executing the above method, can spontaneously offer the user assistance by indicating to him (for example in a “popup” window displayed on a work screen connected to the computer) that recommendations for carrying out the step that is the subject of his message are available. If the user clicks on the window for example, then the computer can for example control the display on the screen of a page of recommendations for carrying out the current step of the process.

[0032] The estimation of the similarity of the current graph with respect to the graphs in the database (and its possible storage in the database) may, on the other hand, correspond to a background task, executed permanently by the aforementioned computer, without requiring any particular intervention from the user.

[0033] In at least one embodiment, the method may further comprise, prior to establishing the current semantic graph, a structuring of the text data contained in the current message. Such an embodiment may in particular help to filter graphs from the database similar to the current graph.

[0034] This structuring of data attached to the steps of a process aims at a so-called "Process Mining" treatment. By attaching a current graph to a given step, if the similarity measure reveals that many graphs in the database are attached to the same step A, process mining can help to identify in addition another characteristic in the message which can be for example a particular geographical area. Thus, it may be possible to filter the graphs attached to step A on the one hand and which have the same value 'geographic area' on the other hand. It may then be possible to measure again the semantic similarity of the graphs thus filtered with the current graph. This is then a possible optimization to identify similar graphs and contexts, in order to suggest better actions (or at least relevant actions) to a user.

[0035] In at least one embodiment, the current message may be a voice message, and the method may include transcribing the voice message into text data, prior to establishing the current semantic graph.

[0036] This description also relates to a computer device comprising a processing circuit for implementing the method according to the present invention.

[0037] According to another aspect, there is provided a computer program comprising instructions for implementing all or part of a method as defined herein when such program is executed by a processor. According to another aspect, there is provided a non-transitory, computer-readable recording medium on which such a program is recorded. Brief description of the drawings

[0038] Other features, details and advantages will become apparent upon reading the detailed description below, and upon analyzing the attached drawings, in which: Fig. 1

[0039] [Fig. 1] shows an example of processing a message according to one embodiment. Fig. 2

[0040] [Fig. 2] shows an example of similarity search between semantic graphs according to one embodiment. Fig. 3

[0041] [Fig. 3] shows an example method for processing a message and updating a BDD database storing a corpus of semantic graphs according to one embodiment. Fig. 4

[0042] [Fig. 4] shows an example of a device for implementing the method of Figure 3, according to one embodiment. Description of the embodiments

[0043] In one embodiment, the method set out below proposes a combination of computer techniques known as “Process Mining” and representation by semantic graphs such as for example “AMR” representation (for “Abstract Meaning Representation”) to determine feedback from the user and possibly other actors in connection with a current situation, in order to improve the conduct of a process using a plurality of computer applications, in particular office applications.

[0044] Process Mining is used to identify process steps that are very close to a step currently being executed (in an application, for example) with their characteristics and / or their efficiency index according to chosen criteria (for example, a company's key performance indicators or KPIs).

[0045] The AMR technique based on semantic representation graphs is used to model ongoing information exchanges and persist them at different stages of the process to complete an experience engine. This makes it possible, for example, to subsequently perform inter-instance graph comparisons of processes (in the same computer application or in different computer applications).

[0046] The aforementioned method implements data processing of the applications currently running to offer a current user of these applications recommendations for problem solving and / or optimizing the way to carry out a process step. Such recommendations may be provided in particular during a voice or text exchange between the user and another user or an agent of the chatbot type or other. To this end, the method relies on a set of previous experiences of the current user and other users, by suggesting the display of business knowledge data on the process in progress and taking into account criteria such as in particular the objectives of a company employing the current user (for example a customer satisfaction index, a cost reduction, a carbon footprint reduction, an employee satisfaction index, or other elements of a company's strategy).

[0047] In a first example, the current user may receive a voice call related to the execution of a process step (as part of an execution of one or more applications The conversation during the voice call is then transcribed and the AMR module synthesizes the ongoing conversation in the form of a graph.

[0048] On the one hand, graph analysis can identify a question or problem to be resolved in the conversation. On the other hand, other applications, such as a so-called "sentiment analysis" component, can, for example, detect dissatisfaction among at least one of the speakers.

[0049] Thus, at least one AMR graph analysis component (possibly in combination with another component such as a sentiment analyzer) can trigger the execution of an experience engine (if for example a questioning index and / or a dissatisfaction index are above a threshold).

[0050] A possible implementation of the method can thus rely on such a sentiment analysis component which, upon detection of a negative sentiment, activates the triggering of the experience engine.

[0051] The experience engine filters similar process instances (e.g., those with the same path of operations up to a current operation) and takes into account the AMR graph to compare it with other AMR graphs present in the experience database. For example, it is possible to search for processes with similar portions, with identical sequences in terms of steps and actions performed.

[0052] An output signal may then include business knowledge data associated with the process and in particular with a current operation of the process. These elements may be selected according to their effectiveness index (in particular according to the number of samples of similar situations and the effectiveness of the recommendation given in said similar situations).

[0053] In a second example, the current user may receive a voice call from someone related to the execution of a process step and the conversation is transcribed. The AMR module synthesizes the conversation in the form of a graph. The analysis of the graph makes it possible to filter and retain only the branches related to business knowledge associated with the current activity. The experience engine then keeps in memory the graph linked to the process instance, even after the process has ended, which is hereinafter called “graph persistence at the process instance”.

[0054] Furthermore, "instance" here means the actual realization of a process (or at least part of this process). Thus, each particular implementation of the process is an instance of the process. Each instance therefore has the same steps (or a subset of these steps) as the process itself. Indeed, the overall process can be defined by a succession of several steps, according to a notion that is thus more global than an isolated step. An example of a process can thus be "connecting to the fiber," containing a certain number of distinct steps. An instance of this process typically expresses the connection of a particular person who has followed the steps in an order specific to them.

[0055] The method described here can thus deliver recommendations to users, operators, or process technical assistants. It can also provide them with visibility into possible paths for carrying out the process. It should be noted that these recommendations can be offered to ordinary, human users, or to robot operators such as chatbots or others.

[0056] The process can thus allow for consideration, in real or near-real time, of solutions from similar experiments with a good efficiency index, thanks in particular to the analysis of AMR graphs in a conversation.

[0057] The operation of the aforementioned experience engine in certain embodiments is detailed below. This engine aggregates all the 'experiences' on a single process. To ensure the operation of this engine, two main components can be provided: - A first component responsible for the ongoing collection of significant data from the experiences of business actors in the same process, and - A second component delivering business knowledge suggestions (or “recommendations” hereinafter) for a given operation of the process, this business knowledge having the best efficiency index.

[0058] For the implementation of the first component, business experience data is collected in two types of format: - audio and / or text data reflecting an exchange between business actors and capturing business expertise expressed in the exchange, and - business process execution analysis data relating to the experience of business actors in terms of sequencing and sequence of operations of said process. The first component may have the objective of storing these experiences in a specific format in an experience base having a structure facilitating their querying: by exploiting the specific model of the aforementioned process mining and a semantic representation of the AMR type.

[0059] The second component mentioned above, for its part, can filter data from experiments having the same path up to the current step of the process (for example Step 1 of Figure 2 discussed later). The second component identifies possible 'future' paths, for example those satisfying certain criteria, as in the example of Figure 2 those having a good efficiency index (Step 2 - Figure 2). The second component analyzes the AMR graphs of the associated experiments for comparison with the current graph in order to identify knowledge not present in the current graph and to propose this knowledge not present to the current user (Steps 4 and 5 - Figure 2). Furthermore, a step (Step 0) of construction of the knowledge database, including semantic graph data, is planned to be carried out before and / or in parallel with steps 1 to 4 (as a continuous background task for example).

[0060] In the following, we detail each component.

[0061] For the first component carrying out the on-the-fly collection, it is planned, in this example of implementation, the implementation of at least four stages (illustrated as an example in Figure 1). The first stage aims to collect the data resulting from the exchanges between the business actors (steps 1 and 1' of figure 1). This data can be in audio format (for example, data from a telephone call recording) and / or in text format (character data in an email, chat, or other) for example. In a second step 2, the audio data is transcribed in text format. The third step 3 generates, from the text format of an exchange, an AMR graph that summarizes the exchange (step 3, figure 1). For example, each "collected" exchange can thus be transmitted in AMR form.

[0062] Recall that an AMR graph is intended to abstract from syntactic representations, in the sense that sentences with similar meaning are assigned the same AMR graph, even if they are not formulated identically. An AMR analysis component can typically identify words in a sentence to define them as nodes of the graph, to assign them respective values ​​called "arguments" (e.g. ARGO, ARG1 , etc., in the AMR graph example below).

[0063] For example, the following sentences are identified in a conversation: # ::snt There are several possibilities for extending the network. Have you considered using powerline?

[0064] The corresponding AMR graph is then written:

[0065] (m / multi-sentence

[0066] ... :snt1 (ii / intend-01

[0067] . :ARG0 (w / we)

[0068] . :ARG1 (r / regard-01

[0069] . :ARG0 w

[0070] . :ARG1 (r2 / network)))

[0071] ... :snt2 (p / possible-01

[0072] . :ARG1 (u / utilize-01

[0073] . :ARG0 (y / you)

[0074] . :ARG1 (p2 / carrier

[0075] . :mod (c / current)))

[0076] . :mod (p3 / more-ieur))

[0077] ,...:snt3 (t / think-01

[0078] . :ARG0 (y2 / you)

[0079] . :ARG1 (u2 / use-01

[0080] . :ARG0 y2

[0081] . :ARG1 p2)))

[0082] As illustrated in Figure 1, the method may comprise (Step 4 - Figure 1) an attachment of the graph to a current process step, by identifying similar graphs. It may notably comprise an analysis of the graph obtained for the current exchange to compare it to existing graphs. For example, similarity scores of a current graph (for the current step of a business process) with reference graphs of this business process may be calculated. For example, the reference graph(s) having a high similarity score(s) relative to the current graph may indicate the process step that is the subject of a user's message.

[0083] Determining similarity and estimating a similarity score between semantic graphs, for example of the AMR type, with respect to nodes and / or edges and / or a graph structure, may be implemented in different ways depending on the embodiments. For example, they may be implemented in a similar manner to what is described in this document:

[0084] “AMR Similarity Metrics from Principles,” by Juri Opitz, Letitia Parcalabescu, and Anette Frank, in Transactions of the Association for Computational Linguistics (Sept. 2020) 8: pp 522-538.

[0085] For example, the step of an ongoing process can be identified in a message exchange between two business actors, based on the best similarity score between: - the current AMR graph, taken from the exchanged messages, and - reference AMR graphs, listed in a database in which each AMR graph is specific to a process step.

[0086] This database can be built by defining, for each process step, an AMR graph that can be quite exhaustive (as exhaustive as possible for example), to characterize this process step by similarity with a current AMR graph (for example to characterize it with a high similarity score with a current AMR graph). In such an implementation, a human operator for example can define appropriate (and for example exhaustive) questions / answers allowing to define the corresponding process step.

[0087] In an implementation described below, the aforementioned database is built over time, according to the interactions between business actors. The reference AMR graphs are then stored in the database according to the aforementioned principle of "graph persistence". The trigger for this persistence can be the variability of the current graph. For example, if a similar graph is present in the database, then instance (instance ID), step (step ID) and graph (graph ID) references can be added to it. It is then possible to find the instances that have exploited the same graphs, while thus preserving as much business information as possible. Thus, a process step can have several reference graphs and it is possible to process a process step according to different dimensions and different degrees of granularity of reference graphs. The database of reference graphs can then be exhaustive and in particular representative of the variability of solutions, actions, responses provided by operators. For example, a set of semantic graphs referring to a set of exchanges that took place to carry out the same process instance can be stored in the database for the same instance. The different sets of reference graphs can be listed in particular according to an instance identifier (instance ID), for example.

[0088] Furthermore, efficiency scores specific to each process step may be assigned, for example depending on the importance of this step in the process, depending on the business rules that this step takes into account, etc. Thus, an embodiment of the present description may go further than a simple estimation of a similarity score between AMR graphs. In the example illustrated in Figure 2, the reference semantic graph having the best similarity score is the one linked to step B of the identified process (illustrated in gray). However, an efficiency score for a given step of a process may be further estimated, as described above. Typically, in the example of Figure 2, if it is estimated that several recommendations that can be given to the user on steps A and B have similar efficiency scores of 0.8, relative to the current process, then two recommendations concerning steps A and B can be given to the user. These efficiency scores are constructed from efficiency indices which can be a function of the characteristics of the step and given business objectives (such as for example the safety of the intervention teams, customer satisfaction, carbon footprint, etc.). These efficiency scores can thus weight similarity scores of reference graphs of the same process to offer the user relevant information on one or more steps of the process. These efficiency scores (or “performance indices”) can be calculated on the fly for a current message to be processed and, in an exemplary embodiment, they may not be persisted in the graphs of the database, in order to dynamically take into account possible new rules which define the aforementioned “performance”.

[0089] Figure 3 illustrates an example of a succession of steps according to a method within the meaning of the description. During a first step S31, a MESS message exchanged between business actors is processed. A first processing of the message may include, for example, a transcription into text data if it is an audio message. The processing S31 may further include, for example: - a search for an interrogative form in the message to interpret the message as a question posed (intended for another actor for example), and / or - an identification of emotion in the message (for example fear, anger or sadness, characterizing stress of the user). In either of these cases, the recommendation engine may interpret the MESS message as requiring a recommendation, with the implementation of the following steps described below.

[0090] Typically, the MESS message may be further processed to be represented by structured data in step S32. This step S32 may for example comprise an identification of keywords in the text data (like a search engine) to filter in step S33 irrelevant messages, typically those not related to a potential business process. If keywords in this message are identified as potentially related to a business process (ok arrow at the output of test S33), the current semantic graph of this message is constructed in step S34, from the text data of the message (not necessarily in structured form).

[0091] In step S35, the corpus of reference AMR graphs relating to all the process steps (or possibly to a part of the process instances if particular keywords are taken into account in the message in step S33) is searched in the database BDD, in order to estimate a similarity score of each of these reference graphs with the current graph obtained in step S34. In step S36, the reference graph having the best similarity score is retained. Alternatively, it is possible to retain a chosen number N of reference graphs having the N best similarity scores with the current graph. These reference graphs are deemed to be attached to the same process step, according to their similarity score with the current graph. They can thus indicate an instance of the process directly linked to the message processed (step S37).For example, it is possible to search for the graphs attached to the current step and possibly to the following steps of the process and to calculate in step S38 the associated performance indices to give the user a list of recommendations, ordered according to these performance index estimates (step S39). Thus, the user is guided in these recommendations and can even visualize the path that the process can take subsequently, according to the chosen orientation. These recommendations are thus:. - on the one hand, in connection with the message formulated at the initial step S31, and - on the other hand, provide relevant informative content in relation to business rules and / or predefined business knowledge.

[0092] Such embodiments can therefore help to offer the user recommendations that are relevant to their current activity.

[0093] In step S40, the benefit of storing the semantic graph of the current message as a reference graph of the database BDD can be assessed. For example, it may be planned to store this current graph if the associated message expresses in particular a variation of experience, and / or given ideas, compared to other similar reference graphs. In such a case, for example, persistence can be triggered for storage in the database BDD. Furthermore, if the experience in the form of the current graph is already present in the database, a relationship between instance and graph can be created to avoid increasing the number of data to be stored in the database BDD. A similarity calculation can also be used to decide on the persistence of a graph in the database BDD.If the similarity score is very close to another reference graph attached to the same step, it may be decided not to need to store the current graph in the database and to just persist the link between the process instance, the identified step and the reference graph.

[0094] Messages exchanged by a user can be tracked to determine whether they have progressed through the process steps. Thus, the recommendation engine can track user implementations of processes, and therefore the evolution of exchanges over the course of these implementations. Semantic graphs can be updated, on the fly, in the BDD database, to store information on the execution of a process and thus allow capitalization on other instances of this process.

[0095] Thus, a "business knowledge" recommendation engine aims to suggest business knowledge to business actors during their interactions with other actors in order to help improve the effectiveness of these interactions (to help optimize them, for example). This engine can exploit on the fly (in real time): - data on the progress of a process instance on which these actors are working, - data from other ongoing or completed instances, - data from message exchanges or audio conversations transcribed into text, this data being “structured” initially to be able to filter out irrelevant exchanges (S33) and which are not linked to a process, this data not necessarily then being “structured” for the establishment of semantic graphs (S34, S40), and - data for estimating efficiency indices linked to process instances.

[0096] In the embodiment of Figure 3, the recommendation engine can for example present the experiences according to their effectiveness index (step S39). The engine can present reference semantic graphs (or sets of predicates), which are relevant to concrete business processes linked to the graph of the current exchange (audio and / or written conversation). The engine can in particular perform an ordering of the recommendations according to their effectiveness score to propose to the user associated business knowledge according to a relevant presentation order. This business knowledge can present for example predicates which are not present in the current conversation and which can suggest new ideas to the user.

[0097] Effectiveness indices can be estimated based on coefficients corresponding to company objectives, such as sustainability, customer satisfaction, employee satisfaction, etc. For example, if a predicate in the graph illustrating one of the possible recommendations is about replacing equipment with recycled equipment, a high weighting (e.g. 1) can be assigned to the coefficient relating to sustainability. Thus, a high performance index can be estimated for this recommendation. Similarly, if a predicate is about changing a piece of equipment without completely replacing the equipment, a high weighting can be retained. If a predicate in the recommendation graph is about personalized follow-up of a problem to be addressed or an analysis of a positive customer sentiment in a conversation, a high weighting can be assigned to the coefficient relating to customer satisfaction, etc. If the company's strategy prioritizes sustainability over other objectives, the weighting of this sustainability coefficient may have a greater value.

[0098] Furthermore, these recommendations can concern the current step to be executed but also subsequent steps. Thus, the BDD database can be searched for similarity up to the current step and then graphs at one or more subsequent steps can be searched to suggest actions likely to advance the process overall.

[0099] Such a method can be executed by a processor PROC (figure 4) cooperating with a working memory MEM and with an interface INT for accessing the database BDD, typically of a processing circuit CT of a computer device such as a workstation (computer, tablet, or other) for example available to tele-advisors or members of business process support teams, for example on virtualized technical assistance platforms or others. The aforementioned processing circuit CT then further comprises an output interface OUT controlled by the processor PROC to generate a signal SIG (a visual signal on a screen ECR and / or an audio signal on a speaker HP, the screen and the speaker connected to the output user interface OUT). This signal SIG then comprises relevant recommendation data, relating to the identified step (step S37), or more generally to the process whose step has been identified.

Claims

Claims

1. Assistance method implemented by a computer device, comprising: - establishing a current semantic graph of a text obtained from a current message (S34), requesting assistance, which said computer device receives, - searching, in a business knowledge database (BDD) listing process steps in correspondence with semantic graphs, at least one graph of which at least one part is similar to the current semantic graph (S36), to identify at least one step of at least one process (S37) corresponding to the current message, - generate a signal (SIG) comprising recommendation data (S39) relating at least to the identified step.

2. The method of claim 1 wherein said database lists steps of multiple business processes.

3. Method according to claim 1 or claim 2, in which similarity scores with the current semantic graph are estimated for at least part of the graphs of the database, to search for at least one graph having the similarity score representative of the highest similarity with said current semantic graph (S36), and in which the generated signal comprises recommendation data relating at least to the step identified as corresponding to the graph having said similarity score representative of the highest similarity.

4. Method according to at least one of the preceding claims, in which a plurality of graphs having similarity scores representative of the highest similarities with said current semantic graph (S36) are searched in the database, the method further comprising an estimation of an effectiveness score for respecting business rules, assigned to each of the graphs of said plurality of graphs.

5. A method according to claim 4, comprising ordering recommendation data associated with the graphs of said plurality of graphs, according to the respective effectiveness scores assigned to the graphs of the plurality of graphs, the generated signal comprising the recommendation data ordered according to said effectiveness scores.

6. A method according to claim 4 or claim 5 wherein said similarity scores are weighted by said effectiveness scores.

7. Method according to at least one of claims 1 to 6, comprising storing the data of the current semantic graph in the database (BDD), said storage being carried out conditionally, as a function of a similarity between said current semantic graph and the graphs of the database corresponding to the at least one identified step, so as to store in the database only the graphs least similar to each other.

8. Method according to claim 7, wherein said storing of the data of the current graph in the database (BDD) is triggered if the similarity scores of the graphs of the database corresponding to the at least one identified step are representative of a similarity lower than a similarity associated with a first similarity score value.

9. Method according to claim 8, wherein the similarity scores are increasing as the similarity increases, and wherein said storing of the data of the current graph in the database (BDD) is triggered if the highest similarity score of the graphs in the database corresponding to the at least one identified step is lower than said first similarity score value.

10. Method according to at least one of claims 3 to 9, wherein said similarity score is estimated in the background by said computing device.

11. A method according to at least one of claims 1 to 10, wherein said generated signal is rendered on an output user interface of said computing device.

12. Method according to at least one of the preceding claims, wherein the generated signal comprises recommendation data (S39) relating to at least one step which follows the identified step in the process whose step is identified.

13. Method according to at least one of the preceding claims, in which the generation of the signal (SIG) comprising the recommendation data (S39) is triggered at least as a function of a mood detected in said current message.

14. Method according to at least one of the preceding claims, in which the generation of the signal (SIG) comprising the recommendation data (S39) is triggered at least upon detection of an interrogative form, specific to a question, in the current message.

15. Method according to one of the preceding claims, comprising, prior to the establishment of the current semantic graph, a structuring of the text data contained in the current message (S32), to filter graphs from the database similar to the current graph.

16. Method according to one of the preceding claims, in which the current message is a voice message, the method comprising a transcription of the voice message into text data, before the establishment of the current semantic graph.

17. Method according to at least one of the preceding claims, where the method comprises filtering the current semantic graph, said filtering taking into account a presence in said text obtained of keywords linked to at least one of said processes, to retain only the branches of the current semantic graph linked to said keywords.

18. Computer device comprising a processing circuit (CT) for implementing the method according to at least one of the preceding claims.

19. Computer program comprising instructions for implementing the method according to at least one of claims 1 to 17 when this program is executed by a processor (PROC) of a processing circuit (CT).

20. Non-transitory recording medium, readable by a computer, on which is recorded a computer program comprising instructions for implementing the method according to at least one of claims 1 to 17 when this program is executed by a processor (PROC) of a processing circuit (CT).

Images