Method for communicating with a computing device comprising a knowledge base

Abstract

The invention relates to a method for communicating with a computing device (DI) comprising a knowledge base (BC) modelling data in the form of a knowledge graph (GC), said method comprising the following, carried out by said device: - receiving (S2) a question posed using a natural language interface, - generating (S4-S5) a response to the question using a language model (LNT), the generated response comprising text data (TXT) from a text conversion (S4) of a data table, and during the text conversion data of the table being enriched (S3) using data of the knowledge graph.

Description

[0001] DESCRIPTION

[0002] Method of communication with a computer system comprising a knowledge base

[0003] Scope of the invention

[0004] The field of the invention is that of querying databases. More specifically, the present invention relates to a method of communicating with a computing device (computer, server, platform, etc.) comprising a knowledge base modeling data in the form of a knowledge graph. During communication, the computing device is queried by a user using a natural language interface and responds to the user in natural language, using knowledge such as tabular data that can be either provided by the user, acquired through training, or obtained from the Web. The present invention also relates to the computing device, a corresponding computer program, and a storage medium.

[0005] Previous art

[0006] With the development of natural language communication based on artificial intelligence techniques, numerous applications now allow users to converse with machines in natural language. One such application is the ChatGPT™ conversational agent, which is specifically designed to receive a question posed by a user in natural language, understand that question, and provide a natural language answer as accurately as possible. To generate such an answer, a conversational agent of this type sometimes needs to convert data tables into text. However, current techniques for converting data tables into text are limited by an insufficiently precise conceptualization of tabular data, making the natural language conversion less accurate and detailed, and generally less informative.Indeed, some implicit information in a table is not transcribed into textual form, which impairs the quality of the conversion and sometimes results in text content with undesirable ambiguities. Furthermore, the complexity of the relationships between the data in a table is not always adequately addressed by prior techniques, given a limited understanding of table structure. Object and summary of the invention.

[0007] One of the aims of the invention is to remedy at least one of the drawbacks of the aforementioned prior art by proposing a new technique for converting a data table into text used in the context of a natural language dialogue between a user and a machine, based on the use of data from a knowledge graph.

[0008] To this end, an object of the present invention relates to a method of communication with a computer device comprising a knowledge base modeling data in the form of a knowledge graph, said method comprising the following, at the level of said device:

[0009] - to receive a question posed using a natural language interface,

[0010] - generate an answer to said question using a language model, said generated answer comprising textual data from a textual conversion of a data table, during which data from the table are enriched using data from the knowledge graph.

[0011] Such a communication method has the advantage of contributing to the improvement of the user experience, the efficiency of data processing in order to develop a response, as well as the accuracy of the response provided.

[0012] In particular, such a communication method allows a significant improvement in the accessibility of data by a user, by allowing the injection into the responses provided to the user of data from a textual conversion of a data table, such data being more accessible and more understandable than data from a table which proves to be complex and difficult for a user to interpret.

[0013] Furthermore, using data from a knowledge graph to perform this conversion has the advantage of enriching the table's data, and therefore the content of the responses provided to the user. Indeed, the communication method according to the invention makes it possible to integrate additional information and relationships between the data, thus enriching the textual response.

[0014] The communication method according to the invention also allows users to easily navigate through textual information, which is often more intuitive and easier than browsing data tables. In one particular embodiment, enriching the table data may include the following:

[0015] - semantically annotate at least some of the data in the table using data from the knowledge graph,

[0016] - for at least one column of the table containing semantically annotated data, obtain additional information on said semantically annotated data from the knowledge graph data, and supplement said semantically annotated data in said at least one column with the additional information obtained, and the textual conversion includes the following:

[0017] - query said language model from a prompt generated in natural language by said device, said prompt asking the language model to generate a textual description of the table from the semantically annotated table data that has been completed,

[0018] - receive the aforementioned textual description generated in response to the query.

[0019] Thanks to such an embodiment, it is possible to enrich a data table before its conversion to text from a knowledge graph, so as to obtain a text whose content is, on the one hand, enriched and therefore particularly precise, and, on the other hand, in which ambiguities are minimized.

[0020] This method of implementation also offers the following advantages:

[0021] - The use of an intermediate structure such as the knowledge graph makes it possible to capture and faithfully convert the knowledge contained in the tabular data and to augment this knowledge with new related facts from the knowledge graph;

[0022] - Verbalization based on a language model ensures a high-quality textual description.

[0023] According to another particular embodiment, the semantic annotation of at least some of the data in the table may include, for at least one cell and / or at least one column and / or at least one pair of columns in the data table, an extraction, in the knowledge graph, of at least one identifier of an entity in the knowledge graph.

[0024] Using knowledge graph entity identifiers to annotate a data table allows for the retrieval of additional information, enriching the table data with more complete and precise details. By using specific identifiers, the communication process can access accurate and verified information, thus improving the precision of the annotations. Entity identifiers allow the table data to be linked to a broader network of information within the knowledge graph, revealing relationships and dependencies that would not otherwise be apparent. Such semantically annotated data can enable a user to discover new relationships within the table data, predict trends, identify complex correlations, and so on. This implementation significantly improves the quality of the table's text conversion.

[0025] According to another particular embodiment, the semantic annotation of at least some of the data in the table may include the following:

[0026] - associate at least one cell in the data table with an entity identifier extracted from the knowledge graph,

[0027] - associate at least one column of the data table with an entity type identifier extracted from the knowledge graph,

[0028] - associate at least one pair of columns from the table with a property identifier from the knowledge graph.

[0029] This embodiment allows the data table to be annotated according to several levels of knowledge. The data table is characterized using the entities, types / classes, and properties carried by the knowledge graph, which constitute different types of information within the knowledge graph. This results in a very precise and detailed annotation of the data table, which in turn allows for a similarly precise and detailed conversion of the data table to text. Furthermore, this embodiment improves the quality of the table data, as incomplete or incorrect data is corrected and completed. Thus, the quality of the table's conversion to text, and therefore the response provided to the user, is improved.According to another particular embodiment, obtaining additional information about said column of the semantically annotated data table may include extracting, from said knowledge graph, knowledge related to said entity identifier, said entity type identifier or said property identifier which was associated during the semantic annotation of at least some of the data in the table.

[0030] This implementation allows for the enrichment of at least one column of the table with additional information from the knowledge graph, thereby improving the quality and depth of the data available in the table. Furthermore, by associating specific entities with well-defined properties, the table data becomes more precise and contextualized, facilitating better understanding and use of the information contained within the table, particularly for converting the table into text.

[0031] According to another particular embodiment, obtaining additional information about the column in the semantically annotated data table may further include associating, using a type relationship, the additional information extracted in relation to the entity identifier with the additional information extracted in relation to the entity type identifier. Such an embodiment optimizes the semantic consistency of the data contained in a column by:

[0032] - Data alignment that ensures cell annotations in the column are consistent with the expected data type for that column, thus strengthening the semantic structure of the table,

[0033] - Explicit typing that clarifies the data type for the column, reducing ambiguities and misinterpretations.

[0034] This implementation method also allows for optimized annotation accuracy thanks to:

[0035] - the use of a typing context that provides additional context for cell annotations, improving the accuracy and relevance of these annotations,

[0036] - a reduction in conversion errors, by using as a reference the additional information extracted in connection with said entity type identifier to detect and correct potential errors in cell annotations. Such an implementation thus contributes to improving the quality of the data table's conversion to text, and therefore the quality of the response provided to the user.

[0037] According to another particular embodiment, obtaining additional information about a column of said pair of the semantically annotated data table may include associating additional information extracted in connection with an entity identifier for said column of said pair with additional information extracted in connection with said or another entity identifier for the other column of said pair.

[0038] This method of implementation allows for further optimization of the consistency of the annotations:

[0039] - by ensuring that cell annotations are consistent not only within each column, but also between linked columns,

[0040] - thanks to a strengthening of semantic structures, due to the establishment of explicit relationships between the entities in the columns of the pair.

[0041] This implementation also allows for the optimization of the enrichment of semantic annotations in the table:

[0042] - by providing additional relational context for cell annotations, thus enriching the data with information on the relationships between entities extracted from the knowledge graph,

[0043] - by augmenting the annotations with additional knowledge from the relationships identified between the columns.

[0044] Such an implementation thus helps to improve the quality of the conversion of the data table to text, and therefore the response provided to the user.

[0045] According to another particular embodiment, obtaining additional information about said column or about said column of said pair of the semantically annotated data table may include an association with additional knowledge extracted from the knowledge graph, additional information extracted in connection with said entity identifier and / or additional information extracted in connection with said entity type identifier and / or additional information extracted in connection with said property identifier.

[0046] This implementation method improves the enrichment of the table data:

[0047] - by enriching the semantic annotations with additional information from the knowledge graph, thus offering a more complete and detailed view of the data,

[0048] - by adding contextual knowledge that can improve the understanding and interpretation of the data. Such an implementation also improves the accuracy and relevance of the information contained in the table, thanks in particular to the fact that the additional knowledge from the knowledge graph allows for the refinement and clarification of the annotations, thus increasing their relevance, resulting in a reduction of ambiguities and a clarification of the meanings of entities and relationships.

[0049] This implementation also significantly improves data quality, and therefore the response provided to the user. Indeed, the additional knowledge from the knowledge graph enhances the consistency and reliability of annotations, thus increasing the overall data quality. Furthermore, the addition of this extra knowledge facilitates the detection of inconsistencies and anomalies in the table data, enabling quick and effective corrections.

[0050] Thanks to this method of implementation, the quality of the conversion of the data table to text is optimized, as is the accuracy and level of precision of the textual information resulting from this conversion.

[0051] According to another particular embodiment, at least some of the data in the table that are semantically annotated are character strings.

[0052] This implementation allows semantic annotation to be targeted at the truly relevant data in the table, namely the strings. Indeed, strings often carry rich and varied textual information, making them particularly suitable for semantic annotation. By focusing solely on strings, the communication process avoids the unnecessary complexity associated with annotating less informative data types, such as numbers, dates, identifiers, and so on. The implementation of this communication process is therefore simpler, more computationally efficient, and faster.

[0053] Furthermore, this implementation optimizes the reduction of annotation errors, as strings can be more easily associated with entities in the knowledge graph, since they contain rich textual information, such as descriptions, names, categories, and other textual details essential for semantic annotation. This results in a more accurate mapping to entities in the knowledge graph.

[0054] The various modes or features of embodiment mentioned above can be added independently or in combination with each other, to the communication process as defined above.

[0055] The invention also relates to a computer system comprising a knowledge base modeling data in the form of a knowledge graph, the computer system being configured to implement:

[0056] - receiving a question posed using a natural language interface,

[0057] - a generation of an answer to said question using a language model, said generated answer comprising textual data from a textual conversion of a data table, during which data from the table are enriched using data from the knowledge graph.

[0058] Such a device is specifically configured to implement the aforementioned communication process, according to one or the other of its modes of realization.

[0059] The invention also relates to a computer program comprising instructions for implementing the communication method according to the invention, according to any one of the particular embodiments described above, when said program is executed by a processor.

[0060] Such instructions can be stored permanently in a non-transient memory medium of the computer device implementing the communication method according to the invention.

[0061] This program can use any programming language, and be in the form of source code, object code, or code somewhere between source code and object code, such as in a partially compiled form, or in any other desirable form.

[0062] The invention also relates to a recording medium or information medium readable by a computer, and comprising instructions for a computer program as mentioned above.

[0063] The recording medium can be any entity or device capable of storing the program. For example, the medium can include a storage means, such as a ROM, for example a CD-ROM or a microelectronic circuit ROM, or a magnetic recording means, for example a mobile device, a hard drive or an SSD.

[0064] On the other hand, the recording medium can be a transmissible medium such as an electrical or optical signal, which can be transmitted via an electrical or optical cable, by radio, or by other means, so that the computer program it contains can be executed remotely. The program according to the invention can, in particular, be uploaded to a network, for example, an Internet-type network. Alternatively, the recording medium can be an integrated circuit in which the program is incorporated, the circuit being adapted to execute or to be used in the execution of the aforementioned communication method.

[0065] In an example implementation, the present technique is implemented using software and / or hardware components. In this context, the term "device" or "module" may refer in this document to a software component, a hardware component, or a set of hardware and software components.

[0066] Brief description of the drawings

[0067] Other features and advantages will become apparent upon reading particular embodiments of the invention, given by way of illustrative and non-limiting examples, and the accompanying drawings, among which:

[0068] Figure 1 represents an architecture in which the communication process is implemented, according to a particular embodiment of the invention. Figure 2 represents a computer device according to a particular embodiment of the invention, as implemented in the architecture of Figure 1. Figure 3 represents the main actions implemented in the communication process, according to a particular embodiment of the invention, as implemented in the architecture of Figure 1.

[0069] Figure 4 represents the main actions implemented during the execution of a semantic annotation action in the communication process, according to a particular embodiment of the invention, as implemented in the architecture of Figure 1.

[0070] Figure 5A represents an example data table used in the communication process as implemented in the architecture of Figure 1.

[0071] Figure 5B shows an example of a data table resulting from a semantic annotation of the data table in Figure 5A. Figure 5C shows an example of a data table resulting from a semantic enrichment of the semantically annotated data table in Figure 5B.

[0072] Figure 6 represents an example of a partial representation of an entity in a knowledge graph used in the communication process as implemented in the architecture of Figure 1.

[0073] Figure 7 represents the main actions implemented during the execution of a data enrichment action for the table in Figure 5B, according to a particular embodiment of the invention, as implemented in the architecture of Figure 1.

[0074] Figure 8A shows an example of a prompt generated during the implementation of the communication process as implemented in the architecture of Figure 1. Figure 8B shows an example of a response to the prompt in Figure 8A, which is generated during the implementation of the communication process as implemented in the architecture of Figure 1.

[0075] Figure 9 represents an example of a prompt and a response to that prompt generated using a state-of-the-art communication method.

[0076] Detailed description of an embodiment of the invention

[0077] Figure 1 represents an architecture in which a communication process is implemented according to an embodiment of the invention.

[0078] Such an architecture includes:

[0079] - a computer system DI comprising a knowledge base BC modeling data in the form of a knowledge graph GC;

[0080] - a human-machine interface (HMI) configured to be activated by a user (UT) to communicate with the computer device (DI).

[0081] The human-machine interface (HMI) is contained within a communication terminal (CT) that is configured to communicate with the computing device (CD) via a communication network (RC), such as an internet, intranet, etc. Examples of telecommunication terminals (CTs) include computers, smartphones, tablets, etc.

[0082] The IT system DI can include, for example, a computer, a server, a platform, etc. In Figure 1, the knowledge base BC is integrated into the IT system DI. Of course, such a knowledge base BC can be separate from the IT system DI, the latter then being configured to communicate with the knowledge base BC via any suitable means. The knowledge graph GC is, for example, of the Wikidata™, DBpedia™, Google Knowledge Graph™, Microsoft Concept Graph™, etc. type.

[0083] The UI interface can include, for example, a text-based graphical interface or a sound sensor coupled with a speech recognition interface.

[0084] Such a user interface (UI) communicates with a web-based internet browsing engine (NAV). The NAV includes a web-type application programming interface (INT) that allows the terminal user (UT) of the terminal (TC) to communicate in natural language with the computer device (DI).

[0085] We will now describe, with reference to figure 2, the simplified structure of the DI computer system.

[0086] According to the invention, the DI computer device comprises:

[0087] - a COM communication module configured to, on the one hand, receive requests generated using the UI interface, said request including a question in natural language asked by the user UT, and on the other hand, send responses in natural language to these requests to the communication terminal TC,

[0088] - a CMD command module configured to command the output of natural language responses to generated queries, in the form of a web page, PW,

[0089] - an ENR module for enriching the data of a TD data table using data from the GC knowledge graph, said ENR module being configured to generate an enriched TDE data table,

[0090] - a CNV module for converting TDE enriched data tables into TXT text.

[0091] The renewable energy enrichment module includes:

[0092] - a semantic annotation submodule ANS for the TD data table, said ANS submodule being configured to annotate at least some of the data in the TD table using data from the GC knowledge graph, so as to obtain a semantically annotated data table TDA; - an enrichment submodule EAS for enriching all or part of said at least some semantically annotated data, said EAS submodule being configured to, for at least one column of the TDA semantically annotated data table, obtain additional information on said semantically annotated data, or a part thereof, from the GC knowledge graph data, and supplement said semantically annotated data belonging to said at least one column, or a part thereof, with the additional information obtained. Following such enrichment, an enriched data table TDE is generated.The NVC module for textual data table conversion includes:

[0093] - a GPR prompt generation sub-module configured to automatically generate, i.e., without user intervention, a PRP prompt in natural language,

[0094] - a natural language model (LNT) configured to be queried from the generated prompt. Such a natural language model can be, for example, an n-grams model, a recurrent neural network (RNN), a large language model (LLM), etc. The natural language model (LNT) has been classically trained to acquire knowledge.

[0095] The PRP prompt contains a request to generate a textual description of the enriched TDE data table, based on the semantically annotated data, or a portion thereof, which has been supplemented with additional information obtained. The LNT language model, queried from the PRP prompt, is configured to generate a TXT textual description of the data table, such a description advantageously reproducing the table's content faithfully, while adding further details and clarifications.

[0096] According to the invention, the CMD command module is configured in particular to control the rendering of a web page containing the TXT textual description, via the user UI interface UT.

[0097] Optionally, the EAS submodule is further configured to obtain additional information about the semantically annotated data, or a portion thereof, for at least the specified column of the enriched data table (TDE), from data originating from at least one external knowledge source (SCE), such as, for example, a database, a corpus of documents of various types and / or formats (texts, tables, etc.), a web resource, etc. Obtaining such additional information is implemented using, for example, a Retrieval-Augmented Generation (RAG) technique. Since obtaining this additional information is optional, it is represented by a dashed line in Figure 2.

[0098] At initialization, the code instructions of the computer program PG are, for example, loaded into RAM (not shown) before being executed by the PROC processor. The PROC processor of the UTR processing unit implements the following actions, in particular, within the framework of the communication process described below, according to the instructions of the computer program PG:

[0099] - to receive a question posed using the natural language UI interface,

[0100] - Enriching data in a TD data table using data from the GC knowledge graph, such enrichment including:

[0101] -- an annotation of at least some of the data in the TD table using data from the GC knowledge graph,

[0102] -- an enrichment of all or part of said semantically annotated data, using data from the knowledge graph GC, and possibly using data from said at least one external knowledge source SCE,

[0103] - convert the enriched data table (TDE) into text format, said conversion including the following:

[0104] -- automatically generate a PRP prompt in natural language, requesting the LNT natural language model to generate a textual description of the TDE enriched data table, from the semantically annotated data, or a part thereof, which has been supplemented by the additional information obtained,

[0105] -- generate, using the LNT natural language model, an enriched textual TXT description of the TD data table,

[0106] - to order the rendering of a PW web page containing the enriched text description TXT. We now describe, in relation to figure 3, together figures 1 and 2, the main steps of a communication method with the computer device DI, according to a particular embodiment of the invention.

[0107] During step S1, the user UT sends a request Q to the computer system DI, via the IU interface. This request Q may include a question on a given topic, a request to summarize a data file attached to question Q, a request to convert a data table TD attached to question Q verbatim, etc. This request is received in S2 by the computer system DI, via its COM module.

[0108] The information contained in the Q query can include one or more words. Such a query can be written in natural language or in a specific computer language, for example SQL (Structured Query Language), Python, etc.

[0109] In one example (not exhaustive), the query Q might be: "Could you give me some information about the drug Valsartan™?". In another example (not exhaustive), the query Q might be: "Here is a data table TD containing information about drugs. Could you provide me with a textual description of this table?". During step S3, the EN R enrichment module of the DI computing system enriches the data in the data table TD using data from the knowledge graph GC, and optionally using data from the external knowledge source SCE. An enriched data table TDE is then generated at the end of step S3. If the data table TD was provided as an attachment to question Q, the enrichment is performed directly from this table.In the event that the TD data table was not provided as an appendix to question Q, the DI computer system can retrieve the TD data table from the data learned from the knowledge base BC, from the external knowledge source SCE, or from the RW web resources.

[0110] During step S4, the CNV conversion module of the DI computing device performs a textual conversion of the enriched data table TDE to generate a TXT textual description of it. During step S5, the DI computing device sends a response R to the question Q. To this end, the CMD control module of the DI computing device commands the display, at the user UT's communication terminal TC, of ​​a PW web page containing the TXT textual description.

[0111] The TC terminal of user UT receives this web page in S6 which is then rendered in S7 using the UI interface.

[0112] We now describe, with reference to Figure 4, together with Figures 1 to 5, the detailed sequence of the enrichment step S3, according to one embodiment of the invention.

[0113] An example of a TD data table is shown in Figure 5A. Such a table contains six rows, L0 to L5, and five columns, C1 to C5. In the example shown, the TD table contains information on five drugs listed in rows L1 to L5 of the TD data table. Columns C1 to C5 are labeled, in row L0, respectively, "Name," "Product," "ephMRA Code" (English abbreviation for "European Pharmaceutical Market Research Association"), "ephMRA Class," and "ATC Class" (English abbreviation for "Anatomical Therapeutic Chemical Classification System"). It goes without saying that the number of rows and columns is not limited to those in Figure 5A, and that the TD table could contain information on other, non-exhaustive topics, for example, information on a company's subsidiaries, information on the vehicles in a particular vehicle brand's catalog, and so on.The data in the TD table can be in various possible formats such as, for example, CSV format (English abbreviation for "comma separated values ​​>>"), Excel XLS binary format, TSV format (English abbreviation for "Tabulation-separated values ​​>>"), etc.

[0114] The S3 enrichment step includes a primitive typing substep S30 in which the ANS semantic annotation submodule of the DI computing device associates at least one column of the TD table with a type from a set of primitive types including a string, an identifier, a date, an integer, a floating-point number, etc.

[0115] In the example in Figure 5A, columns C1, C2, C4, and C5 contain character strings, and column C3 contains identifiers. In Tissue of substep S30, only columns C1, C2, C4, and C5 are selected for semantically annotated data.

[0116] During a substep S31, semantic annotation of some or all of the data in columns C1, C2, C4, and C5 is implemented. Such a substep includes, for at least one cell and / or at least one column and / or at least one pair of columns in the TD table, the extraction of at least one identifier from an entity in the Knowledge Graph (GC). An entity could be, for example, the name of an object, a person, a company, etc. In the example shown in Figure 5B, the Knowledge Graph (GC) used is Wikidata™ https: / / wikidata.org.

[0117] More specifically, in the example shown in Figure 5B, such a semantic annotation is of the type:

[0118] - CEA cell-entity annotation (English abbreviation for "Cell Entity Annotation"),

[0119] - Column Type Annotation (CTA) annotation (English abbreviation for "Column Type Annotation"),

[0120] - and column-property annotation CPA (English abbreviation for "Column Property Annotation").

[0121] To this end, the ANS semantic annotation submodule proceeds:

[0122] - in S310, to a CEA association of at least one cell from the TD table to an entity identifier extracted from the GC knowledge graph,

[0123] - in S311, to a CTA association of at least one column from the TD table to an entity type identifier extracted from the GC knowledge graph,

[0124] - in S312, to a CPA association of at least one pair of columns from the TD table to a property identifier of the knowledge graph GC.

[0125] Other types of semantic annotations can be used as an alternative, such as, for example, annotating the rows in their entirety or thematic information on the table.

[0126] During substep CEA S310, each character string, or some of them, present in columns C1, C2, C4, and C5, excluding column headers, is traversed and associated with a value from a corresponding entity in the knowledge graph (GC) via heuristic techniques combining syntactic distance calculations, table / graph context comparisons, etc., or through machine learning. Thus, for example, in the example in Figure 5A:

[0127] - the cell “VALSARTAN HYDROCHLOROTHIAZIDE CRISTERS” is associated with entity identifier Q48566694 “hydrochlorothiazide / valsartan” in the GC knowledge graph and available at the following address: https: / / wikidata.org / wiki / Q48566694,

[0128] - the cell “ESOMEPRAZOLE BIOGARAN” column C1, row L5, is associated with the entity identifier Q553223 “esomeprazole” in the GC knowledge graph and available at the following address: https: / / wikidata.org / wiki / Q553223,

[0129] - The cell "CGP-48933" column C2, row L5, is associated with the entity identifier Q155472 "valsartan" in the GC knowledge graph and available at the following address: https: / / www.wikidata.org / wiki / Q155472,

[0130] - etc.

[0131] With reference to Figure 5B, an extract from a semantically annotated TDA table is shown, in which the CEA annotations are indicated. As is typical, for each added entity identifier, a confidence score for the annotation is indicated. In the example shown in Figure 5B:

[0132] - a confidence score of 12% is associated with the cell “VALSARTAN HYDROCHLOROTHIAZIDE CRISTERS”,

[0133] - a confidence score of 14% is associated with the cell “ESOMEPRAZOLE BIOGARAN”,

[0134] - a confidence score of 12% is associated with the cell “CGP-48933”,

[0135] - etc.

[0136] During substep CTA S311, for at least one of the columns C1, C2, C4, C5, some of them, or a portion thereof, based on the CEA annotations performed in S310, the heading of said at least one of the columns C1, C2, C4, C5 is associated with an identifier of an entity in the knowledge graph GC that represents the column type. To this end, in a manner known per se, majority votes on the types of the CEA entities are implemented, as described, for example, in the document Chabot et al: "DAGOBAH: An End-to-End Context-Free Tabular Data Semantic Annotation System," or proximity calculations in a vector space, as described, for example, in the document Liu et al: "Radar Station: Using KG Embeddings for Semantic Table Interpretation and Entity Disambiguation." Thus, for example, in the example of Figure 5A:

[0137] - The heading "NAME" in column C1 is associated with entity identifier Q12140

[0138] "drug" in the GC knowledge graph and available at the following address: https: / / www.wikidata.org / wiki / Q12140,

[0139] - the heading "PRODUCT" in column C2 is associated with the entity identifier Q28885102 "pharmaceutical product" in the GC knowledge graph and is available at the following address: https: / / www.wikidata.org / wiki / Q28885102,

[0140] - etc.

[0141] Referring to Figure 5B, CTA annotations are indicated in the extract of the semantically annotated TDA table. As with CEA annotations, completeness rates are also indicated for CTA annotations.

[0142] During the CPA S312 substep, the ANS semantic annotation module identifies properties from the knowledge graph (GC) between at least two columns of the TD table, using well-known majority voting techniques mentioned above. For example, in the example in Figure 5A:

[0143] - The property identifier P527 "includes" in the GC knowledge graph and available at the following address https: / / www.wikidata.Org / wiki / Property:P527 is identified between columns C1 "NAME" and C2 "PRODUCT",

[0144] - the property identifier P527 is also identified between columns C1 "NAME >> and C5 "ATC Class >>".

[0145] With reference to Figure 5B, such CPA annotations are represented in the semantically annotated TDA table extract.

[0146] With further reference to Figure 4, the enrichment step S3 includes a substep S32 in which, for at least one of the semantically annotated columns C1, C2, C4, C5 in the TDA table of Figure 5B, the EAS module for enriching semantic annotations implements the following substeps:

[0147] - in S320, obtain, from the knowledge graph GC and possibly from the external knowledge source SCE, additional information IC on at least one of the annotations CEA, CTA, CPA in relation to said at least one column, - in S321, complete said at least one of the annotations CEA, CTA, CPA with the additional information IC obtained.

[0148] At the end of substep S321, an enriched data table (TDE) is generated. More specifically, during the implementation of substep S320, using at least one identifier from the semantically annotated TDA table, additional knowledge triplets describing the entities, entity types, and properties of the TDA table are retrieved from the knowledge graph (GC). The advantage of this operation is that it identifies a greater number and higher quality of additional knowledge triplets from the GC knowledge graph.

[0149] In the example in Figure 5B, for example for column C2, the CEA annotation relating to the identifier Q155472 added to row L5 corresponds to an entity called "Valsartan" in the knowledge graph GC, the term "Valsartan" constituting supplementary information IC to be added to the identifier Q155472. From the identifier Q155472, other supplementary information IC can be retrieved.

[0150] For example, the entity corresponding to identifier Q155472 is partially represented in Figure 6. In this example, additional IC information includes, for example, the knowledge triplet: "Valsartan", "is a >>, "drug >>".

[0151] In order for such an enrichment of semantic annotations to be as complete and detailed as possible, while limiting ambiguities, the enrichment of these annotations is implemented according to at least a first dimension which associates each CEA annotation with the CTA annotation of its column of belonging via a type relationship.

[0152] To optimize this enrichment of semantic annotations, the enrichment of these annotations can be implemented according to a second dimension associating the CEA annotations of one column with the CEA annotations of another column linked to this column by an identified CPA annotation.

[0153] To further optimize this enrichment of semantic annotations, the enrichment of these annotations is implemented along a third dimension that associates CEA / CTA / CPA annotations with knowledge from the GC knowledge graph. In the non-exhaustive example in Figure 5C, several entities from the GC knowledge graph, as supplementary information (CI), are retrieved during the semantic annotation enrichment step. For improved readability and to allow the LNT language model to correctly use this information, the entity labels are retrieved and used instead of identifiers as follows:

[0154] - https: / / wikidata.org / wiki / Q48566694 = “hydrochlorothiazide / valsartan”,

[0155] - https: / / wikidata.org / wiki / Q553223 = “esomeprazole >>,

[0156] - https: / / www.wikidata.org / wiki / Q155472 = “valsartan >>,

[0157] - https: / / www.wikidata.org / wiki / Q12140 = "medicine" >>,

[0158] - https: / / www.wikidata.org / wiki / Q28885102 = "pharmaceutical product"

[0159] - https: / / www.wikidata.Org / wiki / Property:P527 = "includes"

[0160] - etc.

[0161] Thus, the additional IC information obtained during the enrichment of semantic annotations according to the first dimension is used to complete the semantic annotations, for example in the following way:

[0162] - Hydrochlorothiazide / valsartan is a medication.

[0163] - <esomeprazole is a pharmaceutical product>,

[0164] - cvalsartan is a pharmaceutical product

[0165] - etc.

[0166] The additional IC information obtained during the enrichment of semantic annotations according to the second dimension is used to supplement the semantic annotations, for example in the following way:

[0167] - chydrochlorothiazide / valsartan, includes, valsartan>.

[0168] The additional IC information obtained during the enrichment of semantic annotations according to the third dimension is used to supplement the semantic annotations, for example in the following way:

[0169] - <esomeprazole, mass, 345.114712 daltons>,

[0170] - <esomeprazole, chemical formula, Ci7Hi g N3O3S>,

[0171] - cvalsartan, mass, 435.22704 dalton>,

[0172] - cvalsartan, chemical formula, C 24 H 2gN5O3>. We now describe, with reference to figure 7, together figures 1, 2 and 4, the detailed process of the textual conversion step S4, according to an embodiment of the invention.

[0173] The text conversion step S4 includes a substep S40, during which the prompt generation module GPR automatically generates a prompt PRP requesting the language model LNT to generate a textual description of the data table TD. According to the invention, the PRP prompt includes a verbalization instruction that takes as input the semantic annotations that were enriched in S3.

[0174] An example of such a PRP prompt is shown in Figure 8A. The "DATA" tag is intended to be replaced at runtime by the contents of the TD table. The "ANNOTATION" tag is intended to be replaced at runtime by the additional IC information contained in the enriched TDE table.

[0175] The text conversion step S4 includes a substep S41, during which the PRP prompt is subjected to the LNT language model.

[0176] The text conversion step S4 includes a substep S42, during which the LNT language model generates a TXT text description of the TD table. An example of such a text description is shown in Figure 8B.

[0177] Such an S4 step of textual conversion of tabular data proves to be significantly more efficient than a classic textual conversion of tabular data, such as represented in Figure 9, which might be requested by a user prompting a commercial conversational agent.

[0178] The communication method described above represents a significant advancement in the field of natural language dialogue applications between a user and a machine, offering a precise and efficient solution, particularly in terms of the quality of the text conversion performed, for converting tabular data into detailed and unambiguous natural language text. Verbalizing tabular data notably improves data accessibility and comprehension for a wide range of users. The invention has numerous applications, including but not limited to:

[0179] - the generation of automatic reports on tabular data with a rich description of their content,

[0180] - the generation of automatic summaries for presentations or reports, - access to table data for RAG-based question-and-answer systems,

[0181] - etc.

Claims

DEMANDS

1. A method of communicating with a computer device (CD) comprising a knowledge base (KB) modeling data in the form of a knowledge graph (KG), said method comprising the following, at the level of said device: - receive (S2) a question posed using a natural language interface, - generate (S4-S5) an answer to said question using a language model (LNT), said generated answer comprising textual data (TXT) from a textual conversion (S4) of a data table, during which data from the table are enriched (S3) using data from the knowledge graph.

2. A communication method according to claim 1, wherein the enrichment of the table data comprises the following: - semantically annotate (S31) at least some of the data in the table using data from the knowledge graph, - for at least one column of the table (TDA) containing semantically annotated data, obtain (S320) additional information (IC) on said semantically annotated data, from the data in the knowledge graph, and complete (S321) said semantically annotated data of said at least one column with the additional information obtained, and the textual conversion includes the following: - query (S41) said language model from a prompt (PRP) generated (S40) in natural language by said device, said prompt asking the language model to generate a textual description of the table from the semantically annotated table data that have been completed, - receive (S6) said textual description generated in response to the query.

3. Communication method according to claim 2, wherein the semantic annotation (S31) of at least some data from the table comprises, for at least one cell and / or at least one column and / or at least one pair of columns of the data table, an extraction, in the knowledge graph, of at least one identifier of an entity from the knowledge graph.

4. A communication method according to claim 3, wherein the semantic annotation (S31) of at least some of the table data comprises the following: - associate at least one cell in the data table with an entity identifier extracted from the knowledge graph, - associate at least one column of the data table with an entity type identifier extracted from the knowledge graph, - associate at least one pair of columns from the table with a property identifier from the knowledge graph.

5. A communication method according to any one of claims 3 or 4, wherein the obtaining (S320) of additional information on said column of the semantically annotated data table comprises an extraction, in said knowledge graph, of knowledge related to said entity identifier, said entity type identifier or said property identifier which was associated during the semantic annotation of at least some of the data in the table.

6. A communication method according to claim 5, wherein the obtaining (S320) of additional information on said column of the semantically annotated data table further comprises an association, using a type relationship, of the additional information extracted in connection with said entity identifier to the additional information extracted in connection with said entity type identifier.

7. A communication method according to claim 5 or claim 6, wherein the obtaining (S320) of additional information on a column of said pair of the semantically annotated data table comprises an association of the additional information extracted in connection with an entity identifier for said column of said pair with additional information extracted in connection with said or another entity identifier for the other column of said pair.

8. A communication method according to any one of claims 5 to 7, wherein obtaining (S320) additional information about said column or about said column of said pair of the semantically annotated data table includes an association with knowledge additional information extracted from the knowledge graph, additional information extracted in connection with said entity identifier and / or additional information extracted in connection with said entity type identifier and / or additional information extracted in connection with said property identifier.

9. A communication method according to any one of claims 1 to 8, wherein at least some of the data in the table that are semantically annotated are character strings.

10. A computer system (CS) comprising a knowledge base (KB) modeling data in the form of a knowledge graph (KG), the computer system being characterized in that it is configured to implement: - receiving a question posed using a natural language interface, - a generation of an answer to said question using a language model (LNT), said generated answer comprising textual data from a textual conversion of a data table, during which data from the table are enriched using data from the knowledge graph.

11. A computer program comprising program code instructions for implementing the communication method according to any one of claims 1 to 9, when executed on a computer.

12. A computer-readable information carrier comprising instructions for a computer program according to claim 11.

Images