CH711033B1 - Relational search engine. - Google Patents

Abstract

The invention relates to a relational search engine method where the sought results (c) are n-tuples of documents extracted from a search space (a) linked by at least one relation. The relationship sought is expressed by the user (b) using a man-machine interface and an iterative learning algorithm (e). The method learns at each interaction of the user to better model the relationship(s) sought (c) and, with the help of this enriched model, controls the man-machine interface to improve the relevance of the results displayed (d).

Description

TECHNICAL AREA

The invention relates to a relational search engine with an intelligent man-machine interface which allows a user to search for n-tuples composed of n documents linked by one or more desired relations.

STATE OF THE ART

[0002] Information systems and in particular search engines offer the possibility of searching for concepts with a sequence of keywords, with formulations in natural language (for example a question) and sometimes with examples of documents (e.g. an image). They also make it possible to personalize the search according to the interest of the users. These search engines index the data first and constitute an optimized search space to find documents with p. ex. the method of an inverted index. Search engines can find all web pages or other types of documents that contain the keywords or that have a similarity with the search concepts defined by the user.

[0003] The documents sought can be of the text, image or other type. Some engines offer the possibility of looking for a similarity match with an example of an existing document, then simply indicate the document as being the reference and the engine will search for other documents that have a similarity match with this reference.

[0004] Search engines typically offer the possibility of carrying out a search by correspondence characterized by a similarity metric. The characteristics of this internal metric are defined by the search engine and the user cannot redefine it freely. The search results are then listed and presented in order of relevance based on a set of other metrics and filtering options.

[0005] Numerous solutions already exist in the state of the art for carrying out a search which involves several documents at the same time.

[0006] US2011040749 describes a search engine capable of performing a search for a document in several domains at the same time. A result list is returned with tuples of documents that have the highest combined scores.

[0007] US2014214711 describes a search engine between two sets of documents: job offers and candidate profiles, in which a similarity match is sought and improved by modifying the weighting and the set of keywords and, this , depending on the data found.

[0008] US2013024445 describes a universal search engine capable of searching for different types of document, texts, images, etc. which are called categories, the documents found correspond to the search of the user and therefore have a similarity of correspondence between them.

[0009] US 20080147633 describes a search engine that associates a user profile with search results. Considering a user's profile as a document and the search results as a list of documents, the system looks for a similarity match between the two sets of documents.

[0010] Most of these known solutions are therefore based above all on a search for similarity correspondence predefined or adapted by the system according to existing data. A defect of these known solutions is that they do not offer the possibility of searching for one or more relations desired by the user between a n-tuple of documents with an intelligent man-machine interface.

[0011] Another defect of the known solutions is that there is no practical way to find the n-tuples of documents linked by the desired relationship(s) when the documents are of different types (for example: a document text type and an image type document).

[0012] It would therefore be advantageous for the user to be able to define in natural language the relationship or relationships that he wishes to have between the n documents of an n-tuple, or between the underlying concepts of these . It would also be advantageous for the user to have the possibility of defining the different types of documents which make up the tuples.

[0013] To overcome the aforementioned disadvantages of the prior art, while maintaining the widest possible scope of application, a number of other problems must also be solved. Indeed, it often happens that the desired relationships are not easy to express in words; p. eg, how to express the emotional expressiveness of a relationship between music and an image? Moreover, since it is a relationship perceived by a particular user, this subjective relationship is a priori not relevant for another user.

[0014] In other cases of use, the parameters and the concepts which characterize a relationship are present in additional metadata or in those which must be extracted from the documents. Without access to this additional information, the search engine cannot produce relevant results.

[0015] In other cases, the relevant concepts which characterize a perceived relationship or of a semantic nature are not available a priori in the search space, these are then to be discovered by the search engine thanks to interactions with the user. This is typically the case when a relationship is subjective in nature.

[0016] In other cases, a relation comprises relational metrics, p. ex. specifying a date.

[0017] In other cases of use, the desired relationship may relate to a subset of concepts or to structural subspaces of a document, such as, for example, one or more particular topics of a text, one or more specific graphic objects of an image, one or more passages of a piece of music. In the current state of technology, users do not have the means to express a search based on a desired relationship of this type. Moreover, the expression of the characteristics of this relationship must necessarily be expressed by the user in order to be able to carry out such a relational search, in fact, the search engine cannot know this information a priori.

[0018] In the state of the art, there is no relational search engine with an intelligent human-machine interface capable of searching for n-tuples which are composed of n documents linked by a relationship desired by the user. .

[0019] Without having the possibility of searching for relevant relational tuples or of being automatically informed of their existence, users end up being overwhelmed by a continuous flow of information and, for lack of adequate means at their disposal, they renounce all forms of relational research.

BRIEF SUMMARY OF THE INVENTION

An object of the present invention is to remedy the drawbacks of the prior art which have just been mentioned by providing a relational search engine method which allows a user to find or to be notified of the existence of n-tuples composed of n documents linked by at least one relation desired by the user. The present invention achieves this object by providing, on the one hand, an iterative method for supervised or semi-supervised learning in accordance with appended claim 1.

[0021] Definition: The term document is used here in a generic way, the expression "document" designates any digital representation that a computer can store in the form of data. The expression "document" also designates in particular a folder composed of a set of files or an aggregation of any set of other documents.

[0022] Definition: A concept is a measurable or perceived regularity that characterizes a set of entities or abstractions. A document is typically composed of a set of concepts and underlying concepts.

[0023] Definition: A relationship between a pair of documents and its perception by a particular user is characterized by a set of correspondences of an objective and subjective nature between implicit or explicit concepts. A relation can perfectly be expressed in relation to the presence of a concept but also in relation to its absence. A relationship can also represent any metric between measurable entities.

[0024] Definition: An n-tuple is an ordered set of n documents with n greater than or equal to 2. When n=2, we speak of a pair of documents. The n documents are linked by one or more relations and the tuple is associated with at least as many labels as there are relations.

[0025] Definition: a label makes it possible to express an appreciation of the relevance of an n-tuple, as perceived by the user or estimated by the system, from the point of view of the relationship between a pair of documents, or one of the relations, which links the documents of the tuple. In some use cases, a relationship can be characterized by multiple labels.

[0026] Definition: An intelligent (or self-learning) man-machine interface is a man-machine interface using which a user or a set of users can teach a computer system (the machine) to interact better, to better understand their intentions, or even to anticipate their requests while interacting with it. In the context of the invention, the intelligent man-machine interface typically combines the automatic natural language processing of step (b), the iterative learning processes of step (e), and of steps ( c, d, e) and the retro-relevance learning of sub-step (d1). The learning result is characterized by the adaptation of a computer model composed of inputs and outputs, a graph of functions, data, operators and associated variables.

The invention allows the user to make the search engine learn the desired relationships by means of an intelligent man-machine interface. When a relation is well learned by the system, it can then efficiently search for other corresponding and relevant n-tuples.

[0028] The learning of a relationship is carried out by a set of processing and automatic learning algorithms. During a session of use, an iterative method of supervised or semi-supervised learning is implemented according to the results found and the interactions of the user.

According to an advantageous embodiment of the method of the invention, certain automatic learning models, the structure of the dialogue and the strategies of composition of algorithms are optimized by the system by exploiting the statistical data of the interactions on a set of users.

[0030] Furthermore, in accordance with another advantageous embodiment of the method of the invention, once a relationship is understood and saved by the system, the user can, if he wishes, be spontaneously informed of any new appearances of documents that are bound by this relationship(s). In this case, the system itself takes the initiative to inform the user. Typically, if the documents of a tuple are linked by the desired relation or relations, the user is instantly informed by a change in the visualization during an interaction session in progress or simply by the notification of a message.

BRIEF DESCRIPTION OF FIGURES

The invention will be explained in more detail with the aid of several embodiments presented with the aid of the following drawings: FIG. 1 represents a flow diagram illustrating an example of the method of the invention. • Figure 2 illustrates the example of a pair of documents with a relationship between the two documents. • Figure 3 illustrates an example of an iterative learning relational search engine.

DETAILED DESCRIPTION OF THE INVENTION

[0032] The characteristics described in claim 1 allow a relational search engine to find n-tuples of n documents which have between them one or more relations desired by the user. Once the relation(s) are expressed and modeled, corresponding tuples can be searched in the search space.

The invention includes a first step (a) of creating the search space. This space allows the relational search engine to perform the processing, storage and search of tuples.

The relational search engine also has an embodiment where it processes the document search space in the form of a continuous flow, in streaming mode.

[0035] In certain cases of use, the documents are stored in computer systems which require particular access rights, the user can provide these access rights to the system so that the latter can access the documents. For example, some of the documents sought may be part of a document sharing site or be part of a user's personal data, in which case the user may, if desired, grant access rights to the search system. These connections are processed by connectors and these exist for the main existing systems.

[0036] Depending on the needs, the search space can be partitioned into sets of documents. When the documents are partitioned into different sets, these do not necessarily contain an identical number of documents. For example, it is possible that in one of the document sets, there is only one document.

The documents do not necessarily define computer content or media, they also represent people or the digital modeling of a person, such as for example: the types of music that he prefers.

[0038] In certain cases of use, the sets of documents are not automatically available in the search space, to facilitate their acquisition, the Application programming interface (API) of a classic search engine is used. . The user must then define the search criteria which make it possible to collect the documents and, subsequently, the classic search engine provides the Uniform Resource Locators (URLs) of the documents, these and their content are then automatically ingested into the system search space through the use of specific software connectors.

[0039] The search space stores or processes documents with a specific internal format optimized for search.

The documents are processed by all the ad hoc information preprocessing algorithms. Sometimes documents are produced in another language and need to be translated.

The documents that make up a n-tuple can have different types, as in the following examples of pairs of documents: an image-text; a person - a piece of music. The system converts all types of external documents into an internal format and structures to perform, among other things, semantic content enrichment, metadata extraction and metadata addition, internal modeling, and to offer several methods vectorized and intra-media access, etc.

[0042] Intra-media access allows the search engine to have access by introspection. This method is typically used by the search engine to access the internal contents of documents by their elements or by their internal structures. In some cases, a specific description is used, for example, Synchronized Multimedia Integration Language (SMIL).

[0043] The internal representation of the documents is typically extracted according to the media with different algorithms, such as for example with an automatic speech recognition algorithm, an image recognition algorithm, an optical character recognition algorithm, or with a automatic extraction of graphs when, for example, the document is a composition of objects or topics. For the extraction of particular metrics, ad hoc algorithms are used.

[0044] In general, the graphs allow the establishment of links to increase the knowledge of the internal representation of the documents. Knowledge graphs and knowledge bases are typically used to enrich the internal representations of documents. An example of extraction and representation as Resource Description Framework (RDF) is described in the following article: Knowledge Vault: A Web-Scale Approach to Probabilistic Knowledge Fusion, by Kevin Murphy et al. In KDD 2014).

[0045] The internal structure and methods of accessing documents are determined, in part, by the following parameters: the type of documents sought, the size of the resulting search space, the speed with which the documents evolve, and the response time imposed by the user interface.

When the size of the search space is large (> million documents), a first approach, well known in the state of the art, uses Bag-of-Words (BoW) as a vector representation enriched with a semantic markup extracted from an external source of knowledge. Finally, the BoW terms are weighted to improve performance.

In a space of vectorized documents of the BoW type, a relational search is typically carried out in two phases. For each of the documents in the relationship sought, the engine first finds the relevant documents and then filters the tuples according to the relevance of the relationship or relationships between the documents.

[0048] Depending on the context, other approaches are preferably used: • One approach uses graphs, such as, for example, the Graph-based Semi-Supervised Learning algorithm. This approach applies well to large search space sizes. • The following approach is preferably used when the size of the search space is more modest or is to be built on the fly. It consists in creating a continuous semantic space of documents by machine learning, this technique is known as Paragraph Vector, (cf. Distributed Representations of Sentences and Documents. Quoc Le and Tomas Mikolov, ICML Beijing 2014)

[0049] When the documents represent media or are of different types, deep learning is preferably used. This approach makes it possible to extract an internal representation of several different types of documents, such as: text, an audio signal, images, or video signals.

[0050] Once the search space is available, step (a) is performed, the invention then allows the user to express a desired relationship in step (b), and to understand this relationship to the search engine with the help of an iterative learning carried out by following steps (c, d, e).

A relationship is represented by the relational search engine as metrics which impose constraints on measurable values and by semantic models between the concepts.

The constraints are extracted by ad hoc algorithms and the semantic representation is typically extracted by automatic natural language processing, Natural Language Processing (NLP).

In step (b) of the method, the man-machine interface allows the user to express in natural language one or more relationships between documents with identical or different types. Like for example „Look for an image and a text that illustrates the Second World War!“

In some cases, the user already has a document or a set of documents that he designates as being the first document of the tuple. Then, with respect to this first document, it describes the desired relationship in natural language with the second document of the tuple. An example of a usage scenario could be the following: in sub-step (b1), the user designates an image of a country house, the user then expresses the desired relation in natural language: „Looking for a house identical but with a garage!“. The objective of the relational search engine is then to analyze, to understand the request and then to look for images of houses which correspond to the example of the house with, in addition, a garage.

[0055] The natural language queries and the example documents are processed by the same algorithms as those of the search space, in order to have an identical structure and format.

In other cases of use, it is sometimes tedious, or even impossible, to describe in natural language a perceived subjective relationship which exists between a pair of documents and the user. It is perhaps simpler to give one or more examples of tuples. The sub-step (b1) precisely makes it possible to give one or more examples by designating them or by dragging and dropping them in the man-machine interface of the system. This one will use the iterative learning of the steps (c, d, e) to understand by itself what is the relation which links the pair of documents while searching for new relevant n-tuples.

[0057] The flexibility of the natural language interface allows the user to express metric type relationships that require ad hoc pre-processing algorithms for the extraction of constraints. Metrics can be of a very different nature, for example they can be qualitative, quantitative, temporal, geographical, financial, etc. To process a majority of these metrics, more than a hundred ad hoc algorithms are needed.

[0058] The man-machine interface includes a set of controls that allow the user to reinitialize a relationship, to associate two or more relationships, to rename a relationship, to save, to find, or to delete a relationship. The user can also change the scope of understanding of a relationship, e.g. ex. by reversing the meaning of a relation, and in this case, the n-tuples found are those which correspond the least possible to the expressed relation.

The user can be automatically informed, if desired, by a notification from the system if a new relevant relationship has appeared in the search space, for example, when a new n- tuple of documents

Subsequently, when a first search is performed in step (c), the system displays the n-tuples found in step (d). Each n-tuple (10) is visualized with a first algorithmic evaluation of the relevance of its relations. These are typically visualized graphically with the following labels p. ex. : „relevant“ (14), „undefined“ (15) and „not relevant“ (16).

During this same step (d) of display, a sub-step (d1) of a retro-relevance algorithm adds, if necessary, at least one n-tuple to the n-tuples found or replaces a certain number of n-tuples found with n-tuples chosen by itself. The objective being that the user preferably assigns an evaluation label to this or these n-tuples. This retro-relevance learning approach is known in the literature, see p. ex: the reference (cf. On textbased mining with active learning and background knowledge using SVM, Catarina Silva and Bernardete Ribeiro, Published online, Springer-Verlag 2006). It increases the quality of supervised or semi-supervised learning while reducing the number of iterations and interactions required.

[0062] Following step (d), when the first n-tuples are displayed, the user can use the iteration loop of step (e) to help the system improve its understanding of the relationships by indicating to the system which relationships are most relevant and which are not. With this interaction, the user adapts the labels of the relations of the tuples. For example, relations labeled as „relevant“ (14) are those which correspond to a desired relation and relations labeled as „irrelevant“ (16) do not correspond to a sought relation and this from the user's point of view. .

[0063] For each n-tuple, the labels are typically visualized by symbols: a solid star (label = relevant) (14), a hatched star (label = undefined) (15) and an empty star (label = not- relevant) (16).

For some applications, other types of labels are used and sometimes in greater numbers; for example, a set of keywords, where the user's appreciation is made by selecting a certain number of relevant keywords and deselecting other irrelevant keywords. These selections or deselection are then visualized by a simple graphic highlighting of these keywords.

The labels can be grouped at the level of each relation, as in figure 2, they can also be integrated into the visualization of the content of the documents. Depending on the type of document, labels can be represented by phrases and by words that are highlighted in a text, with areas and graphic objects in the case of an image, by musical passages for a piece of music, or by video sequences for a film. To identify the most relevant tuples, the user selects and deselects tags by interacting directly with the content of the documents.

[0066] The user does not have the obligation to label all the n-tuples, if he does not do so, there are then two sets of relations, those which are labeled by the user and those which are labeled by the system. At each interaction with a label, an iteration loop of step (e) is carried out: the system analyzes the n-tuples modified in step (e1), it improves the modeling of the relationship sought in step (e2) and it visualizes the results in step (e3). More precisely, during step (e3) and with the new enriched model, the system re-evaluates the assignment of its own labels on the set of relations of the n-tuples displayed. This algorithmic evaluation of the labeled relations allows the user to appreciate and adjust, if necessary, the labels on all the relations of the n-tuples displayed.

In accordance with an advantageous mode of implementation of the method of the invention, when the user so desires, he can relaunch a search which takes account of the progress in the modeling of the desired relationships.

[0068] Different types of supervised or semi-supervised learning algorithms can be used. Supervised algorithms make it possible to exploit the knowledge of labeled n-tuples by the user and semi-supervised algorithms can also exploit unlabeled n-tuples from previous searches and even unlabeled n-tuples present in the search space.

When the quantity of data is sufficient, the semi-supervised learning algorithms cf. Semi-Supervised Learning (SSL) are preferred.

[0070] Depending on the type of document and the application context, certain semi-supervised algorithms give better results, for example, in the case where the documents are text, Transductive Learning Vast Margin Separators (TSVM) are used and preferably (cf. Transductive inference for text classification using support vector machines by Joachims T. ICML1999). In this case, the vectors necessary for learning are composed of the concatenation of the internal representations of the documents of the n-tuple with a target +1 when the n-tuple is „relevant“ of -1 if the n-tuple is „not relevant“ or 0 if the vector label is not known.

[0071] When the algorithm has modeled a relation for use in a search for n-tuples, the vectors evaluated by the algorithm produce the result +1 for a relation with a "relevant" label and -1 for a relation with an „not relevant“ tag. In the particular case of the TSVM algorithm, no "undefined" label is produced.

A binary response is not always desirable, as for example in the case where it is necessary to determine an order of relevance of the n-tuples. Other algorithms exist which also provide a confidence measure such as p. ex. the supervised learning algorithm of the type: Relevance Vector Machine (RVM). This is to be used preferably in the case where a richer evaluation is desired for a better visualization of the results.

[0073] In other cases, the data can be analyzed in the form of graphs, ideally when a graph can be automatically extracted on the basis of the entities present in other informative contexts; then the Graph-Based Semi-Supervised Learning algorithm will preferably be used (cf. see the reference: Graph-based Semi-Supervised Learning, Amarnag Subramanya and Partha Pratim Talukdar, Morgan & Claypool Publishers).

[0074] Depending on the number of results found by the relational search engine, these will be represented in such a way as to better visually communicate their relevance with respect to the relationship sought. The choice of the visualization mode is always motivated by the quality of the interaction proposed to minimize the number of necessary interactions and by the efficiency of the automatic learning of the relationship sought.

The man-machine interface also allows the user to search at leisure for the n-tuples found and, individually, for each set of documents which enters into the composition of a n-tuple found.

The relevance of the tuples is typically visualized by an order in a list, by a graphic highlighting in a 2D space, by a proximity in a 3D space. In some cases, the system will preferably offer an immersive interface where the user is immersed inside the space for viewing the n-tuples corresponding to his search. In this mode of representation, additional accessories are necessary to allow this navigation, such as for example a virtual reality headset or an immersive room.

The visualization method typically includes an algorithmic step in which the n-tuples are projected into a space of reduced dimension to improve the expressiveness of the visualization.

The visualization of the n-tuples is advantageously adapted to best represent their relevance. For example, when a document is common to several n-tuples of results, they can be grouped together to simplify the visualization. For some embodiments, such as those that process multimedia data, a suitable man-machine interface is used.

[0079] The performance of the iterative learning cycles of the relationship or relationships is greatly increased by the quality and richness of the visualization, it makes it possible to increase the quality of the interaction and therefore to accelerate the speed with which the relational search engine understands the intentions of the user and delivers the expected relevant results.

[0080] Those skilled in the art will understand that the scope of the invention is not limited by a particular type of search space or by internal structures of the particular data, nor by the choice of a set of metrics , or by the choice of a particular learning algorithm.

EXAMPLES OF EMBODIMENTS OF THE INVENTION

[0081] Profitable developments follow from the independent claim in all the cases of applications where the users seek a desired relation between the documents of a n-tuple.

The computer service of a relational search engine method is preferably made available to users at the same time as a conventional search engine or a generic type question-and-answer service. With the help of a natural language-like human-machine interface, the user can invoke a whole set of information services in a transparent manner. For example, a classic search of a search engine with a sequence of keywords „Italian restaurants Lausanne“, a factual question, p. ex. : „What is the second largest city in the Netherlands?“ or one of the embodiments of the invention p. ex. : „Find me a job offer that matches my skills with a higher salary!“. The information system then classifies the various user requests and activates one of the information systems corresponding to the application domain.

[0083] When the relational search service is identified and activated, the user's query is first classified into a subset of service categories that are specific to a particular domain. These fields can be of the type: search for a job offer, real estate search, search for a partner, etc.

[0084] In the case where the user formulates a request of the type: "Look for an image and a piece of music by Neil Young from the 70s!". The relational search service of the media domain is activated. The analysis of the request is continued by searching for a pair of documents consisting of an image and a piece of music. The pair of documents has a relationship made up of two links, that of the proper name (Neil Young) and that of the temporal metric (from the 1970s).

Certain embodiments of the invention exploit deep learning algorithms to be able to analyze different types of documents and to extract an enriched internal description of these documents. Relationships between these descriptions can then be searched efficiently, which allows the invention to be relatively independent with respect to the types of documents to be processed.

A first exemplary embodiment makes it possible to search for a particular relationship between a pair of documents of the same type. The application context is that of a job search where a user wishes to be automatically informed of new opportunities. He can then express his research with an example document: his curriculum vitae and express the relationship in natural language with respect to this first document, p. ex. „Look for a more qualified job!“. In the example of Figure 1, the system analyzes the relationship sought in step (b) and chooses an appropriate processing and learning strategy, over the cycles of search and learning iterations (c, d, e), the user teaches the system to better understand the relationship sought. With the help of learning algorithms, the system gradually and weightedly models all the matching criteria, the model is then stored by the system. If later, the user wants to be automatically notified of a new opportunity on the job market, he can then activate notifications based on the previously learned relationship.

[0087] Other applications are made possible with an embodiment which involves searches for subjective and perceptual relationships, typically: "I like" or "I don't like" a document, whether to express relationships between two people in a social network or in the fields of artistic expression and creation. In this embodiment, the relational search engine can advantageously have access to and exploit all of a user's personal data and the history of the use of files (system log-files) to better model the personality and person's preferences.

Another embodiment is illustrated with the help of Figure 2, the objective is to find an emotional relationship between musical sequences and accompanying images. In this case, the relationship (13) sought is of a subjective nature and between a pair of documents (12) and the user himself. An adapted interface allows the user to label the pairs "musical passage - image" with the label "I like" (14) and those which do not please him with the label "I do not like" ( 16).

The learning strategy is typically based on deep learning to best describe the content of the images. For music, metadata is extracted from available databases and an extraction of musical parameters is performed to better adapt the images to the rhythm and characteristics of the music. To identify the internal structures of music a deep learning approach is preferably used. Then, by the concatenation of all the information extracted from the two types of document and from the labels, the relationship is then learned by a supervised or preferably semi-supervised learning algorithm. It should be noted that the user must label a certain number of pairs: musical passage - image, so that the algorithm can correctly learn the desired relationship.

[0090] An approach with a large number of users is interesting for prefiltering the most popular correspondences (cf. by crowdsourcing), in this case, the relationships and learning must be shared with the community of users.

[0091] With the aid of Figures 1, 2 and 3, another embodiment illustrates the search for an objective relationship. The system has been specially prepared and trained on a large number of examples of uses specific to this application area. Data representations and query analysis are also specially processed and trained to better respond to requests. The relationship and the application domain are selected following the analysis of the user's request: „Looking for a student couple who would like to go on vacation to the Maldives!“. In this embodiment, the documents linked by the relationship sought are on the one hand male students and on the other hand female students. In the search space, the two sets of documents are already specified by the list of their Facebook URL. By analyzing the above request, the application domain is automatically recognized and activated by the system.

In the example of FIG. 2, the relational search engine makes it possible to find tuples (10) with the relation sought by the user (13) in the document search space (12) which represent data published on the web.

The example in FIG. 3 shows the various software modules and the computer hardware necessary to initialize the search space (213) and to analyze user requests (209). More specifically, in the embodiment described, all of the specific data is ingested from the Web (300) by the software module (211) to constitute the search space (213). With the list of URLs, two sets of documents are created. Nevertheless, the raw data cannot be directly exploited without the semantic enrichment software module (212). The software modules (211) and (212) continuously feed the content of the search space (213) with external data that is new or recently modified. The software module (211) typically consults web pages and the software module (212) typically uses data and ontologies derived from available information sources.

In the example of Figure 3, the user of a travel agency connects to the information system (204) through an Internet connection (300) and this with a man-machine interface (100 ) on a laptop computer (101), a smartphone (102), a smartwatch or any other device capable of communicating with it, for example, by viewing HTML content or by executing an application loaded locally in memory. In certain cases of use, the system can be controlled by voice or by a multimodal interface (103) which combines several modes of interaction such as for example: voice, gestures, etc.

[0095] To allow personalization and saving of the relationships sought, the travel agency user must be identified in the customer database (205). He must therefore have registered beforehand. This database also holds the rights to access the data with protected access, such as, for example, to access the personal data of the travel agency or online travel offer sites.

[0096] The user has several ways of expressing a relationship with the man-machine interface (100). He can express said relation in natural language in the field which defines one of the sought relations. A second way is to select or drag and drop one or more examples of existing customers that best represent the relationship(s) sought. A third way is to define more precisely the documents of the n-tuple by an expression which allows a classic search engine to find the different students of the sought n-tuples. A fourth possibility is to simply resume a relationship already sought and learned by the system from the database (205). This database (205) saves all of the history of searches and relationships sought by the user in question. Finally, the user can combine several of these previously described approaches, if desired.

The software module (206) formats the query, analyzes the relationship or relationships sought, selects the algorithm composition strategy. Finally, it passes the information to the software module (208) which is intended to execute the selected strategy to better understand the concept of the relationship expressed by the user. A more detailed example of this process is depicted with the flowchart in Figure 1.

In the example of FIG. 3, as soon as a request is modeled by the algorithms of the software module (208), it is ready to be executed by the correspondence search software module (210). The visualization of the results is carried out in the module (204).

[0099] Subsequently and depending on the results, the travel agency user is engaged in an iterative learning process with the system to improve the relevance of the results.

[0100] To inform the user of the travel agency, if he wishes, the system will notify him of any new realizations of a saved relationship when certain documents present in the search space are linked by the desired relationship. In the example of FIG. 3, it is the software module (207) which sends the notifications according to the parameters defined and the relevance calculated by the correspondence search software module (210). It should be noted that the processing and learning strategies dedicated to notifications are different from relational research. Indeed, for notifications, the algorithmic constraints are different and so are the expectations of users.

In the example of Figure 1, a flowchart illustrates the different steps that allow the embodiment of the invention in the use case described in Figure 3.

[0102] For all new queries, the user of a travel agency defines the criteria of a search and, more particularly, he defines what relationship he wants to find in the n-tuples. In the example of figure 1, the application context is that of a travel agency which is looking for two students for a holiday offer in the Maldives. In the example of Figure 1, the search options are presented by the interface in step (b), using this interface, the user defines his search parameters and gives the command to launch a first search, step (c). In step (b), the relationship(s) are analyzed, categorized and metrics are identified and a first initial model of the relationship is created. Depending on the results of the analysis and for the following steps, an algorithm composition strategy is activated. More concretely, with the query defined by the user of the travel agency, the analysis extracts several concepts: including two entities (two students) that define the document couple, a geographical destination (in the Maldives). A metric is implicitly associated with students (a preferential rate for young adults), this metric induces a search constraint. With this analysis result, a strategy is selected which determines the choice of a semi-supervised learning algorithm and an ad hoc algorithm for the treatment of the constraint on the offers.

[0103] The objective of this first search of step (c) is to have a sufficient "recall" to allow interaction with the user of the travel agency. The "recall" being the ratio between the number of n-tuples returned by the system divided by the number of relevant n-tuples. The n-tuples are then visualized by highlighting their respective relevance and with their labels which are evaluated by the system.

[0104] When describing the desired relationship “Looking for a student couple who would like to go on vacation to the Maldives!”, the user has only very partially defined the relationship and the intention behind his search, with the human- intelligent machine, it will be able to improve the modeling that the system has of this relationship.

[0105] The user of the travel agency can interact in step (e) with the n-tuples by labeling the relations which seem relevant to him with respect to the relation sought and those which are not. Using this information and after the analysis of step (e1), the iterative semi-supervised learning algorithm progresses in modeling the relationship sought in step (e2), and will adapt the evaluation and the visualization of the labels in step (e3). With this label visualization, the user can instantly correct erroneous labels. More concretely, an n-tuple is displayed with an empty star (label = irrelevant) (16), but if the user considers that the n-tuple is relevant, he then just has to select the full star (label = relevant) (14) to improve the modelling. The user can repeat step (e) and the sub-steps (e1, e2, e3), until the rating labels best represent the desired relationship between the pairs. For example, they both share a strong interest in scuba diving.

To improve the relevance of the tuples found, the user can restart a new search step (c) which this time will use the enriched modeling of the relationship sought. Subsequently, the user can repeat the learning and research iterations (c, d, e) at will.

[0107] The quality of the relational search engine of FIG. 1 depends on the speed with which the system learns to model the desired relations. To accelerate this learning, the system calculates the statistics to evolve the choice of learning strategies according to a set of users and uses.

[0108] In step (d) of displaying the pairs of students, and more particularly in sub-step (d1), a retro-relevance algorithm can submit to the user in the form of a request labeling, a n-tuple that the system determines in an optimal way to speed up learning.

[0109] The software architecture of the relational search engine is that of a system capable of processing large quantities of data, it is typically built around the Apache™ open source software ecosystem and in particular that of Hadoop®.

Embodiments of the invention require automatic natural language processing. Some of the processing can typically be done with GATE tools and software libraries https://gate.ac.uk/. (cf. Integrating NLP Using Linked Data, by Sebastian Hellmann et al. In ISWC 2013, Lecture Notes in Computer Science Volume 8219, 2013, pp 98-113)

Claims (5)

1. Iterative method for supervised or semi-supervised learning of one or more desired relations, and for searching for n-tuples of documents linked together by at least one of said one or more desired relations, the method being implemented using a computer system comprising a man-machine interface, and comprising in order the steps of: a) providing the computer system with access to a search space containing an internal representation of the documents; b) expressing in natural language using the man-machine interface an at least approximate formulation of said one or more desired relationships, to provide the computer system with an initial model of said one or more desired relationships; c) searching in the search space for n-tuples of documents linked together by one or more relations corresponding to said initial model of said one or more desired relations; d) causing the man-machine interface to display a set of n-tuples corresponding to the search of step c), the n-tuples of said set of n-tuples being accompanied by their labels, and the labels being, evaluated by an algorithm; e) causing said one or more desired relations to be learned by the computer system using an iterative supervised or semi-supervised learning algorithm by adapting the value of at least one of the labels of step d) to the help of the man-machine interface, and by linking the iterations as long as it seems useful; each iteration comprising, after adapting the value of said at least one label, the three sub-steps of: e1) having the content of the documents forming part of the tuple(s) whose value of at least one label has been adapted, be analyzed; e2) on the basis of the analysis carried out in step e1), enriching the model with the desired relationship; e3) causing the human-machine interface to display labeling evaluations of a plurality of n-tuple relations of the set of n-tuples most recently displayed by the human-machine interface, the evaluations of labeling being based at least partially on the model enriched in step e2). 2. Iterative learning method according to claim 1, in which a new execution of step c), followed by a new execution of step d) are substituted for the execution of the sub-steps of step e) after at least one iteration of the substeps of step e), during such a new execution of step c) of searching for the n-tuples, step c) is executed on the basis of the model enriched the most recent instead of the initial model. 3. Iterative learning method according to claim 1 or 2, wherein step b) of providing the computer system with an initial model of said one or more desired relationships comprises the sub-step of: b1) designating, using the man-machine interface, at least one first document capable of being linked to at least one second document by said one or more desired relationships. 4. Iterative learning method according to claim 1 or 2, wherein step b) of providing the computer system with an initial model of said one or more desired relationships comprises the sub-step of: b1) entering into the system at least one n-tuples of documents linked together by said one or more desired relations, so as to provide at least one example to the system. 5. Method according to one of the preceding claims, in which step d) comprises the sub-step of: d1) determining at least partially the n-tuples displayed according to a retro-relevance algorithm.