RU2556425C1 - Method for automatic iterative clusterisation of electronic documents according to semantic similarity, method for search in plurality of documents clustered according to semantic similarity and computer-readable media - Google Patents

Abstract

FIELD: information technology.

SUBSTANCE: method for automatic iterative clusterisation of electronic documents according to semantic similarity includes converting each electronic document into a corresponding multidimensional vector in multidimensional space, the number of dimensions of which is determined by terms contained in the electronic document; finding the measure of proximity of the obtained vector to each of the vectors already existing in the clusters, which combine semantically similar documents processed previously; supplementing the cluster for which the found proximity measure is minimal with the document to be processed; determining a new vector for the additional cluster; taking as the term of the additional cluster the name of the document in said cluster for which the proximity measure of its vector to the determined new vector is minimal. Thus, when new electronic documents are input, existing clusters are processed as separate documents and not as a set of documents.

EFFECT: simple and faster processing of processing electronic documents and search in a clustered set of documents which are relevant to a search request.

12 cl, 6 dwg

Description

FIELD OF THE INVENTION

The present invention relates to a method for automatically iteratively clustering electronic documents by semantic proximity, a method for searching collectively clustered by semantic proximity documents, and also to computer-readable media with programs for implementing these methods.

State of the art

The huge amount of information on the Internet leads to the fact that the number of objects issued at the request of the user is very large. A common problem that reduces the user’s efficiency with the search engine is the redundancy of information when issuing results on request. This complicates the process of reviewing the results and selecting the most suitable materials (articles, publications, reports, etc.) from the many found.

At present, various methods for clustering documents by their semantic proximity are known, allowing further searches in the aggregate of such documents.

So, in the patent of the Russian Federation No. 2268488 (publ. 20.01.2006) a method is disclosed in which words, phrases, idioms, sentences and even ideas are encoded for subsequent numerical processing, including clustering. In the patent of the Russian Federation No. 2273879 (published on April 10, 2006), a method is given in which morphological and syntactic analysis of the text is carried out, followed by indexation of the found units for assigning the text to a specific class. In the method according to US patent No. 6871174 (publ. 22.03.2005) determine the similarity of texts by text fragments. The EAPO patent No. 002016 (published on January 22, 2001) describes a method in which unique blocks of information are determined in fragments of a text document and used for subsequent clustering and search. The disadvantage of all these methods is that their implementation requires a very large amount of memory, because when new text arrives, processing has to be repeated for each text already processed.

In US patent No. 6189002 (publ. 13.02.2001) disclosed a method in which the text is divided into paragraphs and words, which are converted into vectors of ordered elements. Each element of the vector corresponds to a paragraph found by applying a given function to the number of occurrences in this paragraph of the word corresponding to this element. The text vector is considered as the semantic profile of the document, suitable for comparison in the case of clustering. However, given the variety of paragraphs, this method also requires a huge array of stored data.

Disclosure of invention

The objective of the present invention is to develop such a method of iteratively clustering electronic documents by semantic proximity, which would simplify and accelerate both the corresponding processing of electronic documents and the subsequent search in a clustered collection of those documents that are relevant to the search query.

To solve this problem and achieve the technical result, the first object of the present invention proposes a method for automatic iterative clustering of electronic documents by semantic proximity, which consists in the following: convert each electronic document to be processed into a corresponding multidimensional vector in a multidimensional space, the dimensions of which are determined by those contained in electronic document terms; find a measure of proximity of the obtained multidimensional vector to each of the multidimensional vectors of existing clusters that combine semantically close electronic documents processed earlier; complement the cluster to be processed with an electronic document for which the measure of proximity found is minimal; determine for the augmented cluster its new multidimensional vector; they take as the theme of the augmented cluster the name of that of electronic documents in this cluster for which the measure of proximity of its multidimensional vector to a certain new multidimensional vector is minimal.

A feature of the method according to the first object of the present invention is that they can accumulate a set of electronic documents to be processed as they appear during a predetermined time interval; after which clustering of each of the electronic documents in the accumulated aggregate is carried out.

Another feature of the method according to the first object of the present invention is that the conversion of an electronic document into a multidimensional vector may include stages in which: planarize the text of the electronic document; form arrays of terms for the planarized text of each of the electronic documents, for which they planenized planarized text, resulting in segments in the form of words, punctuation marks, spaces, and stamped tokenized text, highlighting as a result of the word basis using at least one of heuristic algorithms, after which the weight of each term is found in each of the electronic documents, and each of the electronic documents is expressed as a vector in a multidimensional space, the dimensions of which are determined by found by the weights of terms in the text of this electronic document.

Moreover, the calculation of the weight of each term can be performed using the measure TF-IDF, which is a product of the quantity

by the amount

.

.

Another feature of the method according to the first object of the present invention is that finding the proximity measure of multidimensional vectors may include the steps of: calculating the cosine measure of proximity between each pair of multidimensional vectors; divide all multidimensional vectors into subsets, in each of which the calculated cosine measure of proximity between a pair of any multidimensional vectors is less than a predetermined value; calculate the centroid vector of each of the subsets as the arithmetic mean of all multidimensional vectors of the given subset; each multidimensional vector is attributed to a subset with the nearest centroid vector.

Another feature of the method according to the first object of the present invention is that they can additionally carry out the steps in which: they find a measure of the mutual proximity of multidimensional vectors for each pair of clusters; cluster those clusters for which the found measures of mutual closeness of their multidimensional vectors do not exceed a predetermined threshold value; determine its multidimensional vector for the topic; take as a topic topic the theme of that of its clusters, for which the measure of proximity of its multidimensional vector to a certain multidimensional vector of this topic is minimal.

Another feature of the method according to the first object of the present invention is that they can additionally carry out the steps in which: they find a measure of the mutual proximity of multidimensional vectors for each pair of topics; combine those topics into the corresponding supertopic for which the found measures of mutual closeness of their multidimensional vectors do not exceed a predetermined threshold; determine for a supertop its multidimensional vector; take the topic of a supertopic as the topic of that of its topics, for which the measure of proximity of its multidimensional vector to a certain multidimensional vector of this supertopic is minimal.

Another feature of the method according to the first object of the present invention is that they can construct a graph whose nodes are supertopics, and each of the edges is the ratio of the proximity of the supertopics, topics and documents connected by this edge.

At the same time, they can compile a global glossary of terms to enable the subsequent search for fragments of the graph that are relevant to a particular search document.

To solve the same problem and provide the same technical result, a second object of the present invention provides a search method for collectively clustered by semantic proximity documents, which consists in: clustering electronic documents according to the method according to the first object of the present invention, and then search for relevant search query of electronic documents as fragments of a constructed graph.

To solve the same problem and provide the same technical result, a third object of the present invention provides a machine-readable medium intended for direct participation in the work of a computing tool and containing a program that, when executed in a computing tool, provides the execution of the method according to the first object of the present invention.

To solve the same problem and provide the same technical result, a fourth object of the present invention provides a machine-readable medium intended for direct participation in the work of a computing tool and containing a program that, when executed in a computing tool, provides the execution of the method according to the second aspect of the present invention.

Brief Description of the Drawings

The present invention is illustrated by the accompanying drawings.

Figure 1 shows a block diagram of a semantic clustering algorithm for electronic documents in accordance with the present invention.

Figure 2 illustrates the definition of a measure of proximity of vectors.

Figure 3 illustrates the principle of Canopy clustering.

Figure 4 shows a graph depicting the principle embodied in the method of the present invention.

Figure 5 shows a screenshot of part of the graph returned to the query "Mars living conditions" in a system using this invention.

Figure 6 shows an example of a report generated by the system according to some selection criteria.

Detailed Description of Embodiments

The problem of reducing redundancy can be solved in various ways. In most cases, huge amounts of information can be made accessible to perception if you can break down information sources, such as web pages, into thematic groups. Then the user can immediately drop many documents from groups with low relevance. This process of grouping text data is carried out using clustering.

Clustering a sample of documents is an effective means of improving the quality of a user’s dialogue with a search engine, allowing them to be partitioned by topic. The purpose of document clustering is to automatically identify groups of semantically similar documents among a given fixed set of documents when no characteristics of these groups are specified in advance.

The method of automatic iterative clustering of electronic documents by semantic proximity in accordance with the present invention is intended to increase the efficiency of processing information contained in various sources, in order to provide processes for managing data extracted from various electronic sources of scientific and technical information using innovative semantic technologies. This method can be implemented in a system that is or includes a computing tool (server, personal computer, etc.) programmed to perform the steps described below.

To identify the semantic proximity of electronic documents, data mining methods are of particular relevance. The main feature of these methods is to establish the presence and nature of hidden patterns in the data, while traditional statistical methods deal mainly with the parametric assessment of patterns already established. Among the methods of data mining, clustering occupies a special place. Clustering, based on the relationship of similarity of elements, establishes the subsets (clusters) into which the input data is grouped.

Usually, various sets of Internet documents serve as sources of processed documents. The determination of significant sources is carried out by the user, taking into account his informational and analytical needs and capabilities provided by the information space (availability of sources, availability of information about certain objects, etc.), i.e. the circle of the considered sources of information, the content of the subject area.

In this case, the user determines the type of information need and its limitations (by type of publication, publication language, geographical and chronological framework, etc.).

As a rule, users have preferred sets of Internet resources on the subject of their activities. The list of information resources is formed in such a way that resources complementing each other maximally cover information on this topic.

The source data for implementing the methods of the present invention is the flow of time-related documents that comes into the system. Each document is considered as a source information object with a creation (appearance) time and a unique identifier - a unified resource locator (URL, Uniform Resource Locator).

The task of the system that implements the methods of the present invention is to group the incoming document flow by topics and to establish the proximity between the topics. The system operates with instances of objects of several types - documents, topics, supertopics, explained below. Each instance of any of the types of objects in the system has its own unique number.

The initial operation performed by the system is to replenish the system database with the initial information objects. Specialists understand that the text to be clustered must be submitted in electronic form for subsequent automated processing. This step in Fig. 1 is conventionally indicated by the reference number 1 and can be performed in any known manner, for example, by scanning the text and then recognizing it using well-known means such as ABBYY FineReader. If the text is received for processing from an electronic network, for example, from the Internet, then the stage of its submission in electronic form is performed in advance, before this text is posted on the network. Therefore, in the flowchart of FIG. 1, the step of presenting the document in electronic form is shown in dotted lines. For each of these objects, the original document format is saved in the system database in the original database (item 2 in FIG. 1). On its basis, the text of the document is extracted, extracted from the document in the original format, referred to as planarized text (item 3 in FIG. 1).

Next, form arrays of terms for the planarized text of each of the stored electronic documents. To do this, first planarized text is tokenized, resulting in segments in the form of words, punctuation, spaces. Those. text is segmented into elementary units called tokens. A token can be any object of the following: words consisting of each sequence of letters and, possibly, hyphens; sequence of spaces; punctuation marks; numbers (for example, 2012, Tien Shan). Sometimes sequences of symbols such as А300, i150b, etc. are also included here. The allocation of tokens is always carried out according to fairly simple rules, for example, as in the application for US patent No. 2007/0073533 (published on March 29, 2007). After that, the tokenized text is stamped, i.e. isolate the basics of words using any of the well-known heuristic algorithms - for example, using libraries such as Snowball (see http://snowball.tartarus.org/) and Ispell (see http://www.gnu.org/software/ispell /). In figure 1, this step is conventionally indicated by the reference position 4.

After that, the weight of each term is found in the generated array of terms in each of the electronic documents (item 5 in FIG. 1). This calculation of the weight of each term can be performed, for example, using the measure TF-IDF, which is a product of the quantity

by the amount

.

.

In principle, the weights of the terms can be found differently, for example, as described in the aforementioned US patent No. 6189002.

After the weights of all terms included in the text of a particular electronic document are found, this electronic document is expressed as a vector in multidimensional space (item 6 in FIG. 1). For example, if for this document we write in order the weight of all terms, including those that are not in this document, we get a vector, which will be the representation of this document in vector space. The dimensions of this space are determined by the found term weights in the text of this electronic document. Those. the weight of a term in a document is the “importance” of the word source for that term when identifying the given text. If you count the number of terms used in a document, the so-called frequency of a term, then the more often a word appears in a document, the more weight it will have. If the term does not appear in the document, then its weight in this document is equal to zero.

As a result of the steps indicated by the reference numbers 1-6 in FIG. 1, each electronic document to be processed is converted into a corresponding multidimensional vector in multidimensional space. We can say that the generated multidimensional vector is an ordered set of numbers with a length equal to the number of entries in the ordered glossary of terms found in documents, where each term corresponds to TF-IDF. Terms are words referred to in documents in a normal form.

и

весов термов документов:Next, find a measure of the proximity of the resulting multidimensional vector to each of the multidimensional vectors in existing clusters, combining semantically close electronic documents processed earlier (item 7 in FIG. 1). Such a measure can be calculated in the same way as described in the already mentioned US patent No. 6189002. However, it is preferable to use a cosine measure of proximity of vectors

and

document term weights:

,

,

Where

, i=1, 2 - вектор в пространстве термов

, извлекаемых из всего множества документов

.-

, i = 1, 2 - a vector in the space of terms

retrieved from a wide variety of documents

.

- скалярное произведение векторов весов, вычисляемое по формуле-

is the scalar product of weight vectors calculated by the formula

, где x_i, y_i, - i-e координаты векторов,

, where x _i , y _i , - ie the coordinates of the vectors,

, k=1, 2 - евклидова норма вектора, вычисляемая по формуле-

, k = 1, 2 - the Euclidean norm of the vector, calculated by the formula

,

,

a, x _i - coordinate of the vector (weight of the i-th term), calculated by the formula

,

,

in which tf _t is the frequency (number of references) of the term t in this document, N is the total number of documents, and N _t is the number of documents in which at least one term t occurs. Note that tf _t is the TF value defined above for the term t.

Figure 2 illustrates the determination of the proximity measure of vectors d ₁ and d ₂ in some space.

After finding the proximity measure of the vector in the document to be processed, with the vectors in the already existing clusters, perform the step (pos. 8 in FIG. 1), in which the existing clusters for which the measure found in the previous step are supplemented by this electronic document to be processed proximity is minimal. If we are talking about the initial clustering of several documents (when the clusters have not yet been determined) or the preliminary clustering of a set of not yet processed electronic documents (if they accumulate over a predetermined or arbitrary time interval), then cluster analysis can be performed as follows.

In general, cluster analysis is a way of grouping multidimensional objects, based on the presentation of the results of individual observations by points of a suitable geometric space, followed by the allocation of groups as “clumps” of these points (clusters). Cluster analysis involves the selection of compact, distant from each other groups of objects, looking for a “natural” partition of the population into areas of cluster objects. It is used when the source data is presented in the form of matrices of proximity or distance between objects or in the form of points in multidimensional space. The most common are data of the second type, for which cluster analysis is focused on identifying some geometrically remote groups within which objects are close. In the present invention, the points in multidimensional space are conditionally the ends of the found multidimensional vectors.

Thus, clustering is the partitioning of multiple documents into clusters, i.e. subsets whose parameters are not known in advance. The number of clusters can be arbitrary or fixed. All clustering methods can be divided into numerical and non-numerical. Numerical methods use numerical characteristics about documents, and non-numerical methods use directly the words and phrases that make up the text. There are statistical, hierarchical and graph clustering algorithms.

In the present invention, clustering is implemented as a sequential multi-stage process, at the individual stages of which the “classical” clustering of a certain portion of data is performed using any well-established algorithm. For example, in the practical implementation of the system, the clustering algorithm, called the Canopy method (“canopy” method) is preferably used (see http://www.kamalnigam.com/papers/canopy-kdd00.pdf).

The main idea of this method is to perform clustering in two stages, of which the first stage is rough and fast, it divides the data into intersecting subsets called canopies, and then a second, more thorough stage can be applied, where more costly distance measurements only between points that are under a common canopy ("canopy").

The first step uses a non-cost distance measurement in order to create a number of overlapping subsets called “awnings”. A canopy is simply a subset of elements (ie, data points or elements) that, according to an approximate measurement of similarities, are at some distance from the center point. Most importantly, an element can be under more than one “canopy”, and each element must be at least one “canopy”. “Canopies” are created with the intention that points that are not included in any common “canopies” are located at a sufficient distance from each other, and thus cannot be in the same cluster. 3, circles with solid lines show an example of overlapping “awnings” that span data sets.

Thus, after finding the multidimensional vectors and calculating the measure of their pairwise proximity (i.e., after step 8 in Figure 1), all the obtained multidimensional vectors are divided into subsets, in each of which the calculated measure of proximity (for example, the cosine measure of proximity) between a pair any multidimensional vectors are less than a predetermined value. In this case, the centroid vector of each of the obtained subsets is calculated as the arithmetic mean of all multidimensional vectors of this subset (item 9 in FIG. 1).

At the second stage of the canopy method, only one cluster is assigned to each point according to the principle: each point (i.e. each vector) of a multidimensional space is assigned to the closest centroid. The classical algorithm uses a distance metric and two distance thresholds T1> T2. For each of the many points, if the distance from a certain point in a particular cluster is <T1, this point is added to this cluster (solid circles in FIG. 3). Further, if this distance is <T2, then this point is removed from further consideration and is considered to belong to this "canopy" (dashed circles in Figure 3), which allows to reduce the computational complexity of the algorithm. Thus, an object (point), very close to the center of the future cluster, will be excluded (excluded) from further processing. The algorithm is repeated until the initial set is empty.

Then the formation of clusters is carried out on the basis of the obtained centroids (arithmetic mean of all points included in this “canopy”). For each point, the nearest centroid is searched (the center of this “canopy”, the center of the circle in FIG. 3). Each point is assigned to the closest centroid. In this case, the number of clusters becomes equal to the number of “canopies”.

After the stage of adding a cluster to a new document (pos. 8 in Fig. 1), its new multidimensional vector (pos. 10 in Fig. 1) is determined as the arithmetic mean of the previous vector of this cluster and the vector of the added document. Then, they take as the theme of the aforementioned augmented cluster the name of that of electronic documents in this cluster for which the measure of proximity of its multidimensional vector to a certain new multidimensional vector is minimal (pos. 11 in FIG. 1).

Figure 4 presents the graph template, which fit the clusters obtained as a result of the above clustering.

The horizontal edges in this graph are hierarchical generalizing relationships between entities. The edge of the document-topic is the fact that the document is part of a cluster consisting of documents that are close in meaning. The top-super-top edge already generalizes clusters at a higher level. The presented template, depending on the nature or structure of the source documents, can be continued to the right (hypertopics, etc.) to achieve higher degrees of generalization.

The higher the hierarchy, the lower the number of entities at a given level. And the higher the hierarchy, the higher the degree of generalization. These facts are important when using the resulting structure to speed up information retrieval and to present huge amounts of information in a structured and generalized form.

Edge arcs connect entities within each level of the hierarchy. These links are useful for expanding the list of documents found by documents that are close in meaning.

It is the method of obtaining the structure described above that is the main subject of this invention.

The specificity of the method of the present invention is that existing clusters are processed as separate documents. Those. find the multidimensional vector of each cluster (in fact, it has already been found after the cluster has been supplemented with another electronic document); find a measure of the mutual proximity of multidimensional vectors for each pair of clusters; cluster those clusters for which the found measures of mutual proximity of their multidimensional vectors do not exceed a predetermined threshold value; define a multidimensional vector for the topic; take as a topic topic the theme of that of its clusters, for which the measure of proximity of its multidimensional vector to a certain multidimensional vector of this topic is minimal.

This procedure is preferably repeated already for topics, i.e. find a measure of the mutual proximity of multidimensional vectors for each pair of topics; combine those topics for which the found measures of mutual closeness of their multidimensional vectors do not exceed a predetermined threshold into the corresponding supertopic (plot); define a multidimensional vector for a supertop; they take as the topic of such a supertopic the topic of that of its topics, for which the measure of proximity of its multidimensional vector to a certain multidimensional vector of this supertopic is minimal.

The same procedure can be repeated for plots in order to form superplots, etc. For example, in the case of determining topics for supertopics, a graph is constructed in the method of the present invention, the nodes of which are supertopics, and each of the edges represents the proximity ratio of the supertopics connected by this edge. After that, a global glossary is compiled to enable subsequent searches for fragments of the constructed graph that are relevant to a particular search document.

As an illustration, Figure 5 shows a screenshot of the graph returned to the query "Mars living conditions" in a system that practically uses this invention. This image shows part of the graph structure described above at the topic level, i.e. at the level of document clusters, between the clusters are represented arc-links (see the graph template in Figure 4). The system found that the query vector “Mars living conditions” when searching at the topic level is closest to a cluster of 12 documents with a heading taken from the closest to the centroid cluster of the document “NASA: There were all living conditions on ancient Mars”.

To illustrate the hierarchical relationships in Fig.6 shows an example report on the types of road traffic accidents (RTA) for a certain period of time. Here, entities of a high level in the hierarchy contain objects of lower levels. Figure 6 schematically shows the accident reports of three types: truck and bus (general circle at the bottom left), truck and car (general circle at the top), car and automobile (general circle at the bottom right). Here the circles of the smallest diameter (and the darkest) represent documents (news reports) in which the fact of an accident is mentioned. Further, according to the hierarchy, messages about the same event are combined into topic clusters, topics are combined into supertopics, supertopics are combined according to the type of events described in them (car-car, truck-car). Moreover, the circle of the largest diameter (and the lightest) carries a purely aesthetic load.

For the second object of the present invention, the search method for documents collectively clustered by semantic proximity, electronic documents are first clustered according to the method according to the first object of the present invention with the mandatory construction of the aforementioned graph, and then the electronic documents relevant to the search query are searched as fragments of the constructed graph. For example, a query vector is constructed and, based on the proximity measure, the nearest super-peak is searched, then the topic, and then the document.

Those skilled in the art will understand that both methods of the present invention are executed in a suitably programmed system. Therefore, another two objects of the present invention are computer-readable media, each intended for direct participation in the work of the computing means of the specified system and containing each program, which, when executed in the computing means, ensures the execution of the corresponding method. Such computer-readable media can be either hard drives or other devices, such as flash memory, DVDs, magnetic tapes, etc.

Thus, the use of the present invention allows to simplify and accelerate both the clustering of electronic documents by semantic proximity, and the subsequent search in a clustered collection of those documents that are relevant to the search query.

Claims (12)

1. The method of automatic iterative clustering of electronic documents by semantic proximity, which consists in the fact that:
- convert each electronic document to be processed into a corresponding multidimensional vector in a multidimensional space, the dimensions of which are determined by the terms contained in the said electronic document;
- find a measure of the proximity of the resulting multidimensional vector to each of the multidimensional vectors of existing clusters, combining semantically close electronic documents processed previously;
- supplement the said clusters for which the measure of proximity found is minimal;
- determine for the aforementioned augmented cluster its new multidimensional vector;
- take as the theme of the aforementioned augmented cluster the name of that of electronic documents in this cluster for which the measure of proximity of its multidimensional vector to a specific new multidimensional vector is minimal. 2. The method according to claim 1, in which:
- accumulate a set of electronic documents to be processed as they appear during a predetermined time interval;
- after which they carry out clustering of each of the electronic documents in the accumulated totality. 3. The method according to claim 1, in which said conversion of an electronic document into a multidimensional vector includes the steps of:
- planarize the text of said electronic document;
- form arrays of terms for the planarized text of each of the above electronic documents, for which:
- tokenize the planarized text, resulting in segments in the form of words, punctuation, spaces;
- they stamp the tokenized text, highlighting as a result of the word base using at least one of the heuristic algorithms; then:
- find the weight of each term in each of the mentioned electronic documents;
- Express each of the mentioned electronic documents as a vector in a multidimensional space, the dimensions of which are determined by the found term weights in the text of this electronic document. 4. The method according to claim 3, in which the aforementioned calculation of the weight of each term is performed using the measure TF-IDF, which is a product of the quantity

by the amount

. 5. The method according to claim 1, in which said finding a measure of proximity of multidimensional vectors includes the steps in which:
- calculate the cosine measure of proximity between each pair of these multidimensional vectors;
- divide all the mentioned multidimensional vectors into subsets, in each of which the calculated cosine measure of proximity between a pair of any multidimensional vectors is less than a predetermined value;
- calculate the centroid vector of each of the mentioned subsets as the arithmetic mean of all multidimensional vectors of a given subset;
- they attribute each multidimensional vector to a subset with the nearest centroid vector. 6. The method according to claim 1, further comprising stages in which:
- find a measure of the mutual proximity of multidimensional vectors for each pair of the mentioned clusters;
- cluster those clusters for which the found measures of mutual closeness of their multidimensional vectors do not exceed a predetermined threshold value;
- determine for the mentioned topic its multidimensional vector;
- take as the topic of the mentioned topic the topic of one of the clusters included in it for which the measure of proximity of its multidimensional vector to a certain multidimensional vector of this topic is minimal. 7. The method according to claim 6, further comprising stages in which:
- find a measure of the mutual proximity of multidimensional vectors for each pair of the topics mentioned;
- combine those topics into the corresponding supertopic for which the found measures of mutual closeness of their multidimensional vectors do not exceed a predetermined threshold;
- determine for the said supertopic its multidimensional vector;
- they take as the theme of the mentioned super-topic the topic of that of its topics, for which the measure of proximity of its multidimensional vector to a certain multidimensional vector of this super-peak is minimal. 8. The method according to claim 7, in which a graph is constructed whose nodes are the mentioned supertopics, topics and documents, and each of the edges of this graph is a ratio of the proximity of supertopics, topics and documents connected by this edge. 9. The method of claim 8, in which a global glossary of terms is provided to enable subsequent searches for graph fragments relevant to a particular search document. 10. The search method in the aggregate of documents clustered by semantic proximity, which consists in the fact that:
- carry out the clustering of electronic documents according to the method according to claim 9;
- perform a search for relevant electronic search query electronic documents as fragments of the mentioned graph. 11. Machine-readable medium intended for direct participation in the work of a computing tool and containing a program that, when executed in said computing tool, provides the execution of the method according to any one of claims 1 to 9. 12. A machine-readable medium intended for direct participation in the work of a computing tool and containing a program that, when executed in said computing tool, provides the execution of the method of claim 10.

Images