RU2476927C2 - Method of positioning text in knowledge space based on ontology set - Google Patents

Abstract

FIELD: information technology.

SUBSTANCE: result is achieved by using a method of positioning text in knowledge space. In the disclosed method, elements are selected from input data, which correspond to patterns included in taxa which form taxonomia merged into ontology. Significant taxa are determined, which are weighed based on conditions assigned to patterns. A set of weighed vectors which position the input document in the knowledge space is constructed. An ontology set is used for positioning. When constructing sets of vectors, only those elements which correspond to patterns included in one taxon or in taxa having common parent taxa are considered.

EFFECT: high efficiency of search engines, contextual advertisement systems and systems for behavioural targeting on the Internet.

3 cl, 3 dwg

Description

Field of Invention

The present invention relates to a method for identifying objects by their textual or other descriptions and can be used, for example, in situation analysis, in information retrieval, in the construction of search engines, in contextual advertising systems, etc.

Introduction

For a long time, taxonomies were considered a useful tool for classifying objects. Besides the fact that they give the possibility of naming classes of objects, they also give the opportunity to determine the degree of similarity.

In its simplest form, taxonomy is a hierarchical grouping of individual concepts into more general classes. Two concepts in taxonomy have common properties of the level of grouping that includes both of them, and the degree to which the concepts are similar depends on the relative position of the classes in the hierarchy.

Various authors, for example:

- F. Reznik, “Researching information content to evaluate semantic similarity in a taxonomy” (“Using information content to evaluate semantic similarity in a taxonomy”, Proceedings of the 14th International Joint Conference on Artificial Intelligence, 1995, pp. 48-453);

- J. By et al., “Verb semantics and lexical selection” (Proceedings of the 32nd Annual Meeting of the Association for Computational Linguistics, 1994, pp. 133-138);

- D. Lin, “An information-theoretical definition of similarity” (“Proceedings of the 15th International Conference on Machine Learning Proceedings of the 15th International Conference on Machine Learning, 1998, pp. 266-304);

- J. Rice et al., US Patent Application No. 20080027929 "Computer-based method for finding similar objects using a taxonomy"

developed methods for converting the intuitive idea of "similarity" into a numerical measure that can be used to measure the "similarity" of objects.

Finding similar objects according to a given description is used in many areas of technology. For example, someone may want to find a patent similar to a given one, or an object of “similar” clinical trials of drugs. In bioinformatics there is a need to search for genetic products (for example, proteins) that are similar to a given genetic product. In each of these areas (and not only in them), various detailed taxonomies have been developed and are used to classify sets of objects.

Taxonomy classification may be more complex than the examples described above. First, often a certain class of objects refers to more than one “parent” class (that is, a class that is higher in the hierarchy). Secondly, taxonomies often change over time: new specialized groups form; the content of old groups is changing. Thirdly, even with a constant taxonomy, the classification of a particular object can change with a change in knowledge about it, disagreements of users of a taxonomy about the place of specific objects in them are also possible.

Text Analysis Issues

There are a number of problems that should be solved by systems for analyzing texts, in particular:

(a) the problem of word forms - the same word in Russian can have up to 40 different word forms, which complicates the search;

(b) the problem of synonymy - the use of different words to describe the same phenomenon or idea. In the search process, the user, upon request with the keyword “doctor”, may not find the word “doctor”, meaning in this context the same concept;

(c) the problem of polysemy - one word can have a number of meanings that are not related to each other (for example, “onion” - a plant and “onion” - a weapon). In the search process, there is a high probability of finding the document with the wrong word meaning that the user needs.

Also significant is the issue of determining similarity measures when comparing documents.

Similarity measures

The literature describes three different groups of similarity measures that can be applied to taxonomies. The first group, called “measures of similarity of terms”, can be used to calculate the similarity of individual terms. Two other groups of similarity measures can be applied when an object is denoted by several terms.

In the context of machine translation, By and Palmer in the above article entitled “Verb semantics and lexical choice” described a measure of similarity based on the depth of the taxonomy of the closest common ancestor of the two terms relative to the taxonomic depth of the individual terms. The closer the common "ancestor" to the terms, the greater the similarity. The problem with this approach is that some parts of the taxonomy can be highly developed and contain a significant number of terms, while in other parts of the taxonomy the density of terms is less. Such a difference in the “densities” of terms makes this and other similarity measures based on a simple count of the number of edges in the graph insufficiently accurate.

The idea of using information content to measure similarity belongs to Reznik (see above). Using the taxonomies of the WorldNet project and word frequency estimates obtained from a large array of English texts, Reznik calculated the semantic similarity of word pairs by highlighting a common ancestor with the most information content. For words with several meanings, Reznik used the meaning that gave the maximum measure of similarity. Using the estimates made by people as a standard, Reznik found that such a measure works better than the measures described above, based on counting the number of connections. The Reznik measure cannot be used in the case when the object is indicated not by one, but by several terms. In addition, this measure has a number of other disadvantages. Its range of similarity values is not normalized, but more importantly, choosing the parent nodes with the highest information content, Reznik underestimates the similarity of objects, focusing on only one most important aspect of similarity at the cost of ignoring everyone else.

Lin provided an axiomatic definition of similarity and showed how Reznik's approach can be adapted for use in this situation. The measure of Reznik’s similarity was based solely on the generality of the meanings of words, while Lin’s approach takes into account the meaning of the difference in the meanings of words to determine the normalized similarity coefficient. Lin compared his measure of similarity with Reznik and Wu-Palmer and found that it gives similarity indicators that correlate better with human ratings of this criterion than their methods. However, Lin did not describe how his measure of similarity can be used when different terms can be used to describe the object.

The measure of similarity of the second group is described in the literature as a measure of similarity, based on the number of references (frequency of references) of terms that are common in the descriptions of both objects. Measures of this group include Jaccard, Dice and Set Cosine, which are often used in information retrieval systems and differ from each other in the way that the number of general terms in descriptions is normalized, as well as the measure described by Keller and others in the article “Based on Taxonomy” soft measures of similarity in bioinformatics "(" Taxonomy-based soft similarity measures in bioinformatics ", 2004. Proceedings. 2004 IEEE International Conference on Fuzzy Systems, Volume 1, Issue 25-29 July 2004, pp. 23-29). All these similarity measures do not take into account the structure of the taxonomy. Any candidate object that does not have common terms with the object being compared will receive a similarity coefficient of zero, although in reality they can be very similar.

A third approach to measuring similarity is based on a measure of similarity of terms to determine similarity between individual terms, with a further combination of them to obtain a common similarity coefficient. Chalkidi et al. In THESUS: Organizing Web document collections based on link semantics, The VLDB Journal - The International Journal on Very Large Data Bases, Vol.12, Issue 4 (November 2003) p. 320-332) describe a measure of similarity of this type for use in clustering web pages. Using a measure of the similarity of By and Palmer, Halkidi considers each term in the source and base documents separately and finds the most similar terms for each of the documents. Further, for each set of terms (from the source and base documents, respectively), similarity coefficients are determined and averaged. After that, Halkidi combines similarity coefficients from two sets of terms without taking weight functions into account. Since Halkidi also uses the similarity measure of By and Palmer, using it for the taxonomy represented by directed acyclic graphs is not possible.

Wang et al. In the article “Gene expression correlation and gene ontology-based similarity: An assessment of quantitative relationships”, Computational Intelligence in Bioinformatics and Computational Biology, 2004. Proceedings of the 2004 IEEE Symposium on CIBCB, 7-8 Oct. 2004, pp. 25-31) describe a similarity measure that circumvents a problem that cannot be solved by the Chalkidi method by using the generalized form of Lin's informational and theoretical similarity to determine the similarity of each pair of terms.

It should be noted that Wang et al. Generalize Lin's formula for use in the taxonomy described by a directed oriented graph by choosing the closest common ancestor with the most saturated information content. Their approach differs from the approach of the patent application of Rayet et al. In that they apply the measure of similarity only to pairs of terms and do not exactly follow the axiomatic definition of similarity of Lin because they consider only part of the general meanings of the terms. Wang uses a function different from Halkidi to combine the similarity coefficients of the pairs of terms. Instead of averaging the coefficients for the closest values from both sets of terms, they average the similarity values across all sets of term pairs.

All similarity measures in this third group are artificially separated by combinations of terms for which there are already common taxonomic ancestors, and those for which there are none.

Keller et al. (See above) present several ways to determine similarity based on fuzzy measures based on the depth of the information content of the terms. However, these measures require some subjectively determined weighting coefficients or the solution of polynomial systems of equations of high complexity, which makes the use of these methods ineffective for large amounts of information.

Solving the problem of searching for “similar” or “close” documents, all of the above methods do not solve the problem of interpreting search results. For example, we can measure the exact distances between geographical objects, but at the same time do not get any idea about the position of these objects on the earth's surface.

The key difference between the method used in this invention and the above methods is that it allows you to solve the problems of positioning a document or documents in the knowledge space, not limited to calculating the similarity measures of pairs or groups of documents.

Disclosure of invention

Definitions

Before proceeding to the description of the claimed method, it is advisable to give definitions of some of the concepts found in the description and the attached claims.

Pattern - verbal, graphic, numerical and other components in acceptable forms and modifications or a phrase that includes micro-context.

Microcontext is a part of a pattern that has an independent meaning (for example, a set of characters can be assigned the value "date", "price", "name", etc.)

A directed acyclic graph (NAG) is a case of a directed graph in which there are no directed cycles, that is, paths starting and ending at the same vertex.

Taxonomy is a directed acyclic graph that reflects the hierarchy (tree) of concepts in descending order of categories from general to particular, with one or more patterns per category. The nodes of the graph are called taxa. Any taxon, except the highest in a taxonomy, can relate to one or more taxonomies.

Ontology in this application is understood as a set of taxonomy, the nature (meaning) of which allows a similar interpretation and, in particular, can be characterized by a single coefficient of significance (see figure 1).

The space of knowledge is a set of ontologies that describe different subject areas.

A vector is a collection of sets of significant categories to which the document in question belongs and the weights assigned to them. Each document can be defined by a set of vectors.

Categorizer is a hardware-software mechanism designed to obtain a description of the input text in the form of a set of vectors that position the input document in the knowledge space. The categorizer uses a priori knowledge of the connection and hierarchy of concepts, the characteristic features of categories.

Analogs

The known method of “identifying objects according to their descriptions” (RF patent No. 2164450, IPC G06F 17/30, publication date 12/27/2000) using only a statistical measure of similarity based on the frequency of use of the words found in the document. Due to the limitations of these statistical methods, the problems of polysemy and synonymy in the source texts remain unresolved, which significantly reduces the quality of identification compared to the proposed invention.

The known method of Latent Semantic Analysis (LSA), which can be used to determine the similarity of documents (US patent 4839853 "Search for computer information using latent semantic structures" ("Computer information retrieval using latent semantic structure", authors Dirvester and others, International classification G06F 17/21; G06F 17/30; G06F 15/40) The LSA method uses statistical text processing methods in which information on the frequency of use of words and phrases is extracted from text arrays, from which, in turn, statistical methods are used to extract information about “concepts”, which makes it possible to subsequently determine the “similarity” of documents. In this case, unlike the present invention, the LSA method does not use taxonomies as a basis for calculating similarity, but relies only on statistical methods. For example, the weight of the term is back proportional to the frequency of its mention in the text, regardless of the subject of the document.The LSA method does not position documents in the knowledge space.

Descriptions of approaches to constructing a systematic classification of the knowledge space are proposed in the book by S. Kordonsky “Cycles of activity and ideal objects” (Moscow, 2001).

There is a known method of searching for documents using taxonomy (US patent 6442545 "Term-level text with mining with taxonomies", authors Feldman and others, international classification G06F 17/30), which does not provide for the possibility of entering the same category in different taxonomies, the unification of taxonomy in ontology and the application of the concept of “conditionality coefficient” for ontology.

The known method of generating taxonomy described in US patent 6360227 ("System and method for generating taxonomies with applications to content-based recommendations", authors Aggarwal and others, international classification G06F 17/30), as well as the method of creating categorized document bases described in US patent application 20070106662 ("Categorized document bases", authors Kimbrow and others), considering the similarity of pairs of records (documents) in accordance with a given taxonomy. Unlike the present invention, they do not solve the problem of positioning documents in the knowledge space, but relate only to the search for similar documents.

Closest to the proposed method is presented by Rice and other US patent application number 20080027929 "Computer-based method for finding similar objects using taxonomy" ("Computer-based method for finding similar objects using a taxonomy").

The method presented in this application differs from the method of Rice et al. In that it uses the concept of “ontology” as an aggregate of a taxonomy and allows one to take into account cases when a taxon belongs to one or more taxonomies, and also uses the concept of a conditionality coefficient attributed to a taxonomy to determine usage pattern when calculating the weight category.

It should be noted that the documents encountered in real life are rarely on the same topic. Accordingly, a document, as a rule, is positioned in the knowledge space not by one, but by several vectors, each of which can be assigned a weight coefficient.

As indicated in the aforementioned work of S. Kordonsky, the position of the observer determines the degree of significance for him of certain aspects of the document and makes senseless the attempt to determine the only correct “main” topic of the document.

If we admit that for “real texts” there is no “only correct” and comprehensive ontology, it becomes obvious the need to create a positioning apparatus focused on obtaining multiple results, leaving the possibility of choosing the category most relevant for the system user at the moment.

Unlike the present invention, all of the above analogues do not address this problem.

From the current level of technology, objects that would contain a combination of the above characteristics have not been identified. This allows us to consider the claimed method as new.

From the existing level of technology is also not known a set of features other than the features of the aforementioned closest analogue. This allows us to consider the claimed method with an inventive step.

Thus, the creation of a positioning mechanism in the presence of many ontologies, each of which is represented by its own set of taxonomy, seems to be an actual technical result when creating a really functioning system.

The essence of the invention is the method of positioning documents in the knowledge space, represented by many ontologies, which in turn are represented by a set of combining patterns of taxonomies (see figure 2). Unlike the closest analogue (patent application Rice et al.), Taxa are used as a generalization of terms and can describe not only sets of relevant terms, but also phrases and expressions considered as signs of the corresponding categories. As well as in the considered analogue, the same pattern can fall into different categories. Unlike the closest analogue, in this method, each category can simultaneously be the top of one tree and a subcategory of other trees.

Unlike all of the above methods, in this invention, taxonomies are combined in the ontology with the same rules for interpreting taxa, and additional conditions for its application may be imposed on the pattern. For example, the conditional pattern "financial crisis" can be included in the tree of categories, the upper nodes of which are designated as "Russia", "Germany", "USA", but in each of the three taxa it will be considered only if there are other patterns that clearly belong to this tree.

The space of knowledge, ontology, taxonomy and the form of their representation in the form of directed acyclic graphs or others are not the subject of the present invention. The well-known language RDF (Resource Description Framework) and other well-known languages for describing ontologies and taxonomies can be used as an input description of taxonomies with subsequent translation into the language of the internal representation of the categorizer (310, Fig. 3). Other variants of the input description of taxonomies are also possible, implying the creation of special software-converters from the language of the external representation to the internal representation of the categorizer.

Brief Description of the Drawings

Figure 1. A diagram showing an example of an ontology uniting taxonomies presented in the form of directed acyclic graphs.

Figure 2. A diagram showing an example of a taxonomy presented as a directed acyclic graph with taxon-related patterns.

Figure 3. A diagram illustrating an embodiment of the invention.

Detailed description of an embodiment of the invention

The input text 301 (FIG. 3) is broken down into sentences 302. Each sentence is considered for the presence of words and expressions 303 corresponding to patterns known to categorizer 310. Based on patterns P _H for which relevant words or expressions are found in the text, in available taxonomies and ontologies, subtrees (subgraphs) 304 are distinguished that contain concepts that correspond to this pattern and all its parental concepts. The found patterns of P _H confirm the hypothesis that the text belongs to the corresponding categories. In this case, not only categories directly confirmed by the P _H patterns are considered, but also categories including directly confirmed categories. Subtrees, consisting of one taxon leading to one concept, are discarded.

Further, in each of the resulting subtrees, the weight value of each category 305 is determined. The weight of the category is calculated as:

Where

w ₁ - the sum of the weights of all concepts directly related to this category,

w ₂ is the weight of all subtrees related to this node, calculated as

,

,

Where

w is the sum of the weights of all patterns P _H included in the subtree,

c ₁ - the number of branches in the subtree,

o ₁ - the number of patterns in the subtree,

n _c is the number of nodes in the subtree.

, гдеThe weight of the pattern is everywhere

where

n _c is the number of categories with which this pattern P _H is directly related,

K is a coefficient taking values from 0.3 to 3.

The set of vectors leading from the upper categories of taxonomy to confirmed subcategories, arranged in order of category weights, shows the position of this document in the knowledge space.

In this case, subtrees 306 associated with the common parent category or, in its absence, the upper confirmed categories, which have a weight value below the cutoff border for the ontology into which they enter, are discarded, that is, they are considered insignificant.

The significance of a category in taxonomy depends on the ontology and is determined by the cutoff number, which is a characteristic of this ontology. In accordance with the gained weight, we can determine the categories as more or less significant for this document. Moreover, only categories belonging to the taxonomies of one ontology can participate in the comparison. For example, considering the taxonomy “color” and “vehicles” belonging to different ontologies, it seems senseless to try to determine if a given item is more red than a car when the input text “red car” is received.

Only those categories that have a numerical weight value exceeding the cutoff number specified for a given ontology fall into the significant ones. For example, a text containing only fragments corresponding to the patterns of “Viktor Yushchenko,” “Yulia Tymoshenko,” “gas conflict,” “negotiations with Russia,” may give rise to hypotheses about the significance of taxa: “the government of Ukraine,” “international economic relations,” gas transportation agreement "," beekeeping ". The hypothesis that in this document we are talking about beekeepers (since V. Yushchenko is a well-known beekeeper) will be rejected because he does not find any confirmation with other document patterns. That is, "Viktor Yushchenko" will be considered only in the context of his state activities, and not in the context of his personal hobbies.

Before calculating the weighting factors, all patterns that do not meet the specified conditions are excluded from consideration. So, for example, if the condition is the obligatory confirmation of the presence of at least a certain number of patterns confirming this category, then patterns that do not gain the indicated number are excluded from consideration.

Patterns that are identical in spelling can have different specific attributes (usually the name of one of the parent categories) that indicate the boundaries of the condition coefficient 307. This allows you to specify different conventions for the same spelling patterns. Then, for example, a change in the conditionality of the pattern “tire (motor vehicle)” will not affect the value of the conditionality of the pattern “tire (traumatology)”. Using conditional patterns, in particular, properties can be described that themselves do not exactly determine an object or phenomenon, but can belong to other classes of objects or phenomena, but in certain circumstances they allow you to make the choice of the desired object. For example, the pattern corresponding to the value of “one berth” or “two berths” does not in itself indicate that the document refers to a hotel or train, or a business plane, but in combination with patterns that define the theme "automobile transport" allows us to conclude that we are not talking about all cars, but only truck tractors or mobile homes.

As a result, a set of vectors is defined for the source text that indicate the categories that determine the context of this document, which is considered as the position of the document in the knowledge space.

Documents having similar vectors ("documents about the same") can be considered as similar; a document whose vectors are included as a subset of the vectors of another document (an article about the Po-2 aircraft can be included as a subset in the extensive document "History of USSR Aviation").

All the actions discussed above are performed using well-known software operations - comparisons, counting repetitions, calculating statistical values, working with matrices, etc. The specific type of relevant programs will be determined by the specific type of hardware and the operating system installed on it and is not the subject of patent claims of applicants.

Thus, from the above description it can be seen that the present invention allows to identify (locate in the space of knowledge) various objects, taking into account their similarity.

The present invention can be used in various fields of information technology, for example, in information retrieval, situation assessment, contextual advertising, can be the basis for building search engines, which, unlike traditional search by pattern, combine search by pattern with “concept search”, that is search by position of documents in the knowledge space. For example, in a conventional search, the results of the “crisis in Russia” query will contain all documents containing the words mentioned in this query, while when using the search algorithm in the knowledge space, the search results will be relevant to the semantic content of the query. For example, the “mass disobedience action in Dalnegorsk” is related to the “crisis in Russia”, but this search will not be obtained with a regular search, because the words “crisis” and “Russia” are not directly mentioned in the document about the “mass disobedience in Dalnegorsk” ".

In addition, this algorithm can be used as a means of matching documents in contextual advertising; behavioral advertising (the presence of a knowledge space allows you to introduce the concept of "customer behavior" as the history of his movements in documents positioned in this space, or as the history of his movements in documents corresponding to categories of space); when tying relevant documents to a document, that is, documents that have a similar set of knowledge space vectors.

Using the present invention allows to obtain a technical result, consisting in increasing the financial efficiency of contextual advertising systems by increasing the relevance of the displayed advertisements; increasing the effectiveness of information retrieval systems through more accurate, relevant, and continent output of results; improving the functioning of search engines through clearer targeting of information and advertising.

The examples of implementation of the present invention serve only as illustrations and in no way limit the scope of patent claims defined by the following claims.

Claims (3)

1. The method of positioning texts in the knowledge space, which consists in the fact that (a) elements corresponding to patterns included in taxa forming taxonomies combined in ontology are extracted from the input data; (b) identify significant taxa that are weighted based on the conditions assigned to the patterns; (c) compose a set of weighted vectors that position the input document in the knowledge space, characterized in that it uses a lot of ontologies for positioning, and also in the fact that when compiling sets of vectors, only those elements that correspond to patterns included in one taxon are considered or in taxa having common parent taxa. 2. The method according to claim 1, characterized in that all taxa, except for the top taxon of the taxonomy, can be subsidiaries simultaneously for several parent taxa. 3. The method according to claim 1, characterized in that it uses the concept of a coefficient of conditionality as a condition indicating the possibility of using the pattern only in a given context.

Images