WO2011033457A1 - Système et procédé de classification de contenu - Google Patents

Système et procédé de classification de contenu Download PDF

Info

Publication number
WO2011033457A1
WO2011033457A1 PCT/IB2010/054156 IB2010054156W WO2011033457A1 WO 2011033457 A1 WO2011033457 A1 WO 2011033457A1 IB 2010054156 W IB2010054156 W IB 2010054156W WO 2011033457 A1 WO2011033457 A1 WO 2011033457A1
Authority
WO
WIPO (PCT)
Prior art keywords
senses
category
metadata
metadata field
semantic
Prior art date
Application number
PCT/IB2010/054156
Other languages
English (en)
Inventor
Fulvio Corno
Paolo Pellegrino
Alberto Ciaramella
Original Assignee
Intellisemantic Srl
Politecnico Di Torino
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by Intellisemantic Srl, Politecnico Di Torino filed Critical Intellisemantic Srl
Publication of WO2011033457A1 publication Critical patent/WO2011033457A1/fr

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00—Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/50—Information retrieval; Database structures therefor; File system structures therefor of still image data
    • G06F16/58—Retrieval characterised by using metadata, e.g. metadata not derived from the content or metadata generated manually

Definitions

  • the present invention relates, in general, to the management of contents, in particular multimedia contents.
  • the present invention relates to the automated classification of such contents.
  • the largest multimedia public archives over the Internet (e.g., YouTube, Flickr) use a socially based approach to content description and classification.
  • the description of the resource is composed of some textual fields (such as title, description, and most importantly, tags) specified by the uploader, and may be enriched by other information or resources provided by "viewers" of the resource (e.g., comments, additional tags, new linked contents, bookmarks to the resource, etc).
  • tags may comprise words inserted by users that have uploaded the resource or by other users and may contain English words or words in other languages than English, slang, acronyms, spelling errors, proper nouns, or anything else, it is apparent that such a prior art suffers at least two problems:
  • Resource - a resource is any atomic unit of content, that can be identified, classified and searched on a given, preferably on-line, archive.
  • the present description refers to multimedia resources, i.e., non-textual resources, but resource may be intended more generally as referring to any kind of resource.
  • Resource Description - description of a resource is the set of metadata associated to the resource.
  • metadata may be, for instance, in the form of text, or tags, or categories.
  • metadata may be provided by the original author, or by some website visitor.
  • resource description can comprise one or more of the following elements or element types or metadata types or metadata fields, mostly of textual nature:
  • Site Category one or more categories selected among a list of categories; it means that the author must select, from a predefined (by archive or website creators) list of categories, the one or the ones most relevant to the resource. This information is not always reliable, since it comes from the user, and the predefined list of categories is very often ambiguous and vague (not following information architecture state-of-the-art principles);
  • Tags a set of words (uncontrolled keywords) describing the resource.
  • Tags may contain English words or words in other languages than English, slang, acronyms, spelling errors, proper nouns, or anything else.
  • Such tags are selected by the uploader or user, often by picking among the "most popular" tags in a given domain.
  • tags may contain internal spaces (such as "Deep Purple"), in other systems they can't (and then one would have "Deep” and "Purple” as separate tags).
  • tags may contain internal spaces (such as "Deep Purple"), in other systems they can't (and then one would have "Deep” and "Purple” as separate tags).
  • other users can add their own (personal) tags to a resource that somebody else has previously uploaded;
  • Comments are, usually, a paragraph or more of text, and may contain textual information useful to identify the resource; in general they are a totally uncontrolled source of information because they are added by other users to the resource;
  • - Bookmarks are, usually, personal or favorite resources; for instance, other users can add a resource to their bookmarks.
  • adding a resources to one's favorites requires the user to select some (personal) tags to classify it.
  • Target categories - target categories are application dependent conceptual classes, used to identify and group sets of resources with similar contents.
  • target categories comprise a set of predetermined categories; such a set, according to the preferred embodiment is an input information for the method as disclosed in present invention.
  • Present invention intends to solve the above problem.
  • the goal of the present invention is a system and method arranged for automatically identifying the correct category or a set of correct categories C(i) corresponding to a given resource, in particular a multimedia resource, by analyzing only the textual metadata of the resource.
  • the present invention also relates to a method for content management as well as to a computer program product loadable in the memory of at least one computer unit and including software code portions for performing the steps of the method of the invention when the product is run on at least one computer unit.
  • a computer program product is meant as equivalent to the reference to computer readable medium containing instructions for controlling a system or a device so as to co-ordinate execution of the method according to the invention.
  • Reference to "at least one computer unit” is meant to highlight the possibility for the method of the invention to be carried out in a decentralized manner over a plurality of computer units. Claims are an integral part of the teaching of the present invention.
  • present invention discloses a system for automatic classification of resource descriptions or metadata wherein a computer system is arranged, by using a set of semantic processes, to recognise senses associated to the metadata and to associate category weights to the metadata.
  • the set of semantic processes comprises at least one pre-processor block arranged to find the widest possible senses representing all the possible meanings of the metadata, at least one expander block arranged to identify senses that are recurring in the widest possible senses as expanded senses, and to isolate and delete senses that are marginal, and at least one matching block arranged to compare the expanded senses to constant sets of senses corresponding to target categories.
  • the system comprises an inhibition process arranged for inhibiting the output of the matching block, on the basis of statistical information, so that no classification is made of the metadata field.
  • Fig. 1 shows a block diagram of a system according to present invention
  • Fig. 2 shows a general architecture of modules implemented in the system of Fig. 1;
  • Fig. 3 shows the architecture of Fig. 2 in more detail
  • Fig. 4 shows an architecture of a block of Fig. 3 according to a first embodiment
  • Fig. 5 shows an architecture of a block of Fig. 3 according to a second embodiment
  • Fig. 6 shows one step of an internal process of Fig. 4 or 5.
  • a system for content identification (system) 10 comprises one (or more) computer servers 12, that host software elements arranged for exposing, for instance, Web applications to a plurality of users or end users connected to the servers 12 through a network 16 by means of computer terminals 14, as for instance personal computers.
  • the invocation of such applications is triggered by the servers 12 that access to services provided by at least one classification server 20a for automatic recognition of the correct information to be used for enriching searching and navigation of the end users.
  • the system 10 is arranged for supplying a reliable and semantically validated information about a multimedia resource and is used in real-time.
  • server 20a comprises a set of computer modules or a package 20 (Fig. 1, Fig. 2) having an architecture arranged for receiving:
  • metadata 21 of a single resource d comprises, for instance, a plurality of atomic units of content 31a, 31b, 3 In (Fig. 2, Fig. 3, Fig. 4) as for instance: title, tags, description, comments, or a subset thereof (as already reported); and
  • a constant information not depending on specific inputs comprises, for example, one or more of the following inputs:
  • semantic lexical network 23b for instance the semantic lexical network WordNet as described in "Introduction to WordNet: an on-line lexical database" by G.A. Miller, R. Beckwith, C. Fellbaum, D. Gross, K. J. Miller (Int. J Lexicography, vol. 3, pp. 235-244, January 1990;
  • mappings (Category senses) 23 c of each category onto a lexical network; - a set of additional pre-processing tables (Dictionary) 23 d.
  • the package is further arranged for generating in output 25a:
  • each of such numbers represents, for instance, a number or category weight CW(i) 25 that, for each of the target categories, estimates the relevance of the resource to the corresponding category.
  • Numbers 25, according to present embodiment are in a range from 0 to 1 and represent the automatic recognised correct information to be used for enriching searching and navigation of the end users.
  • the output 25a of the system 10 comprises a category weight CW(i) for each target category C(i).
  • Each of such weights 25 is a real number in the range from 0 to 1 and a higher value of the category weight means that resource d is more relevant to that specific target category.
  • the output of the overall classification system simply consists of a different weight associated to each category.
  • An example of the output is shown in Table 1 , where each category is given a weight by means of a real number, from 0.000 to 1.000, that estimates the relevance of the resource being classified with each possible target category.
  • the weights are such that 0.000 means “totally not relevant", 1 .000 means "with maximum relevance" and intermediate values represent intermediate degrees of relevance.
  • the adopted strategy implemented with the package (classification package) 20 is to process a plurality of atomic units of content (metadata) 31a, 31b, 3 In (Fig. 3) independently in a set of parallel "Estimator" blocks 33a, 33b, 33n, and to compute category weights 35a, 35b, 35n independently for each metadata (title or tags or description etc).
  • category weights 35a, 35b, 35n are combined by a final "Merge” module 37 so as to classify the multimedia content d into at least one category or to infer that the multimedia content is not relevant to any category.
  • Each of the parallel Estimator blocks (Estimators) 33a, 33b, 33n works in the same way, by analyzing one field of input resource (metadata field) 31a, 31b, 31n, and by producing a complete list of category weights for each metadata field.
  • Each Estimator 33a, 33b, 33n will give the measure of relevance stemming from the input information 31a, 31b, 3 In it received. If the various metadata fields 31a, 31b, 31n are consistent with each other, then it is likely that the estimated target category weights 35a, 35b, 35n would be in agreement, too. Otherwise, for badly assorted or inconsistent metadata, the category weights would be inconsistent as well.
  • the Merge module 37 measures the similarity among the category weights and outputs a final relevance weight 25 and category or relevance weights and corresponding categories, only in those cases where a sufficient agreement is found.
  • Each of the Estimators 33a, 33b, 33n respectively, has, preferably, the same external interface and does operate in the same way but on a different metadata field.
  • Estimator 33a (Fig. 4) is arranged for operating as follows:
  • Such a field is represented as a text string, interpreted as a set of words.
  • the Estimator extracts the individual words from the text string (by breaking it at separator characters, such as spaces, commas, etc);
  • the estimator 33a might also refuse to make a classification, whenever the characteristics of the input string are not sufficient for unambiguously identifying the relevant categories, as will be explained in the following; in such a case, no category weights are produced, but just an unknown token or an information corresponding to an unknown token.
  • the Estimator 33 a is arranged for using the list of target categories 23a, at least the semantic lexical network 23b and the associated set of mappings 23c of each category onto the lexical network (Category senses).
  • Estimator 33a uses also the set of additional pre-processing tables (Dictionary) 23 d.
  • the internal working of the Estimator 33a is based on a semantic representation of information.
  • the semantic lexical network 23b is used to represent the semantic senses related to the words of a certain resource description in the input string.
  • semantic lexical network WordNet is used, but, as easily comprehensible to a technician in the field, other semantic lexical networks might also be used without departing from the invention as claimed.
  • semantic lexical network lexical network or WordNet 23b consists of a long list of senses. Each sense is associated to a list of lemmas, i.e., words that might signify that sense in English or in a language used by the semantic lexical network; such association is called, as known, semantic mapping.
  • each lemma or word may be associated to multiple senses (polysemy, i.e., multiple meanings for a single word), and each sense may be associated to multiple lemmas (synonymy, i.e., multiple words with the same meaning).
  • senses are connected to each other by means of semantic (binary) relationships r.
  • semantic (binary) relationships r specify the various kinds of similarities, affinities, and connections among the various senses.
  • semantic relationships are: hypernym (a sense is more general than another) or its inverse hyponym (a sense is more specific), part meronym (a sense describes a subpart, or component, of the other sense) and its inverse part holonym (a sense describes the whole of another part), etc.
  • Each relationship r may be represented as a set of triples: Oi ⁇ r ⁇ Oj, where Oi and Oj are numerical offsets of two senses, and r is one of the semantic relationships supported by WordNet 23b.
  • Estimator block 33a comprises a plurality of semantic processes, 41 , 43, 45 and 63 that rely on a common representation of a given entity.
  • This common representation of the semantic information is encoded in a common form, i.e., a weighted set of relevant senses or SenseSet.
  • a SenseSet is a table listing a set of relevant senses, each mapped to a weight value (a real positive number) that represents the "importance" of that sense to influence the meaning of the entity.
  • a SenseSet SS is represented as an incomplete mapping (offset ⁇ weight):
  • mapping may be represented by means of a table as synthetically reported below in Table 2:
  • - Sense Offset is a constant value read from WordNet
  • the SenseSet may also be interpreted as a vector on a multi-dimensional space, whose components are the various sense offsets, and whose projections along each component are represented by the sense weights. In such interpretation, all Sense Offsets that do not appear in the SenseSet are implicitly assumed to have a null (zero) projection on the corresponding component. With this vector interpretation, the usual vector operations (such as sum, multiplication by a constant, scalar product, etc) may be trivially extended to operate on SenseSets, too.
  • Each semantic process inside the estimator blocks receives and generates in input and output, respectively, one or more SenseSet.
  • the Estimator block 33a elaborates the possible meanings of the input words by means of semantic processes, 41, 43, 45 63, and tries to determine the most relevant senses associated with the input and then relates them to the target categories.
  • the Estimator 33a proceeds in three subsequent steps:
  • Pre-Processor block 41 such a block is arranged to find the widest possible SenseSet SS F representing all the possible meanings of the input text.
  • Expander block 43 such a block is arranged to navigate the semantic network 23b to identify senses that are recurring in the input SenseSet SS F , and to isolate and delete senses that are marginal (i.e., possible interpretations of the text that are not supported by the context in which that text appears), and compute an updated "expanded" SenseSet SS E encoding this new (disambiguated) knowledge.
  • Matching block 45 such a block is arranged to compare the expanded SenseSet SS E computed for this input text to the SenseSets SS EC representing SenseSets corresponding to target categories C(i); the matching block 45 computes a similarity measure between SS E and each of the SS EC - Such computation comprises the category weights CW(i).
  • the Pre-Processor block (Pre-Processor) 41 provides the following architecture:
  • a resource metadata field 31 a comprising a set of words
  • the Pre-Processor is arranged to use the Word-Net semantic network (used for mapping lemmas to senses) 23b and, preferably, at least one of the other auxiliary inputs as for instance the user-generated set of pre-processing instructions (Dictionary) 23d or the SenseSets corresponding to each of the target categories 23c.
  • Word-Net semantic network used for mapping lemmas to senses
  • auxiliary inputs as for instance the user-generated set of pre-processing instructions (Dictionary) 23d or the SenseSets corresponding to each of the target categories 23c.
  • the goal of the Pre-Processor 41 is to extract the possible (semantic) interpretations from the input words.
  • the problems solved by the Pre-Processor are of dual nature:
  • the Pre-Processor 41 works on (unordered) word pairs, instead of single words.
  • individual words are encoded as the pair of a word with itself.
  • the input string "live madonna concert” is transformed to the following set of word pairs: "live live”, “live madonna”, “live concert”, “madonna madonna”, “madonna concert”, “concert concert”.
  • Each word pair is then, preferably, looked up into the Dictionary of pre-defined actions (Dictionary) 23d. According to the action stored in the Dictionary 23d for the current word pair, a new SenseSet may be generated, and used for computing the final output SenseSet SS F .
  • the list of possible Actions that may be specified in the Dictionary is shown in Table 3, where the effect of each action is described, and an Actionlnfo additional information is specified.
  • This word pair generates an output SenseSet weighted list of equal to the weighted sum of the SenseSet(s) Category IDs: list of of some specified target categories (C(i),AWC(i))
  • actions of the pre-processor block are divided in three groups:
  • the first action is used to discard frequently used but semantically useless words (often called “stopwords”), such as "the”, “for”, “of, etc.
  • Remapping is used to normalize a given word pair to an equivalent way expressing the same information (e.g., remapping the word “hiphop”, that appears as the word pair "hiphop hiphop", to the pair "hip hop”).
  • SenseSet(s) of the selected Category/ies are returned.
  • SenseSet corresponds to the meaning(s) that WordNet 23b already assigns to those words; in this case, the returned SenseSet is computed by looking up the two words in WordNet.
  • - Senses covers words whose meaning is not exactly matching a category, nor exactly corresponding to the WordNet senses; in this case, a custom SenseSet is stored in the dictionary, and is returned as a result.
  • the different types of actions of the Pre-Processor block are of different semantic relevance, since they represent different degrees of probability that the "right" meaning of the pair of words has been identified, due to the different quality of source information. Therefore, the four types of actions that may return one or more SenseSets (Category, Semantic, Senses, Default) are associated to an Action Weight (AW) used to influence the contribution of the considered action onto the final returned SenseSet SS F .
  • AW Action Weight
  • Action Weight may be a table of coefficients as listed in the following Table
  • the final SenseSet S S F returned by the Pre-Processor block is computed, for instance according to the expression (2), as the weighted sum of all SenseSets returned by the Actions triggered by the Dictionary or by applying the default action, where weights, as listed in Table 4, depend on the type of the i-th Action:
  • one or more SenseSets are returned, depending on how many categories are listed for the word pair in the Dictionary.
  • the SenseSet corresponding to the Category ID may be multiplied by an "Action- Category Weight” (ACW(j)) coefficient specified in the Dictionary for the Category C(j), for instance according to the expression (3).
  • ACW(j) Ad- Category Weight
  • ⁇ Category ⁇ Awc ⁇ j) . ⁇
  • a SenseSet is built by taking all the relevant senses in WordNet associated with the two words of the pair, and a constant conventional weight 1.0 is assigned to each sense in the SenseSet SSi (a sort of "diagonal SenseSet," since there is no additional information for ranking the relevance of various possible senses). Therefore these two action types behave in the same way, except that, later on, the resulting SenseSet will be multiplied by a different Action Weight AW according to (1) and (2).
  • a SenseSet SS F computed according to expression (2), is returned by the Pre-Processor and is given to the Sense Expander block 43.
  • the Sense Expander block (Expander block) 43 provides the following architecture:
  • the Expander block 43 uses the semantic lexical network 23b as auxiliary input, as for instance the WordNet semantic network, for navigating semantic relationships among senses.
  • the goal of the Sense Expander is to extract the most likely coherent interpretation of the input words, as interpreted by the Pre-Processor, and tackling the following problems:
  • the input SenseSet SS F gives a "wide" breadth of interpretation since it contains any sense that might be associated with the input words. This means that, if a word has multiple meanings, all these meanings will be present in SS F , even if only one of these meanings will be really relevant to the resource being classified. Therefore, senses that are out of the "dominant context" must be 'penalised'. In other words, interpretations that would be plausible if a word is taken out of context, but that become “isolated” in the current context defined by the other words, must be penalized.
  • SenseSet SS F Two or more words with extremely similar meanings might have generated a SenseSet SS F that does not contain identical senses, but only very similar senses. This is due to the very high number of senses in WordNet 23b and to the fact that even small nuances in meaning are represented as different sense offsets. It is, therefore, important to recognize which senses in SS F are "near" enough to be considered having essentially the same meaning, and 'strengthen' them, as well as including in the SenseSet SS F new senses that are strongly connected with existing relevant senses.
  • the Expander block 43 tries to determine the best interpretation, by navigating the semantic relationships defined in WordNet 23b, and by exploiting the following assumption:
  • the internal process in the Expander block 43 computes a new "expanded" SenseSet SS E in which the weights of the relevant senses are greatly increased, and the weights of the non relevant senses have a much lower value. To do this, the weight of each sense is recursively "propagated" to neighbouring senses, as explained below, by adding to each of these neighbours a fraction of the weight of the considered sense.
  • RW is chosen to favour the generalization of the current senses, in order to find (and weight more) general senses that are common ancestors to most senses in SS F .
  • RW is called "Generalizing Relationship Weights" or GRW and the process is illustrated through the following expression (4): (4)
  • GRW is a table of values as shown, for instance, in the following Table 5 wherein L means “low weight” and H means “higher weight” than L.
  • L may be in the range of 0.0 - 0.5.
  • L has a value of 0.2 and H a value of 0.5.
  • the expanded SenseSet SS E is determined by computing the fixed-point of an equation (5), that analyzes the semantic network as illustrated in Figure 6.
  • W E (O) is the weight of the sense with offset o in the SenseSet SS E ;
  • - o' is the offset of a sense in WordNet from which it is possible to reach the sense with offset o by traversing a relationship r in the selected subset.
  • the expression (5) discloses a recursive process arranged for stating the weight W E (o) that each sense with offset (o) should have in the expanded SenseSet SS E , by considering all selected WordNet relationships r pointing to the node corresponding to the sense with offset o.
  • o' is the offset of the node at the other end of the relationship and the weight W E (O') is summed up, weighted according to the weight assigned to the involved relationship RW(r).
  • Weights W F (O) of the initial SenseSet SS F are the external starting points of the recursive process and are added as an additional term in expression (5).
  • Equation (5) is computed by repeatedly computing all W E (O), and re-computing each of them when at least one of the related weights W E (O') changes. The process is therefore repeatedly executed until no further changes occur (thus reaching an exact solution), or by stopping the computation when an approximate solution is reached.
  • the approximate solution is defined by a predefined threshold ⁇ , and the process is stopped when all weight variations are below the threshold: 3 ⁇ 4 new (o) - 3 ⁇ 4 old (o) ⁇ ⁇ wherein the threshold T may be, for instance, in a range of 0.001 - 0.1, more preferably a value of O.01.
  • the resulting values W E (O) are used as components (o ⁇ W E (O)) of the resulting expanded SenseSet SS E , and are returned as the output of the Sense Expander block 43
  • the Matching block 45 is the last step of the Evaluator block 33a and comprises the process of measuring the similarity between the expanded SenseSet SS E and the predefined target categories.
  • Each category C(i) is semantically described by a suitable Category SenseSet SSe ⁇ (Category Senses 23c in Figure 4).
  • the Matching block 45 provides the following architecture:
  • the Category SenseSets SSe ⁇ are manually defined, by choosing in WordNet 23b the most relevant general terms that subsume the actual meaning of the category C(i).
  • the definition of such Category SenseSets is not repeated for each classified resource, but it needs to be defined only once in a configuration phase, when the system is personalized to a specific target domain and target application.
  • SenseSets are, in general more abstract and much more concise (only a handful of sense offsets listed) than SS E . This means that a direct comparison is, in general, not feasible, unless the Category SenseSets SSe ⁇ are processed with an Expansion process similar to the one already disclosed.
  • each Category SenseSet SSc ( i ) is first "expanded” through a Category Sense Expander block 63, yielding an "Expanded Category SenseSet” SSEQ ).
  • each SenseSet is a vector in a high-dimensional space (with as many dimensions as sense offsets in WordNet, i.e., over 100,000 dimensions), and by computing the "cosine of the angle" between the vector corresponding to SS E and each of the vectors corresponding to the various SS E C( I ).
  • Each Category C(i) is therefore assigned to a Category Weight CW(i) computed according to cosine-similarity. Computation is done, for instance, according to equation (7), where the symbol x stands for the scalar product of two vectors. The scalar product, and the modulus operator, are computed according to the vector interpretation of SenseSets.
  • the final category weights CW(i) are returned as the final result of the Matching block 45, and indirectly, as the result of the whole Estimator block 33 a.
  • system 10 is able to automatically extract the most reasonable category weights, starting from the information implicit in the input set of words.
  • the input words are too semantically dispersed, i.e., they are not related to each other in any way. This means that the input lacks internal coherence. No useful information can be derived in such a situation;
  • each Evaluator block, 33a, 33b, 33n is forced anyway to make a prediction about the relevance of categories, but such prediction is very likely to be wrong, due to the lack of meaningful and/or usable and/or coherent information.
  • the Evaluator blocks according to the second embodiment are alternative to blocks 33a,
  • Each of said alternative blocks further comprises a control or inhibition process or block 48, that monitors the progress of the processes within the Evaluator block, as for instance 133a, and estimates whether the computed classification would be correct.
  • inhibition block 48 suspects that the classification would be wrong, then such a process 48 "inhibits" the output of the matching block 45, so that no classification is made (which is better than a wrong classification).
  • the inhibition process or logic (inhibitor) 48 relies on purely statistical information about the inputs and outputs of the pre-processing block 41 , and the Expander block 43 as reported in Table 7 as reported below, and never considers the actual value of the input words nor of the various SenseSets.
  • the inhibitor block 48 works by comparing, from a statistical point of view, the statistical indicators of Table 7, with a model representing 'right' and 'wrong' predictions, trained on a sufficiently large set of manually classified documents.
  • the inhibitor block 48 is trained on a statistically significant sample of resources and is able to predict (with a margin of error) whether the identified Categories would be correct, or not.
  • the inhibitor block 48 receives in input the set of fields listed in Table 7, and produces in output a Boolean value: inhibit/don't inhibit.
  • Table 7 The classification process adopted here is based on a Support Vector Machine known per se in the field of Data Mining.
  • the Merge block 37 (Fig. 3) has the goal of comparing the information generated by the various Estimator blocks, for instance 33a or 133a, and of determining whether such information is consistent.
  • Estimator blocks for instance 33a or 133a (Fig. 3, Fig. 4, Fig. 5), agree on a common interpretation of the resource metadata (meaning that the Target Categories that receive a high ranking are nearly the same, and the Target Categories receiving a low or null ranking are nearly the same), then the Merge block 37 averages the relevant category weights and returns the overall result.
  • Processes used in the Merge block may be different.
  • the process defined by D. Wolfran and H.A. Olson in "A method for comparing large scale inter- indexer consistency using ir modelling" in Canadian Association for Information Science, Canadian Association for Information Science, 2007 may be used.
  • the category weights can be considered as type three classifiers as defined in the known document. It is also apparent that, in the case in which there is only one Evaluator block (because there is only one kind of metadata input, or all metadata fields have been concatenated into one string), the Merge block is not needed.
  • a resource d and an associated resource description 21 into a computer server 12, such server 12 accesses the classification server 20a and requests an on-line automatic classification of the resource d and of the associated resource description 21.
  • the classification server 20a by using the classification package 20 and auxiliary input 23 is arranged to automatically generate in output an unbiased classification associated to the evaluated resource.
  • resources may be searched by end users of the system 10 without any loss of time.
  • wrongly classified resources are avoided.
  • the input string considered by the Estimator block 33a is "Britney Spears - Baby One More Time Pop Music Video”.
  • the pre-processor block 41 identifies the following words: 1. Britney
  • Table 9 All the actions marked SEMANTICS, CATEGORY, DEFAULT in Table 9 generate a SenseSet to be added to the result S S F of the pre-processor block 41. Some of these SenseSets may be identical (in particular when multiple CATEGORY actions point to the same Category).
  • SS F is composed of 14 distinct SenseSets SSi.
  • the weighted sum of all SSi gives the SS F , that in this case is composed of 69 different sense offsets.
  • Table 10 shows a portion (less than half) of SS F , from which we may already appreciate that all the possible meanings of the words in the title are taken into account.
  • the Sense Expander computes the Expanded SenseSet SS E , that is much larger
  • This Expanded SenseSet SS E just computed by the Sense Expander Block 43 is then given as input to the Matching Block 45, that computes the cosine-similarity of SS E with all the Categories Expanded Senses SS EC O ) returned by the Category Sense Expander block 63.
  • Such similarities are the final Category Weights CW(i) returned by the Matching Block 45.
  • a subset resulting from these comparisons are shown in the following Table 12:
  • the Category Weight CW(i) is null (0.0), i.e., they are totally non-relevant categories. Other categories are assigned a higher or lower level, depending on the stronger or weaker similarity with the resource expanded SenseSet.
  • the final result of title classification process may therefore be presented as the "ranked" list of top-relevant categories for the resource analysed, as in the present example a video:

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Library & Information Science (AREA)
  • Data Mining & Analysis (AREA)
  • Databases & Information Systems (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Machine Translation (AREA)
  • Transition And Organic Metals Composition Catalysts For Addition Polymerization (AREA)

Abstract

La présente invention concerne un système de classification automatique de ressources, en particulier de ressources multimédia, comprenant au moins un système informatique (12, 20a) dans lequel sont stockées des ressources et des métadonnées associées auxdites ressources, lesdites métadonnées comprenant plusieurs éléments, chacun desdits éléments comprenant un champ de métadonnées différent. Le système comprend en outre au moins un ensemble de processus sémantiques conçus pour la gestion d'un champ de métadonnées et pour la production en sortie, par l'utilisation d'informations de constante en tant que référence qui comprennent au moins un réseau lexical sémantique, de plusieurs poids de catégorie représentant la classification du champ de métadonnées. La présente invention concerne également un procédé de classification automatique de ressources.
PCT/IB2010/054156 2009-09-16 2010-09-15 Système et procédé de classification de contenu WO2011033457A1 (fr)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
ITTO2009A000704 2009-09-16
IT000704A ITTO20090704A1 (it) 2009-09-16 2009-09-16 Sistema e metodo per la classificazione di contenuti

Publications (1)

Publication Number Publication Date
WO2011033457A1 true WO2011033457A1 (fr) 2011-03-24

Family

ID=42138938

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/IB2010/054156 WO2011033457A1 (fr) 2009-09-16 2010-09-15 Système et procédé de classification de contenu

Country Status (2)

Country Link
IT (1) ITTO20090704A1 (fr)
WO (1) WO2011033457A1 (fr)

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2014097000A1 (fr) * 2012-12-20 2014-06-26 Koninklijke Philips N.V. Système et procédé de recherche d'un article étiqueté principalement non textuel
CN117454892A (zh) * 2023-12-20 2024-01-26 深圳市智慧城市科技发展集团有限公司 元数据管理方法、装置、终端设备以及存储介质

Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6285999B1 (en) 1997-01-10 2001-09-04 The Board Of Trustees Of The Leland Stanford Junior University Method for node ranking in a linked database

Patent Citations (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6285999B1 (en) 1997-01-10 2001-09-04 The Board Of Trustees Of The Leland Stanford Junior University Method for node ranking in a linked database

Non-Patent Citations (8)

* Cited by examiner, † Cited by third party
Title
CHIN O S ET AL: "Automatic discovery of concepts from text", PROCEEDINGS OF THE IEEE/WIC/ACM INTERNATIONAL CONFERENCE ON WEB INTELLIGENCE IEEE COMPUT. SOC LOS ALAMITOS, CA, USA, 2006, pages 4 PP., XP002582242, ISBN: 0-7695-2747-7 *
D. WOLFRAN; H.A. OLSON: "A method for comparing large scale inter- indexer consistency using ir modelling", CANADIAN ASSOCIATION FOR INFORMATION SCIENCE, CANADIAN ASSOCIATION FOR INFORMATION SCIENCE, 2007
G.A. MILLER; R. BECKWITH; C. FELLBAUM; D. GROSS; K. J. MILLER: "Introduction to WordNet: an on-line lexical database", INT. J LEXICOGRAPHY, vol. 3, January 1990 (1990-01-01), pages 235 - 244
GHARIB T F ET AL: "Web document clustering approach using wordnet lexical categories and fuzzy clustering", COMPUTER AND INFORMATION TECHNOLOGY, 2008. ICCIT 2008. 11TH INTERNATIONAL CONFERENCE ON, IEEE, PISCATAWAY, NJ, USA, 24 December 2008 (2008-12-24), pages 48 - 55, XP031443129, ISBN: 978-1-4244-2135-0 *
L. XU; A. KRZYZAK; C.Y.SUEN: "Methods of Combining Multiple Classifiers and Their Applications to Handwriting Recognition", IEEE TRANSACTIONS ON SYSTEMS, MAN AND CYBERNETICS, vol. 22, no. 3, May 1992 (1992-05-01)
PANDYA A ET AL: "Text similarity measurement using concept representation of texts", PATTERN RECOGNITION AND MACHINE INTELLIGENCE. FIRST INTERNATIONAL CONFERENCE, PREMI 2005. PROCEEDINGS (LECTURE NOTES IN COMPUTER SCIENCE VOL.3776) SPRINGER-VERLAG BERLIN, GERMANY, 2005, pages 678 - 683, XP002582240, ISBN: 3-540-30506-8 *
WOLFRAM D ET AL.: "A Method for Comparing large Scale Inter-indexer Consistency Using IR Modeling", CANADIAN ASSOCIATION FOR INFORMATION SCIENCE, 2007, pages 1 - 8, XP002582270, Retrieved from the Internet <URL:http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.83.8621> [retrieved on 20100511] *
YING LIU ET AL: "Using WordNet to disambiguate word senses for text classification", COMPUTATIONAL SCIENCE-ICCS 2007. 7TH INTERNATIONAL CONFERENCE. PROCEEDINGS, PART III (LECTURE NOTES IN COMPUTER SCIENCE VOL.4489) SPRINGER BERLIN, GERMANY, 2007, pages 781 - 789, XP002582243, ISBN: 3-540-72587-3 *

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
WO2014097000A1 (fr) * 2012-12-20 2014-06-26 Koninklijke Philips N.V. Système et procédé de recherche d'un article étiqueté principalement non textuel
CN104854588A (zh) * 2012-12-20 2015-08-19 皇家飞利浦有限公司 用于搜索标记的主要为非文本的项目的系统和方法
JP2016509703A (ja) * 2012-12-20 2016-03-31 コーニンクレッカ フィリップス エヌ ヴェＫｏｎｉｎｋｌｉｊｋｅ Ｐｈｉｌｉｐｓ Ｎ．Ｖ． ラベル付けされた主に非テキストのアイテムを検索するためのシステム及び方法
US9940382B2 (en) 2012-12-20 2018-04-10 Koninklijke Philips N.V. System and method for searching a labeled predominantly non-textual item
CN117454892A (zh) * 2023-12-20 2024-01-26 深圳市智慧城市科技发展集团有限公司 元数据管理方法、装置、终端设备以及存储介质
CN117454892B (zh) * 2023-12-20 2024-04-02 深圳市智慧城市科技发展集团有限公司 元数据管理方法、装置、终端设备以及存储介质

Also Published As

Publication number Publication date
ITTO20090704A1 (it) 2011-03-17

Similar Documents

Publication Publication Date Title
US9171081B2 (en) Entity augmentation service from latent relational data
US8219593B2 (en) System and method for measuring the quality of document sets
Hoffart et al. Discovering emerging entities with ambiguous names
US9589208B2 (en) Retrieval of similar images to a query image
JP5727512B2 (ja) 検索提案のクラスタ化及び提示
US8090724B1 (en) Document analysis and multi-word term detector
US20110231347A1 (en) Named Entity Recognition in Query
CN110390044B (zh) 一种相似网络页面的搜索方法及设备
EP2307951A1 (fr) Procédé et appareil pour associer des ensembles de données à l aide de vecteurs sémantiques et d&#39;analyses de mots-clés
US20140101162A1 (en) Method and system for recommending semantic annotations
Olieman et al. Entity linking by focusing DBpedia candidate entities
JP5146108B2 (ja) 文書重要度算出システム、文書重要度算出方法およびプログラム
WO2011033457A1 (fr) Système et procédé de classification de contenu
Panagiotou et al. A general framework for first story detection utilizing entities and their relations
Zhang et al. Topic level disambiguation for weak queries
KR20100080100A (ko) 정보 검색 방법 및 이를 수행할 수 있는 프로그램이 수록된컴퓨터로 읽을 수 있는 기록 매체
Yengui et al. Towards a Statistical Approach to the Analysis, the Indexing, and the Semantic Search of Medical Videoconferences
CN115186065A (zh) 一种目标词语的检索方法及装置
Jeevananthan et al. Structured Data Extraction for Multiple Web Data Region Using Bayes Classifier

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 10760087

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 10760087

Country of ref document: EP

Kind code of ref document: A1