WO2007061451A1 - A knowledge correlation search engine - Google Patents

A knowledge correlation search engine Download PDF

Info

Publication number
WO2007061451A1
WO2007061451A1 PCT/US2006/025101 US2006025101W WO2007061451A1 WO 2007061451 A1 WO2007061451 A1 WO 2007061451A1 US 2006025101 W US2006025101 W US 2006025101W WO 2007061451 A1 WO2007061451 A1 WO 2007061451A1
Authority
WO
WIPO (PCT)
Prior art keywords
search engine
function
search
serial
subject
Prior art date
Application number
PCT/US2006/025101
Other languages
English (en)
French (fr)
Inventor
Mark Bobick
Carl Wimmer
Original Assignee
Make Sence, Inc.
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
Priority claimed from US11/273,568 external-priority patent/US8108389B2/en
Priority claimed from US11/314,835 external-priority patent/US8126890B2/en
Application filed by Make Sence, Inc. filed Critical Make Sence, Inc.
Priority to JP2008541146A priority Critical patent/JP4864095B2/ja
Priority to CN200680042357XA priority patent/CN101310274B/zh
Priority to EP06774152A priority patent/EP1974292A4/en
Publication of WO2007061451A1 publication Critical patent/WO2007061451A1/en

Links

Classifications

    • GPHYSICS
    • G06COMPUTING; CALCULATING OR COUNTING
    • G06FELECTRIC DIGITAL DATA PROCESSING
    • G06F16/00Information retrieval; Database structures therefor; File system structures therefor
    • G06F16/90Details of database functions independent of the retrieved data types
    • G06F16/95Retrieval from the web
    • G06F16/951Indexing; Web crawling techniques

Definitions

  • the invention is related to information technology and, more particularly, to a search engine that utilizes the results of knowledge correlation to identify network and/or Internet resources significant to any given user question, subject, or topic of a digital information object.
  • IR methods are directed to locating resources (typically documents) that are relevant to a question called a query. That query can take forms ranging from a single search term to a complex sentence composed in a natural language such as English.
  • the collection of potential resources that are searched is called a corpus (body), and different techniques have been developed to search each type of corpus. For example, techniques used to search the set of articles contained in a digitized encyclopedia differ from the techniques used by a web search engine.
  • the core issue in IR is relevance - that is, the relevance of the documents retrieved to the original query.
  • Formal metrics are applied to compare the effectiveness of the various IR methods.
  • Common IR effectiveness metrics include precision, which is the proportion of relevant documents retrieved to all retrieved documents; recall, which is the proportion of relevant documents retrieved to all relevant documents in the corpus; and fall-out, which is the proportion of irrelevant documents retrieved to all irrelevant documents in the corpus.
  • Post retrieval documents deemed relevant are (in most IR systems) assigned a relevance rank, again using a variety of techniques, and results are returned. Although most commonly the query is submitted by — and the results returned to - a human being called a user, the user can be another software process.
  • Text retrieval is a type of IR that is typically concerned with locating relevant documents which are composed of text, and document retrieval is concerned with locating specific fragments of text documents, particularly those documents composed of unstructured (or "free") text.
  • the related knowledge domain of data retrieval differs from IR in that data retrieval is concerned with rapid, accurate retrieval of specific data items, such as records from a SQL database.
  • Information extraction is another type of IR which is has the purpose of automatic extraction of information from unstructured (usually text) documents into data structures such as a template of name/value pairs. From such templates, the information can subsequently correctly update or be inserted into a relational database.
  • search engines that have been described in the literature or released as software products use a number of forms of input, ranging from individual keywords, to phrases, sentences, paragraphs, concepts and data objects.
  • keyword, sentence, and paragraph conform to the common understanding of the terms
  • phrase, concept, and data object varies by implementation.
  • word phrase is defined using its traditional meaning in grammar.
  • types of phrases include Prepositional Phrases (PP), Noun Phrases (NP), Verb Phrases (VP), Adjective Phrases, and Adverbial Phrases.
  • PP Prepositional Phrases
  • NP Noun Phrases
  • VP Verb Phrases
  • Adjective Phrases Adjective Phrases
  • Adverbial Phrases Adjective Phrases
  • the wo ⁇ dphrase may be defined as any proper name (for example "New York City").
  • phrase Most definitions require that a phrase contain multiple words, although at least one definition permits even a single word to be considered a phrase.
  • Some search engine implementations utilize a lexicon (a pre-canned list) of phrases.
  • the WordNet Lexical Database is a common source of phrases.
  • the word concept When used in conjunction with search engines, the word concept generally refers to one of two constructs.
  • the first construct is concept as a cluster of related words, similar to a thesaurus, associated with a keyword. In a number of implementations, this cluster is made available to a user - via a Graphic User Interface (GUI) for correction and customization. The user can tailor the cluster of words until the resulting concept is most representative of the user's understanding and intent.
  • the second construct is concept as a localized semantic net of related words around a keyword. Here, a local or public ontology and taxonomy is consulted to create a semantic net around the keyword.
  • Some implementations of concept include images and other non-text elements.
  • Topics in general practice need to be identified or "detected" from a applying a specific set of operations against a body of text.
  • Different methodologies for identification and/or detection of topics have been described in the literature.
  • Use of a topic as input to a search engine therefore usually means that a body of text is input, and a required topic identification or topic detection function is invoked.
  • an appropriate relevancy function can then be invoked by the search engine.
  • Data objects as input to a search engine can take forms including a varying length set of free form sentences, to full-length text documents, to meta-data documents such as XML documents.
  • the Object Oriented (00) paradigm dictates that OO systems accept objects as inputs. Some software function is almost always required to process the input object so that the subsequent relevance function of the search engine can proceed.
  • Ranked result sets have been the key to marketplace success for search engines.
  • the current dominance of the Google search engine (a product of Google, Inc.) is due to far more to the PageRank system used in Google that lets (essentially) the popularity of a given document dictate result rank.
  • Popularity in the Google example applies to the number of links and to the preferences of Google users who input any given search term or phrase. These rankings permit Google to optimize searches by returning only those documents with ranks above a certain threshold (called k).
  • Other methods used by web search engines to rank results include "Hubs & authorities" which counts links into and out of a given web page or document, Markov chains, and random walks.
  • the present invention discloses a new and novel form of search engine which utilizes a computer implemented method to identify at least one resource, referenced by that resource's unique URI (Uniform Resource Identifier) or referenced by that resource's URL (Uniform Resource Locator), such resource being significant to any given user question, subject, or topic of a digital information object.
  • the user question or subject or topic acts as input.
  • the input is utilized by a software function which attempts to construct or discover logical structures within a collection of data objects, each data object being associated with the resource that contributed the data object, and the constructed or discovered logical structures being strongly associated with the input.
  • that software function is a knowledge correlation function as described in said Serial No.
  • the logical structure is a form of directed acyclic graph termed a quiver of paths. If such logical structures strongly associated with the input are in fact constructed or discovered, the data object members of such logical structures become an answer space. Using the answer space, another software function is then able to determine with a high degree of confidence which of the resources that contributed to the answer space are the most significant contributors to the answer space, and thereby identify URLs and URIs most significant to the input question, subject or topic. Finally, a software function is used to rank in significance to the input each of the URL and URI referenced resources that contributed data objects to the answer space.
  • the present invention differs from existing search engines because the Knowledge Correlation process as described in said Serial No. 11/273,568, which is used in this invention, attempts to construct an exhaustive collection of paths describing all connections - called correlations - between one term, phrase, or concept referred to as X (or "origin") and a minimum of a second term, phrase or concept referred to as Y (or "destination"). If one or more such correlations can in fact be constructed, the present invention identifies as relevant all resources which contributed to constructing the correlation(s).
  • relevancy in the present invention applies not to individual terms, phrases or concepts in isolation but instead to the answer space of correlations that includes not only the X and the Y, but to all the terms, phrases and concepts encountered in constructing the correlations. Because of these novel characteristics, the present invention is uniquely capable of satisfying user queries for which cannot be answered using the content of a single web page or document.
  • Input to the present invention differs from current uses because all input modes of the present invention must present a minimum of two (2) non-identical terms, phrases, or concepts.
  • "Non-identical” in this usage means lexical or semantic overlap or disjunction is required.
  • the minimum two terms, phrases, or concepts are referred to as X and Y (or "origin” and "destination”).
  • No input process can result in synonymy, identity, or idempotent X and Y term, phrases or concepts.
  • text objects and data objects can be accepted (in the present invention, as either X or Y) and the topics and/or concepts can be extracted prior to submission to the Knowledge Correlation process.
  • the form of the input (term, phrase, concept, or object) is not constrained in the present invention. This is possible because the relevancy function (Knowledge Correlation) does not utilize similarity measures to establish relevancy. This characteristic will allow the present invention to be seamlessly integrated with many existing IR applications.
  • the purpose of Knowledge Correlation in the present invention is to establish document relevancy.
  • relevancy is established in IR using three general approaches: set-theoretic models which represent documents by sets; algebraic models which represent documents as vectors or matrices; and probabilistic models which use probabilistic theorems to learn document attributes (such as topic).
  • Each model provides a means of determining if one or more documents are similar and thereby, relevant, to a given input.
  • the most basic set-theoretic model uses the standard Boolean approach to relevancy — does an input word appear in the document? If yes, the document is relevant. If no, then the document is not relevant.
  • Algebraic models utilize techniques such as vector space models where documents represented as vectors of terms are compared to the input query represented as a vector of terms. Similarity of the vectors implies relevancy of the documents. For probabilistic models, relevancy is determined by the compared probabilities of input and document.
  • the present invention establishes relevancy by an entirely different process, using an entirely different criteria than any existing search engine.
  • the present invention is dependent upon Discovery and Acquisition of "relevant" sources within the corpus (especially if that corpus is the WWW).
  • any form of the existing art can be utilized without restriction during the Discovery phase as described in said Serial No. 11/273,568 to assist in identifying candidate resources for input to the Knowledge Correlation process.
  • search engines simply determining relevancy of a given document to a given input is necessary but not sufficient. After all - using the standard Boolean approach to relevancy as an example - for any query against the WWW, which contained the word "computer”, tens of millions of documents would qualify as relevant. If the user was actually interested only in documents describing a specific application of "computer”, such a large result set would prove unusable.
  • users require that search engines rank their results from most relevant to least relevant. Typically, users prefer to have the relevant documents presented in order of decreasing relevance — with the most relevant result first. Because most relevance functions produce real number values, a natural way to rank any search engine result set is to rank the members of the result set by their respective relevance scores.
  • the present invention utilizes a ranking method that is novel because it is a function of the degree to which a given document or resource contributed to the correlation "answer space".
  • that answer space is constructed from data structures called nodes, which in turn are created by decomposition of relevant resources.
  • Even the most na ⁇ ve ranking function of the present invention which counts the frequency of node occurrence in the answer space — can identify documents that uniquely or strongly relevant to the original user query. More sophisticated ranking mechanisms of the current invention as described more hereinafter improve that outcome.
  • Figure 1 is a block diagram showing functional components of a search engine in accordance with one aspect of the invention.
  • Figure 2 is a clock diagram of the pre-search block of Figure 1.
  • Figure 2 A is a block diagram of part of an exemplary subject evaluation function for keywords, phrases, sentences and concepts in accordance with one aspect of the invention.
  • Figure 2B is a block diagram of the remaining part of an exemplary subject evaluation function for compound, complex or orthogonal subjects and for a simple web query in accordance with one aspect of the invention.
  • FIG. 2C is a block diagram of an exemplary topic detection module and related adapter(s) in accordance with one aspect of the invention.
  • Figure 2D is a block diagram of a question generation function in accordance with one aspect of the invention.
  • Figure 3 is a copy of Figure IA of Serial No. 11/273,568.
  • Figure 4 is a copy of Figure IB of Serial No. 11/273,568.
  • Figure 5 is a copy of Figure 1C of Serial No. 11/273,568.
  • Figure 6 is a copy of Figure 2A of Serial No. 11/273,568.
  • Figure 7 is a copy of Figure 2E of Serial No. 11/273,568.
  • Figure 8 is a block diagram of the post search block 120 of Figure 1.
  • Figure 1 is a block diagram of three examples of input accepted by the correlation function 110.
  • a subject 200 is evaluated by the subject evaluation function 220.
  • a digital information object 230 is examined for a topic by an adapter 235 of the topic detection module 240.
  • a canonical form question generation function 250 generates a question 260 as input.
  • a minimum of two inputs in any form and from any source as described more hereinafter must be submitted to the correlation function 110.
  • acceptable inputs may include any combination of two subjects 200, digital information objects 230, or questions 260.
  • a minimum of one X input which is a subject 200, digital information object 230 or question 260 is submitted to the correlation function 110.
  • a second input called a stop correlation condition, is passed to the correlation function 110.
  • the actual value of the required Y which is a correlation destination remains unknown until the stop correlation function is satisfied by the correlation function 110. No actual Y input need be processed as input, but the requirement for a correlation destination is satisfied.
  • the subject 200 may be an individual keyword, a phrase, a sentence, or a concept.
  • the subject 200 is passed by the subject evaluation function 220 directly to the correlation function 110 without further processing.
  • the subject 200 is passed by the subject evaluation function 220 directly to the correlation function 110 without further processing.
  • a natural language parser (NLP) 133 will be invoked to perform a syntactic analysis of the sentence to extract the actual subject 200 of the sentence in the form of words and/or phrases. Such words or phrases will then be passed to the correlation function 110.
  • NLP natural language parser
  • Additional words or phrases may be extracted from the sentence and submitted to the correlation function 110 as context.
  • any number of context words or phrases which are in addition to the X or Y words or phrases can be submitted to the correlation function 110 to improve said function.
  • the selection of what words or phrases (if any) that are to be extracted from a sentence is based upon the membership of the word or phrase in any lexicon of the NLP 133, and the absence of the word from a common list of stop words. Stop words are well known in IR. Such words cannot be used to establish relevance in set-theoretic models of IR, so are never added to the indexes built for such models.
  • the subject evaluation function 220 will extract from the sentence both the X and Y words or phrases and submit them to the correlation function 110.
  • the concept word or phrase will be submitted to the correlation function 110 as either X or Y, and the remaining terms in the concept cluster or map will be submitted to the correlation function 110 as context words or phrases.
  • the subject will be provided by a user by means of Graphical User Interface such as Serial No. 11/273,568 Figure 2A.
  • Graphical User Interface such as Serial No. 11/273,568 Figure 2A.
  • any well known input interface will be utilized (e.g. text input field, spoken input, etc.).
  • the subject 200 shall take the form of a complex subject, that is, a subject that consists of one independent clause, and one or more dependent clauses.
  • a complex subject that is, a subject that consists of one independent clause, and one or more dependent clauses.
  • the subject 200 shall take the form of a compound subject, that is, a subject that consists of two or more independent clauses connected using logical operators such as "and" "or” "not".
  • the subject 200 shall take the form of a multi-part orthogonal subject, that is, a subject that consists of two or more independent clauses which are not connected, and which may be orthogonal with respect to each other.
  • the subject 200 shall take the form of a multipart orthogonal subject, that is, a subject that consists of two or more independent clauses which are not connected, and which may be orthogonal with respect to each other.
  • a multipart orthogonal subject that is, a subject that consists of two or more independent clauses which are not connected, and which may be orthogonal with respect to each other.
  • advanced NLP methods for clause recognition see Hachey, B.C. 2002. Thesis: Recognising Clauses Using Symbolic and Machine Learning Approaches. University of Edinburgh) will be applied to the subject 200 to first decompose the subject 200 into clauses and from there, by means of syntactic analysis, into keywords and phrases.
  • Clause recognition techniques will be used to discriminate between X, Y, and context inputs to the correlation function 110.
  • the subject evaluation function 220 will determine if the user- provided subject 200 would produce as response from the present invention a listing as the most appropriate response. For example, referring to FIGURE IB, is the user- provided subject is "Italian restaurants Dover DE", the subject evaluation function 220 will recognize that a listing of Italian restaurants in Dover, Delaware is sought. In this event, the subject evaluation function 220 will either direct the user to use one of the well known simple web search engines such as Google (a product of Google, Inc.) or Yahoo (a product of Yahoo, Inc.), or will directly invoke one of those simple search engines. Alternatively, the subject evaluation function 220 will determine if the user-provided subject would produce as response a single web page as the most appropriate response.
  • Google a product of Google, Inc.
  • Yahoo a product of Yahoo, Inc.
  • the subject evaluation function 220 will recognize that the web site for the Rialto Theatre is sought. In this event, the subject evaluation function 220 will either direct the user to use one of the well known simple web search engines such as Google or Yahoo, or will directly invoke the web site of Rialto Theatre, or will directly invoke one of the simple search engines named above. This is achieved by an automatic phrase recognition techniques (see Kelledy, F., Smeaton, A.F. 1997. Automatic Phrase Recognition and Extraction from Text. Proceedings of the 19 th Annual BCS-IRSG Colloquium on IR Research) using the rule that when precisely two perfect phrases comprise the subject 220 and one of the phrases is a proper geographical name (e.g.
  • the third input mode illustrated in Figure 1 is more fully illustrated in Figure 2A of patent application Serial No. 11/273,568 wherein the input to the correlation function 110 is a user question, and the user question shall be composed of an incomplete question in canonical form and, in addition, one or more key words, wherein the key words complete the question [comparable to the well known paradigm of "fill in the blanks"].
  • the incomplete question will be explicitly selected by the user.
  • the incomplete question will be explicitly selected by the user from a list or menu of supported canonical form questions.
  • the list or menu of incomplete supported canonical form questions will be "static" - that is, the list will not vary at each invocation.
  • the list or menu of incomplete supported canonical form questions will be "dynamic" - that is, the list varies at each invocation.
  • the dynamic list or menu of incomplete supported canonical form questions will be generated at each invocation by means of a software function, the canonical form question generation function 250, a software program component, written in a computer programming language (e.g. Java, a product of Sun Microsystems, Inc.).
  • the incomplete question will be implicit, the question being selected by a software program component, the canonical form question generation function 250.
  • the incomplete implicit question that will be selected by the canonical form question generation function 250 will be "static" — that is, it will not vary at each invocation.
  • the static implicit selected question is "What are the connections between [keyword 1] and [keyword 2]?”
  • the static implicit selected question is "What are the connections between [keyword 1] and [keyword 2] in the context of [keyword 3] and/or [keyword 4] and/or [keyword 5]?”
  • the incomplete implicit question that will be selected by the canonical form question generation function 250 will be "dynamic” — that is, it will vary at each invocation.
  • the digital information object 230 will be provided by a user.
  • the digital information object 230 will include, but not be limited to the forms: (i) text (plain text) files, (ii) Rich Text Format (RTF) (a standard developed by Microsoft, Inc.).
  • RTF Rich Text Format
  • An alternative method is to first obtain clean text from RTF by the intermediate use of a RTF-to-text conversion utility (e.g. RTF-Parser-1.09, a product of Pete Sergeant), (iii) Extended Markup Language (XML) (a project of the World Wide Web
  • RTF-to-text conversion utility e.g. RTF-Parser-1.09, a product of Pete Sergeant
  • XML Extended Markup Language
  • any dialect of markup language files including, but not limited to: HyperText Markup Language (HTML) and Extensible HyperText Markup Language (XHTMLTM) (projects of the World Wide Web Consortium), RuIeML (a project of the RuIeML Initiative), Standard Generalized Markup Language (SGML) (an international standard), and Extensible Stylesheet Language (XSL) (a project of the World Wide Web Consortium).
  • HTML HyperText Markup Language
  • XHTMLTM Extensible HyperText Markup Language
  • RuIeML a project of the RuIeML Initiative
  • Standard Generalized Markup Language SGML
  • XSL Extensible Stylesheet Language
  • PDF Portable Document Format
  • MS WORD files e.g. DOC files used to store documents by MS WORD (a word processing software product of Microsoft, Inc.
  • This embodiment programmatically utilizes a MS Word-to-text parser (e.g. the Apache POI project, a product ofApache.org).
  • the POI project API also permits programmatically invoked text extraction from Microsoft Excel spreadsheet files (XLS).
  • An MS Word file can also be processed by a NLP as a plain text file containing special characters, although XLS files cannot, (vii) event-information capture log files, including, but not limited to: transaction logs, telephone call records, employee timesheets, and computer system event logs.
  • the topic of the digital information object 230 will be determined by a software function, the topic detection function 240, a software program component. Examples of such topic detection software have been well described in the literature (see Chen, K. 1995. Topic Identification in Discourse. Morgan Kaufman).
  • the topic detection function 240 will be implemented with software adapters 235 that handle each form of digital information object 230. Such software adapters 235 are well known (for an example, see http://www-306.ibm.com/software/integration/wbiadapters/framework ).
  • the output of the topic detection function will be keywords and/or phrases which will then be submitted to the correlation function 110.
  • Figure 8 is a flow chart of the search engine process initiated by the knowledge correlation function 110 upon inputs as described in Figure 1, and continuing through to presentation of results to a user in accordance with one aspect of the invention.
  • the correlation function 110 places relevant data structure objects 830, triples 835 and associated objects 837 into an answer space 885.
  • the significance of the objects in the answer space 885 is determined by a significance calculation function 840 which sets up data for the ranking function 845 to rank by significance.
  • Output is then displayed to the user.
  • the correlation function creates any kind of directed acyclic graph
  • the graph can be displayed to the user after being organized for layout by the hierarchical layout function 850.
  • Figure IA a user enters at least one term via using a GUI interface.
  • Figure 2 A is a screen capture of the GUI component intended to accept user input. Significant fields in the interface are “X Term", “Y Term” and “Tangents”. As described more hereinafter, the user's entry of between one and five terms or phrases has a significant effect on the behavior of the present invention. In a preferred embodiment as shown in Figure 2 A, the user is required to provide at least two input terms or phrases. Referring to FIGURE IA, the user input 100, "GOLD" is captured as a searchable term or phrase 110, by being entered into the "X Term" data entry field of FIGURE 2 A.
  • the user input 100 "INFLATION" is captured as a searchable term or phrase 110 by being entered into the "Y Term " data entry field of FIGURE 2A.
  • a search 120 is undertaken to identify actual and potential sources for information about the term or phrase of interest. Each actual and potential source is tested for relevancy 125 to the term or phrase of interest.
  • sources searched are computer file systems, the Internet, Relational Databases, email repositories, instances of taxonomy, and instances of ontology. Those sources found relevant are called resources 128.
  • the search 120 for relevant resources 128 is called "Discovery”.
  • the information from each resource 128 is decomposed 130 into digital information objects 138 called nodes.
  • nodes 180A and 180B are data structures which contain and convey meaning. Each node is self contained. A node requires nothing else to convey meaning.
  • nodes 180A, 180B from resources 128 that are successfully decomposed 130 are placed into a node pool 140.
  • the node pool 140 is a logical structure for data access and retrieval. The capture and decomposition of resources 128 into nodes 180A, 180B is called "Acquisition ".
  • a correlation 155 is then constructed using the nodes 180A, 180B in the node pool 140, called member nodes. Referring to FIGURE IB, the correlation is started from one of the nodes in the node pool that explicitly contains the term or phrase of interest.
  • Such a node is called a term- node.
  • the term-node When used as the first node in a correlation, the term-node is called the origin 152 (source).
  • the correlation is constructed in the form of a chain (path) of nodes.
  • the path begins at the origin node 152 (synonymously referred to as path root).
  • the path is extended by searching among node members 151 of the node pool 140 for a member node 151 that can be associated with the origin node 152. If such a node (qualified member 151H) is found, that qualified member node is chained to the origin node 152, and designated as the current terminus of the path.
  • the path is further extended by means of the iterative association with and successive chaining of qualified member nodes of the node pool to the successively designated current terminus of the path until the qualified member node associated with and added to the current terminus of the path is deemed the final terminus node (destination node 159), or until there are no further qualified member nodes in the node pool.
  • the association and chaining of the destination node 159 as the final terminus of the path is called a success outcome (goal state), in which case the path is thereafter referred to as a correlation 155, and such correlation 155 is preserved.
  • a completed correlation 155 associates the origin node 152 with each of the other nodes in the correlation, and in particular with the destination node 159 of the correlation.
  • the name for this process is "Correlation ".
  • the correlation 155 thereby forms a knowledge bridge that spans and ties together information from all sources identified in the search. The knowledge bridge is discovered knowledge. "
  • the relevancy tests (Serial No. 11/278,568 Figure IA Item 125) applied to potential sources are of interest.
  • the Discovery phase of the correlation function 110 as described above utilizes relevancy tests (Serial No. 11/278,568 Figure IA Item 125) to identify resources (Serial No. 11/278,568 Figure IA Item 128) for subsequent Acquisition.
  • These relevancy tests (Serial No. 11/278,568 Figure IA Item 125) are analogous to and intersect with the relevancy approaches described in the Related Art. Note that the fact that a resource (Serial No.
  • 11/278,568 Figure IA Item 128) is deemed sufficiently relevant to warrant Acquisition by the correlation function 110 does not imply or guarantee that the resource (Serial No. 11/278,568 Figure IA Item 128) will be found to contribute in a significant way to the answer space 800.
  • the relevancy tests (Serial No. 11/278,568 Figure IA Item 125) enumerated in Serial No. 11/278,568 are listed following. In one embodiment of the present invention, all the relevancy tests (Serial No. 11/278,568 Figure IA Item 125) enumerated in Serial No. 11/278,568 as well as all relevancy approaches described in the Related Art will be utilized to select resources (Serial No. 11/278,568 Figure IA Item 128) for Acquisition by the correlation function 110.
  • Tests for relevancy in accordance with Serial No. 11/278,568 can include, but are not limited to:
  • (xi) use of a taxonomy to determine that a term contained in the potential source has a parent, child or sibling relation to the term or phrase of interest.
  • the vertex containing the term or phrase of interest is located in the taxonomy. This is the vertex of interest.
  • the parent, siblings and children vertices of the taxonomy are searched by tracing the relations (links) from the vertex of interest to parent, sibling, and children vertices of the vertex of interest. If any of the parent, sibling or children vertices contain the wordfi'om the content of the potential source, a match is declared, and the source is considered an actual source of information about the term or phrase of interest.
  • a software function called a graph traversal function, is used to locate and examine the parent, sibling, and child vertices of term or phrase of interest.
  • (xiv) use of an ontology to determine that a degree (length) one semantic distance separates the source from the term or phrase of interest.
  • the vertex containing the term or phrase of interest is located in the ontology. This is the vertex of interest.
  • the ontology is searched by tracing the relations (links) from the vertex of interest to all adjacent vertices. If any of the adjacent vertices contain the word from the content of the potential source, a match is declared, and the source is considered an actual source of information about the term or phrase of interest, (xv) uses an ontology to determine that a degree (length) two semantic distance separates the source from the term or phrase of interest.
  • the vertex containing the term or phrase of interest is located in the ontology. This is the vertex of interest.
  • the relevancy test for semantic degree one is performed for each word located in the contents of the potential source. If this fails, the ontology is searched by tracing the relations (links) from the vertices adjacent to the vertex of interest to all respective adjacent vertices.
  • Such vertices are semantic degree two from the vertex of interest. If any of the semantic degree two vertices contain the word from the content of the potential source, a match is declared, and the source is considered an actual source of information about the term or phrase of interest.
  • (xvi) uses a universal ontology such as the CYC Ontology (a product ofCycorp, Inc) to determine the degree (length) of semantic distance from one of the terms and/or phrases of interest to any content of a potential source located during a search
  • (xvii) uses a specialized ontology such as the Gene Ontology (a project of the Gene
  • Ontology Consortium to determine the degree (length) of semantic distance from one of the terms and/or phrases of interest to any content of a potential source located during a search. uses an ontology and for the test, the ontology is accessed and navigated using an Ontology Language (e.g. Web Ontology Language) (OWL) (a project of the World Wide Web Consortium).
  • OWL Web Ontology Language
  • Computers that may be searched in this way include individual personal computers, individual computers on a network, network server computers, network ontology server computers, network taxonomy server computers, network database server computers, network email server computers, network file server computers.
  • Network ontology servers are special typically high performance computers which are dedicated to the task of supporting semantic search functions for a large group of users.
  • Network taxonomy servers are special typically high performance computers which are dedicated to the task of supporting taxonomic search functions for a large group of users.
  • Network database servers are special typically high performance computers which are dedicated to the task of supporting database functions for a large group of users.
  • Network email servers are special typically high performance computers which are dedicated to the task of supporting email functions for a large group of users.
  • Network file servers are special typically high performance computers which are dedicated to the task of supporting file persistence and retrieval functions for a large group of users.
  • the computer network has a minimum of two network nodes and the maximum number of network nodes is infinity.
  • the computer file system has a minimum of two files and the maximum number of files is infinity.
  • an answer space 800 Upon successful completion of the correlation function 110, an answer space 800 will exist. As described in said Serial No. 11/273,568, and illustrated in Figure 8 of this application, the answer space 800 is composed of correlations (Serial No. 11/278,568 Figure IB Item 155).
  • the correlations (Serial No. 11/278,568 Figure IB Item 155) are in turn composed of nodes Figure 5 (Serial No. 11/278,568 Figure 1C Items 180A and 180B).
  • the successful correlations Figure 4 (Serial No. 11/278,568 Figure IB Item 155) produced by the correlation function 110 are together modeled as a directed graph (also called a digraph) of correlations in one preferred embodiment.
  • the successful correlations Figure 4 (Serial No.
  • 11/278,568 Figure IB Item 155) produced by the correlation function 110 are together modeled as a quiver of paths of successful correlations.
  • Successful correlations Figure 4 (Serial No. 11/278,568 Figure IB Item 155) produced by the correlation function 110 are together called, with respect to correlation, the answer space 800.
  • the correlation function 110 constructs a quiver of paths where each path in the quiver of paths is a successful correlation, all successful correlations share as a starting point the origin node (Serial No. 11/278,568 Figure IB Item 152), and all possible correlations (Serial No. 11/278,568 Figure IB Item 155) from the origin node (Serial No. 11/278,568 Figure IB Item 152) are constructed.
  • the answer space 800 is stored in a computer digital memory, or stored on a computer digital storage media (e.g. a hard drive). Such digital memory and digital storage devices are well known.
  • the answer space 800 transiently resides or is persisted on a computing device, a computer network-connected device, or a personal computing device.
  • Well known computing devices include, but are not limited to super computers, mainframe computers, enterprise-class computers, servers, file servers, blade servers, web servers, departmental servers, and database servers.
  • Well known computer network-connected devices include, but are not limited to internet gateway devices, data storage devices, home internet appliances, set-top boxes, and in-vehicle computing platforms.
  • Well known personal computing devices include, but are not limited to, desktop personal computers, laptop personal computers, personal digital assistants (PDAs), advanced display cellular phones, advanced display pagers, and advanced display text messaging devices.
  • the answer space 800 contains or associates a minimum of two nodes (Serial No. 11/278,568 Figure 1C Items 180A and 180B) and the maximum number of nodes (Serial No. 11/278,568 Figure 1C Items 180A and 180B) is infinity.
  • the nodes are the products of a decomposition function (Serial No. 11/278,568 Figure IB Item 130) applied against the resources (Serial No. 11/278,568 Figure IA Item 128) identified by the Discovery phase of the correlation function 110, the nodes (Serial No. 11/278,568 Figure 1C Items 180A and 180B) are strongly associated with the resources (Serial No. 11/278,568 Figure IA Item 128) from which the nodes (Serial No. 11/278,568 Figure 1C Items 180A and 180B) were derived. Such resources (Serial No.
  • 11/278,568 Figure IA Item 128) are here called contributing resources.
  • the answer space 800 is strongly associated with a user query (manifested as input subjects 200, digital information objects 230, or questions 250) because a successful correlation (Serial No. 11/278,568 Figure IB Item 155) is an existential proof (existential quantification) that the user query can be satisfied from the contents of corpus.
  • the present invention is based upon the fact that the strong association of the user query to the answer space 800 is transitive to the resources (Serial No. 11/278,568 Figure IA Items 128) which contributed nodes (Serial No.
  • a requirement of the present invention is that the resources (Serial No. 11/278,568 Figure IA Item 128) which contributed nodes (Serial No. 11/278,568 Figure 1C Items 180A and 180B) to the answer space 185 must be identified (i.e what are the contributing resources 000?).
  • Serial No. 11/278,568 Figure 1C, Item 180B a member of node Serial No. 11/278,568 Figure 1C Item 180B is the Sequence (source) (Serial No. 11/278,568 Figure 1C Item 188).
  • the sequence (Serial No. 11/278,568 Figure 1C Item 188) contains the URI of the resource (Serial No.
  • the present invention can identify contributing resources 128 which are relevant to the user query by simply enumerating the URIs of all resources (Serial No. 11/278,568 Figure IA Item 128) found in all nodes (Serial No. 11/278,568 Figure 1C Item 188) in the answer space 185.
  • each correlation (Serial No. 11/278,568 Figure IB Item 155) can be examined, and the frequency of occurrence of a contributing resource 128 in the correlation (Serial No. 11/278,568 Figure IB Item 155) can be captured in a histogram. The cumulative counts for the occurrence of all contributing resources 128 can then be sorted. The URIs for all contributing resources 000 can then be presented to the user in order of descending frequency of occurrence. For this embodiment and referring to Figure 2, the examination of the correlations (Serial No.
  • the significance calculation function 842 is a statistical function that is based upon the number of unique nodes (Serial No. 11/278,568 Figure 1C Item 180B) contributed to the answer space 885 by each contributing resource 128. In this embodiment, any correlations (Serial No. 11/278,568 Figure IB Item 155) in the answer space 885 are not considered.
  • the significance calculation function 842 first lists the unique nodes (Serial No. 11/278,568 Figure 1C Item 180B) in the answer space 885, with one entry in the list for each node (Serial No. 11/278,568 Figure 1C Item 180B). Then, the frequency of reference to each contributing resource 128 is counted. Using standard and well-known statistical criteria and methods to measure statistical significance, the k threshold to be used by the ranking function 845 is established, and the most significant contributing resources 128 can be identified and presented to the user.
  • the significance calculation function 842 correlates the simple occurrence frequency to the simple contribution frequency value, resulting in a rudimentary significance score. If a scatter plot were used to display this data, the significant resources 128 with highest occurrence frequency and the highest contribution frequency would place farthest to the right and closest to the top. Again, as for all the varied embodiments of the significance calculation function 842 described more hereinafter, standard and well known statistical significance measures are utilized to provide appropriate k threshold information for the ranking function 845.
  • a node significance score can be computed by using measures such as the ratio, frequency of occurrence over number of nodes (Serial No. 11/278,568 Figure 1C Item 180B) contributed by that specific node's (Serial No. 11/278,568 Figure 1C Item 180B) contributing resource 128, or the ratio, frequency of occurrence over the average number of nodes (Serial No. 11/278,568 Figure 1C Item 180B) contributed by all contributing resources 128.
  • node significance scores can be normalized (0,1) or (-1,1), with the possibility thereby to rapidly determine if a given contributing resource 128 was significant or not significant to the answer space.
  • the significance calculation function 842 is a link analysis function, the link analysis function 842 taking the correlation (Serial No. 11/278,568 Figure IB Item 155) as input. This exploits the differences between the correlation (Serial No. 11/278,568 Figure IB Item 155) created by the correlation function 110 compared to a web graph.
  • the significance calculation function 842 as link analysis function establishes a link popularity score on each of node (Serial No. 11/278,568 Figure 1C Item 180B) in the answer space 128.
  • the link popularity score is determined by means of the number of in-degree links to each node (Serial No. 11/278,568 Figure 1C Item 180B) in the answer space 885.
  • the popularity score values of all nodes (Serial No. 11/278,568 Figure 1C Item 180B) contributed by a contributing resource 128 are then summed.
  • the aggregate popularity scores of all nodes (Serial No. 11/278,568 Figure 1C Item 180B) contributed by a contributing resource 128 are transit to the contributing resource 128 itself.
  • the significance calculation function 842 as link analysis function establishes an importance score on each of the nodes (Serial No. 11/278,568 Figure 1C Item 180B).
  • the importance score is determined by means of the well known Kleinberg Hubs and authorities algorithm. Hub or Authority scores for all nodes (Serial No. 11/278,568 Figure 1C Item 180B) contributed by a contributing resource 128 are then summed.
  • the aggregate Hub and Authority scores of all nodes (Serial No. 11/278,568 Figure 1C Item 180B) contributed by a contributing resource 128 are transit to the contributing resource 128.
  • the importance score is determined by means of the well-known 2nd version of the PageRank algorithm.
  • PageRank scores for all nodes (Serial No. 11/278,568 Figure 1C Item 180B) contributed by a contributing resource 128 are then summed.
  • the aggregate PageRank scores of all nodes (Serial No. 11/278,568 Figure 1C Item 180B) contributed by a contributing resource 128 are transit to the contributing resource 128.
  • the results of resource significance calculation function 842 will be ranked by means of a software function, the ranking function 845, a software program component.
  • the ranking function 845 implements a simple descending sort, with the contributing resource 128 given the highest value by the significance calculation function 842 awarded the number one rank by the ranking function 845, and the ordinal rank of the other contributing resources 128 being assigned based upon their relative position in the sorted list of significance values.
  • the significance calculation function 842 is a statistical function that is based upon the number of discrete nodes (Serial No. 11/278,568 Figure 1C Item 180B) contributed to the answer space 885 by each contributing resource 128, and when the ranking function 845 implements a simple descending sort, the ranking function is called rank by contribution.
  • the significance calculation function 842 is a statistical function that will calculate the sum of the relevance scores for all nodes (Serial No. 11/278,568 Figure 1C Item 180B) contributed to the answer space 885 by each contributing resource 128, and when the ranking function 845 implements a simple descending sort, the ranking function is called rank by relevance.
  • the significance calculation function 842 is a statistical function that will calculate the sum of the popularity scores, Hub and Authority scores, or PageRank scores for all nodes (Serial No. 11/278,568 Figure 1C Item 180B) contributed to the answer space 885 by each contributing resource 128, and when the ranking function 845 implements a simple descending sort, the ranking function is called rank by significance.
  • At least two categories of contributing resources 128 contribute nodes (Serial No. 11/278,568 Figure 1C Item 180B) to the answer space 885.
  • the two categories of contributing resources are here designated topical resources, and reference resources.
  • Topical resources provide nodes (Serial No. 11/278,568 Figure 1C Item 180B) with explicit reference to a topic, for example the triple GLOBAL WARMING-AFFECTS-GLACIERS.
  • Reference resources provide nodes (Serial No. 11/278,568 Figure 1C Item 180B) which anchor the foundations in knowledge that support topical resource nodes (Serial No. 11/278,568 Figure 1C Item 180B), for example the triple GLOBAL WARMING-IS-CLIMATE CHANGE, or GLOBAL WARMING-FROM-EMISSIONS.
  • a Support Vector Machine is created to classify and rank contributing resources.
  • SVM Support Vector Machine
  • variations of SVM are utilized, including SVM- RFE (SVM-Recursive Feature Elimination), and R-SVM (Reduced-SVM).
  • SVM- RFE SVM-Recursive Feature Elimination
  • R-SVM Reduced-SVM
  • the correlation (Serial No. 11/278,568 Figure IB Item 155) constructed or discovered by the correlation function 110 can be displayed to a user. This display is called a presentation.
  • the presentation of the answer space 128 will be implemented using a hierarchical layout 890.
  • the hierarchical layout 890 will be created using a software function, the hierarchical layout function 850, a software program component.
  • the hierarchical layout function 850 assigns the nodes of graphs on different layers in such a way that most edges in the graph flow in the same direction and the number of intersecting edges are minimized.
  • hierarchical layout function 850 uses the Sugiyama-layout algorithm.
PCT/US2006/025101 2005-11-14 2006-06-28 A knowledge correlation search engine WO2007061451A1 (en)

Priority Applications (3)

Application Number Priority Date Filing Date Title
JP2008541146A JP4864095B2 (ja) 2005-11-14 2006-06-28 知識相関サーチエンジン
CN200680042357XA CN101310274B (zh) 2005-11-14 2006-06-28 知识相关性搜索引擎
EP06774152A EP1974292A4 (en) 2005-11-14 2006-06-28 KNOWLEDGE CORRELATION PORTAL

Applications Claiming Priority (4)

Application Number Priority Date Filing Date Title
US11/273,568 2005-11-14
US11/273,568 US8108389B2 (en) 2004-11-12 2005-11-14 Techniques for knowledge discovery by constructing knowledge correlations using concepts or terms
US11/314,835 2005-12-21
US11/314,835 US8126890B2 (en) 2004-12-21 2005-12-21 Techniques for knowledge discovery by constructing knowledge correlations using concepts or terms

Publications (1)

Publication Number Publication Date
WO2007061451A1 true WO2007061451A1 (en) 2007-05-31

Family

ID=38067518

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/US2006/025101 WO2007061451A1 (en) 2005-11-14 2006-06-28 A knowledge correlation search engine

Country Status (4)

Country Link
EP (1) EP1974292A4 (un)
JP (1) JP4864095B2 (un)
CN (1) CN101310274B (un)
WO (1) WO2007061451A1 (un)

Families Citing this family (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US9104779B2 (en) * 2005-03-30 2015-08-11 Primal Fusion Inc. Systems and methods for analyzing and synthesizing complex knowledge representations
CN102662923A (zh) * 2012-04-23 2012-09-12 天津大学 一种基于机器学习的本体实例学习方法
CN112650838B (zh) * 2020-12-31 2023-07-14 南京视察者智能科技有限公司 一种基于历史案件大数据的智能问答方法及系统

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20020059220A1 (en) * 2000-10-16 2002-05-16 Little Edwin Colby Intelligent computerized search engine
US20020152202A1 (en) * 2000-08-30 2002-10-17 Perro David J. Method and system for retrieving information using natural language queries

Family Cites Families (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH04252375A (ja) * 1991-01-28 1992-09-08 Nippon Telegr & Teleph Corp <Ntt> 情報提供方法
JP3981170B2 (ja) * 1996-10-22 2007-09-26 富士通株式会社 情報検索装置
GB9821969D0 (en) * 1998-10-08 1998-12-02 Canon Kk Apparatus and method for processing natural language
US6654740B2 (en) * 2001-05-08 2003-11-25 Sunflare Co., Ltd. Probabilistic information retrieval based on differential latent semantic space
JP2003228580A (ja) * 2002-02-04 2003-08-15 Celestar Lexico-Sciences Inc 文献知識管理装置、文献知識管理方法、プログラム、および、記録媒体
US7257530B2 (en) * 2002-02-27 2007-08-14 Hongfeng Yin Method and system of knowledge based search engine using text mining
JP3960530B2 (ja) * 2002-06-19 2007-08-15 株式会社日立製作所 テキストマイニングプログラム、方法、及び装置
GB2406399A (en) * 2003-09-23 2005-03-30 Ibm Seaching within a computer network by entering a search term and optional URI into a web browser
JP4242794B2 (ja) * 2004-03-10 2009-03-25 日本電信電話株式会社 メタデータ生成装置
US8126890B2 (en) * 2004-12-21 2012-02-28 Make Sence, Inc. Techniques for knowledge discovery by constructing knowledge correlations using concepts or terms
WO2006053306A2 (en) * 2004-11-12 2006-05-18 Make Sence, Inc Knowledge discovery by constructing correlations using concepts or terms

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20020152202A1 (en) * 2000-08-30 2002-10-17 Perro David J. Method and system for retrieving information using natural language queries
US20020059220A1 (en) * 2000-10-16 2002-05-16 Little Edwin Colby Intelligent computerized search engine

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
See also references of EP1974292A4 *

Also Published As

Publication number Publication date
CN101310274A (zh) 2008-11-19
EP1974292A4 (en) 2009-04-01
CN101310274B (zh) 2013-04-17
JP2009528581A (ja) 2009-08-06
EP1974292A1 (en) 2008-10-01
JP4864095B2 (ja) 2012-01-25

Similar Documents

Publication Publication Date Title
US10496722B2 (en) Knowledge correlation search engine
US8140559B2 (en) Knowledge correlation search engine
Rahman et al. Effective reformulation of query for code search using crowdsourced knowledge and extra-large data analytics
US5933822A (en) Apparatus and methods for an information retrieval system that employs natural language processing of search results to improve overall precision
US7562074B2 (en) Search engine determining results based on probabilistic scoring of relevance
Bernardini et al. A WaCky introduction
US11216520B2 (en) Knowledge correlation search engine
Dima Intui2: A Prototype System for Question Answering over Linked Data.
Stratica et al. Using semantic templates for a natural language interface to the CINDI virtual library
Yusuf et al. Query expansion method for quran search using semantic search and lucene ranking
Yang et al. Ontology-supported FAQ processing and ranking techniques
JP4864095B2 (ja) 知識相関サーチエンジン
Abebe et al. Supporting concept location through identifier parsing and ontology extraction
Khedr et al. Ontology based semantic query expansion for searching queries in programming domain
Buey et al. An approach for automatic query expansion based on NLP and semantics
Belyaev et al. Solution of the answer formation problem in the question-answering system in Russian
Alashti et al. Parsisanj: an automatic component-based approach toward search engine evaluation
Krishnan et al. Retrieval of semantic concepts based on analysis of texts for automatic construction of ontology
Keyvanpour et al. A Useful Framework for Identification and Analysis of Different Query Expansion Approaches based on the Candidate Expansion Terms Extraction Methods.
El-Ansari et al. Personalized question-answering over linked data
Stojanovic On the conceptualisation of the query refinement task
Vargas-Vera et al. AQUA: hybrid architecture for question answering services
Lei et al. An infrastructure for building semantic web portals
Di Sciullo A reason to optimize information processing with a core property of natural language
Ahmad Ontological Approach for Semantic Modelling of Malay Translated Qur’an

Legal Events

Date Code Title Description
WWE Wipo information: entry into national phase

Ref document number: 200680042357.X

Country of ref document: CN

DPE2 Request for preliminary examination filed before expiration of 19th month from priority date (pct application filed from 20040101)
121 Ep: the epo has been informed by wipo that ep was designated in this application
WWE Wipo information: entry into national phase

Ref document number: 764/MUMNP/2008

Country of ref document: IN

WWE Wipo information: entry into national phase

Ref document number: 2006774152

Country of ref document: EP

WWE Wipo information: entry into national phase

Ref document number: 2008541146

Country of ref document: JP

NENP Non-entry into the national phase

Ref country code: DE