WO2015170963A1 - System and method for automatically generating a knowledge base - Google Patents

System and method for automatically generating a knowledge base Download PDF

Info

Publication number
WO2015170963A1
WO2015170963A1 PCT/MY2015/050025 MY2015050025W WO2015170963A1 WO 2015170963 A1 WO2015170963 A1 WO 2015170963A1 MY 2015050025 W MY2015050025 W MY 2015050025W WO 2015170963 A1 WO2015170963 A1 WO 2015170963A1
Authority
WO
WIPO (PCT)
Prior art keywords
entity
knowledge base
component
concept
entities
Prior art date
Legal status (The legal status 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 status listed.)
Ceased
Application number
PCT/MY2015/050025
Other languages
French (fr)
Inventor
Benjamin Min Xian CHU
Simon Liu QIANG
Dickson Lukose
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Mimos Bhd
Original Assignee
Mimos Bhd
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 Mimos Bhd filed Critical Mimos Bhd
Publication of WO2015170963A1 publication Critical patent/WO2015170963A1/en
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/20—Natural language analysis
    • G06F40/279—Recognition of textual entities
    • G06F40/289—Phrasal analysis, e.g. finite state techniques or chunking
    • G06F40/295—Named entity recognition
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F40/00—Handling natural language data
    • G06F40/30—Semantic analysis

Definitions

  • the present invention relates to a system and method for generating knowledge base automatically. More particularly, the present invention relates to a system and method for generating knowledge base automatically from structured and unstructured texts. BACKGROUND OF THE INVENTION
  • a knowledge base is a special database for knowledge management. It is an information repository that provides a means for information to be collected, organised, shared, searched and utilized.
  • a knowledge base can either be machine- readable or intended for human use. Human-readable knowledge bases are used to enable people to retrieve and use the knowledge they contain.
  • a human-readable knowledge base can also be coupled with a machine-readable one.
  • Machine- readable knowledge bases store knowledge in a computer-readable form which usually has automated deductive reasoning applied to them.
  • These knowledge bases contain a set of data, often in the form of rules that describe the knowledge in a logically consistent manner, whereby an ontology is used to define the structure of the stored data.
  • Knowledge base generation is a complex task as it is laborious and is usually manual. It needs to be updated and enriched continuously. Besides that, the mining-based approach implemented in natural language processing (NLP) techniques still requires human assistance.
  • NLP natural language processing
  • the present invention relates to a system and method for generating knowledge base automatically from structured and unstructured texts.
  • the system (1000) for generating knowledge base (111) automatically comprises of a Knowledge Base Generator (100); a Linked Data (LD) (108); a Linguistic Resources (LR) (109); and an Extraction Patterns component (110) to support a series of process and techniques for the Knowledge Base Generator (100).
  • a Knowledge Base Generator 100
  • LD Linked Data
  • LR Linguistic Resources
  • Extraction Patterns component (110) to support a series of process and techniques for the Knowledge Base Generator (100).
  • the Knowledge Base Generator (100) includes a Concept Pruning component (101) to identify salient entities or the right concepts to be included; a Concept Taxonomy Derivation component (102) to identify how a particular concept relates to its parent concept by referring to Linked Data (108) or Linguistic Resource (109); a Property Taxonomy Derivation component (103) to arrange all derived relations into a hierarchy; an Individuals Identification component (104) to detect individual and its concept type that is involved in a sentence; a Generic Relation Extraction component (105) to extract generic relations which can be discovered from Linguistic Resource (109) or the Extraction Patterns (110); a Semantic Relation Discovery component (106) to discover implicit relations that the Generic Relation Extraction component (105) is not able to discover; and a Resolution component (107) to resolve a best relation candidate that can be applied to a context of a text as there are many possible relation candidates.
  • a Concept Pruning component (101) to identify salient entities or the right concepts to be included
  • the method for generating knowledge base (111) automatically is characterised by the steps of tokenising a paragraph into sentences; identifying if a sentence has an appositive phrase type by a Generic Relation Extraction component (105); identifying if the appositive phrases are essential or non-essential; referring to a vector list to determine the current position for each of the appositive phrases; retrieving the entity prior to the appositive phrase; identifying all salient entities from sentences to extract generic relations; performing a check on all the pairs; and adding possible pairs which represent the relations of a particular concept to another concept and to a knowledge base (111).
  • the step of performing a check on all the pairs includes determining if either one of the entities or the prior entity of the pairs is a name; identifying the entity as an "instance” if the entity is a name; and adding the "instance" into an individual index.
  • the steps include determining if either one of the entities or the prior entity is an animate; and identifying the entity with a generic relation of "is-a" between subject and object of the entity if the entity is an animate.
  • the entity is identified with a generic relation of "hasType" between subject and object of the entity.
  • the steps include retrieving all salient entities that are identified by the Generic Relation Extraction component (105), wherein these entities are managed by a vector list; selecting each entry in the vector list and checking if the entities exist in a knowledge base (111); checking the entity if it is defined in a Linguistic Resource (109) or a Linked Data (108) if the entity is not in the knowledge base (111); selecting a pattern entry if the pattern can be applied, wherein each pattern has a corresponding property type; and adding a triple to the knowledge base (111) if the pattern can be applied.
  • the steps include extracting sense descriptions of the salient entity from the Linguistic Resource (109); transforming the sense description into context semantic structures; retrieving all context semantic structures form context index in selecting the best candidate by a Resolution component (107); performing graph similarity matching between identified sentence and the context semantic structures; and choosing and adding the best candidate triple information into the knowledge base (111)
  • the step of choosing and adding the best candidate triple information includes retrieving all context semantic structures from the context index by the Resolution component (107); performing semantic graph matching for each of the context semantic structures; identifying all sub graphs attached to the verb for each of the candidate context semantic structures; checking if the concept conforms to the semantic constraint for each of the concepts; merging the concept to the semantic structure once the concept conforms to the semantic constraint to produce a new sub graph; adding a sub graph count with a value of 1 ; consolidating all the sub graph counts; merging all the sub graphs once the iteration is complete; selecting a top context entry with the highest total sub graph counts
  • the steps include extracting an abstract description that defines the entity; and transforming the abstract description into context semantic structure.
  • FIG. 1 illustrates a system (1000) for generating knowledge base automatically according to an embodiment of the present invention.
  • FIGS. 2 (a) - (c) illustrate flow charts for automated knowledge base generation process according to an embodiment of the present invention.
  • FIG. 3 (a) illustrates an example of a whole graph structure.
  • FIGS. 3 (b - d) illustrate sub graphs examples which are only a part of the whole graph structure of FIG. 3 (a). DESCRIPTION OF THE PREFFERED EMBODIMENT
  • FIG. 1 illustrates a system (1000) for generating knowledge base (111) automatically according to an embodiment of the present invention.
  • the system (1000) which generates a knowledge base (111) automatically from natural language texts, comprises of a Knowledge Base Generator (100) which utilises external resources such as Linked Data (LD) (108) Linguistic Resource (LR) (109) and an Extraction Patterns component (110) to support a series of process and techniques for the Knowledge Base Generator (100).
  • LD Linked Data
  • LR Linguistic Resource
  • Extraction Patterns component (110)
  • the Knowledge Base Generator (100) further comprises a Concept Pruning component (101), a Concept Taxonomy Derivation component (102), a Property Taxonomy Derivation component (103), an Individuals Identification component (104), a Generic Relation Extraction component (105), a Semantic Relation Discovery component (106) and a Resolution component (107).
  • the Concept Pruning component (101) is used to identify salient entities or the right concepts to be included by filtering all the irrelevant ones.
  • the Concept Taxonomy Derivation component (102) is used to identify how a particular concept relates to its parent concept by referring to Linked Data (108) or Linguistic Resource (109). It then eventually builds up the whole taxonomy for all the concepts.
  • the Property Taxonomy Derivation component (103) is then used to arrange all derived relations into a hierarchy.
  • the Individuals Identification component (104) is used to detect the individual and concept type in the sentence that is said to be appositive.
  • the Generic Relation Extraction component (105) is used to extract generic relations discovered from Linguistic Resource (109) or from the Extraction Patterns component (110).
  • Linguistic Resource (109) comprises the standard relations of linking of a particular concept, wherein the relations are considered as taxonomic type of relations.
  • the Semantic Relation Discovery component (106) is used to discover implicit relations that Generic Relation Extraction component (105) is not able to discover.
  • FIGS. 2 (a - c) there are illustrated flow charts for a method for generating knowledge base automatically according to an embodiment of the present invention. Initially, any tokenisation method is used to tokenize a paragraph into sentences, wherein these sentences are stored into a sentence vector. These sentences are fetched from the sentence vector to be processed. Once all the sentences are processed as in decision 201 , entities of the sentences are checked if the entities are defined in the Linguistic Resource (109) or Linked Data (108) as in decision 217 shown in FIG. 2 (b).
  • the Generic Relation Extraction component (105) selects the next sentence from the sentence vector which also stores the semantic structure that represents each sentence and the type of phrase of the sentence as in step 202.
  • the Generic Relation Extraction component (105) then identifies if the sentence has an appositive phrase type as in step 203.
  • appositive phrases can be defined as essential or non-essential, wherein essential appositives are not set off with commas, and wherein non-essential appositives are defined with the use of commas.
  • An appositive phrase is said to be essential when a sentence is made clear without the use of commas.
  • a sentence "The author J. K. Rowling is a British novelist," wherein "J.K.Rowling" is referred as appositive, the sentence is said to be essential as a comma is not used to make it clear.
  • a nonessential appositive phrase is when commas are used to make the meaning of a sentence clear.
  • An example of a non-essential appositive phase is "My teacher, Miss Marshall is teaching us history," wherein the sentence uses commas to make the meaning of the sentence clearer.
  • the Generic Relation Extraction component (105) refers to a vector list of E 1t E 2 , E n to determine the current position for each of the appositive phrase and retrieves the entity prior to it. For example, if the appositive phrase found is E 2 , the Generic Relation Extraction component (105) will retrieve Ei as in step 204. The Generic Relation Extraction component (105) then identifies all salient entities from sentences to extract generic relations using any recognition method as in step 212 and these entities are stored and managed by the vector list as in step 213.
  • the Generic Relation Extraction component (105) continues with the next appositive phrase. All appositive phrases are processed and all possible pairs are identified to be checked by the Generic Relation Extraction component (105) as in steps 205 and 206. These possible pairs which represent the relations of a particular concept to another concept and will finally be added to the knowledge base (1 11) as in steps 210 and 211.
  • these possible pairs Before these possible pairs are being added to the knowledge base (111 ), these possible pairs first need to be checked to determine if either one of the entities or the prior entity of the pairs is a name that refers to a person or an organisation as in decision 207. If it is, the entity is then identified as an "instance" and the "instance” is then added into the individual index as in step 208.
  • An individual index is a list of index to manage all individuals. Examples of individual index are Roger Federer, tennis-player, Joshua Bell, violinist; and Isaac Newton, mathematician.
  • the Generic Extraction component (105) then directly checks if either one of the entities or its prior entity is an animate as in decision 209 without adding anything into the individual index. If either one of the entities or its prior entity is an animate, the entity is identified with a generic relation of "is-a" between the subject and object as in step 210.
  • the Generic Relation Extraction component (105) retrieves all salient entities that are identified as in step 213. These entities which are managed by a vector list, are then determined if these entities all already processed as in decision 214. If all these entities are already processed, these entities are checked if these entities are defined in the Linguistic Resource (109) or Linked Data (108) as in decision 217 shown in FIG. 2 (b).
  • each entity in the list is selected and checked if the entities exist in the knowledge base (111) as in step 214 and decision 216. If an entity is not in the knowledge base (111), the entity is then checked if it is defined in the Linguistic Resource (109) or Linked Data (108) as in decision 217 shown in FIG. 2 (b). If it is found in the Linguistic Resource (109), all extraction patterns are iterated and a pattern entry is selected by the Generic Relation Extraction component (105) as in decision 218 to check if the pattern can be applied, wherein each pattern has a corresponding property type.
  • the Generic Relation Extraction component (105) adds the triple of the ⁇ subject, property, object> to the knowledge base (111) as in step 223 and step 227a shown in FIG. 2 (c). For example, when a sentence is identified as non-appositive and 4 salient entities are retrieved from the sentence, wherein the entities are Entity 1, Entity 2, Entity 3 and Entity 4, and one of the entities which is Entity 4 is not found in the knowledge base, the Generic Relation Extraction component (105) then checks if Entity 4 is identified in the Linguistic Resource (109). If Entity 4 is identified in the Linguistic Resource (109), the Generic Relation Extraction component (105) retrieves the description.
  • Entity 4 has the concept "basement' with a description of "the lowest portion of a structure partly or wholly below ground level; often used for storage", the Generic Relation Extraction component (105) then iterates through all the extraction patterns.
  • An example of a selected pattern entry is shown below:
  • the Generic Relation Extraction component (105) adds a triple of ⁇ basement, part-of, building> into the knowledge base (111) when the pattern can be applied to the description of the concept found in the Linguistic Resource (109).
  • sense descriptions of the salient entity can be extracted from the Linguistic Resource (109) and transformed into context semantic structures as in step 219a.
  • These context semantic structures which are the conceptual structures or the graph representations, are then stored into a context index, which is an index that stores all possible descriptions, with the following information which is ⁇ entity, context semantic structure, hypernym> as in step 220a.
  • the Resolution component (107) retrieves all context semantic structures from the context index and determines if all the entries are already processed as in decision 222. If all context semantic structures from the context index are already processed in decision 222, the process ends.
  • the Resolution component (107) performs graph similarity matching between the identified sentence and its context semantic structures to finally add the best candidate triple information into the knowledge base (111) as in steps 224 to 227b.
  • the Generic Relation Extraction component (105) refers to the context index to get 2 sense descriptions which are ⁇ "Joshua Bell', dbpedia- owl: abstract, "Joshua David Bell (born December 9, 1967) is an American Grammy Award-winning violinisf> and ⁇ " Joshua Bell', dbpedia-owl.abstract, "Joshua L. Bell (born November 14, 1986) is an American professional baseball player who is a free agent.
  • the Semantic Relation Discovery component (106) checks in the Linked Data (108) and extracts an abstract description that defines the entity as in step 219b. Similar to step 220a, the abstract description extracted from the Linked Data (108) is then transformed into context semantic structure and is stored into the context index with the following information which is ⁇ entity, context semantic structure, type property> as in step 220b and step 221.
  • the information for the entity "Joshua Bell' can be retrieved from the Linked Data (108) and the information related to a violin player may be represented as ⁇ " Joshua Bell', dpedia-owl.occupation, " ' musician”>; ⁇ "Joshua Bell', rdfiype, "Artisf>; ⁇ "Joshua Bell', dbpedia-owl.abstract, "Joshua David Bell (bornDecember 9, 1967) is an American Grammy Award-winning violinisf> and so forth.
  • information related to an athlete who plays the baseball can also be retrieved from the Linked Data (108) and may be represented as ⁇ "Joshua Bell', rdfiype, dbpedia-owl:BaseballPlayer>; ⁇ "Joshua Bell', dbpedia- owl: abstract, "Joshua L. Bell (born November 14, 1986) is an American professional baseball player who is a free agent.
  • a third baseman, Base has played in Major League Baseball for the Baltimore Orioles and Arizona Diamondbacks"> and so forth.
  • the Resolution component (107) distinguishes the correct sense between "baseball player” or “violinist” to extract the abstract description of the correct sense before the abstract information is being transformed into a semantic structure as in steps 224 to 227b.
  • FIG. 3 (a) shows a whole graph structure with its sub graphs illustrated in FIGS. 3 (b - d).
  • FIG. 3 (b) shows a sub graph of the sentence "Joshua Bell who is a celebrity violinist has recently taken up conducting and is now a music director of the Academy of St. Martin in the Fields.” 2 descriptions are found for the entity "Joshua Bell' which are "Joshua David Bell (born December 9, 1967) is an American Grammy Award-winning violinist” and "Joshua L. Bell (born November 13, 1986) is an American professional baseball player who is a free agent. A third baseman, Bell has played in Major League Baseball for the Baltimore Orioles and Arizona Diamondbacks.” These 2 descriptions are then transformed into complete graphs.
  • FIG. 3 (b) and FIG. 3 (c) show a sub graph of a first sense which is "violinist' with a description of "Joshua David Bell (born December 9, 1967) is an American Grammy Award-winning violinist," and wherein FIG. 3 (c) shows a sub graph of a second sense which is "baseball player” with a description of "Joshua L. Bell (born November 13, 1986) is an American professional baseball player who is a free agent. A third baseman, Bell has played in Major League Baseball for the Baltimore Orioles and Arizona Diamondbacks.” From the examples given, the sentence sub graph as shown in FIG. 3 (b) and the first sense sub graph as shown in FIG.
  • the Resolution component (107) To choose the best candidate and to add the triple information into the knowledge base (111), the Resolution component (107) initially retrieves all context semantic structures from the context index which is represented by ⁇ subject, context semantic structure, property, object> as in step 224. Next, the Resolution component (107) performs semantic graph matching for each of the context semantic structures and returns the best match candidate triple as in step 225. To perform semantic graph matching, all sub graphs attached to the verb for each of the candidate context semantic structures are identified. All the concepts from the sentence are iterated and checked if the concept conforms to the semantic constraint for each of the concepts in the sub graph. An example of the semantic structure for a concept "musician" is [animate- being] ⁇ -(agnt) ⁇ -[play]->(thme)->[violin].
  • a sub graph count, SG is added with a value of 1 , wherein eventually, all context entries have their own total sub graph count. Thereon, all the sub graph counts are consolidated and all the sub graphs are merged once the iteration has been completed. Finally, a top context entry, T c with the highest total sub graph counts, G is selected as in step 226, before the triple ⁇ subject, property, object> is added into the knowledge base (111) as in step 227b.
  • T 2 which is the top context entry will be selected before its triple of ⁇ subject, property, object> is added into the knowledge base.

Landscapes

  • Engineering & Computer Science (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Artificial Intelligence (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Computational Linguistics (AREA)
  • General Health & Medical Sciences (AREA)
  • Physics & Mathematics (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Machine Translation (AREA)
  • Information Retrieval, Db Structures And Fs Structures Therefor (AREA)

Abstract

The present invention relates to a system (1000) and method for generating knowledge base (111) automatically from structured and unstructured texts. The system (1000) which generates a knowledge base (111) automatically from natural language texts, comprises of a Knowledge Base Generator (100) which utilises a Linked Data (LD) (108); a Linguistic Resource (LR) (109); and an Extraction Patterns component (110) to support a series of process and techniques for the Knowledge Base Generator (100). The Knowledge Base Generator (100) further comprises a Concept Pruning component (101), a Concept Taxonomy Derivation component (102), a Property Taxonomy Derivation component (103), an Individuals Identification component (104), a Generic Relation Extraction component (105), a Semantic Relation Discovery component (106) and a Resolution component (107).

Description

SYSTEM AND METHOD FOR AUTOMATICALLY GENERATING A
KNOWLEDGE BASE
FIELD OF INVENTION
The present invention relates to a system and method for generating knowledge base automatically. More particularly, the present invention relates to a system and method for generating knowledge base automatically from structured and unstructured texts. BACKGROUND OF THE INVENTION
A knowledge base is a special database for knowledge management. It is an information repository that provides a means for information to be collected, organised, shared, searched and utilized. A knowledge base can either be machine- readable or intended for human use. Human-readable knowledge bases are used to enable people to retrieve and use the knowledge they contain. A human-readable knowledge base can also be coupled with a machine-readable one. Machine- readable knowledge bases store knowledge in a computer-readable form which usually has automated deductive reasoning applied to them. These knowledge bases contain a set of data, often in the form of rules that describe the knowledge in a logically consistent manner, whereby an ontology is used to define the structure of the stored data. Knowledge base generation is a complex task as it is laborious and is usually manual. It needs to be updated and enriched continuously. Besides that, the mining-based approach implemented in natural language processing (NLP) techniques still requires human assistance.
An example of a method for building knowledge structures consisting of concepts and interrelations and representing knowledge extracted from the input text is disclosed in the International Patent Publication Number WO 2005/062202. The method involves reading an input, using semantic based means for extracting concepts and the interrelations from said input text, transforming said input text into a machine understandable knowledge representation so as to provide knowledge libraries from said documents and optionally storing said libraries. However, the common existing system and method for managing or generating a knowledge base is laborious and troublesome as it continuously needs to be updated and enriched. There is no disclosure of the techniques of generating a knowledge base automatically from both structured and unstructured collection of texts. Therefore, there is a need to provide a system and method that addresses the above mentioned drawbacks of the existing system and method for knowledge base generation when both entities and relations are given.
SUMMARY OF INVENTION
The present invention relates to a system and method for generating knowledge base automatically from structured and unstructured texts. The system (1000) for generating knowledge base (111) automatically comprises of a Knowledge Base Generator (100); a Linked Data (LD) (108); a Linguistic Resources (LR) (109); and an Extraction Patterns component (110) to support a series of process and techniques for the Knowledge Base Generator (100). Additionally, the Knowledge Base Generator (100) includes a Concept Pruning component (101) to identify salient entities or the right concepts to be included; a Concept Taxonomy Derivation component (102) to identify how a particular concept relates to its parent concept by referring to Linked Data (108) or Linguistic Resource (109); a Property Taxonomy Derivation component (103) to arrange all derived relations into a hierarchy; an Individuals Identification component (104) to detect individual and its concept type that is involved in a sentence; a Generic Relation Extraction component (105) to extract generic relations which can be discovered from Linguistic Resource (109) or the Extraction Patterns (110); a Semantic Relation Discovery component (106) to discover implicit relations that the Generic Relation Extraction component (105) is not able to discover; and a Resolution component (107) to resolve a best relation candidate that can be applied to a context of a text as there are many possible relation candidates. The method for generating knowledge base (111) automatically is characterised by the steps of tokenising a paragraph into sentences; identifying if a sentence has an appositive phrase type by a Generic Relation Extraction component (105); identifying if the appositive phrases are essential or non-essential; referring to a vector list to determine the current position for each of the appositive phrases; retrieving the entity prior to the appositive phrase; identifying all salient entities from sentences to extract generic relations; performing a check on all the pairs; and adding possible pairs which represent the relations of a particular concept to another concept and to a knowledge base (111). Preferably, the step of performing a check on all the pairs includes determining if either one of the entities or the prior entity of the pairs is a name; identifying the entity as an "instance" if the entity is a name; and adding the "instance" into an individual index. Preferably, if the entity is not identified as an "instance", the steps include determining if either one of the entities or the prior entity is an animate; and identifying the entity with a generic relation of "is-a" between subject and object of the entity if the entity is an animate. Preferably, if either one of the entities or the prior entity is not an animate the entity is identified with a generic relation of "hasType" between subject and object of the entity.
Preferably, if a sentence is identified as a non-appositive phrase type, the steps include retrieving all salient entities that are identified by the Generic Relation Extraction component (105), wherein these entities are managed by a vector list; selecting each entry in the vector list and checking if the entities exist in a knowledge base (111); checking the entity if it is defined in a Linguistic Resource (109) or a Linked Data (108) if the entity is not in the knowledge base (111); selecting a pattern entry if the pattern can be applied, wherein each pattern has a corresponding property type; and adding a triple to the knowledge base (111) if the pattern can be applied.
Preferably, if there is no pattern can be applied, the steps include extracting sense descriptions of the salient entity from the Linguistic Resource (109); transforming the sense description into context semantic structures; retrieving all context semantic structures form context index in selecting the best candidate by a Resolution component (107); performing graph similarity matching between identified sentence and the context semantic structures; and choosing and adding the best candidate triple information into the knowledge base (111) Preferably, the step of choosing and adding the best candidate triple information includes retrieving all context semantic structures from the context index by the Resolution component (107); performing semantic graph matching for each of the context semantic structures; identifying all sub graphs attached to the verb for each of the candidate context semantic structures; checking if the concept conforms to the semantic constraint for each of the concepts; merging the concept to the semantic structure once the concept conforms to the semantic constraint to produce a new sub graph; adding a sub graph count with a value of 1 ; consolidating all the sub graph counts; merging all the sub graphs once the iteration is complete; selecting a top context entry with the highest total sub graph counts; and adding the best candidate triple information into the knowledge base.
Preferably, if the entity is found in the Linked Data (109), the steps include extracting an abstract description that defines the entity; and transforming the abstract description into context semantic structure.
BRIEF DESCRIPTION OF THE DRAWINGS
The accompanying drawings, which are incorporated in and constitute a part of the specification, illustrate embodiments of the invention and, together with the description, serve to explain the principles of the invention.
FIG. 1 illustrates a system (1000) for generating knowledge base automatically according to an embodiment of the present invention.
FIGS. 2 (a) - (c) illustrate flow charts for automated knowledge base generation process according to an embodiment of the present invention.
FIG. 3 (a) illustrates an example of a whole graph structure.
FIGS. 3 (b - d) illustrate sub graphs examples which are only a part of the whole graph structure of FIG. 3 (a). DESCRIPTION OF THE PREFFERED EMBODIMENT
A preferred embodiment of the present invention will be described herein below with reference to the accompanying drawings. In the following description, well known functions or constructions are not described in detail since they would obscure the description with unnecessary detail.
Reference is made initially to FIG. 1 which illustrates a system (1000) for generating knowledge base (111) automatically according to an embodiment of the present invention. The system (1000) which generates a knowledge base (111) automatically from natural language texts, comprises of a Knowledge Base Generator (100) which utilises external resources such as Linked Data (LD) (108) Linguistic Resource (LR) (109) and an Extraction Patterns component (110) to support a series of process and techniques for the Knowledge Base Generator (100). The Knowledge Base Generator (100) further comprises a Concept Pruning component (101), a Concept Taxonomy Derivation component (102), a Property Taxonomy Derivation component (103), an Individuals Identification component (104), a Generic Relation Extraction component (105), a Semantic Relation Discovery component (106) and a Resolution component (107). The Concept Pruning component (101) is used to identify salient entities or the right concepts to be included by filtering all the irrelevant ones. The Concept Taxonomy Derivation component (102) is used to identify how a particular concept relates to its parent concept by referring to Linked Data (108) or Linguistic Resource (109). It then eventually builds up the whole taxonomy for all the concepts. The Property Taxonomy Derivation component (103) is then used to arrange all derived relations into a hierarchy. The Individuals Identification component (104) is used to detect the individual and concept type in the sentence that is said to be appositive. The Generic Relation Extraction component (105) is used to extract generic relations discovered from Linguistic Resource (109) or from the Extraction Patterns component (110). Linguistic Resource (109) comprises the standard relations of linking of a particular concept, wherein the relations are considered as taxonomic type of relations. The Semantic Relation Discovery component (106) is used to discover implicit relations that Generic Relation Extraction component (105) is not able to discover. This can be done by referring to Linked Data (108) to know how a particular concept relates to another concept, wherein these relations are considered as non-taxonomic type of relations. Finally, the Resolution component (107) is used to resolve the best relation candidate that can be applied to a context of the text as there are many possible relation candidates. Referring to FIGS. 2 (a - c), there are illustrated flow charts for a method for generating knowledge base automatically according to an embodiment of the present invention. Initially, any tokenisation method is used to tokenize a paragraph into sentences, wherein these sentences are stored into a sentence vector. These sentences are fetched from the sentence vector to be processed. Once all the sentences are processed as in decision 201 , entities of the sentences are checked if the entities are defined in the Linguistic Resource (109) or Linked Data (108) as in decision 217 shown in FIG. 2 (b).
However, if there are still sentences not processed as in decision 201 , the Generic Relation Extraction component (105) selects the next sentence from the sentence vector which also stores the semantic structure that represents each sentence and the type of phrase of the sentence as in step 202. The Generic Relation Extraction component (105) then identifies if the sentence has an appositive phrase type as in step 203. These appositive phrases can be defined as essential or non-essential, wherein essential appositives are not set off with commas, and wherein non-essential appositives are defined with the use of commas.
An appositive phrase is said to be essential when a sentence is made clear without the use of commas. For example, in a sentence "The author J. K. Rowling is a British novelist," wherein "J.K.Rowling" is referred as appositive, the sentence is said to be essential as a comma is not used to make it clear. On the other hand, a nonessential appositive phrase is when commas are used to make the meaning of a sentence clear. An example of a non-essential appositive phase is "My teacher, Miss Marshall is teaching us history," wherein the sentence uses commas to make the meaning of the sentence clearer.
If appositive phrases are found, the Generic Relation Extraction component (105) refers to a vector list of E1t E2, En to determine the current position for each of the appositive phrase and retrieves the entity prior to it. For example, if the appositive phrase found is E2, the Generic Relation Extraction component (105) will retrieve Ei as in step 204. The Generic Relation Extraction component (105) then identifies all salient entities from sentences to extract generic relations using any recognition method as in step 212 and these entities are stored and managed by the vector list as in step 213. An example of a vector list is V1 = {tennis-player, tournament, ball, grand-slam}, wherein tennis-player, tournament, ball and grand- slam are salient entities that are extracted from a text.
It then iterates all the entities in the vector list until it completes for one appositive phrase. Once it has completed iterating for one appositive phrase, the Generic Relation Extraction component (105) continues with the next appositive phrase. All appositive phrases are processed and all possible pairs are identified to be checked by the Generic Relation Extraction component (105) as in steps 205 and 206. These possible pairs which represent the relations of a particular concept to another concept and will finally be added to the knowledge base (1 11) as in steps 210 and 211.
Before these possible pairs are being added to the knowledge base (111 ), these possible pairs first need to be checked to determine if either one of the entities or the prior entity of the pairs is a name that refers to a person or an organisation as in decision 207. If it is, the entity is then identified as an "instance" and the "instance" is then added into the individual index as in step 208. An individual index is a list of index to manage all individuals. Examples of individual index are Roger Federer, tennis-player, Joshua Bell, violinist; and Isaac Newton, mathematician. On the other hand, if the entity is not identified as an "instance" as in decision 207, the Generic Extraction component (105) then directly checks if either one of the entities or its prior entity is an animate as in decision 209 without adding anything into the individual index. If either one of the entities or its prior entity is an animate, the entity is identified with a generic relation of "is-a" between the subject and object as in step 210.
For example, "My uncle, a lawyer is visiting us." From this sentence, the salient entities are identified as "uncle" and "lawyer." The entity "uncle" is identified as an "instance" and is referred as an animate concept by the Generic Relation Extraction component (105). Therefore, a triple with the generic relation of "is-a" is put between the subject "uncle" and object "lawyer." Hence, a triple of <uncle, is-a, lawyer> is added into the knowledge base (111).
However, if either one of the entities or its prior entity is not an animate as in decision 209, the entity is then identified with a generic relation of "hasType" between the subject and object as in step 211.
For example, in the sentence "Our house, a brick bungalow is located at Oak Street," the salient entities are identified as "house" and "brick bungalow." From this example, both entities are not identified as "instance" and are referred as inanimate concepts by the Generic Relation Extraction component (105), therefore a triple with the generic relation "hasType" is identified between the subject "house" and object "brick bungalow". Hence, a triple of <house, hasType, brick bungalow> is added into the knowledge base (111).
On the other hand, if a sentence is identified as a non-appositive phrase type in decision 203, the Generic Relation Extraction component (105) retrieves all salient entities that are identified as in step 213. These entities which are managed by a vector list, are then determined if these entities all already processed as in decision 214. If all these entities are already processed, these entities are checked if these entities are defined in the Linguistic Resource (109) or Linked Data (108) as in decision 217 shown in FIG. 2 (b).
However, if there are still some entities that have yet to be processed as in decision 214, each entity in the list is selected and checked if the entities exist in the knowledge base (111) as in step 214 and decision 216. If an entity is not in the knowledge base (111), the entity is then checked if it is defined in the Linguistic Resource (109) or Linked Data (108) as in decision 217 shown in FIG. 2 (b). If it is found in the Linguistic Resource (109), all extraction patterns are iterated and a pattern entry is selected by the Generic Relation Extraction component (105) as in decision 218 to check if the pattern can be applied, wherein each pattern has a corresponding property type. If a pattern can be applied, the Generic Relation Extraction component (105) adds the triple of the <subject, property, object> to the knowledge base (111) as in step 223 and step 227a shown in FIG. 2 (c). For example, when a sentence is identified as non-appositive and 4 salient entities are retrieved from the sentence, wherein the entities are Entity 1, Entity 2, Entity 3 and Entity 4, and one of the entities which is Entity 4 is not found in the knowledge base, the Generic Relation Extraction component (105) then checks if Entity 4 is identified in the Linguistic Resource (109). If Entity 4 is identified in the Linguistic Resource (109), the Generic Relation Extraction component (105) retrieves the description. If Entity 4 has the concept "basement' with a description of "the lowest portion of a structure partly or wholly below ground level; often used for storage", the Generic Relation Extraction component (105) then iterates through all the extraction patterns. An example of a selected pattern entry is shown below:
Extracti o n- patte rn#22 :
Pattern Property
" ...basement of a building..." part-of wherein from the selected pattern entry, the Generic Relation Extraction component (105) adds a triple of <basement, part-of, building> into the knowledge base (111) when the pattern can be applied to the description of the concept found in the Linguistic Resource (109).
However, if there is no pattern that can be applied as in decision 218, sense descriptions of the salient entity can be extracted from the Linguistic Resource (109) and transformed into context semantic structures as in step 219a. These context semantic structures, which are the conceptual structures or the graph representations, are then stored into a context index, which is an index that stores all possible descriptions, with the following information which is <entity, context semantic structure, hypernym> as in step 220a. In selecting the best candidate, which is a context index with the best sense description, the Resolution component (107) retrieves all context semantic structures from the context index and determines if all the entries are already processed as in decision 222. If all context semantic structures from the context index are already processed in decision 222, the process ends. If not, the Resolution component (107) performs graph similarity matching between the identified sentence and its context semantic structures to finally add the best candidate triple information into the knowledge base (111) as in steps 224 to 227b. For example in a sentence, "Joshua Bell who is a superstar violinist has recently taken up conducting and is now a music director of the Academy of St. Martin in the Fields," the Generic Relation Extraction component (105) refers to the context index to get 2 sense descriptions which are <"Joshua Bell', dbpedia- owl: abstract, "Joshua David Bell (born December 9, 1967) is an American Grammy Award-winning violinisf> and <" Joshua Bell', dbpedia-owl.abstract, "Joshua L. Bell (born November 14, 1986) is an American professional baseball player who is a free agent. A third baseman, Base has played in Major League Baseball for the Baltimore Orioles and Arizona Diamondbacks"> , wherein these 2 sense descriptions can then be transformed into context semantic structures.
On the other hand, if the entity is not found in the Linguistic Resource (109) as in decision 217, the Semantic Relation Discovery component (106) then checks in the Linked Data (108) and extracts an abstract description that defines the entity as in step 219b. Similar to step 220a, the abstract description extracted from the Linked Data (108) is then transformed into context semantic structure and is stored into the context index with the following information which is <entity, context semantic structure, type property> as in step 220b and step 221.
For instance, in the same example given previously which is "Joshua Bell who is a superstar violinist has recently taken up conducting and is now a music director of the Academy of St. Martin in the Fields," the information for the entity "Joshua Bell' can be retrieved from the Linked Data (108) and the information related to a violin player may be represented as <" Joshua Bell', dpedia-owl.occupation, " ' musician">; <"Joshua Bell', rdfiype, "Artisf>; <"Joshua Bell', dbpedia-owl.abstract, "Joshua David Bell (bornDecember 9, 1967) is an American Grammy Award-winning violinisf> and so forth. Furthermore, information related to an athlete who plays the baseball can also be retrieved from the Linked Data (108) and may be represented as <"Joshua Bell', rdfiype, dbpedia-owl:BaseballPlayer>; <"Joshua Bell', dbpedia- owl: abstract, "Joshua L. Bell (born November 14, 1986) is an American professional baseball player who is a free agent. A third baseman, Base has played in Major League Baseball for the Baltimore Orioles and Arizona Diamondbacks"> and so forth.
Based on the previous example, the Resolution component (107) distinguishes the correct sense between "baseball player" or "violinist" to extract the abstract description of the correct sense before the abstract information is being transformed into a semantic structure as in steps 224 to 227b.
A detailed explanation of steps 224 to 227b is provided herein below with respect to FIGS 3 (a - d). FIG. 3 (a) shows a whole graph structure with its sub graphs illustrated in FIGS. 3 (b - d). In FIG. 3 (b), it shows a sub graph of the sentence "Joshua Bell who is a superstar violinist has recently taken up conducting and is now a music director of the Academy of St. Martin in the Fields." 2 descriptions are found for the entity "Joshua Bell' which are "Joshua David Bell (born December 9, 1967) is an American Grammy Award-winning violinist" and "Joshua L. Bell (born November 13, 1986) is an American professional baseball player who is a free agent. A third baseman, Bell has played in Major League Baseball for the Baltimore Orioles and Arizona Diamondbacks." These 2 descriptions are then transformed into complete graphs.
One sub graph for each description is shown in FIG. 3 (b) and FIG. 3 (c) respectively, wherein FIG. 3 (b) shows a sub graph of a first sense which is "violinist' with a description of "Joshua David Bell (born December 9, 1967) is an American Grammy Award-winning violinist," and wherein FIG. 3 (c) shows a sub graph of a second sense which is "baseball player" with a description of "Joshua L. Bell (born November 13, 1986) is an American professional baseball player who is a free agent. A third baseman, Bell has played in Major League Baseball for the Baltimore Orioles and Arizona Diamondbacks." From the examples given, the sentence sub graph as shown in FIG. 3 (b) and the first sense sub graph as shown in FIG. 3 (c) are said to be almost similar because the concepts "superstar-violinist" and "American-violinist" have a smaller distance in concept hierarchy if compared to the concept "baseball player," wherein a concept hierarchy refers to the taxonomy of the Linguistic Resource (109). Due to that, the first sense is said to be a better candidate than the second sense, wherein all the triple information that belongs to the first sense is added into the knowledge base (111).
To choose the best candidate and to add the triple information into the knowledge base (111), the Resolution component (107) initially retrieves all context semantic structures from the context index which is represented by <subject, context semantic structure, property, object> as in step 224. Next, the Resolution component (107) performs semantic graph matching for each of the context semantic structures and returns the best match candidate triple as in step 225. To perform semantic graph matching, all sub graphs attached to the verb for each of the candidate context semantic structures are identified. All the concepts from the sentence are iterated and checked if the concept conforms to the semantic constraint for each of the concepts in the sub graph. An example of the semantic structure for a concept "musician" is [animate- being]<-(agnt)<-[play]->(thme)->[violin]. For the agnt role, there are 2 semantic constraints, wherein one is on its left and one is on its right. The one on the left is an animate being and the one on the right is an action verb. Since the concept violin- player conforms to the agnt left constraint, which means the concept falls under the concept hierarchy of the animate being concept, this concept can be merged to the semantic structure to produce [musician]<-agnt<-[play]->(thme)->[violin].
Once the semantic constraint is met and once the concept can be merged to the semantic structure to produce a new sub graph, a sub graph count, SG is added with a value of 1 , wherein eventually, all context entries have their own total sub graph count. Thereon, all the sub graph counts are consolidated and all the sub graphs are merged once the iteration has been completed. Finally, a top context entry, Tc with the highest total sub graph counts, G is selected as in step 226, before the triple <subject, property, object> is added into the knowledge base (111) as in step 227b. For example, if context entry 1 , T1t context entry 2, T2 and context entry 3, T3 have their own total sub graph count of 5, 7 and 3 respectively, T2, which is the top context entry will be selected before its triple of <subject, property, object> is added into the knowledge base. While embodiments of the invention have been illustrated and described, it is not intended that these embodiments illustrate and describe all possible forms of the invention. Rather, the words used in the specifications are words of description rather than limitation and various changes may be made without departing from the scope of the invention.

Claims

1. A system (1000) for generating knowledge base (111) automatically comprises of:
a) a Knowledge Base Generator (100);
b) a Linked Data (LD) (108);
c) a Linguistic Resources (LR) (109); and
d) an Extraction Patterns component (110) to support a series of process and techniques for the Knowledge Base Generator (100), wherein the system (1000) is characterised in that the Knowledge Base Generator (100) includes:
i. a Concept Pruning component (101) to identify salient entities or the right concepts to be included;
ii. a Concept Taxonomy Derivation component (102) to identify how a particular concept relates to its parent concept by referring to the Linked Data (108) or the Linguistic Resource (109);
iii. a Property Taxonomy Derivation component (103) to arrange all derived relations into a hierarchy;
iv. an Individuals Identification component (104) to detect individual and its concept type that is involved in a sentence; v. a Generic Relation Extraction component (105) to extract generic relations which can be discovered from the Linguistic Resource (109) or the Extraction Patterns (110);
vi. a Semantic Relation Discovery component (106) to discover implicit relations that the Generic Relation Extraction component (105) is not able to discover; and
vii. a Resolution component (107) to resolve a best relation candidate that can be applied to a context of a text as there are many possible relation candidates.
2. A method for generating knowledge base (111) automatically is characterised by the steps of:
a) tokenising a paragraph into sentences;
b) identifying if a sentence has an appositive phrase type by a Generic Relation Extraction component (105);
c) identifying if the appositive phrases are essential or non-essential; d) referring to a vector list to determine the current position for each of the appositive phrases;
e) retrieving the entity prior to the appositive phrase;
f) identifying all salient entities from sentences to extract generic relations;
g) performing a check on all the pairs; and
h) adding possible pairs which represent the relations of a particular concept to another concept and to a knowledge base (111).
The method as claimed in claim 2, wherein the step of performing a check on all the pairs includes:
a) determining if either one of the entities or the prior entity of the pairs is a name;
b) identifying the entity as an "instance" if the entity is a name; and c) adding the "instance" into an individual index.
The method as claimed in claim 3, wherein if the entity is not identified as an "instance", the steps include:
a) determining if either one of the entities or the prior entity is an animate; and
b) identifying the entity with a generic relation of "is-a" between subject and object of the entity if the entity is an animate.
The method as claimed in claim 4, wherein if either one of the entities or the prior entity is not an animate the entity is identified with a generic relation of "hasType" between subject and object of the entity.
6. The method as claimed in claim 2, wherein if a sentence is identified as a non-appositive phrase type, the steps include:
a) retrieving all salient entities that are identified by the Generic Relation
Extraction component (105), wherein these entities are managed by a vector list;
b) selecting each entry in the vector list and checking if the entities exist in a knowledge base (111); c) checking the entity if it is defined in a Linguistic Resource (109) or a Linked Data (108) if the entity is not in the knowledge base (111); d) selecting a pattern entry if the pattern can be applied, wherein each pattern has a corresponding property type; and
e) adding a triple to the knowledge base (111) if the pattern can be applied.
The method as claimed in claim 6, wherein if there is no pattern can be applied, the steps include:
a) extracting sense descriptions of the salient entity from the Linguistic Resource (109);
b) transforming the sense description into context semantic structures; c) retrieving all context semantic structures form context index in selecting the best candidate by a Resolution component (107);
d) performing graph similarity matching between identified sentence and the context semantic structures; and
e) choosing and adding the best candidate triple information into the knowledge base (111). The method as claimed in claim 7, wherein the step of choosing and adding the best candidate triple information includes:
a) retrieving all context semantic structures from the context index by the Resolution component (107);
b) performing semantic graph matching for each of the context semantic structures;
c) identifying all sub graphs attached to the verb for each of the candidate context semantic structures;
d) checking if the concept conforms to the semantic constraint for each of the concepts;
e) merging the concept to the semantic structure once the concept conforms to the semantic constraint to produce a new sub graph; f) adding a sub graph count with a value of 1 ;
g) consolidating all the sub graph counts;
h) merging all the sub graphs once the iteration is complete; i) selecting a top context entry with the highest total sub graph counts; and
j) adding the best candidate triple information into the knowledge base.
9. The method as claimed in claim 6, wherein if the entity is found in the Linked Data (109), the steps include:
a) extracting an abstract description that defines the entity; and b) transforming the abstract description into context semantic structure.
PCT/MY2015/050025 2014-05-05 2015-04-22 System and method for automatically generating a knowledge base Ceased WO2015170963A1 (en)

Applications Claiming Priority (2)

Application Number Priority Date Filing Date Title
MYPI2014701129 2014-05-05
MYPI2014701129A MY181677A (en) 2014-05-05 2014-05-05 System and method for generating knowledge base automatically

Publications (1)

Publication Number Publication Date
WO2015170963A1 true WO2015170963A1 (en) 2015-11-12

Family

ID=53398171

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/MY2015/050025 Ceased WO2015170963A1 (en) 2014-05-05 2015-04-22 System and method for automatically generating a knowledge base

Country Status (2)

Country Link
MY (1) MY181677A (en)
WO (1) WO2015170963A1 (en)

Cited By (11)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10037362B1 (en) 2017-07-24 2018-07-31 International Business Machines Corpoation Mining procedure dialogs from source content
WO2018174815A1 (en) * 2017-03-24 2018-09-27 Agency For Science, Technology And Research Method and apparatus for semantic coherence analysis of texts
US10361981B2 (en) 2015-05-15 2019-07-23 Microsoft Technology Licensing, Llc Automatic extraction of commitments and requests from communications and content
CN111191044A (en) * 2019-12-25 2020-05-22 湖北大学 A method of knowledge extraction and fusion based on big data
CN111461322A (en) * 2020-03-13 2020-07-28 中国科学院计算技术研究所 A deep neural network model compression method
CN111930856A (en) * 2020-07-06 2020-11-13 北京邮电大学 Method, device and system for constructing domain knowledge graph ontology and data
US10984387B2 (en) 2011-06-28 2021-04-20 Microsoft Technology Licensing, Llc Automatic task extraction and calendar entry
CN113792157A (en) * 2021-09-14 2021-12-14 哈尔滨工业大学 Domain mechanism-oriented knowledge base construction method
CN116340414A (en) * 2023-05-31 2023-06-27 北京华云安信息技术有限公司 Knowledge graph-based attack surface visual modeling method and device
US11995612B2 (en) 2019-04-08 2024-05-28 Phenom People Knowledge engine using machine learning and predictive modeling for optimizing recruitment management systems
CN120256590A (en) * 2025-06-04 2025-07-04 湖南工程学院 Innovative entrepreneurship counseling question-answer matching method and system based on semantic understanding

Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6076088A (en) * 1996-02-09 2000-06-13 Paik; Woojin Information extraction system and method using concept relation concept (CRC) triples
WO2002063493A1 (en) * 2001-02-08 2002-08-15 2028, Inc. Methods and systems for automated semantic knowledge leveraging graph theoretic analysis and the inherent structure of communication
WO2005062202A2 (en) 2003-12-23 2005-07-07 Thomas Eskebaek Knowledge management system with ontology based methods for knowledge extraction and knowledge search

Patent Citations (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6076088A (en) * 1996-02-09 2000-06-13 Paik; Woojin Information extraction system and method using concept relation concept (CRC) triples
WO2002063493A1 (en) * 2001-02-08 2002-08-15 2028, Inc. Methods and systems for automated semantic knowledge leveraging graph theoretic analysis and the inherent structure of communication
WO2005062202A2 (en) 2003-12-23 2005-07-07 Thomas Eskebaek Knowledge management system with ontology based methods for knowledge extraction and knowledge search

Non-Patent Citations (1)

* Cited by examiner, † Cited by third party
Title
SERGEI NIRENBURG ET AL: "Learning by Reading by Learning to Read", SEMANTIC COMPUTING, 2007. ICSC 2007. INTERNATIONAL CONFERENCE ON, IEEE, PISCATAWAY, NJ, USA, 1 September 2007 (2007-09-01), pages 694 - 701, XP031138706, ISBN: 978-0-7695-2997-4 *

Cited By (14)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US10984387B2 (en) 2011-06-28 2021-04-20 Microsoft Technology Licensing, Llc Automatic task extraction and calendar entry
US10361981B2 (en) 2015-05-15 2019-07-23 Microsoft Technology Licensing, Llc Automatic extraction of commitments and requests from communications and content
WO2018174815A1 (en) * 2017-03-24 2018-09-27 Agency For Science, Technology And Research Method and apparatus for semantic coherence analysis of texts
US10169428B1 (en) 2017-07-24 2019-01-01 International Business Machines Corporation Mining procedure dialogs from source content
US10185753B1 (en) 2017-07-24 2019-01-22 International Business Machines Corporation Mining procedure dialogs from source content
US10037362B1 (en) 2017-07-24 2018-07-31 International Business Machines Corpoation Mining procedure dialogs from source content
US11995612B2 (en) 2019-04-08 2024-05-28 Phenom People Knowledge engine using machine learning and predictive modeling for optimizing recruitment management systems
CN111191044A (en) * 2019-12-25 2020-05-22 湖北大学 A method of knowledge extraction and fusion based on big data
CN111461322A (en) * 2020-03-13 2020-07-28 中国科学院计算技术研究所 A deep neural network model compression method
CN111461322B (en) * 2020-03-13 2024-03-08 中国科学院计算技术研究所 A deep neural network model compression method
CN111930856A (en) * 2020-07-06 2020-11-13 北京邮电大学 Method, device and system for constructing domain knowledge graph ontology and data
CN113792157A (en) * 2021-09-14 2021-12-14 哈尔滨工业大学 Domain mechanism-oriented knowledge base construction method
CN116340414A (en) * 2023-05-31 2023-06-27 北京华云安信息技术有限公司 Knowledge graph-based attack surface visual modeling method and device
CN120256590A (en) * 2025-06-04 2025-07-04 湖南工程学院 Innovative entrepreneurship counseling question-answer matching method and system based on semantic understanding

Also Published As

Publication number Publication date
MY181677A (en) 2020-12-31

Similar Documents

Publication Publication Date Title
WO2015170963A1 (en) System and method for automatically generating a knowledge base
CN108073569B (en) Law cognition method, device and medium based on multi-level multi-dimensional semantic understanding
US10120861B2 (en) Hybrid classifier for assigning natural language processing (NLP) inputs to domains in real-time
KR101136007B1 (en) System and method for anaylyzing document sentiment
CN111694823B (en) Institutional standardization method, device, electronic device and storage medium
RU2591175C1 (en) Method and system for global identification in collection of documents
CN112256845B (en) Intention recognition method, apparatus, electronic device, and computer-readable storage medium
RU2491622C1 (en) Method of classifying documents by categories
CN112307364B (en) Character representation-oriented news text place extraction method
JP6729095B2 (en) Information processing device and program
KR102091633B1 (en) Searching Method for Related Law
US20110320490A1 (en) Named entity database or mining rule database update apparatus and method using named entity database and mining rule merged ontology schema
CN109446313B (en) Sequencing system and method based on natural language analysis
Kanapala et al. Passage-based text summarization for legal information retrieval
JP5117744B2 (en) Word meaning tag assigning device and method, program, and recording medium
US20240070175A1 (en) Method for Determining Company Related to News Based on Scoring and Apparatus for Performing the Method
Habib et al. A hybrid approach for robust multilingual toponym extraction and disambiguation
WO2015053607A1 (en) System and method for semantic-level sentiment analysis of text
Lai et al. An unsupervised approach to discover media frames
KR20150051969A (en) Method for evaluate value of technology
JP5832869B2 (en) Keyword extraction system and keyword extraction method using category matching
CN117874230A (en) Method, device, and electronic device for obtaining case category
CN116069948A (en) Content risk control knowledge base construction method, device, equipment and storage medium
KR101590962B1 (en) Database system for evaluate technical value of key patent
CN121615637B (en) Crime name identification method and system integrating rule engine and knowledge graph

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: 15729251

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: 15729251

Country of ref document: EP

Kind code of ref document: A1