Systems and methods for applying statistical inference techniques to knowledge representations
Techniques for analyzing and synthesizing complex knowledge representations (KRs) may utilize an atomic knowledge representation model including both an elemental data structure and knowledge processing rules stored as machine-readable data and/or programming instructions. One or more of the knowledge processing rules may be applied to analyze an input complex KR to deconstruct its complex concepts and/or concept relationships to elemental concepts and/or concept relationships to be included in the elemental data structure. One or more of the knowledge processing rules may be applied to synthesize an output complex KR from the stored elemental data structure in accordance with context information. Methods of populating an elemental data structure and methods of synthesizing a complex KR from the elemental data structure may rely on statistical inference techniques.
1. A method of modifying a computer-readable elemental data structure, the method comprising:
estimating, using at least one processor executing stored program instructions, a relevance associated with an elemental component, wherein the elemental component comprises an elemental concept or an elemental concept relationship between two or more elemental concepts, and wherein estimating the relevance comprises estimating a frequency of occurrence in reference data of one or more labels associated with the elemental component,
wherein the elemental component includes the elemental concept, wherein the relevance associated with the elemental component comprises a concept relevance associated with the elemental concept, wherein the one or more labels associated with the elemental component comprises a label associated with the elemental concept, and wherein estimating the frequency of occurrence in the reference data of the one or more labels associated with the elemental component comprises estimating a term frequency of the label within at least a portion of the reference data; and
in response to determining that the relevance associated with the elemental component exceeds a relevance threshold, modifying the elemental data structure, wherein modifying the elemental data structure comprises:
adding the elemental component to the elemental data structure, and
storing the relevance in data associated with the elemental component,
wherein the elemental component is encoded as a computer-readable data structure storing the data associated with the elemental component.
2. The method of claim 1 , further comprising applying, using at least one processor executing stored program instructions, one or more rules to deconstruct a knowledge representation into one or more elemental components comprising the elemental component.
3. The method of claim 1 , wherein estimating the frequency of occurrence of the one or more labels comprises obtaining statistics associated with reference data.
4. The method of claim 1 , wherein:
the elemental component includes the elemental concept relationship, and
the relevance associated with the elemental component comprises a relationship relevance associated with the elemental concept relationship.
5. The method of claim 1 , wherein estimating the frequency of occurrence in the reference data of the one or more labels associated with the elemental component comprises estimating a term-document frequency of the label within at least a portion of the reference data.
6. The method of claim 1 , wherein estimating the frequency of occurrence in the reference data of the one or more labels associated with the elemental component comprises estimating an inverse document frequency of the label within at least a portion of the reference data.
7. The method of claim 4 , wherein:
estimating the frequency of occurrence in the reference data of the one or more labels associated with the elemental component comprises estimating a term-document frequency according to a formula Docs(L1 and L2)/NumDocs,
the elemental concept relationship relates a first elemental concept and a second elemental concept,
L1 represents a first label associated with the first elemental concept,
L2 represents a second label associated with the second elemental concept;
Docs(L1 and L2) represents a number of documents in at least a portion of the reference data that contain the first label and the second label, and
NumDocs represents a number of documents in the at least a portion of the reference data.
8. The method of claim 7 , wherein a document comprises at least one of a sentence, a plurality of sentences, a paragraph, or a plurality of paragraphs.
9. The method of claim 7 , wherein NumDocs represents the number of documents in the at least a portion of the reference data that contain at least one of the first label or the second label.
10. The method of claim 1 , wherein estimating the frequency of occurrence in the reference data of the one or more labels associated with the elemental component comprises using a search engine to search the reference data for documents containing the one or more labels associated with the elemental component.
11. A method of modifying a computer-readable graphical model associated with an elemental data structure, the elemental data structure comprising one or more elemental components, the one or more elemental components comprising one or more elemental concepts and one or more elemental concept relationships, the method comprising:
obtaining the graphical model associated with the elemental data structure, the graphical model comprising one or more graphical components, the one or more graphical components comprising:
one or more nodes corresponding to the respective one or more elemental concepts of the elemental data structure, and
one or more edges incident on nodes of the one or more nodes, the one or more edges corresponding to the respective one or more elemental concept relationships of the elemental data structure;
estimating, using at least one processor executing stored program instructions, a semantic coherence of an elemental component; and
modifying the graphical model by assigning a probability corresponding to the semantic coherence to a graphical component of the graphical model,
wherein the one or more elemental components are encoded as one or more respective computer-readable data structures storing data associated with the one or more elemental components,
wherein the one or more graphical components are encoded as one or more respective computer-readable data structures storing data associated with the one or more graphical components,
wherein the elemental component is encoded as a computer readable data structure storing data associated with the elemental component,
wherein the graphical component is encoded as a computer readable data structure storing data associated with the graphical component,
wherein the method further comprises modifying the elemental data structure,
wherein the elemental component comprises an elemental concept relationship, the elemental concept relationship being between a first elemental concept and a second elemental concept of the one or more elemental concepts,
wherein one or more labels comprise a first label associated with the first elemental concept and a second label associated with the second elemental concept,
wherein the graphical component comprises an edge of the graphical model, the edge corresponding to the elemental concept relationship, and
wherein modifying the elemental data structure comprises adding the elemental concept relationship to the elemental data structure in a case where the semantic coherence exceeds a threshold.
12. The method of claim 11 , wherein estimating the semantic coherence comprises estimating a frequency of occurrence in reference data of one or more labels associated with the elemental component.
13. The method of claim 11 , wherein:
estimating the semantic coherence of the elemental concept relationship comprises calculating a joint probability associated with the first elemental concept and the second elemental concept.
14. The method of claim 11 , wherein the elemental concept relationship is a first elemental concept relationship, and wherein:
the elemental component further comprises second elemental concept relationship, the second elemental concept relationship being between a third elemental concept and a fourth elemental concept of the one or more elemental concepts,
the elemental data structure comprises the second elemental concept relationship, and
modifying the elemental data structure further comprises removing the second elemental concept relationship from the elemental data structure in response to determining that the semantic coherence of the elemental concept relationship is less than a threshold semantic coherence.
15. The method of claim 11 , wherein:
the elemental component further comprises an elemental concept,
the one or more labels comprise a third label associated with the elemental concept,
the graphical component comprises a node of the graphical model, the node corresponding to the elemental concept, and
modifying the elemental data structure further comprises adding the elemental concept to the elemental data structure in response to determining that the semantic coherence exceeds a threshold semantic coherence.
16. The method of claim 11 , wherein:
the elemental component further comprises an elemental concept,
the one or more labels comprise a third label associated with the elemental concept,
the graphical component comprises a node of the graphical model, the node corresponding to the elemental concept,
the elemental data structure comprises the elemental concept, and
modifying the elemental data structure further comprises removing the elemental concept from the elemental data structure in response to determining that the semantic coherence of the elemental concept is less than a threshold semantic coherence.
17. A knowledge representation apparatus for modifying a computer-readable elemental data structure, the apparatus comprising: one or more processors including
a probabilistic unit configured to:
estimate a relevance associated with an elemental component, wherein the elemental component comprises an elemental concept or an elemental concept relationship between two or more elemental concepts, and wherein estimating the relevance comprises estimating a frequency of occurrence in reference data of one or more labels associated with the elemental component,
wherein the elemental component includes the elemental concept, wherein the relevance associated with the elemental component comprises a concept relevance associated with the elemental concept, wherein the one or more labels associated with the elemental component comprises a label associated with the elemental concept, and wherein estimating the frequency of occurrence in the reference data of the one or more labels associated with the elemental component comprises estimating a term frequency of the label within at least a portion of the reference data; and
in response to determining that the relevance associated with the elemental component exceeds a relevance threshold, modify the elemental data structure by:
adding the elemental component to the elemental data structure, and
storing the probability in data associated with the elemental component;
wherein the elemental component is encoded as a computer-readable data structure storing the data associated with the elemental component.
18. The method of claim 1 , wherein the relevance associated with the elemental component comprises a relevance of the elemental component to a user, and wherein modifying the elemental data structure comprises modifying the elemental data structure to be specific to the user.