Apparatus and method for building and using inference engines based on representations of data that preserve relationships between objects

• US 9,563,847 B2
• Filed: 06/04/2014
• Issued: 02/07/2017
• Est. Priority Date: 06/05/2013
• Status: Active Grant

• Current Assignee: Multimodel Research, LLC
• Original Assignee: Multimodel Research, LLC
• Inventors: Gallant, Stephen I.
• Primary Examiner(s): Hill, Stanley K
• Assistant Examiner(s): Bharadwaj, Kalpana
• Application Number:
  US14/295,880
• Publication Number:
  US 20140365410A1
• Time in Patent Office: 979 Days
• US Class Current: 1/1
• CPC Class Codes: G06N 5/04 Inference or reasoning models

Abstract

This disclosure describes, among other things, an apparatus for generating an inference engine about a document. The apparatus includes at least one processor and a memory with instructions. The memory including instructions that, when executed by the at least one processor, cause the at least one processor to perform a number of processes. The processor accesses a set of documents. Each document has a corresponding inference. The processor also generates a vector representation for each document in the set of documents. First, the processor parses text of the document into groups of words, and generates a vector representation for each group.

Claims

1. An apparatus for generating an inference engine about a document, the apparatus comprising:

at least one processor;

a memory with instructions, the memory including instructions that, when executed by the at least one processor, cause the at least one processor to;

access a set of documents, each document having a corresponding inference;

generate a vector representation for each document in the set of documents by

(i) parsing text of the document into groups of words;

(ii) for each group of words, generating a vector representation of the group by

(a) obtaining a binding operator corresponding to a type of the group, the binding operator comprising a matrix of values,

(b) obtaining vector representations of the words in the group,

(c) summing the vector representations of the words, and

(d) multiplying the binding operator and the sum of the vector representations to generate a vector representation of the group; and

(iii) summing the vector representations of the groups to generate a vector representation of the document; and

train an inference engine on the vector representations of the documents and the corresponding inferences for each of the documents.

2. The apparatus according to claim 1, wherein the memory further includes instructions that, when executed by the processor, cause the processor to:

access a new document;

generate a vector representation of the new document; and

process the vector representation of the new document according to the trained inference engine to determine an inference about the new document.

3. The apparatus according to claim 2, wherein the inferences comprise a positive or negative sentiment.

4. The apparatus according to claim 2, wherein the inferences comprise a probability that a classification label is appropriate for a document.

5. The apparatus according to claim 2, wherein the inferences comprise a prediction about a movement of a security price.

6. The apparatus according to claim 2, wherein the inferences comprise a prediction about purchase behavior of a consumer.

7. The apparatus according to claim 1, wherein the types of groups include a group that pertains to a subject of a sentence, a group that pertains to a verb of the sentence, and a group that pertains to an object of the sentence.

8. The apparatus according to claim 1, wherein the binding operator corresponding to any given type of group is a unique two-dimensional matrix of values.

9. The apparatus according to claim 1, wherein the vector representation of any given word is a unique one-dimensional array of values.

10. The apparatus according to claim 1, wherein the memory further includes instructions that, when executed by the processor, cause the processor to:

normalizing the vector representation of the group.

11. A method for generating an inference engine, the method comprising:

accessing, by at least one processor, a set of documents, each document having a corresponding inference;

generating, by the at least one processor, a vector representation for each document in the set of documents by

(i) parsing text of the document into groups of words,

(ii) for each group of words, generating a vector representation of the group by

(a) obtaining a binding operator corresponding to a type of the group, the binding operator comprising a matrix of values,

(b) obtaining vector representations of the words in the group,

(c) summing the vector representations of the words, and

(d) multiplying the binding operator and the sum of the vector representations to generate a vector representation of the group; and

(iii) summing the vector representations of the groups to generate a vector representation of the document; and

training, by the at least one processor, an inference engine on the vector representations of the documents and the corresponding inferences for each of the documents.

12. The method of claim 11, further comprising:

accessing, by the at least one processor, a new document;

generating, by the at least one processor, a vector representation of the new document; and

processing, by the at least one processor, the vector representation of the new document according to the trained inference engine to determine an inference about the new document.

13. The method of claim 12, wherein processing the new document comprises:

determining whether a sentiment of the new document is positive or negative.

14. The method of claim 12, wherein processing the new document comprises:

determining whether a classification label is applicable to the new document.

15. The method of claim 12, wherein processing the new document comprises:

predicting a movement of a security price.

16. The method of claim 12, wherein processing the new document comprises:

predicting a purchase behavior of a consumer.

17. The method of claim 11, wherein organizing words into groups comprises:

grouping a first set of the words that pertain to a subject of a sentence;

grouping a second set of the words that pertain to a verb of the sentence; and

grouping a third set of the words that pertain to an object of the sentence.

18. The method of claim 11, wherein generating the vector representation of the group comprises

normalizing the vector representation of the group.

19. An engine for generating an inference about an object, the engine comprising:

at least one processor;

a memory including instructions that, when executed by the at least one processor, cause the at least one processor to;

access a set of data, each datum having a corresponding inference;

generate a vector representation for each datum in the data by

(i) parsing contents of the datum into objects,

(ii) organizing objects into groups;

(iii) for each group of objects, generating a vector representation of the group by

(a) obtaining a binding operator corresponding to a type of the group, the binding operator comprising a matrix of values,

(b) obtaining vector representations of the objects in the group,

(c) summing the vector representations of the objects, and

(d) multiplying the binding operator and the sum of the vector representations to generate a vector representation of the group; and

(iv) summing the vector representations of the groups to generate a vector representation of the datum; and

train an inference engine on the vector representations of the data and the corresponding inferences for each of the data.

20. The engine of claim 19, wherein each datum is an image.