Deep symbolic validation of information extraction systems

Patent metadata

Field	Value
Publication number	US-2020218968-A1
Application number	US-201916241569-A
Country	US
Kind code	A1
Filing date	Jan 7, 2019
Priority date	Jan 7, 2019
Publication date	Jul 9, 2020
Grant date	—

Abstract

A system comprises a memory that stores computer-executable components; and a processor, operably coupled to the memory, that executes the computer-executable components. The system includes a receiving component that receives a corpus of data; a relation extraction component that generates noisy knowledge graphs from the corpus; and a training component that acquires global representations of entities and relation by training from output of the relation extraction component.

First claim

What is claimed is:

1 . A system, comprising:
a memory that stores computer-executable components; a processor, operably coupled to the memory, that executes the computer-executable components stored in the memory, wherein the computer-executable components comprise:
a receiving component that receives a corpus of data;
a relation extraction component that generates noisy knowledge graphs from the corpus; and a training component that acquires global representations of entities and relation by training from output of the relation extraction component.

2 . The system of claim 1 , wherein the relation extraction component generates a set of quads from the corpus of data, wherein the quads have form q=<e1, r; e2; s> where eiϵV are entities found in the corpus of data, rϵR is a finite set of relations and sϵ[0, 1].

3 . The system of claim 2 , further comprising a perception component that implements function RelEx(e1; e2, Ø(e1; e2)) for each relation rϵR, returns a set of quads assessing their confidence from analysis of textual evidence.

4 . The system of claim 3 , wherein the textual evidence is: RelEx(e1, e2, Ø(e1, e2))=<e1, ri, e, si>riϵR, where si is a confidence score for relation ri.

5 . The system of claim 3 , further comprising a validation component that returns a confidence score for any possible triple such that e1, e2ϵV and rϵR.

6 . The system of claim 1 , wherein a mathematical loss function is implemented to account for confidence associated with triples

7 . The system of claim 1 , wherein the training is dependent upon noisy output of relation extraction.

8 . The system of claim 2 , wherein the loss function is dependent upon a specific mathematical function defines as:

ℒ
=

-

1



O
′





∑

i
∈

O
′



∑

i
=
1


ɛ




q
i

h
,
r


log




v
i

h
,
r

.

9 . The system of claim 1 , wherein the relation triples can identify threats in cybersecurity.

10 . The system of claim 3 , wherein the validation is implemented by using a deep net where a loss function is modified to account for fuzzy truth values provided by output of the perception component.

11 . A computer-implemented method, comprising:
receiving, by a processor operatively coupled to a memory, a corpus of data; generating via relation extraction, by the processor, noisy knowledge graphs from the corpus of data; and acquiring, by the processor, global representations of entities and relation by training from output of the relation extraction.

12 . The method of claim 11 , wherein the relation extraction generates a set of quads from the corpus of data, wherein the quads have form q=<e1, r; e2; s> where eiϵV are entities found in the corpus of data, rϵR is a finite set of relations and sϵ[0, 1].

13 . The method of claim 12 , further comprising performing a perception act that implements function RelEx(e1; e2, Ø(e1; e2)) for each relation rϵR, and returns a set of quads assessing their confidence from analysis of textual evidence.

14 . The method of claim 13 , wherein the textual evidence is: RelEx(e1, e2, Ø(e1, e2))=<e1, ri, e, si>riϵR, where si is a confidence score for relation ri.

15 . The method of claim 13 , further comprising performing a validation act that returns a confidence score for any possible triple such that e1, e2ϵV and rϵR.

16 . The method of claim 11 , wherein a mathematical loss function is implemented to account for confidence associated with triples

17 . The method of claim 1 , wherein the training is dependent upon noisy output of the relation extraction.

18 . The method of claim 12 , wherein the loss function is dependent upon a specific mathematical function defines as:

ℒ
=

-

1



O
′







?





∑

i
=
1


ɛ




q
i

h
,
r


log




v
i

h
,
r

.



Assignees

• IBM

Inventors

• Gliozzo Alfio Massimiliano
• Dash Sarthak
• Glass Michael Robert
• Canim Mustafa

Classifications

• G06N3/043

  based on fuzzy logic, fuzzy membership or fuzzy inference, e.g. adaptive neuro-fuzzy inference systems [ANFIS] · CPC title

• G06N3/0895

  Weakly supervised learning, e.g. semi-supervised or self-supervised learning · CPC title

• G06N3/0499

  Feedforward networks · CPC title

• G06N3/042

  Knowledge-based neural networks; Logical representations of neural networks · CPC title

• G06F21/554

  involving event detection and direct action · CPC title