Text-based inference chaining

Patent metadata

Field	Value
Publication number	US-10438119-B2
Application number	US-201213679618-A
Country	US
Kind code	B2
Filing date	Nov 16, 2012
Priority date	Oct 12, 2012
Publication date	Oct 8, 2019
Grant date	Oct 8, 2019

Abstract

A method, system and computer program product for generating inference graphs over content to answer input inquiries. First, independent factors are produced from the inquiry, and these factors are converted to questions. The questions are then input to a probabilistic question answering system (PQA) that discovers relations which are used to iteratively expand an inference graph starting from the factors and ending with possible answers. A probabilistic reasoning system is used to infer the confidence in each answer by, for example, propagating confidences across relations and nodes in the inference graph as it is expanded. The inference graph generator system can be used to simultaneously bi-directionally generate forward and backward inference graphs that uses a depth controller component to limit the generation of both paths if they do not meet. Otherwise, a joiner process forces the discovery of relations that may join the answers to factors in the inquiry.

First claim

What is claimed is:

1. A system for inferring answers to inquiries comprising:
one or more unstructured content sources providing searchable content; and
a computer system having one or more hardware processor devices configured to:
receive an input inquiry submitted via a user interface;
decompose the input inquiry to obtain one or more factors, each factor being a logically independent statement contained in the received input inquiry and forming an initial current node of an inference graph; and
from said initial current nodes, iteratively construct said inference graph, wherein at each iteration, the hardware processor device is configured to:
determine, based on a text of a statement at an initial current node of said inference graph, a type of relation to define a new path between a statement of a respective said initial current node and a statement of a new inference node for a next successive level of nodes of said inference graph, a relation type selected from a group consisting of: a relation associating a statement of said current factor node with a statement of the new inference node, a relation indicating a statement of the new inference node by the statement of a current factor node; a relation associating a statement of the new inference node as a cause of the statement of a current factor node, a relation associating a statement of the new inference node as an affect caused by the statement of the current factor node; and a relation associating a statement of the new inference node that is contained within the statement of said current factor node;
apply one or more templates corresponding to a respective relation type to said statement text; and
generate natural language questions based on said applied templates;
search, using parallel Question-Answer systems, said one or more unstructured content sources using said generated natural language questions to generate said statement for said new inference node at said next successive level of said graph, said statement for said new inference node at said next successive level potentially comprising an inferred answer to said input inquiry;
extend said inference graph by adding said inference nodes having said statement at said next successive level of said inference graph, an added inference node connected to one or more said nodes at a prior level of said graph by a weighted edge, each said weighted edge representing an associated confidence level computed as a probability value that a relation is believed true based on a justifying passage obtained from the search,
form a Bayesian network from nodes and relations of said inference graph and each associated computed probability value of belief that a supporting passage justifies the answer for the node; and in each answer, and
propagate associated confidence values across said relations and nodes represented in said Bayesian network;
determine if the extended inference graph meets a criteria for terminating said iterative constructing, and based on the determining, one of:
terminate said iterative constructing if said criteria is met and provide an answer to said inquiry from said extended inference graph based on a confidence score associated with an inference node at a next successive level of nodes indicating a likelihood the answer is correct; otherwise

repeat, for a statement text of each respective added new inference node at a current level of said inference graph, and at each next successive level thereafter, the generating natural language questions based on respective templates corresponding to relation types applied to the statement text, the searching said unstructured content sources to obtain said statement for said inference nodes at said next successive level of nodes, and the extending of said inference graph by adding said inference nodes at said next successive level, until said criteria for terminating has been met, and
output to said user interface of said computer system the answer to said input inquiry from an added inference node of said extended inference graph.

2. The system as claimed in claim 1 , wherein said terminating criteria comprises:
identifying a node of said extended inference graph having an inferred confidence value exceeding a predetermined threshold; or,
performing a predetermined number of iterations.

3. The system as claimed in claim 1 , wherein said factors or current nodes comprise statements, said processor device further configured to, at each iteration:
prioritize selected statements as factors for corresponding question generation; or
filter selected statements and removing them as factors for corresponding question generation.

4. The system as claimed in claim 1 , wherein to extend the inference graph, the one or more hardware processor devices are further configured to:
compare a computed confidence against a threshold level, and prune certain inference nodes as a relation confidence of an inference node drops below said threshold level.

5. The system as claimed in claim 1 , wherein to generate said questions, the one or more hardware processor devices are further configured to, for a first iteration:
initially determine a set of seed relations based on said input inquiry and said one or more decomposed factors.

6. A computer program product for inferring answers to inquiries, the computer program product comprising a non-transitory storage device readable by a processing circuit and storing instructions run by the processing circuit for performing a method, the method comprising:
receiving an input inquiry submitted via a user interface;
decomposing the input inquiry to obtain one or more factors, each factor being a logically independent statement contained in the received input inquiry and forming an initial current node of an inference graph; and
from said initial current nodes, iteratively constructing said inference graph, wherein at each iteration, the processing circuit performs:
determining, based on a text of a statement at an initial current node of said inference graph, a type of relation to define a new path between a statement of a respective said initial current node and a statement of a new inference node for a next successive level of nodes of said inference graph, a relation type selected from a group consisting of: a relation associating a statement of said current factor node with a statement of the new inference node, a relation indicating a statement of the new inference node by the statement of a current factor node; a relation associating a statement of the new inference node as a cause of the statement of a current factor node, a relation associating a statement of the new inference node as an affect caused by the statement of the current factor node; and a relation associating a statement of the new inference node that is contained within the statement of said current factor node;
applying one or more templates corresponding to a respective relation type to said statement text; and
generating natural language questions based on said applied templates;
searching, using parallel Question-Answer systems, one or more unstructured content sources using said generated natural language questions to generate said statement for said new inference node at said next successive level of said graph, said statement for said new inference node at said next successive level potentially comprising an inferred answer to said input inquiry;
extending said inference graph by adding said inference nodes having said statement at said next successive level of said inference graph, an added inference node connected to one or more said nodes at a prior level of said graph by a weighted edge, each said weighted edge representing an associated confidence level computed as a probability value that a relati

Assignees

• IBM

Inventors

• Buchanan David W
• Ferrucci David A
• Lally Adam P

Classifications

• G06F40/30

  Semantic analysis · CPC title

• G06N5/02 Primary

  Knowledge representation; Symbolic representation · CPC title