Browse Prior Art Database

Projection based scoring of candidate answers in a question answering system

IP.com Disclosure Number: IPCOM000239281D
Publication Date: 2014-Oct-27
Document File: 3 page(s) / 142K

Publishing Venue

The IP.com Prior Art Database

Abstract

Question answering systems (like watson) initially generate a lot of possible candidate answers in order to increase the recall. The candidate answers are, then, scored using various techniques to increase the precision. Many times the single document may contain the evidences for the candidate answer, but it may not be contiguous thus resulting into low score given to the answer. The proposal is to score a candidate answer using the projection of a document with respect to a given set of concepts (derived from question analysis and candidate answers).

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 52% of the total text.

Page 01 of 3

Projection based scoring of candidate answers in a question answering system
Question answering systems (like watson) initially generate a lot of possible candidate answers in order to increase the recall. The candidate answers are, then, scored using various techniques to increase the precision.

Many times the single document may contain the evidences for the candidate answer, but it may not be contiguous thus resulting into low score given to the answer. proposal is to score a candidate answer using the projection of a document with respect to a given set of concepts (derived from question analysis and candidate answers)

Typically the existing methods score an answer based on contiguous segments of text.

Projections are considered which are a sequence of segments of text that need not be contiguous.

The projections are summaries of documents with respect to a given set of concepts.

Described below is a high level diagram to illustrate the overall object of scoring based on projections. High Level Diagram:

Score calculation:


The score giving to a candidate answer is the function of overlap of the projections from concepts of question and candidate answer.

Score = f (overlap of projections from question and answer)

The way overlap could have multiple features and attributes like token overlap, named entity overlap, similar noun, verb phrases and so on. All these features would be assigned weight to ultimately come up with an overall score for the candidate answer.

Thus, lead t...