System and method for domain adaptation in question answering

• US 9,805,613 B2
• Filed: 09/15/2015
• Issued: 10/31/2017
• Est. Priority Date: 05/14/2008
• Status: Active Grant

• Current Assignee: International Business Machines Corporation
• Original Assignee: International Business Machines Corporation
• Inventors: Bagchi, Sugato, Ferrucci, David A., Gondek, David C., Levas, Anthony T., Zadrozny, Wlodek W.
• Primary Examiner(s): Chbouki, Tarek
• Application Number:
  US14/854,579
• Publication Number:
  US 20160005325A1
• Time in Patent Office: 777 Days
• CPC Class Codes: G09B 7/00 Electrically-operated teaching apparatus or devices working with questions and answers mechanically operated G09B3/00; computing arrangements G06F

Abstract

A method for providing adaptation to a question answering (QA) system having an associated plurality of trace data and a question-answer set. The method includes submitting a set of questions to the QA system; receiving back from the QA system a set of answers; comparing the set of answers to answers in the question-answer set; and generating the plurality of trace data based on comparison, and an estimate of how much more training data is needed. The generating comprises (a) for each question type, successively sample an increasing number of question-answer pairs from the question-answer set; (b) automatically train the QA system using the sampled question-answer pairs; (c) automatically compute a functional dependence of the QA system performance on the remaining questions relative to the size of the sample; and (d) extrapolate from the functional dependence a number of training question-answer pairs of each question type.

Claims

1. A method for providing adaptation to a question answering system, wherein the question answering system has associated therewith a plurality of trace data and a question-answer set, the method comprising the steps of:

submitting a set of questions to the question answering system;

receiving back from the question answering system a set of answers generated in response to the set of questions;

comparing the set of answers received back from the question answering system to answers in the question-answer set; and

generating, based on the comparison and the plurality of trace data, an estimate of how much more training data is needed for each of a plurality of question types or answer types;

wherein the training data comprises a plurality of training questions along with a corresponding plurality of correct training answers;

wherein the trace data is generated during the generation of the set of answers; and

wherein the step of generating further comprises the following steps;

(a) for each question type, successively sample an increasing number of question-answer pairs from the question-answer set;

(b) automatically train the question answering system using the sampled question-answer pairs;

(c) automatically compute a function relating the question answering system performance on the remaining questions relative to the number of successively sampled question-answer pairs; and

(d) extrapolate from the function a number of training question-answer pairs of each question type.

2. The method of claim 1, wherein the plot of performance relates to at least one of:

(a) accuracy; and

(b) recall.

3. The method of claim 1, wherein the steps are carried out in the order recited.