Browse Prior Art Database

A method of finding historical answers to questions in a QA system

IP.com Disclosure Number: IPCOM000242183D
Publication Date: 2015-Jun-23
Document File: 3 page(s) / 117K

Publishing Venue

The IP.com Prior Art Database

Abstract

A method for finding historical answers to questions in a question answering (QA) system is disclosed.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 43% of the total text.

Page 01 of 3

A method of finding historical answers to questions in a QA system

Disclosed is a method for finding historical answers to questions in a question answering (QA) system. A complete history of the corpus and models of a QA system

are maintained, allowing users to ask questions of the system as it existed at any point in time. Support is provided to ask time based questions and to show how an answer to a question would have changed over the life of the system.

Within a QA system, the corpus changes over time. New documents are ingested, other documents are replaced with newer versions and so on. As the corpus matures,
it may be necessary/useful to see how the system would have answered a question at a previous point in time. This article discloses a method for tracking corpus changes over time so that the system can answer time based questions from previous versions of the corpus.

Take the example of a financial customer, where the following questions were asked: "What is the current market forecast for CompanyA" vs "What was the market forecast for CompanyA yesterday".

If the QA system were ingesting a document that provided a market forecast for CompanyA on a daily basis, the answer to "What was the market forecast for CompanyA yesterday" would come from a previous version of that document and not

the latest version. The QA system in general would want to run with the latest documents only, so that a question such as: "What is the market forecast for CompanyA" wouldn't return an answer from an older version of the document. So the system should provide a way to access the older version of the documents only when necessary.

When a system is first set up and a corpus is created, the initial set of documents becomes the base corpus for the system. By way of example, a QA system ingests three documents at first. The base corpus is shown with only the three documents in Figure 1.

DocA_v1

DocB_v1 DocC_v1

Figure 1

Subsequently in the example, a new document is ingested and one of the existing documents is replaced without retraining as depicted in Figure 2. In this example DocA_v1 is replaced with DocA_v2 and DocD_v1 is added.

DocA_v1

DocB_v1 DocC_v1

DocA_v2
DocB_v1
DocC_v1

06/09/2014

Trained_Model_

1

06/09/2014

Trained_Model_

1

06/10/2014

1


Page 02 of 3

DocD_v1

Figure 2

The example continues where the system deletes DocB_v1 and is retrained as depicted in Figure 3.

DocA_v1 DocB_v1 06/09/2014

DocC_v1 DocA_v2 DocB_v1 DocC_v1

Trained_Model_

1

06/10/2014

DocD_v1

06/11/2014

DocD_v1

DocA_v2 DocC_v1

Trained_Model_

2

Figure 3

The disclosed system does not maintain 3 complete versions of the corpus, instead it only store the differences from the base. However, all versions of the docs are kept (in ingested form), even the deleted ones. Various use cases may take advantage of the

access to the information from different versions of the documents with restrictions to time.

Scenario 1: System provides answers to a given question based on the previous...