Browse Prior Art Database

Treating Dates differently in QA systems

IP.com Disclosure Number: IPCOM000250305D
Publication Date: 2017-Jun-26
Document File: 2 page(s) / 103K

Publishing Venue

The IP.com Prior Art Database

Abstract

Disclosed is a process for a QA system that, given the binary nature of the date criteria and its almost universal applicability, uses the date as a filter to select/eliminate candidate answers that satisfy other parts of the question.

This text was extracted from a PDF file.
This is the abbreviated version, containing approximately 51% of the total text.

1

Treating Dates differently in QA systems

A common and very frustrating problem in QA systems is that a system can provide an answer that is correct in every dimension except time frame.

For example, the question is, “Who is the Republican Governor elected to the White House in 1980?” The answer is Ronald Reagan, who had been Governor of California. By all indications, it should be a simple answer to return. The user carefully worded the question to eliminate classic problems with “wh” words; it clearly seeks a person as the answer type. Extracting evidence about political party identification, status as a governor, elected, and office (i.e., White House) should allow the system to return the correct answer.

However, the same criteria could suggest incorrect answers: George W Bush (R, Texas 2000), Spiro Agnew (R, MD, and Vice President), Jimmy Carter (D, GA, 1976), or a host of candidates for President who fit all the criteria, but did not win. The current system might assign a higher score to those answers than it does to the correct response based on how it scores a given passage for language and context. In each of these cases, the main discriminator between correct and incorrect answers is the date.

A method is needed to differently treat dates from other features extracted from a question in an open domain QA system. The system cannot treat time as simply another feature to be used in the Machine Learning (ML) models to calculate P(Y=1). The date is a powerful discriminator between correct and incorrect answers in almost any substantive question. Unlike more nuanced indicators, it is binary; the case either does or does not fit the criterion.

The novel contribution is a process for a QA system that, given the binary nature of the date criteria and its almost universal applicability, uses the date as a filter to select/eliminate candidate answers that satisfy other parts of the question.

The steps for implementing the date filtering process in an open domain QA system follow:

1. Extract and score candidate answers 2. Separately, extract and retain date-time specific information from the question 3. Calculate P(Y=1) for each candidate answer, and rank the value in descending

order of probability (i.e., answer evaluated most likely to be correct is listed first) 4. In a new and separate step, submit each answer to a date-specific filter based on

temporal information extracted from the question 5. Remove answers not meeting the date criterion (with some tolerance pos...