Predicting Entailment – Gavin Hackeling

Predicting Entailment – Gavin Hackeling

1 1


A slide deck based on reveal.js

On Github gavinmh / itp-thesis-presentation

Predicting Entailment

Gavin Hackeling

In the first part of this presentation, I will describe a language-processing capability.

In the second part I will demonstrate one application of this capability.

Part I:

Predicting Entailment in Unstructured Text

Entailment is the relationship between sentences that holds when one sentence follows from another.

I ate food.

follows from

I ate pizza.

Carolina beat Duke.

follows from

North Carolina defeated Duke after crushing State.

A student presented a project.

does not follow from

No student presented.

I have built a system that predicts the entailment relations that exists between pairs of plain English texts.

Predicting entailment can allow applications to understand what a text means in relation to another.


change url. need to setup post route and do ajax response and unclude examples First Text: Second Text:

Part II:

An Application of Entailment Prediction: Polar Question Answering

First, a note about search engine usage:

People often use search engines to find answers, not documents.

Many people use English sentences as their queries rather than keywords.

quotes and statistics here about search engine usage

Search engines, and assistants like Google Now and Siri, sometimes surface answers to a type of question called factoids.

These systems use structured data sources to answer factoid questions.

Producing such structured data is expensive and time-consuming.

Users like this feature, and have become habituated to it.

Statistics about habituation

The problem is that not all questions are factoids, and search engines do not answer many of these types of questions.

These other types include polar, or yes-or-no questions.

Some polar questions are difficult to rephrase as factoid questions.

Are cats crepuscular?

Did Ariel Sharon have a stroke?

Is malt an ingredient in beer?

Can Bitcoin transactions be tracked?

In my tests, users often attempted to rephrase polar questions as factoid questions.

When the polar question could not be rephrased, they used the polar question verbatim as their search query.

The answers to polar questions can sometimes be found in the descriptions of the top results, but no search engine directly surfaces an answer.

In the mobile, wearable future of computing, many interfaces will not be capable of supporting long pages of search results.

Users want services that answer their questions with as little work as possible.

Entailment prediction can be used to automatically answer polar questions using only unstructured text.

A system that answers polar questions takes one text, the user's question, as input.

It then uses a search engine to retrieve similar texts from the web.

The system predicts the entailment relations between these retrieved texts and the user's question. It uses these relations to predict a yes or no answer. High-scoring texts can be returned as justification.

If I ask "was Millard Fillmore the last Whig President?" the search engine might retrieve the following sentences:

Millard Fillmore was the last President who was neither a Democrat nor a Republican

Millard Fillmore was the 13th president.

Millard Fillmore was the 13th president and the final Whig president.

Millard Fillmore was the last President who was neither a Democrat nor a Republican

Millard Fillmore was the 13th president.

Millard Fillmore was the 13th president and the final Whig president.

The system predicts relations for each of these pairs, and uses the relations to predict a yes or no answer.

Automatically answering polar questions from unstructured text lets people use search engines intuitively; it solves a problem that will be exacerbated as mobile and wearable interfaces become more prevalent.

Entailment prediction systems will be used in other applications as they become more robust.