This repository contains python code for experiments relating to RTE. At the moment, the main thing of interest here is the Guardian Headlines Entailment Training Dataset.
Recognising Textual Entailment (RTE) is the task of determining, given two sentences whether the first (called the "text") entails or implies the second (called the "hypothesis"). In this task, the term "entailment" is generally fairly loosely defined, and datasets are normally built by asking human subjects whether they consider entailment to hold.
RTE as a field of study really kicked off with the Recognising Textual Entailment Challenge.
Textual Entailment is a generalisation of many tasks in natural language processing. If you have a system that is good at recognising textual entailment, it should be easier to build good systems for information retrieval, question answering, paraphrase recognition, information extraction and summarisation.
Because of the expense of manually constructing entailment datasets, they are normally fairly small, which means machine learning approaches to the task perform sub-optimally.
As Hickl et al. showed, automatically constructed datasets can improve the performance of systems using machine learning by up to ten percent.
The dataset consists of around 32,000 pairs of sentences (16,233 for which entailment does hold and 16,249 for which it doesn't) automatically extracted from The Guardian newspaper using their API. We follow a similar methodology to Hickl et al.: we treat headlines as being entailed by the first sentence, and adjacent sentences in the remainder of the text as non-entailing.
Each pair must pass a number of criteria (arrived at in a fairly ad-hoc manner):
- The pair must share a named entity (or part of a named entity) in common (as required by Hickl et al.)
- Each sentence in the pair must not be too short or long
- Each sentence must not contain too many new line characters
- Each sentence must contain an even number of quotation marks
The source of the data is 78,696 Guardian articles from 1st January 2004 onwards obtained through the Guardian API.
No analysis has yet been performed on the dataset, so use it at your own risk! The intention is eventually to manually analyse a sample of the data.
The data is in XML format and is the same as the RTE-1 dataset:
<pair id="1" value="TRUE">
<t>
Italian authorities yesterday blocked mail from Bologna addressed
to EU institutions as they tried to end a letter-bomb assault that
has been aimed at European targets.
</t>
<h>
Mail block to catch EU book bombs
</h>
</pair>
Permission is hereby granted, free of charge, to any person obtaining a copy of this software and associated documentation files (the "Software"), to deal in the Software without restriction, including without limitation the rights to use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, and to permit persons to whom the Software is furnished to do so, subject to the following conditions:
The above copyright notice and this permission notice shall be included in all copies or substantial portions of the Software.
THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.