Reference QA Benchmarking Dataset
This repository contains factoid-curated, a reference question dataset for benchmarking Question Answering systems, as used e.g. by the YodaQA system.
The complete dataset is a combination of two sub-datasets (irc/ and trec/) and consists of three files:
curated-full.tsvis the complete set of questions that are part of this dataset. If you modify this set, you should also tweak the appropriate sub-dataset.
curated-train.tsvis the standard train split of the dataset, for primary development, detailed performance analysis and training (and if possible also testing) machine learning algorithms.
curated-test.tsvis the standard test split of the dataset, for benchmarking and reporting system performance. Everyone should attempt to treat this dataset as "blind" and do not analyze or optimize performance for individual questions in this dataset!
trecnew-raw-test.tsvis a variant of the test split that is not curated, also containing questions we have removed from the dataset during curation and their original phrasings and validation regexes; it contains the TREC 2002, TREC 2003 year datasets. This dataset is meant for reporting system performance in the context of papers that do not use our curated data.
(Some small portion of questions may be left out of the splits; these are included in large2470-train, though.)
Ideally, humans should be doing all stages of evaluation instead of just using regex matches, as time by time an unconcieved legitimate answer pops up and on the other hand, sometimes the regex is unintentionally over-permissive. For a similar reason (to allow e.g. leading the/a and other variations), we declare the answer as correct when the regex matches any substring.
Using This Dataset
As explained above, please use the test dataset only for performance reporting, not for question-by-question error analysis. Always report the version of the dataset you used - this is the v2 dataset
To make results comparable, it is not enough to use the same set of questions, we should strive to use the same or similar knowledge bases as well; we realize that especially as time goes, this might not be practical, but some degree of effort would be appreciated. We also expect the primary public YodaQA endpoints to track the latest version of this dataset, so these could be reused. We use:
- enwiki-20150112 (archived at http://v.or.cz/~pasky/brmson/)
- Freebase RDF dump from Jan 11 2015 (but it's probably not very volatile)
- DBpedia 2014
- WordNet 3.1
Aims of This Dataset
We want to build a dataset of questions, which are:
Factoid. This means that there is only a single answer (not a list or a sequence [of steps]), the answer is typically a simple phrase (not a story or topic summary; definition query like "Who is Obama?" is borderline and does not occur much) and the question is not a boolean yes/no (which requires quite different techniques to answer).
Wikipedia focused. Wikipedia is a vast, unstructured repository of knowledge, so the bias is not terribly high, but we would like to think that a human with Wikipedia in hand would be able to answer all of these questions. Most of them should require just a straightforward information retrieval, some are more challenging and might need some inference.
We may want to relax either requirement, but at that point we should start tagging the questions to still keep a set of "simpler" ones. The motivation is to give a chance even to simple, focused systems.
Large Variant of the Dataset
An extra variant of the dataset is provided, which is a strict superset of the curated dataset. It is not curated, i.e. it is a lot more noisy and may not fulfill the above constraints. The aim of this dataset is to check behavior of QA systems when given larger, noisy training data, exploring generalization capabilities of systems.
The large2470 dataset is built by adding TREC 1999, 2000, 2001 data to the curated dataset. We also added questions asked on live.ailao.eu with user feedback until Feb 25, 2016. Unmodified questions were shuffled and 80% were added to the train split, 20% to the test split. Gold standard was revised by Mechanical Turk.
The dataset is called large2470, which refers to the number of questions in the dataset. We may build even larger datasets (e.g. including the WebQuestions, TREC years 2004+, QALD challenges or such) of similar nature in the future.
Discussion: How much data is appropriate?
It might be obvious to just use the largest dataset possible, but there are three issues with that:
Development convenience. Many QA systems are pretty slow, and just providing a big dataset may tempt system authors to create custom splits of the datasets for their development, something we would like to discourage.
Noise. The large2470 dataset is already relatively low quality, some questions are really arbitrary and not answerable using public knowledge bases. The problem is even worse when we also consider the answer patterns, which may unfairly disqualify many correct answers, and worse this adds question type bias (probably almost no numerical quantity answers will pass the pattern muster, but years or names typically will). Beyond the large2470 dataset, e.g. the WebQuestions dataset also contains typos in questions or downright mistakes; this is not unrealistic, but adds extra burden to authors of early systems; furthermore, the answer patterns based on Freebase will create an unfair bias towards FB in multi-source systems. So the noise issue is not just about noisy measurements, but also introducing biases, not all of them obvious.
Realistic setting. In the end, one aspect of QA systems we consider important is domain adaptability; but when moving the QA system to a particular domain, it is unlikely there will be a large dataset available to train it. So observing the scaling behavior of a system and its small-data performance is completely reasonable.