Question and Answer Test-Train Overlap in Open Domain Question Answering Datasets

What happens when your test and training datasets overlap too much? How bad does it affect model generalization? This paper provides an empirical measurement of its impact.

Open-domain question answering is a popular research area, There are a lot of strong performing models that can achieve human-level performance in this domain. However, there seems to be a problem with those frequently used research datasets. This paper explores the test-train data overlap problem in three most popular open-domain QA datasets, and measures how much it inflates model performance.

0:00 - Intro 2:10 - Open domain question answering 3:39 - Datasets 5:27 - Random splitting 10:10 - Question overlap 12:48 - Statistics of overlap 13:52 - Implications for modelling 15:39 - Question memorization 17:23 - Simple duplicates 17:41- Sophisticate duplicates 19:31 - Experiments 23:57 - Nearest neighbor models 28:51 - Summary

Learn more about open-domain question answering https://youtu.be/JQ-bxQT5Qsw

Paper: Question and Answer Test-Train Overlap in Open Domain Question Answering Datasets https://arxiv.org/abs/2008.02637

Code: https://github.com/facebookresearch/qa-overlap

Abstract The dominant sequence transduction models are based on complex recurrent or convolutional neural networks that include an encoder and a decoder. The best performing models also connect the encoder and decoder through an attention mechanism. We propose a new simple network architecture, the Transformer, based solely on attention mechanisms, dispensing with recurrence and convolutions entirely. Experiments on two machine translation tasks show these models to be superior in quality while being more parallelizable and requiring significantly less time to train. Our model achieves 28.4 BLEU on the WMT 2014 English- to-German translation task, improving over the existing best results, including ensembles, by over 2 BLEU. On the WMT 2014 English-to-French translation task, our model establishes a new single-model state-of-the-art BLEU score of 41.0 after training for 3.5 days on eight GPUs, a small fraction of the training costs of the best models from the literature. We show that the Transformer generalizes well to other tasks by applying it successfully to English constituency parsing both with large and limited training data. ... https://www.youtube.com/watch?v=Cb5sj4_Ztfo