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

By Patrick Lewis, Pontus Stenetorp, Sebastian Riedel. 2020

198 citations recorded by the source corpus.

Ideally Open-Domain Question Answering models should exhibit a number of competencies, ranging from simply memorizing questions seen at training time, to answering novel question formulations with answers seen during training, to generalizing to completely novel questions with novel answers. However, single aggregated test set scores do not show the full picture of what capabilities models truly have. In this work, we perform a detailed study of the test sets of three popular open-domain benchmark datasets with respect to these competencies. We find that 30% of test-set questions have a near-duplicate paraphrase in their corresponding train sets. In addition, we find that 60-70% of answers in the test sets are also present in the train sets. Using these findings, we evaluate a variety of popular open-domain models to obtain greater insight into what extent they can generalize, and what drives their overall performance. We find that all models perform substantially worse on questions that cannot be memorized from train sets, with a mean absolute performance difference of 61% between repeated and non-repeated data. Finally we show that simple nearest-neighbor models outperform a BART closed-book QA model, further highlighting the role that train set memorization plays in these benchmarks

Subfields: Artificial Intelligence, Computer Vision and Pattern Recognition, Natural Language Processing, Evaluation & Benchmarks

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Artificial Intelligence

In Computer Vision and Pattern Recognition

In Natural Language Processing

In Evaluation & Benchmarks

Citations in the corpus

Improving language models by retrieving from trillions of tokens (2021, background): “However, significant leakage between train and test datasets (Lee et al., 2021; Lewis et al., 2021) makes comparing and evaluating large models trained on large datasets difficult, especially once retrieval capabilities over the training dataset are added.”
Ethical and social risks of harm from Language Models (2021, background): “For example, (Lewis et al., 2020) report that the Exact Match score of a Q&A T5-11B model dropped from 44.7 to 22.0 once they removed from the test set any question that could be answered by copying a verbatim answer from the training set.”
WebGPT: Browser-assisted question-answering with human feedback (2021, methodology): “Previous work on question-answering such as REALM [Guu et al., 2020] and RAG [Lewis et al., 2020a] has focused on improving document retrieval for a given query.”
Autoregressive Search Engines: Generating Substrings as Document Identifiers (2022, methodology): “To better understand the generalization capabilities of our retrieval so-lution we use the question/answer overlap split of Lewis et al. (2021b).”
Learning to Retrieve Passages without Supervision (2021, methodology): “However, their approach utilizes PAQ (Lewis et al., 2021b), a dataset which was generated using models trained on NQ, and is therefore not unsupervised.”
R2-D2: A Modular Baseline for Open-Domain Question Answering (2021, methodology): “In addition to evaluation on the TQ-Open and NQOpen shown in Table 2, we also report results on subsets of these datasets in Table 12, as split by Lewis et al. (2021).”
Time-Aware Language Models as Temporal Knowledge Bases (2021, background): “Splitting along subject entities ensures that none of the facts required to answer the test queries are seen during training on T EMP LAMA (Lewis et al., 2021).”
Teaching language models to support answers with verified quotes (2022, methodology): “To keep the cost of human evaluation manageable, and avoid train-test overlap (Lewis et al., 2021b), we use subsets of test questions from these standard question-answering benchmarks.”
Dynabench: Rethinking Benchmarking in NLP (2021, neutral): “…competencies (Geiger et al., 2018; Naik et al., 2018; Glockner et al., 2018; White et al., 2018; Warstadt et al., 2019; Gauthier et al., 2020; Hossain et al., 2020; Jeretic et al., 2020; Lewis et al., 2020; Saha et al., 2020; Schuster et al., 2020; Sugawara et al., 2020; Warstadt et al., 2020).”
PAQ: 65 Million Probably-Asked Questions and What You Can Do With Them (2021, methodology): “It has recently been shown that current closed-book models mostly memorise training QA-pairs, and can struggle to answer questions that do not overlap with training data (Lewis et al., 2020b).”
Hurdles to Progress in Long-form Question Answering (2021, neutral): “Simple baselines such as just repeatedly copying the question, or choosing a random training set answer, can outperform LFQA systems such as RAG (Lewis et al., 2020c) in terms of ROUGE-L.”
Mind the Gap: Assessing Temporal Generalization in Neural Language Models (2021, background): “…both LM and downstream task evaluations more broadly, e.g., in widely used QA datasets like Natural Questions (Kwiatkowski et al., 2019) and TriviaQA (Joshi et al., 2017a), which have been shown to contain alarmingly high proportions of overlapping training and test data (Lewis et al., 2020c).”
Improving Question Answering Model Robustness with Synthetic Adversarial Data Generation (2021, background): “Many approaches have been proposed to generate question-answer pairs given a passage (Du et al., 2017; Du and Cardie, 2018; Zhao et al., 2018; Lewis and Fan, 2019; Alberti et al., 2019; Puri et al., 2020; Lewis et al., 2021b).”
Entity-Based Knowledge Conflicts in Question Answering (2021, methodology): “For simplicity we use the MRQA Workshop Shared Task’s versions for each of these datasets where the same tokenization and pre-processing are used (Fisch et al., 2019).3 Lewis et al. (2021) show the Natural Questions training and development sets contain many similar queries and answers.”
UnitedQA: A Hybrid Approach for Open Domain Question Answering (2021, methodology): “Table 6: Breakdown evaluation on NaturalQuestions (NQ) and TriviaQA based on test splits defined in (Lewis et al., 2021).”