A Qualitative Comparison of CoQA, SQuAD 2.0 and QuAC

A Qualitative Comparison of CoQA, SQuAD 2.0 and QuAC

Mark Yatskar
Allen Institute for Artificial Intelligence

We compare three new datasets for question answering: SQuAD 2.0, QuAC, and CoQA, along several of their new features: (1) unanswerable questions, (2) multi-turn interactions, and (3) abstractive answers. We show that the datasets provide complementary coverage of the first two aspects, but weak coverage of the third. Because of the datasets’ structural similarity, a single extractive model can be easily adapted to any of the datasets and we show improved baseline results on both SQuAD 2.0  and CoQA. Despite the similarity, models trained on one dataset are ineffective on another dataset, but we find moderate performance improvement through pretraining. To encourage cross-evaluation, we release code for conversion between datasets at https://github.com/my89/co-squac.


noitemsep,topsep=5pt,parsep=5pt,partopsep=5pt \mdfdefinestyledialogue backgroundcolor=yellow!20, innermargin=5pt \aclfinalcopy

Dataset Entity Salad False Premise Topic Error Missing Information Content Negation Answerable Questions Total Questions
CoQA 0.0 0.0 0.0 60.0 0.0 40.0 5 (0.5%)
SQuAD 2.0 21.3 21.3 13.5 16.1 16.1 10.9 230 (50.1%)
QuAC 5.5 0.0 16.4 71.2 0.0 6.8 73 (20.2%)
Table 1: Comparison of unanswerable questions on 50 random contexts from the development set of each dataset. SQuAD 2.0 contains a diverse set of circumstances that make questions unanswerable, QuAC focuses on information that could plausibly be in context material and CoQA does not significantly cover unanswerable questions.

1 Introduction

Question answering on textual data has served as a challenge problem for the NLP community Voorhees (2001); Richardson et al. (2013). With the development of large scale benchmarks and sufficiently simple evaluations  Trischler et al. (2016); Nguyen et al. (2016); Hermann et al. (2015) progress has been rapid. In recent evaluation on SQuAD  Rajpurkar et al. (2016), performance exceeded that of annotators  Wang et al. (2018); Hu et al. (2017); Wang et al. (2017).

In response to this development, there have been a flurry of new datasets. In this work, we analyze three such new proposed datasets, SQuAD 2.0 Rajpurkar et al. (2018), QuAC Choi et al. (2018), and CoQA Reddy et al. (2018).111A review of other new datasets is in the related work. In each of these datasets, crowd workers are asked to (1) produce questions about a paragraph of text (context) and (2) produce a reply by either indicating there is no answer, or providing an extractive answer from the context by highlighting one contiguous span. QuAC and CoQA contain two other features: questions are asked in the form of a dialog, where co-reference to previous interactions is possible and directly answering yes/no is possible. CoQA also allows workers to edit the spans to provide abstractive answers.222Also, SQuAD 2.0 and QuAC  cover only Wikipedia text, CoQA  covers six other domains and QuAC is the only one of these datasets that doesn’t allow the questioner to see the context before formulating a question.

We compare these three datasets along several of their new features: (1) unanswerable questions, (2) multi-turn interactions, and (3) abstractive answers. Unanswerable question coverage is complementary among datasets; SQuAD 2.0 focuses more on questions of extreme confusion, such as false premise questions, while QuAC primarily focuses on missing information. QuAC and CoQA  dialogs simulate different types of user behavior: QuAC dialogs often switch topics while CoQA dialogs include more queries for details. Unfortunately, no dataset provides significant coverage of abstractive answers beyond yes/no answers, and we show that a method can achieve an extractive answer upper bound of 100 and 97.8 F1 on QuAC  and CoQA , respectively.

Motivated by the above analysis, we apply the baseline presented in QuAC Choi et al. (2018), BiDAF++, a model based on BiDAF Seo et al. (2016), augmented with self attention Clark and Gardner (2018) and ELMo contextualized embeddings Peters et al. (2018) to all datasets. Experiments show that this extractive baseline outperforms existing extractive and abstractive baselines on CoQA by 14.2 and 2.7 F1 respectively. Finally, we show models can transfer between datasets with pretraining yielding moderate gains.333To facilitate easy future cross-evaluation, we release tools for conversion between these dataset.

Dataset Topic Shift Drill Down Return to Topic Clarification Question Definition Question Sentence Coverage Total Questions
CoQA 21.6 72.0 2.9 0.0 0.7 63.3 722
QuAC 35.4 55.3 5.6 0.7 3.0 28.4 302
Table 2: Comparison of dialog features in 50 random contexts from the development set of each dataset. CoQA  contains questions that drill into details about topics and cover 60% of sentences in the context while in QuAC  dialog switch topic more often and cover less than 30% of sentences. Neither dataset has a significant number of returns to previous topics, clarifications, or definitional interactions.
Dataset Yes/No Coref Counting Picking Fluency Max F1
CoQA 21.4 3.2 1.3 0.6 4.2 97.8
QuAC 21.1 0.0 0.0 0.0 0.0 100.0
Table 3: Comparison of abstractive features in 50 random contexts in the develoment set of each dataset. Both QuAC and CoQA contain yes/no questions while CoQA also contains answers that improve fluency through abstractive behavior. The extractive upper bound from CoQA is high because most absractivive answers involve adding a pronoun (Coref) or inserting prepositions and changing word forms (Fluency) to existing extractive answers, resulting in extremely high overlap with possible extractive answers.

2 Dataset Analysis

In this section we analyze unanswerable questions, dialog features, abstractive answers in SQuAD 2.0, QuAC, and CoQA. All analysis was performed by the authors, on a random sample of 50 contexts (300-700 questions) from the development set of each dataset.

2.1 Unanswerable Questions

In Table 1 we compare types of unanswerable questions across dataset. We identify five types of questions found between the datasets:

1. Entity Salad

A nonsensical reference to entities found in the context or made-up entities (e.g. “What infinite hierarchy implies that the graph isomorphism problem s NQ-complete?”). Such questions are unanswerable for any context.

2. False Premise

A fact that contradicts the context is asserted in the question (e.g. “When is the correlation positive?” but in the context says “the correlation is strictly negative”).

3. Topic Error

A questions that references an entity in the context but the context does not focus on that entity (e.g “How many earthquakes occur in California?” when the article focus is actually about “Southern California” ). Such questions potentially have answers, but it would be unlikely for the answer to be found in the context.

4. Missing Information

A question who’s answer could be plausibly in the context but is not (e.g. “What is the record high in January?” and the article is about temperature extremes). Such questions have an answer but it is not mentioned.

5. Content Negation

A question which asks for the opposite information of something mentioned in the context (e.g. “Who didn’t cause the dissolution of the Holy Roman Empire?”). Such questions either have answers that are the set of all entities other than the one mentioned or answers that could be found in some other context.


SQuAD 2.0 contains the highest diversity of unanswerable questions of all datasets analyzed. Some SQuAD 2.0 questions are unlikely to be asked without significant foreknowledge of the context material and do not occur in QuAC. 444Such questions resemble text from entailment datasets such as SNLI Bowman et al. (2015) and seem more likely to arise if questioners are receiving very complex information and become confused. Both SQuAD 2.0  and QuAC  cover a significant number of unanswerable questions that could be plausibly in the article. The difference in settings and distributions of unanswerable questions in SQuAD 2.0  and QuAC  appear to be complementary: SQuAD 2.0  focuses more on questions simulating questioner confusion, while QuAC  primarily focuses on missing information. 555CoQA  does not contain a significant number of unanswerable questions, and many of the ones that do exist are erroneously marked.

2.2 Dialog Features

In Table 2 we analyze five dialog behaviors:

1. Topic Shift

A question about something previously discussed (e.g. “Q: How does he try to take over? … Q: Where do they live?”).

2. Drill Down

A request for more information about a topic being discussed (e.g. “A: The Sherpas call Mount Everest Chomolungma. Q: Is Mt. Everest a holy site for them?”)

3. Topic Return

Asking about a topic again after it had previously been shifted away from.

4. Clarification

Reformulating a question that had previously been asked.

5. Definition

Asking what is meant by a term (e.g. “What are polygenes?”)


QuAC and CoQA  contain many similar features but at very different rates, offering complementary coverage of types of user behavior. CoQA  dialogs drill down for details significantly more frequently and cover more than 60% of sentences in the context material (Sentence Coverage). QuAC dialogs shift to new topics frequently and cover less than 30% of sentences in the context. Both datasets contain only a small numbers of definition questions and returns to previous topics and few requests for clarification.

2.3 Abstractive Answers

Table 3 compares abstractive behavior in CoQA and QuAC. We observed five phenomena:

1. Yes/No

Questions annotated with yes/no. In QuAC such questions and their corresponding yes or no are marked in addition to an extractive answer. In CoQA, the single token “yes” or “no” is simply asserted as the abstractive answer, with an extractive answer provided in the rationale (e.g. “Q: Is atmosphere one of them? A: yes”).

2. Coref

Coreference is added to previously mentioned entities in either context or question (e.g. “Q: How was France’s economy in the late 2000s? A: it entered the recession”).

3. Count

Counting how many entities of some type were mentioned (e.g. “Q: how many specific genetic traits are named? A: five”)

4. Picking

A question that requires the answer to pick from a set defined in the question (e.g. “Q: Is this a boy or a girl? A: boy)

5. Fluency

Adding a preposition, changing the form of a word, or merging two non-contiguous spans (e.g. “Q: how did he get away? A: by foot)


Both QuAC and CoQA have a similar rate of yes/no questions. QuAC contains no other abstractive phenomena while CoQA contains a small number of predominately insertions, often at the beginning of an extractive span, for coreference and or other fluency improvements. Because abstractive behavior in CoQA  includes mostly small modifications to spans in the context, the maximum achievable performance by a model that predicts spans from the context is 97.8 F1. 666To compute the upper bound, if abstractive answer is exactly “yes”, “no”, or “unknown”, we consider the upper bound to be 100. Otherwise, we use the CoQA evaluation script to find a span in the context that has maximum F1 with respect to the abstractive answer.

Overall F1
DrQA (Extractive) 54.7
DrQA + PGNet (Abstractive) 66.2
BiDAF++ w/ 0-ctx 63.4
BiDAF++ w/ 3-ctx 69.2
Table 4: Development set performance by training BiDAF++ Choi et al. (2018) models (extractive) on CoQA data with handling yes/no and no-answer questions as in QuAC. Despite being extractive, these models significantly outperform reported baselines, DrQA and DrQA + PGNet  Reddy et al. (2018).
in-F1 out-F1 F1
DrQA 54.5 47.9 52.6
DrQA + PGNet 67.0 60.4 65.1
BiDAF++ w/ 3-ctx 69.4 63.8 67.8
Table 5: Test set results on CoQA. We report in domain F1 (in-F1), out of domain F1 on two held out domains, Reddit and Science (out-F1) and the overall F1 (F1).

3 New Extractive Baseline for CoQA

Our analysis strongly implies that beyond yes/no questions, abstractive behavior is not a significant component in either QuAC  or CoQA. As such, QuAC models can be trivially adapted to CoQA.

We train a set of BiDAF++ baselines from the original QuAC  dataset release Choi et al. (2018) by optimizing the model to predict the span with maximum F1 overlap with respect to annotated abstractive answers.777We use the implementation on http://allennlp.org, and do not modify any hyper-parameters except the the maximum dialog length and that models were allowed to train up to 65 epochs. If the abstractive answer is exactly “yes” or “no”, we train the model to output the whole rationale span, and classify the question as yes/no with the appropriate answer. At evaluation time, if the model predicts a question is a yes/no question, instead of returning the extracted span, we simply return “yes” or “no”.


Table 4 and Table 5 summarize our results for training BiDAF++ with varying contexts on CoQA. Beyond the difference of underlying base question-answer models (DrQA Chen et al. (2017) vs. BiDAF Seo et al. (2016) with self attention Clark and Gardner (2018)), BiDAF++ has two core differences with respect to DRQA+PGNet: (1) instead of appending previous questions and answers to input question tokens, BiDAF++ marks answers of previous questions directly on the context, and (2) BiDAF++ uses contextualized word embeddings through ELMo Peters et al. (2018). These differences, in combination with appropriate handling of yes/no and unanswerable questions significantly improves on the existing extractive baseline (+14.2 F1) and even on the existing abstractive baseline (+2.7 F1).

BiDAF++ w/ 2-ctx 60.6 55.7 4.0
Train SQuAD 2.0 34.3 18.0 0.3
Train CoQA 31.2 19.2 0.0
Ft from SQuAD 2.0 62.6 58.3 5.9
Ft from CoQA 63.3 59.2 5.1
Table 6: Cross dataset transfer to QuAC development set. Models do not transfer directly (rows 3 and 4), but after fine tuning improve performance (rows 5 and 6).

4 Cross-Dataset Experiments

In this section we consider whether models can benefit from transfer between SQuAD 2.0, QuAC, and CoQA, and show that the datasets, while ineffective for direct transfer, can be used as pretraining. In all experiments, we use BiDAF++, either with two context or no context, depending on if we are training for dialog settings or not, with default configurations. Models are trained by initializing from other models trained on different datasets and we do not decrease initial learning rates from just training directly on the target dataset. When SQuAD 2.0 is used to initialize models that use context, we randomly order questions in SQuAD 2.0  and train as if questions were asked in the form of a dialog. 888Likely a better strategy exists but we would like to demonstrate transfer in the simplest way. We only report development numbers as these experiments are meant to be exploratory.

In Domain F1
DrQA + PGNet 66.2
BiDAF++ w/ 2-ctx 67.6
SQuAD 2.0 41.4
QuAC 29.1
Ft from SQuAD 2.0 69.2
Ft from QuAC 68.0
Table 7: Cross dataset transfer to CoQA development set. Models do not transfer directly (rows 3 and 4), but after fine tuning improve performance (rows 5 and 6). For an explanation of why BiDAF++ outperforms DrQA + PGNet, see Section 3.
Baseline 67.6 65.1
BiDAF++ 70.5 67.4
CoQA 38.1 32.4
QuAC 25.4 16.8
Ft from CoQA 72.5 69.4
Ft from QuAC 69.5 66.8
Table 8: Cross dataset transfer to SQuAD 2.0 development set. BiDAF++ Choi et al. (2018) outperforms the baseline, a different implementation of the same model Rajpurkar et al. (2018) likely because of better hyper parameter tuning.


Tables 6-8 summarize our results. Across all of the datasets, BiDAF++ outperforms other baselines, and there exists at least one other dataset that significantly improves performance on a target dataset on average +2.1 F1. Experiments do not support that direct transfer is possible.

5 Related Work

Other proposals exist other than the three we analyzed that expand on features in SQuAD Rajpurkar et al. (2016). For example, maintaining question independence of context to reduce the role of string matching and having long context length Joshi et al. (2017); Kociský et al. (2017), higher level reasoning  Khashabi et al. (2018); Clark et al. (2018); Yang et al. (2018), multi-turn information seeking interactions, in either table settings Iyyer et al. (2017); Talmor and Berant (2018); Saha et al. (2018), regulation settings Saeidi et al. (2018), or Quiz Bowl settings Elgohary et al. (2018). Other work considers multi-modal contexts where interactions are a single turn  Tapaswi et al. (2016); Antol et al. (2015); Lei et al. (2018) or multi-turn Das et al. (2017); Pasunuru and Bansal (2018). These efforts contain alternative challenges than ones we analyze in this paper.


We thank Eunsol Choi, Hsin-Yuan Huang, Mohit Iyyer, He He, Yejin Choi, Percy Liang, and Luke Zettlemoyer for their helpful discussions in formulating this work. Also, Siva Reddy and Danqi Chen for help evaluating on CoQA and all reviewers for their comments.


  • S. Antol, A. Agrawal, J. Lu, M. Mitchell, D. Batra, C. Lawrence Zitnick and D. Parikh (2015) Vqa: visual question answering. In Proceedings of the IEEE international conference on computer vision, pp. 2425–2433. Cited by: §5.
  • S. R. Bowman, G. Angeli, C. Potts and C. D. Manning (2015) A large annotated corpus for learning natural language inference. arXiv preprint arXiv:1508.05326. Cited by: footnote 4.
  • D. Chen, A. Fisch, J. Weston and A. Bordes (2017) Reading Wikipedia to answer open-domain questions. In Association for Computational Linguistics (ACL), Cited by: §3.
  • E. Choi, H. He, M. Iyyer, M. Yatskar, W. Yih, Y. Choi, P. Liang and L. Zettlemoyer (2018) QuAC: question answering in context. In Empirical Methods in Natural Language Processing, Cited by: §1, §1, Table 4, §3, Table 8.
  • C. Clark and M. Gardner (2018) Simple and effective multi-paragraph reading comprehension. acl. Cited by: §1, §3.
  • P. Clark, I. Cowhey, O. Etzioni, T. Khot, A. Sabharwal, C. Schoenick and O. Tafjord (2018) Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv preprint arXiv:1803.05457. Cited by: §5.
  • A. Das, S. Kottur, K. Gupta, A. Singh, D. Yadav, J. M. Moura, D. Parikh and D. Batra (2017) Visual dialog. In cvpr, Cited by: §5.
  • A. Elgohary, C. Zhao and J. Boyd-Graber (2018) A dataset and baselines for sequential open-domain question answering. In emnlp, Cited by: §5.
  • K. M. Hermann, T. Kocisky, E. Grefenstette, L. Espeholt, W. Kay, M. Suleyman and P. Blunsom (2015) Teaching machines to read and comprehend. In Advances in Neural Information Processing Systems, pp. 1693–1701. Cited by: §1.
  • M. Hu, Y. Peng, Z. Huang, X. Qiu, F. Wei and M. Zhou (2017) Reinforced mnemonic reader for machine reading comprehension. arXiv preprint arXiv:1705.02798. Cited by: §1.
  • M. Iyyer, W. Yih and M. Chang (2017) Search-based neural structured learning for sequential question answering. In acl, Cited by: §5.
  • M. Joshi, E. Choi, D. S. Weld and L. Zettlemoyer (2017) TriviaQA: a large scale distantly supervised challenge dataset for reading comprehension. In acl, Cited by: §5.
  • D. Khashabi, S. Chaturvedi, M. Roth, S. Upadhyay and D. Roth (2018) Looking beyond the surface:a challenge set for reading comprehension over multiple sentences. In Proceedings of North American Chapter of the Association for Computational Linguistics (NAACL), Cited by: §5.
  • T. Kociský, J. Schwarz, P. Blunsom, C. Dyer, K. M. Hermann, G. Melis and E. Grefenstette (2017) The narrativeqa reading comprehension challenge. tacl abs/1712.07040. Cited by: §5.
  • J. Lei, L. Yu, M. Bansal and T. L. Berg (2018) TVQA: localized, compositional video question answering. In EMNLP, Cited by: §5.
  • T. Nguyen, M. Rosenberg, X. Song, J. Gao, S. Tiwary, R. Majumder and L. Deng (2016) MS marco: a human generated machine reading comprehension dataset. arXiv abs/1611.09268. Cited by: §1.
  • R. Pasunuru and M. Bansal (2018) Game-based video-context dialogue. In EMNLP, Cited by: §5.
  • M. E. Peters, M. Neumann, M. Iyyer, M. Gardner, C. Clark, K. Lee and L. Zettlemoyer (2018) Deep contextualized word representations. naacl. Cited by: §1, §3.
  • P. Rajpurkar, R. Jia and P. Liang (2018) Know what you don’t know: unanswerable questions for squad. In acl, Cited by: §1, Table 8.
  • P. Rajpurkar, J. Zhang, K. Lopyrev and P. Liang (2016) Squad: 100,000+ questions for machine comprehension of text. In emnlp, Cited by: §1, §5.
  • S. Reddy, D. Chen and C. D. Manning (2018) CoQA: a conversational question answering challenge. ArXiv. Cited by: §1, Table 4.
  • M. Richardson, C. J. Burges and E. Renshaw (2013) Mctest: a challenge dataset for the open-domain machine comprehension of text. In Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pp. 193–203. Cited by: §1.
  • M. Saeidi, M. Bartolo, P. Lewis, S. Signh, T. Rocktäschel, M. Sheldon, G. Bouchard and S. Riedel (2018) Interpretation of natural language rules in conversational machine reading. In emnlp, Cited by: §5.
  • A. Saha, V. Pahuja, M. M. Khapra, K. Sankaranarayanan and S. Chandar (2018) Complex sequential question answering: towards learning to converse over linked question answer pairs with a knowledge graph. In aaai, Cited by: §5.
  • M. Seo, A. Kembhavi, A. Farhadi and H. Hajishirzi (2016) Bidirectional attention flow for machine comprehension. iclr. Cited by: §1, §3.
  • A. Talmor and J. Berant (2018) The web as knowledge-base for answering complex questions. In naacl, Cited by: §5.
  • M. Tapaswi, Y. Zhu, R. Stiefelhagen, A. Torralba, R. Urtasun and S. Fidler (2016) Movieqa: understanding stories in movies through question-answering. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 4631–4640. Cited by: §5.
  • A. Trischler, T. Wang, X. Yuan, J. Harris, A. Sordoni, P. Bachman and K. Suleman (2016) NewsQA: a machine comprehension dataset. arXiv preprint arXiv:1611.09830. Cited by: §1.
  • E. M. Voorhees (2001) The trec question answering track. Natural Language Engineering 7 (4), pp. 361–378. Cited by: §1.
  • W. Wang, M. Yan and C. Wu (2018) Multi-granularity hierarchical attention fusion networks for reading comprehension and question answering. In Proceedings of the 56th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Vol. 1, pp. 1705–1714. Cited by: §1.
  • W. Wang, N. Yang, F. Wei, B. Chang and M. Zhou (2017) Gated self-matching networks for reading comprehension and question answering. In Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), Vol. 1, pp. 189–198. Cited by: §1.
  • Z. Yang, P. Qi, S. Zhang, Y. Bengio, W. W. Cohen, R. Salakhutdinov and C. D. Manning (2018) HotpotQA: a dataset for diverse, explainable multi-hop question answering. In Proceedings of the Conference on Empirical Methods in Natural Language Processing (EMNLP), Cited by: §5.
Comments 0
Request Comment
You are adding the first comment!
How to quickly get a good reply:
  • Give credit where it’s due by listing out the positive aspects of a paper before getting into which changes should be made.
  • Be specific in your critique, and provide supporting evidence with appropriate references to substantiate general statements.
  • Your comment should inspire ideas to flow and help the author improves the paper.

The better we are at sharing our knowledge with each other, the faster we move forward.
The feedback must be of minimum 40 characters and the title a minimum of 5 characters
Add comment
Loading ...
This is a comment super asjknd jkasnjk adsnkj
The feedback must be of minumum 40 characters
The feedback must be of minumum 40 characters

You are asking your first question!
How to quickly get a good answer:
  • Keep your question short and to the point
  • Check for grammar or spelling errors.
  • Phrase it like a question
Test description