View on GitHub

NLP-progress

Repository to track the progress in Natural Language Processing (NLP), including the datasets and the current state-of-the-art for the most common NLP tasks.

Question answering

Question answering is the task of answering a question.

Table of contents

ARC

The AI2 Reasoning Challenge (ARC) dataset is a question answering, which contains 7,787 genuine grade-school level, multiple-choice science questions. The dataset is partitioned into a Challenge Set and an Easy Set. The Challenge Set contains only questions answered incorrectly by both a retrieval-based algorithm and a word co-occurrence algorithm. Models are evaluated based on accuracy.

A public leaderboard is available on the ARC website.

ShARC

ShARC is a challenging QA dataset that requires logical reasoning, elements of entailment/NLI and natural language generation.

Most work in machine reading focuses on question answering problems where the answer is directly expressed in the text to read. However, many real-world question answering problems require the reading of text not because it contains the literal answer, but because it contains a recipe to derive an answer together with the reader’s background knowledge. We formalise this task and introduce the challenging ShARC dataset with 32k task instances.

The goal is to answer questions by possibly asking follow-up questions first. We assume that the question does not provide enough information to be answered directly. However, a model can use the supporting rule text to infer what needs to be asked in order to determine the final answer. Concretely, The model must decide whether to answer with “Yes”, “No”, “Irrelevant”, or to generate a follow-up question given rule text, a user scenario and a conversation history. Performance is measured with Micro and Macro Accuracy for “Yes”/”No”/”Irrelevant”/”More” classifications, and the quality of follow-up questions are measured with BLEU.

The public data, further task details and public leaderboard are available on the ShARC Website.

Reading comprehension

Most current question answering datasets frame the task as reading comprehension where the question is about a paragraph or document and the answer often is a span in the document. The Machine Reading group at UCL also provides an overview of reading comprehension tasks.

AdversarialQA

AdversarialQA provides three Reading Comprehension datasets constructed using an adversarial model-in-the-loop, where the annotator has to ask questions that the model fails to answer successfully. The passages are identical to the ones used in SQuADv1.1.

AdversarialQA uses three different models; BiDAF, BERT-Large, and RoBERTa-Large in the annotation loop to create three datasets; D(BiDAF), D(BERT), and D(RoBERTa), each with 10,000 training examples, 1,000 validation, and 1,000 test examples.

The adversarial human annotation paradigm ensures that these datasets consist of questions that current state-of-the-art models (at least the ones used as adversaries in the annotation loop) find challenging.

Examples:

Dataset Passage Question Answer
D(BiDAF) Martin Luther married Katharina von Bora, one of 12 nuns he had helped escape from the Nimbschen Cistercian convent in April 1523, when he arranged for them to be smuggled out in herring barrels. “Suddenly, and while I was occupied with far different thoughts,” he wrote to Wenceslaus Link, “the Lord has plunged me into marriage.” At the time of their marriage, Katharina was 26 years old and Luther was 41 years old. In a letter who did Luther credit for his union with Katharina? the Lord
D(BERT) This combination of cancellations and σ and π overlaps results in dioxygen’s double bond character and reactivity, and a triplet electronic ground state. An electron configuration with two unpaired electrons as found in dioxygen (see the filled π* orbitals in the diagram), orbitals that are of equal energy—i.e., degenerate—is a configuration termed a spin triplet state. Hence, the ground state of the O2 molecule is referred to as triplet oxygen.[b] The highest energy, partially filled orbitals are antibonding, and so their filling weakens the bond order from three to two. Because of its unpaired electrons, triplet oxygen reacts only slowly with most organic molecules, which have paired electron spins; this prevents spontaneous combustion. What are in the orbitals of atoms? electrons
D(RoBERTa) Between 1991 and 2000, the total area of forest lost in the Amazon rose from 415,000 to 587,000 square kilometres (160,000 to 227,000 sq mi), with most of the lost forest becoming pasture for cattle. Seventy percent of formerly forested land in the Amazon, and 91% of land deforested since 1970, is used for livestock pasture. Currently, Brazil is the second-largest global producer of soybeans after the United States. New research however, conducted by Leydimere Oliveira et al., has shown that the more rainforest is logged in the Amazon, the less precipitation reaches the area and so the lower the yield per hectare becomes. So despite the popular perception, there has been no economical advantage for Brazil from logging rainforest zones and converting these to pastoral fields. Of the two countries that produce soybeans, which country is clearing rain forest in order to increase production? Brazil

More details on the task and the dataset can be found in the TACL paper. The public leaderboard is available through Dynabench as the QA round 1 task.

CliCR

The CliCR dataset is a gap-filling reading comprehension dataset consisting of around 100,000 queries and their associated documents. The dataset was built from clinical case reports, requiring the reader to answer the query with a medical problem/test/treatment entity. The abilities to perform bridging inferences and track objects have been found to be the most frequently required skills for successful answering.

The instructions for accessing the dataset, the processing scripts, the baselines and the adaptations of some neural models can be found here.

Example:

Document Question Answer
We report a case of a 72-year-old Caucasian woman with pl-7 positive antisynthetase syndrome. Clinical presentation included interstitial lung disease, myositis, mechanic’s hands and dysphagia. As lung injury was the main concern, treatment consisted of prednisolone and cyclophosphamide. Complete remission with reversal of pulmonary damage was achieved, as reported by CT scan, pulmonary function tests and functional status. […] Therefore, in severe cases an aggressive treatment, combining ____ and glucocorticoids as used in systemic vasculitis, is suggested. cyclophoshamide
Model F1 Paper
Gated-Attention Reader (Dhingra et al., 2017) 33.9 CliCR: A Dataset of Clinical Case Reports for Machine Reading Comprehension
Stanford Attentive Reader (Chen et al., 2016) 27.2 CliCR: A Dataset of Clinical Case Reports for Machine Reading Comprehension

CNN / Daily Mail

The CNN / Daily Mail dataset is a Cloze-style reading comprehension dataset created from CNN and Daily Mail news articles using heuristics. Close-style means that a missing word has to be inferred. In this case, “questions” were created by replacing entities from bullet points summarizing one or several aspects of the article. Coreferent entities have been replaced with an entity marker @entityn where n is a distinct index. The model is tasked to infer the missing entity in the bullet point based on the content of the corresponding article and models are evaluated based on their accuracy on the test set.

  CNN Daily Mail
# Train 380,298 879,450
# Dev 3,924 64,835
# Test 3,198 53,182

Example:

Passage Question Answer
( @entity4 ) if you feel a ripple in the force today , it may be the news that the official @entity6 is getting its first gay character . according to the sci-fi website @entity9 , the upcoming novel “ @entity11 “ will feature a capable but flawed @entity13 official named @entity14 who “ also happens to be a lesbian . “ the character is the first gay figure in the official @entity6 – the movies , television shows , comics and books approved by @entity6 franchise owner @entity22 – according to @entity24 , editor of “ @entity6 “ books at @entity28 imprint @entity26 . characters in “ @placeholder “ movies have gradually become more diverse @entity6
Model CNN Daily Mail Paper / Source
GA Reader(Dhingra et al., 2017) 77.9 80.9 Gated-Attention Readers for Text Comprehension
BIDAF(Seo et al., 2017) 76.9 79.6 Bidirectional Attention Flow for Machine Comprehension
AoA Reader(Cui et al., 2017) 74.4 - Attention-over-Attention Neural Networks for Reading Comprehension
Neural net (Chen et al., 2016) 72.4 75.8 A Thorough Examination of the CNN/Daily Mail Reading Comprehension Task
Classifier (Chen et al., 2016) 67.9 68.3 A Thorough Examination of the CNN/Daily Mail Reading Comprehension Task
Impatient Reader (Hermann et al., 2015) 63.8 68.0 Teaching Machines to Read and Comprehend

CODAH

CODAH is an adversarially-constructed evaluation dataset with 2.8k questions for testing common sense. CODAH forms a challenging extension to the SWAG dataset, which tests commonsense knowledge using sentence-completion questions that describe situations observed in video.

The dataset and more information can be found here

CoQA

CoQA is a large-scale dataset for building Conversational Question Answering systems. CoQA contains 127,000+ questions with answers collected from 8000+ conversations. Each conversation is collected by pairing two crowdworkers to chat about a passage in the form of questions and answers.

The data and public leaderboard are available here.

HotpotQA

HotpotQA is a dataset with 113k Wikipedia-based question-answer pairs. Questions require finding and reasoning over multiple supporting documents and are not constrained to any pre-existing knowledge bases. Sentence-level supporting facts are available.

The data and public leaderboard are available from the HotpotQA website.

MS MARCO

MS MARCO aka Human Generated MAchine Reading COmprehension Dataset, is designed and developed by Microsoft AI & Research. Link to paper

The leaderboards for multiple tasks are available on the MS MARCO leaderboard page.

MultiRC

MultiRC (Multi-Sentence Reading Comprehension) is a dataset of short paragraphs and multi-sentence questions that can be answered from the content of the paragraph. We have designed the dataset with three key challenges in mind:

The leaderboards for the dataset is available on the MultiRC website.

Natural Questions

The Natural Questions corpus contains questions from real users issued to the Google search engine. It requires QA systems to read and comprehend an entire Wikipedia article that may or may not contain the answer to the question. Questions are presented along with a Wikipedia page and an extracted long answer (typically a paragraph) and short answer (one or more entities) if present on the page, or marked null if no long/short answer was present.

Example:

Question Wikipedia Page Long Answer Short Answer
who lives in the imperial palace in tokyo Tokyo_Imperial_Palace The Tokyo Imperial Palace (, Kkyo, literally “Imperial Residence”) is the primary residence of the Emperor of Japan. It is a large park-like area located in the Chiyoda ward of Tokyo and contains buildings including the main palace (, Kyden), the private residences of the Imperial Family, an archive, museums and administrative offices. The Imperial Family

The leaderboard and the dataset are available in the Google’s Natural Question website

NewsQA

The NewsQA dataset is a reading comprehension dataset of over 100,000 human-generated question-answer pairs from over 10,000 news articles from CNN, with answers consisting of spans of text from the corresponding articles. Some challenging characteristics of this dataset are:

Example:

Story Question Answer
MOSCOW, Russia (CNN) – Russian space officials say the crew of the Soyuz space ship is resting after a rough ride back to Earth. A South Korean bioengineer was one of three people on board the Soyuz capsule. The craft carrying South Korea’s first astronaut landed in northern Kazakhstan on Saturday, 260 miles (418 kilometers) off its mark, they said. Mission Control spokesman Valery Lyndin said the condition of the crew – South Korean bioengineer Yi So-yeon, American astronaut Peggy Whitson and Russian flight engineer Yuri Malenchenko – was satisfactory, though the three had been subjected to severe G-forces during the re-entry. […] Where did the Soyuz capsule land? northern Kazakhstan

The dataset can be downloaded here.

Model F1 EM Paper / Source
DecaProp (Tay et al., 2018) 66.3 53.1 Densely Connected Attention Propagation for Reading Comprehension
AMANDA (Kundu et al., 2018) 63.7 48.4 A Question-Focused Multi-Factor Attention Network for Question Answering
MINIMAL(Dyn) (Min et al., 2018) 63.2 50.1 Efficient and Robust Question Answering from Minimal Context over Documents
FastQAExt (Weissenborn et al., 2017) 56.1 43.7 Making Neural QA as Simple as Possible but not Simpler

QAngaroo

QAngaroo is a set of two reading comprehension datasets, which require multiple steps of inference that combine facts from multiple documents. The first dataset, WikiHop is open-domain and focuses on Wikipedia articles. The second dataset, MedHop is based on paper abstracts from PubMed.

The leaderboards for both datasets are available on the QAngaroo website.

QuAC

Question Answering in Context (QuAC) is a dataset for modeling, understanding, and participating in information seeking dialog. Data instances consist of an interactive dialog between two crowd workers: (1) a student who poses a sequence of freeform questions to learn as much as possible about a hidden Wikipedia text, and (2) a teacher who answers the questions by providing short excerpts (spans) from the text.

The leaderboard and data are available on the QuAC website.

RACE

The RACE dataset is a reading comprehension dataset collected from English examinations in China, which are designed for middle school and high school students. The dataset contains more than 28,000 passages and nearly 100,000 questions and can be downloaded here. Models are evaluated based on accuracy on middle school examinations (RACE-m), high school examinations (RACE-h), and on the total dataset (RACE).

The public leaderboard is available on the RACE leaderboard.

Model RACE-m RACE-h RACE Paper Code
XLNet (Yang et al., 2019) 85.45 80.21 81.75 XLNet: Generalized Autoregressive Pretraining for Language Understanding Official
OCN_large (Ran et al., 2019) 76.7 69.6 71.7 Option Comparison Network for Multiple-choice Reading Comprehension  
DCMN_large (Zhang et al., 2019) 73.4 68.1 69.7 Dual Co-Matching Network for Multi-choice Reading Comprehension  
Finetuned Transformer LM (Radford et al., 2018) 62.9 57.4 59.0 Improving Language Understanding by Generative Pre-Training Official
BiAttention MRU (Tay et al., 2018) 60.2 50.3 53.3 Multi-range Reasoning for Machine Comprehension  

SQuAD

The Stanford Question Answering Dataset (SQuAD) is a reading comprehension dataset, consisting of questions posed by crowdworkers on a set of Wikipedia articles. The answer to every question is a segment of text (a span) from the corresponding reading passage. Recently, SQuAD 2.0 has been released, which includes unanswerable questions.

The public leaderboard is available on the SQuAD website.

Story Cloze Test

The Story Cloze Test is a dataset for story understanding that provides systems with four-sentence stories and two possible endings. The systems must then choose the correct ending to the story.

More details are available on the Story Cloze Test Challenge.

Model Accuracy Paper / Source Code
Reading Strategies Model (Sun et al., 2018) 88.3 Improving Machine Reading Comprehension by General Reading Strategies  
Finetuned Transformer LM (Radford et al., 2018) 86.5 Improving Language Understanding by Generative Pre-Training Official
Liu et al. (2018) 78.7 Narrative Modeling with Memory Chains and Semantic Supervision Official
Hidden Coherence Model (Chaturvedi et al., 2017) 77.6 Story Comprehension for Predicting What Happens Next  
val-LS-skip (Srinivasan et al., 2018) 76.5 A Simple and Effective Approach to the Story Cloze Test  

SWAG

SWAG (Situations With Adversarial Generations) is a large-scale dataset for the task of grounded commonsense inference, unifying natural language inference and physically grounded reasoning. The dataset consists of 113k multiple choice questions about grounded situations. Each question is a video caption from LSMDC or ActivityNet Captions, with four answer choices about what might happen next in the scene. The correct answer is the (real) video caption for the next event in the video; the three incorrect answers are adversarially generated and human verified, so as to fool machines but not humans.

The public leaderboard is available on the AI2 website.

RecipeQA

RecipeQA is a dataset for multimodal comprehension of cooking recipes. It consists of over 36K question-answer pairs automatically generated from approximately 20K unique recipes with step-by-step instructions and images. Each question in RecipeQA involves multiple modalities such as titles, descriptions or images, and working towards an answer requires (i) joint understanding of images and text, (ii) capturing the temporal flow of events, and (iii) making sense of procedural knowledge.

The public leaderboard is available on the RecipeQA website.

NarrativeQA

NarrativeQA is a dataset built to encourage deeper comprehension of language. This dataset involves reasoning over reading entire books or movie scripts. This dataset contains approximately 45K question answer pairs in free form text. There are two modes of this dataset (1) reading comprehension over summaries and (2) reading comprehension over entire books/scripts.

The results for the first, summary mode are below.

Model BLEU-1 BLEU-4 METEOR Rouge-L Paper / Source Code
DecaProp (Tay et al., 2018) 44.35 27.61 21.80 44.69 Densely Connected Attention Propagation for Reading Comprehension official
BiAttention + DCU-LSTM (Tay et al., 2018) 36.55 19.79 17.87 41.44 Multi-Granular Sequence Encoding via Dilated Compositional Units for Reading Comprehension  
BiDAF (Seo et al., 2017) 33.45 15.69 15.68 36.74 Bidirectional Attention Flow for Machine Comprehension  

The results for the second mode (question answering over entire books or movie scripts) are below.

Model BLEU-1 BLEU-4 METEOR Rouge-L Paper / Source Code
Retriever + Reader (Izacard and Grave, 2020) 35.3 7.5 11.1 32.0 Distilling Knowledge from Reader to Retriever for Question Answering Official
Summary + Reader (UnifiedQA) (Wu et al., 2021) 21.82 3.87 10.52 21.03 Recursively Summarizing Books with Human Feedback  
ReadTwice (Zemlyanskiy et al., 2021) 21.1 4.0 7.0 23.2 ReadTwice: Reading Very Large Documents with Memories Official

DuoRC

DuoRC contains 186,089 unique question-answer pairs created from a collection of 7680 pairs of movie plots where each pair in the collection reflects two versions of the same movie.

DuoRC pushes the NLP community to address challenges on incorporating knowledge and reasoning in neural architectures for reading comprehension. It poses several interesting challenges such as:

DROP

DROP is a crowdsourced, adversarially-created, 96k-question benchmark, in which a system must resolve references in a question, perhaps to multiple input positions, and perform discrete operations over them (such as addition, counting, or sorting). These operations require a much more comprehensive understanding of the content of paragraphs than what was necessary for prior datasets.

Cosmos QA

Cosmos QA is a large-scale dataset of 35.6K problems that require commonsense-based reading comprehension, formulated as multiple-choice questions. It focuses on reading between the lines over a diverse collection of people’s everyday narratives, asking questions concerning on the likely causes or effects of events that require reasoning beyond the exact text spans in the context.

ReClor (logical reasoning)

The ReClor dataset is a reading comprehension dataset requiring logical reasoning, which is extracted from standardized exams GMAT (Graduate Management Admission Test) and LSAT (Law School Admission Test). This dataset is very challenging and even graduate students can only achieve 63% accuracy. It has various logical reasoning types, ie, Necessary/Sufficient Assumptions, Strengthen/Weaken, Evaluation, Implication, Conclusion/Main Point, Most Strongly Supported, Explain or Resolve, Principle, Dispute, Technique, Role, Identify a Flaw, Match Flaws, Match the Structure and others.

The dataset, public leaderboard, and code are available on the project page ReClor (logical reasoning).

Open-domain Question Answering

DuReader

DuReader is a large-scale, open-domain Chinese machine reading comprehension (MRC) dataset, designed to address real-world MRC. Link to paper

DuReader has three advantages over other MRC datasets:

To help the community make these improvements, both the dataset of DuReader and baseline systems have been posted online.

The leaderboard is avaiable on DuReader page.

Quasar

Quasar is a dataset for open-domain question answering. It includes two parts: (1) The Quasar-S dataset consists of 37,000 cloze-style queries constructed from definitions of software entity tags on the popular website Stack Overflow. (2) The Quasar-T dataset consists of 43,000 open-domain trivia questions and their answers obtained from various internet sources.

Model EM (Quasar-T) F1 (Quasar-T) Paper / Source Code
Denoising QA (Lin et al. 2018) 42.2 49.3 Denoising Distantly Supervised Open-Domain Question Answering official
DecaProp (Tay et al., 2018) 38.6 46.9 Densely Connected Attention Propagation for Reading Comprehension official
R^3 (Wang et al., 2018) 35.3 41.7 R^3: Reinforced Ranker-Reader for Open-Domain Question Answering official
BiDAF (Seo et al., 2017) 25.9 28.5 Bidirectional Attention Flow for Machine Comprehensio official
GA (Dhingra et al., 2017) 26.4 26.4 Gated-Attention Readers for Text Comprehension  

SearchQA

SearchQA was constructed to reflect a full pipeline of general question-answering. SearchQA consists of more than 140k question-answer pairs with each pair having 49.6 snippets on average. Each question-answer-context tuple of the SearchQA comes with additional meta-data such as the snippet’s URL.

Model Unigram Acc N-gram F1 EM F1 Paper / Source Code
DecaProp (Tay et al., 2018) 62.2 70.8 56.8 63.6 Densely Connected Attention Propagation for Reading Comprehension official
Denoising QA (Lin et al. 2018) - - 58.8 64.5 Denoising Distantly Supervised Open-Domain Question Answering official
R^3 (Wang et al., 2018) - - 49.0 55.3 R^3: Reinforced Ranker-Reader for Open-Domain Question Answering official
Bi-Attention + DCU-LSTM (Tay et al., 2018) 49.4 59.5 - - Multi-Granular Sequence Encoding via Dilated Compositional Units for Reading Comprehension  
AMANDA (Kundu et al., 2018) 46.8 56.6 - - A Question-Focused Multi-Factor Attention Network for Question Answering official
Focused Hierarchical RNN (Ke et al., 2018) 46.8 53.4 - - Focused Hierarchical RNNs for Conditional Sequence Processing  
ASR (Kadlec et al, 2016) 41.3 22.8 - - Text Understanding with the Attention Sum Reader Network  

Knowledge Base Question Answering

Knowledge Base Question Answering is the task of answering natural language question based on a knowledge base/knowledge graph such as DBpedia or Wikidata.

QALD-9

QALD-9 is a manually curated superset of the previous eight editions of the Question Answering over Linked Data (QALD) challenge published in 2018. It is constructed by human experts to cover a wide range of natural language to SPARQL conversions based on DBpedia 2016-10 knowledge base. Each question-answer-pair has additional meta-data. QALD-9 is best evaluated using the GERBIL QA platform for repeatability of the evaluation numbers.

Annotator Macro P Macro R Macro F1 Error Count Average Time/Doc ms Macro F1 QALD Paper (including links to webservices/source code)
Elon (WS) 0.049 0.053 0.050 2 219 0.100  
QASystem (WS) 0.097 0.116 0.098 0 1014 0.200  
TeBaQA (WS) 0.129 0.134 0.130 0 2668 0.222  
wdaqua-core1 (DBpedia) 0.261 0.267 0.250 0 661 0.289 Diefenbach, Dennis, Kamal Singh, and Pierre Maret. “Wdaqua-core1: a question answering service for rdf knowledge bases.” Companion of the The Web Conference 2018 on The Web Conference 2018. International World Wide Web Conferences Steering Committee, 2018.
gAnswer (WS) 0.293 0.327 0.298 1 3076 0.430 Zou, Lei, et al. “Natural language question answering over RDF: a graph data driven approach.” Proceedings of the 2014 ACM SIGMOD international conference on Management of data. ACM, 2014.

Go back to the README