---
title: "Question answering — papers"
description: "Question answering: 183 papers of the Elamur Research library, the most cited first."
canonical: "https://research.elamur.ai/free/en/topics/question-answering/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/topics/question-answering/)

[home](https://md.research.elamur.ai/free/en/) / [topics](https://md.research.elamur.ai/free/en/topics/) / Question answering

# ❓ Question answering

183 papers · page 1 of 8 · the most cited first · [all topics](https://md.research.elamur.ai/free/en/topics/)

Systems that answer questions from a text, a collection or a knowledge base.

A survey · about 20 minutes to read · every paper named here has its own page in this library

**In this article**

1. [What question answering is](#a-what-question-answering-is)
2. [History and milestones](#a-history-and-milestones)
3. [Main problems and tasks](#a-main-problems-and-tasks)
4. [Methods and schools](#a-methods-and-schools)
5. [Datasets, resources and benchmarks](#a-datasets-resources-and-benchmarks)
6. [Landmark papers in this library](#a-landmark-papers-in-this-library)
7. [People](#a-people)
8. [Institutions and places](#a-institutions-and-places)
9. [Open questions](#a-open-questions)
10. [Neighbouring topics](#a-neighbouring-topics)
11. [Key works in this library](#a-key-works-in-this-library)

Question answering (QA) is the branch of natural language processing that builds systems which take a question posed in ordinary language and return an answer, rather than a list of documents. The answer may be a short span copied from a passage, a yes or no, an entity drawn from a knowledge base, a number computed from a table, or a paragraph written from scratch. What unites these variants is the demand that a machine connect the meaning of a question to evidence and commit to a reply that a person could check.

The field is old by the standards of computing, and since the middle of the 2010s QA has also become the main proving ground for reading comprehension models, for retrieval over very large text collections, and later for large language models that are expected to answer almost anything. Many of the most cited datasets in natural language processing are QA datasets, and the format is now so general that tasks such as relation extraction, event extraction and dialogue state tracking are routinely recast as questions and answers.

The papers here cover that whole arc, from TREC-era pattern systems through the dataset wave and dense retrieval to attribution, bias and evaluation in the era of large models.

## What question answering is

At its core, a QA system receives a question and some source of knowledge and must produce an answer. The field is usually divided along three axes.

**The source of knowledge.** In *reading comprehension* (often called machine reading comprehension, MRC) the system is handed a passage that is known to contain the answer. In *open-domain QA* it must first find relevant text in a large collection such as Wikipedia or the web. In *knowledge-base QA* (KBQA) the answer lives in a structured store of facts, and the question must be mapped to a query over it. Hybrid settings mix tables with text, or images and video with language.

**The form of the answer.** Extractive QA returns a span of the given text; multiple-choice QA picks among options; yes/no QA returns a polar verdict; abstractive or long-form QA writes a free answer, possibly several sentences long. Some benchmarks also require the system to say that no answer exists.

**The interaction.** Most datasets treat each question on its own. Conversational QA instead poses a chain of follow-up questions whose meaning depends on the preceding turns, which brings in ellipsis and coreference.

QA borders several neighbouring areas. It overlaps with [Information retrieval & search](https://md.research.elamur.ai/free/en/topics/information-retrieval/), which ranks documents without having to state the answer; with [Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/), since many recent benchmarks are built to require arithmetic, logic or several inference steps; and with [Semantic parsing & meaning representation](https://md.research.elamur.ai/free/en/topics/semantic-parsing/), which supplies the machinery for turning questions into executable queries. Because so much of the field is organised around shared test sets, it is also closely tied to [Evaluation & benchmarks](https://md.research.elamur.ai/free/en/topics/evaluation-benchmarks/).

## History and milestones

### Before the evaluation campaigns

Restricted-domain QA goes back to systems such as BASEBALL and LUNAR in the 1960s and early 1970s, which turned English questions into database queries but did not generalise beyond their domain.

### The TREC era (1999–2007)

The modern field took shape when the Text REtrieval Conference added a QA track in 1999. [The TREC-8 Question Answering Track](https://md.research.elamur.ai/free/en/papers/voorhees-2000-trec-8-question-answering-track/) set out the task that would dominate the next decade: answer short fact-seeking questions from a fixed newswire collection by returning a short string. Systems of this period were pipelines: classify the question type, retrieve documents, pick candidate sentences, and extract a phrase of the expected type.

[Using Coreference for Question Answering](https://md.research.elamur.ai/free/en/papers/morton-1999-coreference-question-answering/) and [Importance of Pronominal Anaphora Resolution in Question Answering Systems](https://md.research.elamur.ai/free/en/papers/vicedo-2000-importance-pronominal-anaphora-resolution-question/) showed that resolving pronouns lets a system find answers stated across sentences. [Learning surface text patterns for a Question Answering System](https://md.research.elamur.ai/free/en/papers/ravichandran-2002-learning-surface-text-patterns-question/) learned answer patterns automatically from the web. Later in the decade, attention moved to deeper linguistic matching: the quasi-synchronous grammar of [What is the Jeopardy Model? A Quasi-Synchronous Grammar for QA](https://md.research.elamur.ai/free/en/papers/wang-2007-what-is-jeopardy-model-quasi/) and the frame-semantic matching of [Using Semantic Roles to Improve Question Answering](https://md.research.elamur.ai/free/en/papers/shen-2007-semantic-roles-improve-question-answering/).

### Knowledge bases and early neural methods (2010–2015)

With the growth of Freebase and of facts extracted automatically from web text, QA over knowledge bases became a major strand. [Paraphrase-Driven Learning for Open Question Answering](https://md.research.elamur.ai/free/en/papers/fader-2013-paraphrase-driven-learning-open-question/) learned a lexicon from community paraphrase clusters instead of annotated queries, and [Semantic Parsing for Single-Relation Question Answering](https://md.research.elamur.ai/free/en/papers/yih-2014-semantic-parsing-single-relation-question/) replaced its hand-induced rules with learned convolutional similarity. At the same time, [Towards AI-Complete Question Answering: A Set of Prerequisite Toy Tasks](https://md.research.elamur.ai/free/en/papers/weston-2015-towards-ai-complete-question-answering/) proposed the bAbI tasks as unit tests for reasoning, and memory networks were developed to pass them. [WikiQA: A Challenge Dataset for Open-Domain Question Answering](https://md.research.elamur.ai/free/en/papers/yang-2015-wikiqa-challenge-dataset-open-domain/) brought real search queries into answer-sentence selection.

### The dataset wave and neural readers (2015–2019)

The decisive turn came when large training sets made it possible to learn readers end to end. Cloze-style data built from CNN and Daily Mail news (2015) was quickly followed by SQuAD (2016), the Stanford dataset of crowd-written questions over Wikipedia paragraphs whose answers are text spans. Within two years a series of attention-based architectures pushed accuracy on SQuAD past the reported human level, and then pretrained transformers such as BERT (2018) overtook them all.

A flood of new benchmarks followed, each meant to fix weaknesses of the earlier ones; most of the highly cited papers on this page date from this era.

Landmark papers on this topic: the 12 most cited, in time The 12 most cited papers on this topic span 2016–2021; the dot grows with the citations.

**The numbers**

| Paper | Citations |
| --- | --- |
| [2016 — Multimodal Compact Bilinear Pooling for Visual Question Answering and Visual Grounding](https://md.research.elamur.ai/free/en/papers/fukui-2016-multimodal-compact-bilinear-pooling-visual/) | 1,642 |
| [2017 — TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension](https://md.research.elamur.ai/free/en/papers/joshi-2017-triviaqa-large-scale-distantly-supervised/) | 3,204 |
| [2017 — Reading Wikipedia to Answer Open-Domain Questions](https://md.research.elamur.ai/free/en/papers/chen-2017-reading-wikipedia-answer-open-domain/) | 2,250 |
| [2017 — Adversarial Examples for Evaluating Reading Comprehension Systems](https://md.research.elamur.ai/free/en/papers/jia-2017-adversarial-examples-evaluating-reading-comprehension/) | 1,935 |
| [2017 — RACE: Large-scale ReAding Comprehension Dataset From Examinations](https://md.research.elamur.ai/free/en/papers/lai-2017-race-large-scale-reading-comprehension/) | 1,619 |
| [2018 — HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering](https://md.research.elamur.ai/free/en/papers/yang-2018-hotpotqa-dataset-diverse-explainable-multi/) | 3,489 |
| [2019 — Natural Questions: A Benchmark for Question Answering Research](https://md.research.elamur.ai/free/en/papers/kwiatkowski-2019-natural-questions-benchmark-question-answering/) | 4,031 |
| [2019 — GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering](https://md.research.elamur.ai/free/en/papers/hudson-2019-gqa-new-dataset-real-world/) | 1,997 |
| [2019 — CoQA: A Conversational Question Answering Challenge](https://md.research.elamur.ai/free/en/papers/reddy-2019-coqa-conversational-question-answering-challenge/) | 1,541 |
| [2019 — Latent Retrieval for Weakly Supervised Open Domain Question Answering](https://md.research.elamur.ai/free/en/papers/lee-2019-latent-retrieval-weakly-supervised-open/) | 1,234 |
| [2020 — Dense Passage Retrieval for Open-Domain Question Answering](https://md.research.elamur.ai/free/en/papers/karpukhin-2020-dense-passage-retrieval-open-domain/) | 5,370 |
| [2021 — Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering](https://md.research.elamur.ai/free/en/papers/izacard-2021-leveraging-passage-retrieval-generative-models/) | 1,674 |

### Retrieval, generation and large models (2019 onwards)

Once readers became strong, the bottleneck moved to finding the right evidence. From 2019, learned dense retrievers began to replace keyword search, and generative readers that fuse many passages replaced span extractors. Retrieval-augmented generation became a general recipe. Large language models then showed that a great deal of factual knowledge can be answered "closed-book", without any retrieval, while also producing fluent but unsupported claims. Current work therefore stresses attribution, faithfulness, freshness of knowledge, false premises, bias, and evaluation methods suited to free-form answers.

Papers on this topic by year 183 dated papers from 1999 to 2026; the busiest year is 2019 with 26, and 81% of them appeared in 2018 or later.

**The numbers**

| Year | Papers |
| --- | --- |
| 2026 | 20 |
| 2025 | 8 |
| 2024 | 8 |
| 2023 | 14 |
| 2022 | 14 |
| 2021 | 16 |
| 2020 | 24 |
| 2019 | 26 |
| 2018 | 19 |
| 2017 | 11 |
| 2016 | 10 |
| 2015 | 4 |
| 2014 | 2 |
| 2013 | 1 |
| 2007 | 2 |
| 2002 | 1 |
| 2000 | 2 |
| 1999 | 1 |

## Main problems and tasks

**Factoid and open-domain QA.** Answering short questions about the world from a large corpus. The central difficulty is retrieval at scale combined with precise extraction; [Reading Wikipedia to Answer Open-Domain Questions](https://md.research.elamur.ai/free/en/papers/chen-2017-reading-wikipedia-answer-open-domain/) named this setting "machine reading at scale".

**Reading comprehension.** Answering from a given passage. Subtypes include span extraction, cloze completion, multiple choice and free-form answers. [How Much Reading Does Reading Comprehension Require?](https://md.research.elamur.ai/free/en/papers/kaushik-2018-how-much-reading-does-reading/) showed that models can do surprisingly well on several datasets with the passage or the question partly removed.

**Multi-hop QA.** Questions whose answer requires combining facts from two or more documents or sentences, as in HotpotQA, [WikiHop](https://md.research.elamur.ai/free/en/papers/welbl-2018-constructing-datasets-multi-hop-reading/) and 2WikiMultiHopQA. A related problem is decomposing complex questions into simpler ones, studied in [decomposition with rescoring](https://md.research.elamur.ai/free/en/papers/min-2019-multi-hop-reading-comprehension-through/) and [unsupervised decomposition](https://md.research.elamur.ai/free/en/papers/perez-2020-unsupervised-question-decomposition-question-answering/).

**Discrete and numerical reasoning.** Counting, sorting, comparing and doing arithmetic over text or tables, as tested by DROP, TAT-QA, [HybridQA](https://md.research.elamur.ai/free/en/papers/chen-2020-hybridqa-dataset-multi-hop-question/) and [ConvFinQA](https://md.research.elamur.ai/free/en/papers/chen-2022-convfinqa-exploring-chain-numerical-reasoning/).

**Commonsense and logical reasoning.** Benchmarks such as Cosmos QA, ReCoRD, LogiQA and [ReClor: A Reading Comprehension Dataset Requiring Logical Reasoning](https://md.research.elamur.ai/free/en/papers/yu-2020-reclor-reading-comprehension-dataset-requiring/) ask for knowledge or inference that is not stated in the text.

**Conversational QA.** Answering a sequence of dependent questions, as in CoQA, QuAC and [work on conversational open-domain QA](https://md.research.elamur.ai/free/en/papers/anantha-2021-open-domain-question-answering-goes/), which turns context-dependent questions into stand-alone ones before retrieval.

**Long-form QA.** Writing explanatory answers to "why" and "how" questions. [ELI5: Long Form Question Answering](https://md.research.elamur.ai/free/en/papers/fan-2019-eli5-long-form-question-answering/) introduced the task; [Hurdles to Progress in Long-form Question Answering](https://md.research.elamur.ai/free/en/papers/krishna-2021-hurdles-progress-long-form-question/) later showed that its evaluation is fragile.

**Knowledge-base and table QA.** Mapping questions to structured queries or reasoning over graphs and tables, from [Question Answering on Freebase via Relation Extraction and Textual Evidence](https://md.research.elamur.ai/free/en/papers/xu-2016-question-answering-freebase-relation-extraction/) to [FeTaQA: Free-form Table Question Answering](https://md.research.elamur.ai/free/en/papers/nan-2022-fetaqa-free-form-table-question/).

**Question generation.** The inverse task: producing questions from text, used for education, for data augmentation and for evaluation. See [Learning to Ask: Neural Question Generation for Reading Comprehension](https://md.research.elamur.ai/free/en/papers/du-2017-learning-ask-neural-question-generation/) and [paragraph-level question generation](https://md.research.elamur.ai/free/en/papers/zhao-2018-paragraph-level-neural-question-generation/).

**QA as a format for other tasks.** Many papers here use questions as an interface: [relation extraction](https://md.research.elamur.ai/free/en/papers/li-2019-entity-relation-extraction-as-multi/), [event extraction](https://md.research.elamur.ai/free/en/papers/liu-2020-event-extraction-as-machine-reading/), [semantic role labelling](https://md.research.elamur.ai/free/en/papers/fitzgerald-2018-large-scale-qa-srl-parsing/), [discourse relations](https://md.research.elamur.ai/free/en/papers/pyatkin-2020-qadiscourse-discourse-relations-as-qa/), [bridging anaphora](https://md.research.elamur.ai/free/en/papers/hou-2020-bridging-anaphora-resolution-as-question/) and [dialogue state tracking](https://md.research.elamur.ai/free/en/papers/gao-2019-dialog-state-tracking-neural-reading/). QA is also used to measure the factual consistency of summaries, as in [FEQA](https://md.research.elamur.ai/free/en/papers/durmus-2020-feqa-question-answering-evaluation-framework/) and [QAFactEval](https://md.research.elamur.ai/free/en/papers/fabbri-2022-qafacteval-improved-qa-based-factual/).

## Methods and schools

### Pipelines and patterns

Classic open-domain systems were modular: a question classifier predicted the expected answer type, a search engine returned documents, and an extractor chose a phrase of that type. Ravichandran and Hovy bootstrapped answer patterns from a few seed pairs and web search, relying on the redundancy of the web to make short surface patterns effective.

### Syntactic and semantic matching

A second school argued that matching questions to answers needs linguistic structure. Wang, Smith and Mitamura modelled the question as a loose syntactic transformation of the answer sentence, summing over hidden word alignments between dependency trees. Shen and Lapata projected FrameNet roles onto questions and candidates and matched the resulting graphs, finding that semantic roles help but cover only part of the questions. In KBQA, semantic parsing mapped questions to logical forms; learned similarity models and paraphrase-based training later reduced the need for hand-written rules.

### Neural readers

Neural reading comprehension models encode the question and passage, let them attend to each other, and point to where the answer begins and ends. The [Dynamic Coattention Networks For Question Answering](https://md.research.elamur.ai/free/en/papers/xiong-2016-dynamic-coattention-networks-question-answering/) introduced two-way attention and an iterative decoder that could revise its span; [Gated Self-Matching Networks for Reading Comprehension and Question Answering](https://md.research.elamur.ai/free/en/papers/wang-2017-gated-self-matching-networks-reading/) added a relevance gate and self-attention over the passage; [QANet: Combining Local Convolution with Global Self-Attention for Reading Comprehension](https://md.research.elamur.ai/free/en/papers/yu-2018-qanet-combining-local-convolution-global/) removed recurrence entirely. Pretrained transformers then made a fine-tuned BERT the default reader.

A parallel line of work asked what such readers actually learn. [A Thorough Examination of the CNN/Daily Mail Reading Comprehension Task](https://md.research.elamur.ai/free/en/papers/chen-2016-thorough-examination-cnn-daily-mail/) found that much of a popular benchmark could be solved by matching one sentence, and adversarial distractors exposed brittleness in SQuAD models.

### Retriever–reader systems

Open-domain QA settled on a two-stage design: a retriever selects a few passages from millions, and a reader extracts or writes the answer. DrQA used TF-IDF with hashed bigrams; [End-to-End Open-Domain Question Answering with BERTserini](https://md.research.elamur.ai/free/en/papers/yang-2019-end-end-open-domain-question/) paired BM25 from the Anserini toolkit with a BERT reader. Handling many passages at once raised its own problems; [Simple and Effective Multi-Paragraph Reading Comprehension](https://md.research.elamur.ai/free/en/papers/clark-2017-simple-effective-multi-paragraph-reading/) and [Multi-passage BERT](https://md.research.elamur.ai/free/en/papers/wang-2019-multi-passage-bert-globally-normalized/) normalised answer scores across paragraphs so they could be compared.

### Dense retrieval

The main methodological debate of 2019–2021 was whether learned vector retrieval could beat keyword search. [Latent Retrieval for Weakly Supervised Open Domain Question Answering](https://md.research.elamur.ai/free/en/papers/lee-2019-latent-retrieval-weakly-supervised-open/) trained retriever and reader jointly, with a pretraining task to make that feasible. [Dense Passage Retrieval for Open-Domain Question Answering](https://md.research.elamur.ai/free/en/papers/karpukhin-2020-dense-passage-retrieval-open-domain/) showed that a simple dual encoder trained on question–passage pairs suffices. Follow-ups improved training ([RocketQA](https://md.research.elamur.ai/free/en/papers/qu-2021-rocketqa-optimized-training-approach-dense/)), mixed sparse and dense indexes ([a dense–sparse phrase index](https://md.research.elamur.ai/free/en/papers/seo-2019-real-time-open-domain-question/)), expanded queries with generated text ([generation-augmented retrieval](https://md.research.elamur.ai/free/en/papers/mao-2020-generation-augmented-retrieval-open-domain/)) and reranked passages by how likely they make the question ([zero-shot question generation](https://md.research.elamur.ai/free/en/papers/sachan-2022-improving-passage-retrieval-zero-shot/)). [Multi-hop retrieval](https://md.research.elamur.ai/free/en/papers/asai-2019-learning-retrieve-reasoning-paths-over/) followed links between Wikipedia articles.

### Generative readers and unified models

Sequence-to-sequence models changed the reader. Fusion-in-Decoder encoded each retrieved passage separately and let the decoder combine them, and [UNIFIEDQA: Crossing Format Boundaries with a Single QA System](https://md.research.elamur.ai/free/en/papers/khashabi-2020-unifiedqa-crossing-format-boundaries-single/) trained one text-to-text model across extractive, abstractive, multiple-choice and yes/no formats. Earlier, [The Natural Language Decathlon: Multitask Learning as Question Answering](https://md.research.elamur.ai/free/en/papers/mccann-2018-natural-language-decathlon-multitask-learning/) had argued that almost any NLP task can be posed as a question over a context.

### Weak supervision and synthetic data

Weakly supervised training with [hard EM](https://md.research.elamur.ai/free/en/papers/min-2019-discrete-hard-em-approach-weakly/) treats the correct derivation as latent. A [roundtrip filter](https://md.research.elamur.ai/free/en/papers/alberti-2019-synthetic-qa-corpora-generation-roundtrip/) keeps a generated question only if a QA model recovers its answer, and [later work](https://md.research.elamur.ai/free/en/papers/puri-2020-training-question-answering-models-synthetic/) scaled this idea up.

### Large language models

Large models can answer many questions without retrieval and can be prompted with search results. [WebGPT: Browser-assisted question-answering with human feedback](https://md.research.elamur.ai/free/en/papers/nakano-2021-webgpt-browser-assisted-question-answering/) taught a model to browse and cite; [Internet-augmented language models](https://md.research.elamur.ai/free/en/papers/lazaridou-2022-internet-augmented-language-models-through/) conditioned answers on search results through prompting alone; [Leveraging Large Language Models for Multiple Choice Question Answering](https://md.research.elamur.ai/free/en/papers/robinson-2022-leveraging-large-language-models-multiple/) studied how to present options to such models. [ToolQA: A Dataset for LLM Question Answering with External Tools](https://md.research.elamur.ai/free/en/papers/zhuang-2023-toolqa-dataset-llm-question-answering/) separates questions that need tools from those a model may simply remember. Retrieval keeps answers current and checkable, a point sharpened by [Entity-Based Knowledge Conflicts in Question Answering](https://md.research.elamur.ai/free/en/papers/longpre-2021-entity-based-knowledge-conflicts-question/), which examined what happens when retrieved text contradicts what a model memorised.

## Datasets, resources and benchmarks

QA has been driven by its datasets more than most fields:

| Family | Representative datasets |
| --- | --- |
| Extractive reading comprehension | SQuAD, NewsQA, [DuReader: a Chinese Machine Reading Comprehension Dataset from Real-world Applications](https://md.research.elamur.ai/free/en/papers/he-2017-dureader-chinese-machine-reading/), [A Span-Extraction Dataset for Chinese Machine Reading Comprehension](https://md.research.elamur.ai/free/en/papers/cui-2018-span-extraction-dataset-chinese/) |
| Real user questions | Natural Questions, MS MARCO, WikiQA, TyDi QA |
| Trivia and open-domain | TriviaQA, CuratedTREC, WebQuestions |
| Multi-hop | HotpotQA, WikiHop, 2WikiMultiHopQA |
| Conversational | CoQA, QuAC, [DREAM: A Challenge Data Set and Models for Dialogue-Based Reading Comprehension](https://md.research.elamur.ai/free/en/papers/sun-2019-dream-challenge-data-set-models/) |
| Exams and multiple choice | RACE, MultiRC, LogiQA, ReClor, [QuALITY: Question Answering with Long Input Texts, Yes!](https://md.research.elamur.ai/free/en/papers/pang-2022-quality-question-answering-long-input/) |
| Numerical and tabular | DROP, TAT-QA, HybridQA, FeTaQA |
| Biomedical and clinical | PubMedQA, MedQA, MedMCQA, [emrQA: A Large Corpus for Question Answering on Electronic Medical Records](https://md.research.elamur.ai/free/en/papers/pampari-2018-emrqa-large-corpus-question-answering/) |
| Visual and video | VQA v2, GQA, Visual7W, A-OKVQA, TVQA, ActivityNet-QA, ChartQA |
| Multilingual | MLQA, TyDi QA, [XOR QA: Cross-lingual Open-Retrieval Question Answering](https://md.research.elamur.ai/free/en/papers/asai-2021-xor-qa-cross-lingual-open/), [Belebele](https://md.research.elamur.ai/free/en/papers/bandarkar-2023-belebele-benchmark-parallel-reading-comprehension/) |

[the MRQA 2019 shared task](https://md.research.elamur.ai/free/en/papers/fisch-2019-mrqa-2019-shared-task-evaluating/) and [MultiQA](https://md.research.elamur.ai/free/en/papers/talmor-2019-multiqa-empirical-investigation-generalization-transfer/) tested how far a reader fitted to a single dataset transfers elsewhere. [English Machine Reading Comprehension Datasets: A Survey](https://md.research.elamur.ai/free/en/papers/dzendzik-2021-english-machine-reading-comprehension-datasets/) and [A guide to the dataset explosion in QA, NLI, and commonsense reasoning](https://md.research.elamur.ai/free/en/papers/rogers-2020-guide-dataset-explosion-qa-nli/) map the landscape.

The library also holds a notable group of resources for languages other than English, especially Semitic ones: [ArabicaQA: A Comprehensive Dataset for Arabic Question Answering](https://md.research.elamur.ai/free/en/papers/abdallah-2024-arabicaqa-comprehensive-dataset-arabic-question/), [HeQ: a Large and Diverse Hebrew Reading Comprehension Benchmark](https://md.research.elamur.ai/free/en/papers/findings-2023-cohen-heq-large-diverse-hebrew-reading/), [ParaShoot: A Hebrew Question Answering Dataset](https://md.research.elamur.ai/free/en/papers/ud-2021-keren-parashoot-hebrew-question-answering-dataset/), [AmQA: Amharic Question Answering Dataset](https://md.research.elamur.ai/free/en/papers/abedissa-2023-amqa-amharic-question-answering-dataset/), [AmaSQuAD: A Benchmark for Amharic Extractive Question Answering](https://md.research.elamur.ai/free/en/papers/hailemariam-2025-amasquad-benchmark-amharic-extractive-question/), [MedAraBench](https://md.research.elamur.ai/free/en/papers/abu-2026-daoud-medarabench-large-scale-arabic-medical/) and [EDRAC: Benchmarking Arabic Dialect Reading Comprehension](https://md.research.elamur.ai/free/en/papers/mokh-2026-edrac-benchmarking-arabic-dialect-reading/), among others.

## Landmark papers in this library

### Dense Passage Retrieval (2020)

[Dense Passage Retrieval for Open-Domain Question Answering](https://md.research.elamur.ai/free/en/papers/karpukhin-2020-dense-passage-retrieval-open-domain/) by Karpukhin and colleagues at Facebook AI showed that the extra pretraining used by earlier dense retrievers is not needed. Two BERT encoders, one for questions and one for passages, are trained on existing QA pairs, using the correct passages of other questions in a batch as negatives together with one hard negative from BM25. The resulting retriever beat BM25 by a wide margin on most open-domain benchmarks and, with a standard reader, set new end-to-end results. Dense retrieval of this kind became the default first stage of retrieval-augmented systems.

### Natural Questions (2019)

[Natural Questions: A Benchmark for Question Answering Research](https://md.research.elamur.ai/free/en/papers/kwiatkowski-2019-natural-questions-benchmark-question-answering/) from Google built a benchmark from real, anonymised search queries, each paired with a Wikipedia page from the top results. Annotators marked a long answer (a paragraph, list or table), a short answer, or that the page had none. Because the questions were asked by people who did not yet know the answer, they lack the word overlap of crowd-written sets. Its open version became a standard yardstick for retrieval-based QA.

### HotpotQA (2018)

[HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering](https://md.research.elamur.ai/free/en/papers/yang-2018-hotpotqa-dataset-diverse-explainable-multi/) asked crowd workers to write questions linking two Wikipedia paragraphs, either through a bridge entity or by comparing two entities, and to mark the sentences that support the answer. Those supporting facts made the benchmark a test of explanation as well as of accuracy. It was the leading multi-hop dataset for several years and inspired later work such as [Constructing A Multi-hop QA Dataset for Comprehensive Evaluation of Reasoning Steps](https://md.research.elamur.ai/free/en/papers/ho-2020-constructing-multi-hop-qa-dataset/), which added reasoning paths drawn from Wikidata.

### TriviaQA (2017)

[TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension](https://md.research.elamur.ai/free/en/papers/joshi-2017-triviaqa-large-scale-distantly-supervised/) took questions written by trivia enthusiasts and gathered evidence documents afterwards from the web and Wikipedia. Since the questions were not written with a passage in view, they differ in wording from the evidence and often need several sentences.

### DrQA: Reading Wikipedia (2017)

[Reading Wikipedia to Answer Open-Domain Questions](https://md.research.elamur.ai/free/en/papers/chen-2017-reading-wikipedia-answer-open-domain/) by Chen, Fisch, Weston and Bordes revived open-domain QA in the neural era. A fast non-learned retriever picks a handful of Wikipedia articles, and a recurrent reader finds the answer span in them. The paper showed how much accuracy is lost when a reader trained on SQuAD faces all of Wikipedia rather than a gold paragraph, and it fixed the retriever–reader template that later work refined.

### Adversarial examples for reading comprehension (2017)

[Adversarial Examples for Evaluating Reading Comprehension Systems](https://md.research.elamur.ai/free/en/papers/jia-2017-adversarial-examples-evaluating-reading-comprehension/) by Jia and Liang appended a single misleading sentence to SQuAD paragraphs: it resembled the question but did not answer it. The accuracy of published models fell by more than half, while human readers were barely affected. It made adversarial evaluation a standard part of the field.

### Fusion-in-Decoder (2021)

[Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering](https://md.research.elamur.ai/free/en/papers/izacard-2021-leveraging-passage-retrieval-generative-models/) by Izacard and Grave encoded each retrieved passage separately with T5 and let the decoder attend over all of them. Since the expense rises only in proportion to how many passages are read, the reader could take in up to a hundred, and accuracy kept rising as more were added. A mid-sized model with retrieval outperformed a far larger closed-book model, a strong argument for combining retrieval with generation.

### ORQA: Latent retrieval (2019)

[Latent Retrieval for Weakly Supervised Open Domain Question Answering](https://md.research.elamur.ai/free/en/papers/lee-2019-latent-retrieval-weakly-supervised-open/) by Lee, Chang and Toutanova trained retrieval and reading together from question–answer pairs alone, treating the evidence block as a hidden variable and warming up the retriever with the Inverse Cloze Task. It found that learned retrieval helps most when questions come from genuine information seekers, while BM25 remains strong when question writers had seen the answer.

### CoQA and QuAC (2018–2019)

[CoQA: A Conversational Question Answering Challenge](https://md.research.elamur.ai/free/en/papers/reddy-2019-coqa-conversational-question-answering-challenge/) and [QuAC: Question Answering in Context](https://md.research.elamur.ai/free/en/papers/choi-2018-quac-question-answering-context/) appeared at nearly the same time and defined conversational QA. CoQA pairs free-form answers with supporting rationales across several domains; QuAC hides the text from the asker, so the questions are more exploratory and many cannot be answered. Both showed that resolving references to earlier turns is central to the task.

### RACE (2017)

[RACE: Large-scale ReAding Comprehension Dataset From Examinations](https://md.research.elamur.ai/free/en/papers/lai-2017-race-large-scale-reading-comprehension/) collected English exams written by teachers for Chinese secondary-school students. The options are full sentences, often absent from the passage, and many questions ask for a main idea or an author's attitude.

### DROP (2019)

[DROP: A Reading Comprehension Benchmark Requiring Discrete Reasoning Over Paragraphs](https://md.research.elamur.ai/free/en/papers/dua-2019-drop-reading-comprehension-benchmark-requiring/) was collected adversarially against a reading model, so that its questions need counting, sorting, addition or subtraction over several parts of a passage. BERT fell far below its SQuAD level, and the authors' reader with small numerical modules narrowed but did not close the gap to humans. DROP pushed the field toward neural-symbolic and program-based reasoning.

### Making the V in VQA Matter (2016)

[Making the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering](https://md.research.elamur.ai/free/en/papers/goyal-2016-making-v-vqa-matter-elevating/) paired nearly every question with two similar images that require different answers. Models that had exploited regularities in the questions lost accuracy on the balanced data, which became VQA v2.0. Together with [GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering](https://md.research.elamur.ai/free/en/papers/hudson-2019-gqa-new-dataset-real-world/) and [Multimodal Compact Bilinear Pooling for Visual Question Answering and Visual Grounding](https://md.research.elamur.ai/free/en/papers/fukui-2016-multimodal-compact-bilinear-pooling-visual/), it defines the visual side of this topic.

### PubMedQA and MedQA (2019–2020)

[PubMedQA: A Dataset for Biomedical Research Question Answering](https://md.research.elamur.ai/free/en/papers/jin-2019-pubmedqa-dataset-biomedical-research-question/) turned PubMed articles with question titles into yes/no/maybe questions answered from the abstract, and [What Disease does this Patient Have? A Large-scale Open Domain Question Answering Dataset from Medical Exams](https://md.research.elamur.ai/free/en/papers/jin-2020-disease-this-patient-have-large/) drew questions from licensing exams in three regions. Both later served as targets for medical language models such as the one in [Towards Expert-Level Medical Question Answering with Large Language Models](https://md.research.elamur.ai/free/en/papers/singhal-2023-towards-expert-level-medical-question/).

### WebGPT (2021)

[WebGPT: Browser-assisted question-answering with human feedback](https://md.research.elamur.ai/free/en/papers/nakano-2021-webgpt-browser-assisted-question-answering/) let GPT-3 operate a text browser, quote pages and write long answers with references, trained on human demonstrations and comparisons. The work joined long-form QA, retrieval and learning from human feedback, and foreshadowed today's search-backed assistants.

Citations of the papers on this topic, by when they were published Papers published in 2015–2019 gather 66% of the 79,285 citations of the papers on this topic.

**The numbers**

| Published | Papers and their citations |
| --- | --- |
| 2025–2029 | 28 papers, 63 citations |
| 2020–2024 | 76 papers, 23,434 citations |
| 2015–2019 | 70 papers, 52,208 citations |
| 2010–2014 | 3 papers, 1,209 citations |
| 2005–2009 | 2 papers, 1,025 citations |
| 2000–2004 | 3 papers, 1,311 citations |
| 1995–1999 | 1 papers, 35 citations |

## People

Authors with the most papers on this topic Luke Zettlemoyer leads with 10 papers; the 12 authors shown wrote 68 of the papers on this topic between them, counting each paper once per author.

**The numbers**

| Author | Papers here · their citations |
| --- | --- |
| [Luke Zettlemoyer](https://md.research.elamur.ai/free/en/authors/luke-zettlemoyer/) | 10 papers · 5,986 citations |
| [Wen-tau Yih](https://md.research.elamur.ai/free/en/authors/wen-tau-yih/) | 9 papers · 9,226 citations |
| [Hannaneh Hajishirzi](https://md.research.elamur.ai/free/en/authors/hannaneh-hajishirzi/) | 7 papers · 2,142 citations |
| [Danqi Chen](https://md.research.elamur.ai/free/en/authors/danqi-chen/) | 6 papers · 10,458 citations |
| [Christopher D. Manning](https://md.research.elamur.ai/free/en/authors/christopher-d-manning/) | 6 papers · 8,012 citations |
| [Sewon Min](https://md.research.elamur.ai/free/en/authors/sewon-min/) | 5 papers · 6,819 citations |
| [Eunsol Choi](https://md.research.elamur.ai/free/en/authors/eunsol-choi/) | 5 papers · 5,533 citations |
| [Tom Kwiatkowski](https://md.research.elamur.ai/free/en/authors/tom-kwiatkowski/) | 4 papers · 5,074 citations |
| [Yejin Choi](https://md.research.elamur.ai/free/en/authors/yejin-choi/) | 4 papers · 2,100 citations |
| [Alon Talmor](https://md.research.elamur.ai/free/en/authors/alon-talmor/) | 4 papers · 2,071 citations |
| [Ming-Wei Chang](https://md.research.elamur.ai/free/en/authors/ming-wei-chang/) | 4 papers · 2,018 citations |
| [Jonathan Berant](https://md.research.elamur.ai/free/en/authors/jonathan-berant/) | 4 papers · 1,984 citations |

**[Luke Zettlemoyer](https://md.research.elamur.ai/free/en/authors/luke-zettlemoyer/)** (University of Washington and Meta) is the most prolific author on this topic in the library. His work here runs from paraphrase-driven KBQA to TriviaQA, QuAC and question decomposition.

**[Wen-tau Yih](https://md.research.elamur.ai/free/en/authors/wen-tau-yih/)** (University of Washington) links the knowledge-base and open-domain strands: single-relation semantic parsing, the WikiQA answer-selection dataset, QuAC and Dense Passage Retrieval.

**[Danqi Chen](https://md.research.elamur.ai/free/en/authors/danqi-chen/)** (Princeton University) is behind several turning points: the critical study of the CNN/Daily Mail task, DrQA, CoQA and Dense Passage Retrieval.

**[Christopher D. Manning](https://md.research.elamur.ai/free/en/authors/christopher-d-manning/)** (Stanford University) co-authored HotpotQA, GQA, CoQA and the CNN/Daily Mail analysis, tying together reading comprehension, multi-hop reasoning and visual QA.

**[Hannaneh Hajishirzi](https://md.research.elamur.ai/free/en/authors/hannaneh-hajishirzi/)** (University of Washington and the Allen Institute for Artificial Intelligence) works on unified and multi-hop QA, including UnifiedQA, question decomposition, reasoning paths over the Wikipedia graph and weakly supervised training.

**[Sewon Min](https://md.research.elamur.ai/free/en/authors/sewon-min/)** contributed to Dense Passage Retrieval, UnifiedQA, question decomposition and hard-EM weak supervision for QA.

**[Eunsol Choi](https://md.research.elamur.ai/free/en/authors/eunsol-choi/)** (New York University) led QuAC and co-authored TriviaQA, TyDi QA and the MRQA shared task on generalisation.

**[Tom Kwiatkowski](https://md.research.elamur.ai/free/en/authors/tom-kwiatkowski/)** led Natural Questions and co-authored TyDi QA, the dense–sparse phrase index and work on attributed QA for large language models.

**[Kenton Lee](https://md.research.elamur.ai/free/en/authors/kenton-lee/)** co-authored ORQA, Natural Questions and XOR QA, with [Kristina Toutanova](https://md.research.elamur.ai/free/en/authors/kristina-toutanova/) (Google) and [Ming-Wei Chang](https://md.research.elamur.ai/free/en/authors/ming-wei-chang/) as close collaborators on learned retrieval.

**[Jason Weston](https://md.research.elamur.ai/free/en/authors/jason-weston/)** (New York University) proposed the bAbI tasks and co-authored DrQA and ELI5, spanning synthetic reasoning tests, open-domain retrieval and long-form answers.

**[Matt Gardner](https://md.research.elamur.ai/free/en/authors/matt-gardner/)** co-authored DROP, Quoref and TORQUE and the multi-paragraph reading model with Christopher Clark, all aimed at reading that cannot be solved by surface matching.

**[Yejin Choi](https://md.research.elamur.ai/free/en/authors/yejin-choi/)** (Stanford University) brings commonsense into QA through Cosmos QA and DREAM, and freshness of knowledge through RealTime QA.

**[Caiming Xiong](https://md.research.elamur.ai/free/en/authors/caiming-xiong/)** (Salesforce) and **[Richard Socher](https://md.research.elamur.ai/free/en/authors/richard-socher/)** co-authored the Dynamic Coattention Network and the Natural Language Decathlon, which cast many NLP tasks as QA.

**[Patrick Lewis](https://md.research.elamur.ai/free/en/authors/patrick-lewis/)** (Cohere) co-authored DPR and MLQA and worked on unsupervised question decomposition.

[Michael Collins](https://md.research.elamur.ai/free/en/authors/michael-collins/), [Jennimaria Palomaki](https://md.research.elamur.ai/free/en/authors/jennimaria-palomaki/), [Chris Alberti](https://md.research.elamur.ai/free/en/authors/chris-alberti/) and [Jacob Devlin](https://md.research.elamur.ai/free/en/authors/jacob-devlin/) belong to the Google group behind Natural Questions and TyDi QA.

**[Siva Reddy](https://md.research.elamur.ai/free/en/authors/siva-reddy/)** (McGill University and Mila) led CoQA and worked on Freebase QA and learned paraphrasing of questions; **[Mohit Iyyer](https://md.research.elamur.ai/free/en/authors/mohit-iyyer/)** (University of Massachusetts Amherst) worked on sequential QA over tables and on the evaluation of long-form answers; **[Eduard Hovy](https://md.research.elamur.ai/free/en/authors/eduard-hovy/)** co-authored both the TREC-era surface-pattern system and RACE, bridging two generations of the field.

## Institutions and places

Where the authors on this topic work now Google (United States) is the current home of authors of 28 papers on this topic; counted are the authors with a page here, by their latest institution.

**The numbers**

| Institution | Papers |
| --- | --- |
| [Google (United States)](https://md.research.elamur.ai/free/en/companies/google/) | 28 papers |
| [Carnegie Mellon University](https://md.research.elamur.ai/free/en/universities/carnegie-mellon-university/) | 25 papers |
| [University of Washington](https://md.research.elamur.ai/free/en/universities/university-of-washington/) | 25 papers |
| [Stanford University](https://md.research.elamur.ai/free/en/universities/stanford-university/) | 20 papers |
| [Allen Institute for Artificial Intelligence](https://md.research.elamur.ai/free/en/labs/allen-institute-for-artificial-intelligence/) | 19 papers |
| [Meta (United States)](https://md.research.elamur.ai/free/en/companies/meta/) | 15 papers |
| [New York University](https://md.research.elamur.ai/free/en/universities/new-york-university/) | 15 papers |
| [Chinese Academy of Sciences](https://md.research.elamur.ai/free/en/government/chinese-academy-of-sciences/) | 14 papers |
| [Google DeepMind (United Kingdom)](https://md.research.elamur.ai/free/en/companies/google-deepmind/) | 13 papers |
| [University of Chinese Academy of Sciences](https://md.research.elamur.ai/free/en/universities/university-of-chinese-academy-of-sciences/) | 12 papers |

[Google](https://md.research.elamur.ai/free/en/companies/google/) and [Google DeepMind](https://md.research.elamur.ai/free/en/companies/google-deepmind/) produced Natural Questions, TyDi QA and work on attributed and medical QA. [Meta](https://md.research.elamur.ai/free/en/companies/meta/) (formerly Facebook AI Research) produced DrQA, DPR, MLQA, ELI5 and Fusion-in-Decoder. Among universities, the [University of Washington](https://md.research.elamur.ai/free/en/universities/university-of-washington/) and the [Allen Institute for Artificial Intelligence](https://md.research.elamur.ai/free/en/labs/allen-institute-for-artificial-intelligence/) in Seattle form one of the densest clusters, responsible for TriviaQA, QuAC, UnifiedQA, DROP and much multi-hop work. [Stanford University](https://md.research.elamur.ai/free/en/universities/stanford-university/) produced SQuAD, HotpotQA, CoQA and GQA; [Carnegie Mellon University](https://md.research.elamur.ai/free/en/universities/carnegie-mellon-university/) contributed RACE, QANet and HotpotQA co-authors; [New York University](https://md.research.elamur.ai/free/en/universities/new-york-university/) is home to the BBQ bias benchmark and QuALITY. Other active centres include the [Massachusetts Institute of Technology](https://md.research.elamur.ai/free/en/universities/massachusetts-institute-of-technology/), the [University of California, Berkeley](https://md.research.elamur.ai/free/en/universities/university-of-california-berkeley/), [Princeton University](https://md.research.elamur.ai/free/en/universities/princeton-university/), the [University of Maryland, College Park](https://md.research.elamur.ai/free/en/universities/university-of-maryland-college-park/), the [University of Waterloo](https://md.research.elamur.ai/free/en/universities/university-of-waterloo/) (BERTserini), [Mila](https://md.research.elamur.ai/free/en/labs/mila-quebec-artificial-intelligence-institute/), the [Canadian Institute for Advanced Research](https://md.research.elamur.ai/free/en/labs/canadian-institute-for-advanced-research/) and [Amazon](https://md.research.elamur.ai/free/en/companies/amazon/). In China, the [Chinese Academy of Sciences](https://md.research.elamur.ai/free/en/government/chinese-academy-of-sciences/), the [University of Chinese Academy of Sciences](https://md.research.elamur.ai/free/en/universities/university-of-chinese-academy-of-sciences/) and the [University of Science and Technology of China](https://md.research.elamur.ai/free/en/universities/university-of-science-and-technology-of-china/) are prominent, alongside industrial work such as DuReader and RocketQA from Baidu.

Countries of the authors on this topic United States comes first: authors working there now wrote 124 papers on this topic.

**The numbers**

| Country | Papers |
| --- | --- |
| United States | 124 papers |
| China | 61 papers |
| United Kingdom | 32 papers |
| Israel | 17 papers |
| Canada | 16 papers |
| Hong Kong | 11 papers |
| Japan | 11 papers |
| Germany | 10 papers |
| India | 8 papers |
| Australia | 7 papers |

The United States dominates by a wide margin, followed by China, the United Kingdom and Canada.

Where the papers on this topic were published Proceedings of the Fifteenth Language Resources and Evaluation Conference carries the most, 14 of 183.

**The numbers**

| Venue | Papers |
| --- | --- |
| Proceedings of the Fifteenth Language Resources and Evaluation Conference | 14 papers |
| Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) | 8 papers |
| Transactions of the Association for Computational Linguistics | 8 papers |
| Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics | 7 papers |
| Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing | 6 papers |
| Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics | 5 papers |
| Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing | 4 papers |
| Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP) | 4 papers |
| Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies | 4 papers |
| Proceedings of the 55th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) | 4 papers |

Most papers appear at the main computational linguistics venues: the ACL annual meetings, EMNLP, NAACL, EACL and the journal Transactions of the ACL. The Language Resources and Evaluation Conference is prominent too, since so much QA work consists of building new datasets, especially for less-resourced languages.

## Open questions

**Do benchmarks measure understanding?** High scores have repeatedly rested on artifacts such as word overlap and answer-type priors, and each new benchmark tends to be saturated faster than the last.

**Evaluation of free-form answers.** Exact match and token F1 penalise correct answers phrased differently, while ROUGE rewards fluent text that may be wrong. [Evaluating Open-Domain Question Answering in the Era of Large Language Models](https://md.research.elamur.ai/free/en/papers/kamalloo-2023-evaluating-open-domain-question-answering/) found that lexical metrics badly underrate large models, and work such as [MORQA](https://md.research.elamur.ai/free/en/papers/yim-2026-morqa-benchmarking-evaluation-metrics-medical/) tests metrics against expert judgment.

**Attribution and faithfulness.** A useful answer should be traceable to a source. [Attributed question answering](https://md.research.elamur.ai/free/en/papers/bohnet-2022-attributed-question-answering-evaluation-modeling/) frames this as a task of its own, and [HalluScore](https://md.research.elamur.ai/free/en/papers/alansari-2026-halluscore-large-language-model-hallucination/) targets unsupported claims directly.

**False premises and pragmatics.** Real questions often rest on mistaken assumptions. [CREPE](https://md.research.elamur.ai/free/en/papers/yu-2023-crepe-open-domain-question-answering/), [presupposition verification](https://md.research.elamur.ai/free/en/papers/kim-2021-which-linguist-invented-lightbulb-presupposition/) and [IfQA](https://md.research.elamur.ai/free/en/papers/yu-2023-ifqa-dataset-open-domain-question/) ask systems to notice such premises, while [PragmatiCQA](https://md.research.elamur.ai/free/en/papers/qi-2023-pragmaticqa-dataset-pragmatic-question-answering/) and [work on maternal-health questions](https://md.research.elamur.ai/free/en/papers/srikanth-2024-pregnant-questions-importance-pragmatic-awareness/) ask them to answer what the user actually needs.

**Time-sensitive knowledge.** [RealTime QA: What's the Answer Right Now?](https://md.research.elamur.ai/free/en/papers/kasai-2022-realtime-qa-s-answer-right/) evaluates systems on questions about current events.

**Bias and safety.** [BBQ: A hand-built bias benchmark for question answering](https://md.research.elamur.ai/free/en/papers/parrish-2022-bbq-hand-built-bias-benchmark/) showed that QA models fall back on stereotypes when context is ambiguous; medical and legal QA raise further questions of harm.

**Languages and domains.** Most resources remain English and Wikipedia-based. Work on Arabic, Hebrew, Amharic and dialects, on legal QA such as [MizanQA](https://md.research.elamur.ai/free/en/papers/bahaj-2025-mizanqa-benchmarking-large-language-models/) and [CBR-RAG](https://md.research.elamur.ai/free/en/papers/wiratunga-2024-cbr-rag-case-based-reasoning/), and on cross-dialect gaps such as [a study of Cantonese–Mandarin and Bavarian–German QA](https://md.research.elamur.ai/free/en/papers/pei-2026-information-asymmetry-across-language-varieties/) addresses this unevenly.

**Reasoning that is genuinely compositional.** Multi-hop and numerical benchmarks still show that models take shortcuts. Recent sets such as [CRiT-QA](https://md.research.elamur.ai/free/en/papers/yun-2026-crit-qa-evaluating-multi-hop/) and [graph-structured reasoning QA](https://md.research.elamur.ai/free/en/papers/yone-2026-reasoning-graph-structured-question-answering/) try to make shortcuts fail.

## Neighbouring topics

Neighbouring topics: papers shared Reasoning shares the most: 24 of the 183 papers here (13%) stand on both.

**The numbers**

| Topic | Papers shared |
| --- | --- |
| [Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/) | 24 shared · 13% |
| [Evaluation & benchmarks](https://md.research.elamur.ai/free/en/topics/evaluation-benchmarks/) | 23 shared · 13% |
| [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) | 21 shared · 11% |
| [Information retrieval & search](https://md.research.elamur.ai/free/en/topics/information-retrieval/) | 17 shared · 9% |
| [Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) | 14 shared · 8% |
| [Knowledge graphs & ontologies](https://md.research.elamur.ai/free/en/topics/knowledge-graphs/) | 7 shared · 4% |
| [Multilingual & cross-lingual NLP](https://md.research.elamur.ai/free/en/topics/multilingual-nlp/) | 7 shared · 4% |
| [Dialogue systems](https://md.research.elamur.ai/free/en/topics/dialogue-systems/) | 6 shared · 3% |
| [Vision & language](https://md.research.elamur.ai/free/en/topics/multimodal/) | 6 shared · 3% |
| [Coreference resolution](https://md.research.elamur.ai/free/en/topics/coreference-resolution/) | 5 shared · 3% |

Question answering shares the most papers with [Evaluation & benchmarks](https://md.research.elamur.ai/free/en/topics/evaluation-benchmarks/), unsurprising for a field built on leaderboards, and with [Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/), since multi-hop, logical and numerical QA are among the main ways reasoning is measured. The large overlap with [Semitic languages](https://md.research.elamur.ai/free/en/topics/semitic-languages/) and [Hebrew & Aramaic](https://md.research.elamur.ai/free/en/topics/hebrew-aramaic/) reflects the many Arabic and Hebrew QA datasets in this library. [Information retrieval & search](https://md.research.elamur.ai/free/en/topics/information-retrieval/) supplies the first stage of open-domain systems, and [Language models & pretraining](https://md.research.elamur.ai/free/en/topics/language-models/) supplies the readers and, increasingly, the whole system.

Other links run through specific subtasks. Multilingual benchmarks connect to [Multilingual & cross-lingual NLP](https://md.research.elamur.ai/free/en/topics/multilingual-nlp/) and [Low-resource languages](https://md.research.elamur.ai/free/en/topics/low-resource-languages/); conversational QA to [Dialogue systems](https://md.research.elamur.ai/free/en/topics/dialogue-systems/); visual QA to [Vision & language](https://md.research.elamur.ai/free/en/topics/multimodal/); and KBQA to [Knowledge graphs & ontologies](https://md.research.elamur.ai/free/en/topics/knowledge-graphs/) and [Semantic parsing & meaning representation](https://md.research.elamur.ai/free/en/topics/semantic-parsing/). The QA format is reused in [Coreference resolution](https://md.research.elamur.ai/free/en/topics/coreference-resolution/), [Semantic role labeling](https://md.research.elamur.ai/free/en/topics/semantic-role-labeling/), [Information extraction](https://md.research.elamur.ai/free/en/topics/information-extraction/) and [Summarization](https://md.research.elamur.ai/free/en/topics/summarization/), while presupposition and pragmatic QA meet [Implicature](https://md.research.elamur.ai/free/en/topics/implicature/). Domain work touches [Legal AI](https://md.research.elamur.ai/free/en/topics/legal-ai/) and [Bias, fairness & safety](https://md.research.elamur.ai/free/en/topics/bias-fairness-safety/), and question generation borders [Text generation](https://md.research.elamur.ai/free/en/topics/text-generation/).

The shelves this topic stands on Natural language processing: 79 of the 183 papers here.

**The numbers**

| Shelf | Papers |
| --- | --- |
| [Natural language processing](https://md.research.elamur.ai/free/en/shelves/nlp/) | 79 papers |
| [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/) | 70 papers |
| [Other](https://md.research.elamur.ai/free/en/shelves/other/) | 30 papers |
| [AI & machine learning](https://md.research.elamur.ai/free/en/shelves/ai/) | 15 papers |
| [Machine translation](https://md.research.elamur.ai/free/en/shelves/machine-translation/) | 5 papers |
| [Mathematics & statistics](https://md.research.elamur.ai/free/en/shelves/mathematics/) | 3 papers |
| [Computing & GPUs](https://md.research.elamur.ai/free/en/shelves/computing/) | 2 papers |
| [Law](https://md.research.elamur.ai/free/en/shelves/law/) | 2 papers |

The topic stands mainly on the [Natural language processing](https://md.research.elamur.ai/free/en/shelves/nlp/) and [Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/) shelves, with smaller parts on [AI & machine learning](https://md.research.elamur.ai/free/en/shelves/ai/), [Machine translation](https://md.research.elamur.ai/free/en/shelves/machine-translation/), [Mathematics & statistics](https://md.research.elamur.ai/free/en/shelves/mathematics/) and [Law](https://md.research.elamur.ai/free/en/shelves/law/).

## Key works in this library

- [The TREC-8 Question Answering Track](https://md.research.elamur.ai/free/en/papers/voorhees-2000-trec-8-question-answering-track/) — the evaluation that defined open-domain factoid QA.
- [Learning surface text patterns for a Question Answering System](https://md.research.elamur.ai/free/en/papers/ravichandran-2002-learning-surface-text-patterns-question/) — answer patterns bootstrapped from the web.
- [What is the Jeopardy Model? A Quasi-Synchronous Grammar for QA](https://md.research.elamur.ai/free/en/papers/wang-2007-what-is-jeopardy-model-quasi/) — answer selection as soft tree alignment.
- [Towards AI-Complete Question Answering: A Set of Prerequisite Toy Tasks](https://md.research.elamur.ai/free/en/papers/weston-2015-towards-ai-complete-question-answering/) — bAbI, unit tests for reasoning.
- [Making the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering](https://md.research.elamur.ai/free/en/papers/goyal-2016-making-v-vqa-matter-elevating/) — balanced visual QA.
- [Reading Wikipedia to Answer Open-Domain Questions](https://md.research.elamur.ai/free/en/papers/chen-2017-reading-wikipedia-answer-open-domain/) — DrQA's retriever–reader design.
- [TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension](https://md.research.elamur.ai/free/en/papers/joshi-2017-triviaqa-large-scale-distantly-supervised/) — trivia questions with evidence gathered afterwards.
- [Adversarial Examples for Evaluating Reading Comprehension Systems](https://md.research.elamur.ai/free/en/papers/jia-2017-adversarial-examples-evaluating-reading-comprehension/) — distractor sentences that break SQuAD models.
- [HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering](https://md.research.elamur.ai/free/en/papers/yang-2018-hotpotqa-dataset-diverse-explainable-multi/) — multi-hop QA with supporting facts.
- [QuAC: Question Answering in Context](https://md.research.elamur.ai/free/en/papers/choi-2018-quac-question-answering-context/) — information-seeking dialogue over hidden text.
- [CoQA: A Conversational Question Answering Challenge](https://md.research.elamur.ai/free/en/papers/reddy-2019-coqa-conversational-question-answering-challenge/) — free-form conversational answers with rationales.
- [Natural Questions: A Benchmark for Question Answering Research](https://md.research.elamur.ai/free/en/papers/kwiatkowski-2019-natural-questions-benchmark-question-answering/) — real search queries with long and short answers.
- [DROP: A Reading Comprehension Benchmark Requiring Discrete Reasoning Over Paragraphs](https://md.research.elamur.ai/free/en/papers/dua-2019-drop-reading-comprehension-benchmark-requiring/) — arithmetic and counting over text.
- [Dense Passage Retrieval for Open-Domain Question Answering](https://md.research.elamur.ai/free/en/papers/karpukhin-2020-dense-passage-retrieval-open-domain/) — the dual-encoder retriever.
- [TyDi QA: A Benchmark for Information-Seeking Question Answering in Typologically Diverse Languages](https://md.research.elamur.ai/free/en/papers/clark-2020-tydi-qa-benchmark-information-seeking/) — native questions in eleven languages.
- [Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering](https://md.research.elamur.ai/free/en/papers/izacard-2021-leveraging-passage-retrieval-generative-models/) — Fusion-in-Decoder.
- [WebGPT: Browser-assisted question-answering with human feedback](https://md.research.elamur.ai/free/en/papers/nakano-2021-webgpt-browser-assisted-question-answering/) — a browsing, citing answerer trained from human feedback.

## All 183 papers on this topic

**1** [2](https://md.research.elamur.ai/free/en/topics/question-answering/page/2/) … [8](https://md.research.elamur.ai/free/en/topics/question-answering/page/8/) [next →](https://md.research.elamur.ai/free/en/topics/question-answering/page/2/)

- [Dense Passage Retrieval for Open-Domain Question Answering](https://md.research.elamur.ai/free/en/papers/karpukhin-2020-dense-passage-retrieval-open-domain/) — Vladimir Karpukhin, Barlas Oguz, Sewon Min, et al. · 2020 · 5 370 citations
- [Natural Questions: A Benchmark for Question Answering Research](https://md.research.elamur.ai/free/en/papers/kwiatkowski-2019-natural-questions-benchmark-question-answering/) — Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, et al. · 2019 · 4 031 citations
- [HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering](https://md.research.elamur.ai/free/en/papers/yang-2018-hotpotqa-dataset-diverse-explainable-multi/) — Zhilin Yang, Peng Qi, Saizheng Zhang, et al. · 2018 · 3 489 citations
- [TriviaQA: A Large Scale Distantly Supervised Challenge Dataset for Reading Comprehension](https://md.research.elamur.ai/free/en/papers/joshi-2017-triviaqa-large-scale-distantly-supervised/) — Mandar Joshi, Eunsol Choi, Daniel Weld, et al. · 2017 · 3 204 citations
- [Reading Wikipedia to Answer Open-Domain Questions](https://md.research.elamur.ai/free/en/papers/chen-2017-reading-wikipedia-answer-open-domain/) — Danqi Chen, Adam Fisch, Jason Weston, et al. · 2017 · 2 250 citations
- [GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering](https://md.research.elamur.ai/free/en/papers/hudson-2019-gqa-new-dataset-real-world/) — Drew A. Hudson, Christopher D. Manning · 2019 · 1 997 citations
- [Adversarial Examples for Evaluating Reading Comprehension Systems](https://md.research.elamur.ai/free/en/papers/jia-2017-adversarial-examples-evaluating-reading-comprehension/) — Robin Jia, Percy Liang · 2017 · 1 935 citations
- [Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering](https://md.research.elamur.ai/free/en/papers/izacard-2021-leveraging-passage-retrieval-generative-models/) — Gautier Izacard, Edouard Grave · 2021 · 1 674 citations
- [Multimodal Compact Bilinear Pooling for Visual Question Answering and Visual Grounding](https://md.research.elamur.ai/free/en/papers/fukui-2016-multimodal-compact-bilinear-pooling-visual/) — Akira Fukui, Dong Huk Park, Daylen Yang, et al. · 2016 · 1 642 citations
- [RACE: Large-scale ReAding Comprehension Dataset From Examinations](https://md.research.elamur.ai/free/en/papers/lai-2017-race-large-scale-reading-comprehension/) — Guokun Lai, Qizhe Xie, Hanxiao Liu, et al. · 2017 · 1 619 citations
- [CoQA: A Conversational Question Answering Challenge](https://md.research.elamur.ai/free/en/papers/reddy-2019-coqa-conversational-question-answering-challenge/) — Siva Reddy, Danqi Chen, Christopher D. Manning · 2019 · 1 541 citations
- [Latent Retrieval for Weakly Supervised Open Domain Question Answering](https://md.research.elamur.ai/free/en/papers/lee-2019-latent-retrieval-weakly-supervised-open/) — Kenton Lee, Ming-Wei Chang, Kristina Toutanova · 2019 · 1 234 citations
- [Making the V in VQA Matter: Elevating the Role of Image Understanding in Visual Question Answering](https://md.research.elamur.ai/free/en/papers/goyal-2016-making-v-vqa-matter-elevating/) — Yash Goyal, Tejas Khot, Douglas Summers-Stay, et al. · 2016 · 1 169 citations
- [PubMedQA: A Dataset for Biomedical Research Question Answering](https://md.research.elamur.ai/free/en/papers/jin-2019-pubmedqa-dataset-biomedical-research-question/) — Qiao Jin, Bhuwan Dhingra, Zhengping Liu, et al. · 2019 · 1 152 citations
- [CommonsenseQA: A Question Answering Challenge Targeting Commonsense Knowledge](https://md.research.elamur.ai/free/en/papers/talmor-2019-commonsenseqa-question-answering-challenge-targeting/) — Alon Talmor, Jonathan Herzig, Nicholas Lourie, et al. · 2019 · 1 146 citations
- [WikiQA: A Challenge Dataset for Open-Domain Question Answering](https://md.research.elamur.ai/free/en/papers/yang-2015-wikiqa-challenge-dataset-open-domain/) — Yi Yang, Wen-tau Yih, Christopher Meek · 2015 · 1 117 citations
- [MS MARCO: A Human Generated MAchine Reading COmprehension Dataset](https://md.research.elamur.ai/free/en/papers/bajaj-2016-ms-marco-human-generated/) — Payal Bajaj, Daniel Campos, Nick Craswell, et al. · 2016 · 1 103 citations
- [QuAC: Question Answering in Context](https://md.research.elamur.ai/free/en/papers/choi-2018-quac-question-answering-context/) — Eunsol Choi, He He, Mohit Iyyer, et al. · 2018 · 1 057 citations
- [The NarrativeQA Reading Comprehension Challenge](https://md.research.elamur.ai/free/en/papers/kocisky-2018-narrativeqa-reading-comprehension-challenge/) — Tomáš Kočiský, Jonathan Schwarz, Phil Blunsom, et al. · 2018 · 1 026 citations
- [Learning surface text patterns for a Question Answering System](https://md.research.elamur.ai/free/en/papers/ravichandran-2002-learning-surface-text-patterns-question/) — Deepak Ravichandran, Eduard Hovy · 2002 · 970 citations
- [DROP: A Reading Comprehension Benchmark Requiring Discrete Reasoning Over Paragraphs](https://md.research.elamur.ai/free/en/papers/dua-2019-drop-reading-comprehension-benchmark-requiring/) — Dheeru Dua, Yizhong Wang, Pradeep Dasigi, et al. · 2019 · 941 citations
- [Visual7W: Grounded Question Answering in Images](https://md.research.elamur.ai/free/en/papers/zhu-2015-visual7w-grounded-question-answering-images/) — Yuke Zhu, Oliver Groth, Michael Bernstein, et al. · 2015 · 915 citations
- [Learning to Ask: Neural Question Generation for Reading Comprehension](https://md.research.elamur.ai/free/en/papers/du-2017-learning-ask-neural-question-generation/) — Xinya Du, Junru Shao, Claire Cardie · 2017 · 913 citations
- [UNIFIEDQA: Crossing Format Boundaries with a Single QA System](https://md.research.elamur.ai/free/en/papers/khashabi-2020-unifiedqa-crossing-format-boundaries-single/) — Daniel Khashabi, Sewon Min, Tushar Khot, et al. · 2020 · 843 citations
- [Gated Self-Matching Networks for Reading Comprehension and Question Answering](https://md.research.elamur.ai/free/en/papers/wang-2017-gated-self-matching-networks-reading/) — Wenhui Wang, Nan Yang, Furu Wei, et al. · 2017 · 823 citations

**1** [2](https://md.research.elamur.ai/free/en/topics/question-answering/page/2/) … [8](https://md.research.elamur.ai/free/en/topics/question-answering/page/8/) [next →](https://md.research.elamur.ai/free/en/topics/question-answering/page/2/)

- [← previous topic Psycholinguistics & language acquisition](https://md.research.elamur.ai/free/en/topics/psycholinguistics/)
- [next topic → Reasoning](https://md.research.elamur.ai/free/en/topics/reasoning/)
- [↑ back to topics](https://md.research.elamur.ai/free/en/topics/)