Back to home
All Datasets 🏷
All datasets from our datasets repository and community bucket.
Also check out the list of supported Metrics 📉.
106 results
AdversarialQA is a Reading Comprehension dataset, consisting of questions posed by crowdworkers on a set of Wikipedia articles using an adversarial model-in-the-loop. We use three different models; BiDAF (Seo et al., 2016), BERT-Large (Devlin et al., 2018), and RoBERTa-Large (Liu et al., 2019) in the annotation loop and construct 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.
The Amazon reviews dataset consists of reviews from amazon. The data span a period of 18 years, including ~35 million reviews up to March 2013. Reviews include product and user information, ratings, and a plaintext review.
AmbigNQ, a dataset covering 14,042 questions from NQ-open, an existing open-domain QA benchmark. We find that over half of the questions in NQ-open are ambiguous. The types of ambiguity are diverse and sometimes subtle, many of which are only apparent after examining evidence provided by a very large text corpus. AMBIGNQ, a dataset with 14,042 annotations on NQ-OPEN questions containing diverse types of ambiguity. We provide two distributions of our new dataset AmbigNQ: a full version with all annotation metadata and a light version with only inputs and outputs.
Chinese word segmentation (CWS) trained from open source corpus faces dramatic performance drop when dealing with domain text, especially for a domain with lots of special terms and diverse writing styles, such as the biomedical domain. However, building domain-specific CWS requires extremely high annotation cost. In this paper, we propose an approach by exploiting domain-invariant knowledge from high resource to low resource domains. Extensive experiments show that our mode achieves consistently higher accuracy than the single-task CWS and other transfer learning baselines, especially when there is a large disparity between source and target domains. This dataset is the accompanied medical Chinese word segmentation (CWS) dataset. The tags are in BIES scheme. For more details see
It is a large dataset of Android applications belonging to 23 differentapps categories, which provides an overview of the types of feedback users report on the apps and documents the evolution of the related code metrics. The dataset contains about 395 applications of the F-Droid repository, including around 600 versions, 280,000 user reviews (extracted with specific text mining approaches)
A large-scale dataset consisting of approximately 100,000 algebraic word problems. The solution to each question is explained step-by-step using natural language. This data is used to train a program generation model that learns to generate the explanation, while generating the program that solves the question.
This dataset provides the template sentences and relationships defined in the ATOMIC common sense dataset. There are three splits - train, test, and dev. From the authors. Disclaimer/Content warning: the events in atomic have been automatically extracted from blogs, stories and books written at various times. The events might depict violent or problematic actions, which we left in the corpus for the sake of learning the (probably negative but still important) commonsense implications associated with the events. We removed a small set of truly out-dated events, but might have missed some so please email us ( if you have any concerns.
The Bengali Hate Speech Dataset is a collection of Bengali articles collected from Bengali news articles, news dump of Bengali TV channels, books, blogs, and social media. Emphasis was placed on Facebook pages and newspaper sources because they attract close to 50 million followers and is a common source of opinions and hate speech. The raw text corpus contains 250 million articles and the full dataset is being prepared for release. This is a subset of the full dataset. This dataset was prepared for hate-speech text classification benchmark on Bengali, an under-resourced language.
A dataset adopting the FEVER methodology that consists of 1,535 real-world claims regarding climate-change collected on the internet. Each claim is accompanied by five manually annotated evidence sentences retrieved from the English Wikipedia that support, refute or do not give enough information to validate the claim totalling in 7,675 claim-evidence pairs. The dataset features challenging claims that relate multiple facets and disputed cases of claims where both supporting and refuting evidence are present.
The COmmonsense Dataset Adversarially-authored by Humans (CODAH) is an evaluation set for commonsense question-answering in the sentence completion style of SWAG. As opposed to other automatically generated NLI datasets, CODAH is adversarially constructed by humans who can view feedback from a pre-trained model and use this information to design challenging commonsense questions. Our experimental results show that CODAH questions present a complementary extension to the SWAG dataset, testing additional modes of common sense.
This dataset is designed to provide training data for common sense relationships pulls together from various sources. The dataset is multi-lingual. See langauge codes and language info here: This dataset provides an interface for the conceptnet5 csv file, and some (but not all) of the raw text data used to build conceptnet5: omcsnet_sentences_free.txt, and omcsnet_sentences_more.txt. One use of this dataset would be to learn to extract the conceptnet relationship from the omcsnet sentences. Conceptnet5 has 34,074,917 relationships. Of those relationships, there are 2,176,099 surface text sentences related to those 2M entries. omcsnet_sentences_free has 898,161 lines. omcsnet_sentences_more has 2,001,736 lines. Original downloads are available here For more information, see: The omcsnet data comes with the following warning from the authors of the above site: Remember: this data comes from various forms of crowdsourcing. Sentences in these files are not necessarily true, useful, or appropriate.
Named entities are phrases that contain the names of persons, organizations, locations, times and quantities. Example: [PER Wolff] , currently a journalist in [LOC Argentina] , played with [PER Del Bosque] in the final years of the seventies in [ORG Real Madrid] . The shared task of CoNLL-2002 concerns language-independent named entity recognition. We will concentrate on four types of named entities: persons, locations, organizations and names of miscellaneous entities that do not belong to the previous three groups. The participants of the shared task will be offered training and test data for at least two languages. They will use the data for developing a named-entity recognition system that includes a machine learning component. Information sources other than the training data may be used in this shared task. We are especially interested in methods that can use additional unannotated data for improving their performance (for example co-training). The train/validation/test sets are available in Spanish and Dutch. For more details see and
ConvAI is a dataset of human-to-bot conversations labelled for quality. This data can be used to train a metric for evaluating dialogue systems. Moreover, it can be used in the development of chatbots themselves: it contains the information on the quality of utterances and entire dialogues, that can guide a dialogue system in search of better answers.
ConvAI is a dataset of human-to-bot conversations labelled for quality. This data can be used to train a metric for evaluating dialogue systems. Moreover, it can be used in the development of chatbots themselves: it contains the information on the quality of utterances and entire dialogues, that can guide a dialogue system in search of better answers.
The Conv AI 3 challenge is organized as part of the Search-oriented Conversational AI (SCAI) EMNLP workshop in 2020. The main aim of the conversational systems is to return an appropriate answer in response to the user requests. However, some user requests might be ambiguous. In Information Retrieval (IR) settings such a situation is handled mainly through the diversification of search result page. It is however much more challenging in dialogue settings. Hence, we aim to study the following situation for dialogue settings: - a user is asking an ambiguous question (where ambiguous question is a question to which one can return > 1 possible answers) - the system must identify that the question is ambiguous, and, instead of trying to answer it directly, ask a good clarifying question.
53,640 Japanese tweets with annotation if a tweet is related to COVID-19 or not. The annotation is by majority decision by 5 - 10 crowd workers. Target tweets include "COVID" or "コロナ". The period of the tweets is from around January 2020 to around June 2020. The original tweets are not contained. Please use Twitter API to get them, for example.
This dataset contains 14K dialogs (181K utterances) where users and assistants converse about geographic topics like geopolitical entities and locations. This dataset is annotated with pre-existing user knowledge, message-level dialog acts, grounding to Wikipedia, and user reactions to messages.
DART is a large and open-domain structured DAta Record to Text generation corpus with high-quality sentence annotations with each input being a set of entity-relation triples following a tree-structured ontology. It consists of 82191 examples across different domains with each input being a semantic RDF triple set derived from data records in tables and the tree ontology of table schema, annotated with sentence description that covers all facts in the triple set. DART is released in the following paper where you can find more details and baseline results:
Benchmark dataset for low-resource multiclass classification, with 4,015 training, 500 testing, and 500 validation examples, each labeled as part of five classes. Each sample can be a part of multiple classes. Collected as tweets.
The E2E dataset is used for training end-to-end, data-driven natural language generation systems in the restaurant domain, which is ten times bigger than existing, frequently used datasets in this area. The E2E dataset poses new challenges: (1) its human reference texts show more lexical richness and syntactic variation, including discourse phenomena; (2) generating from this set requires content selection. As such, learning from this dataset promises more natural, varied and less template-like system utterances. E2E is released in the following paper where you can find more details and baseline results:
An update release of E2E NLG Challenge data with cleaned MRs and scripts, accompanying the following paper: Ondřej Dušek, David M. Howcroft, and Verena Rieser (2019): Semantic Noise Matters for Neural Natural Language Generation. In INLG, Tokyo, Japan.
ETHOS: onlinE haTe speecH detectiOn dataSet. This repository contains a dataset for hate speech detection on social media platforms, called Ethos. There are two variations of the dataset: Ethos_Dataset_Binary: contains 998 comments in the dataset alongside with a label about hate speech presence or absence. 565 of them do not contain hate speech, while the rest of them, 433, contain. Ethos_Dataset_Multi_Label: which contains 8 labels for the 433 comments with hate speech content. These labels are violence (if it incites (1) or not (0) violence), directed_vs_general (if it is directed to a person (1) or a group (0)), and 6 labels about the category of hate speech like, gender, race, national_origin, disability, religion and sexual_orientation.
When humans read or listen, they make implicit commonsense inferences that frame their understanding of what happened and why. As a step toward AI systems that can build similar mental models, we introduce GLUCOSE, a large-scale dataset of implicit commonsense causal knowledge, encoded as causal mini-theories about the world, each grounded in a narrative context.
This dataset is intended to advance topic classification for German texts. A classifier that is efffective in English may not be effective in German dataset because it has a higher inflection and longer compound words. The 10kGNAD dataset contains 10273 German news articles from an Austrian online newspaper categorized into 9 categories. Article titles and text are concatenated together and authors are removed to avoid a keyword-like classification on authors that write frequently about one category. This dataset can be used as a benchmark for German topic classification.
The GoEmotions dataset contains 58k carefully curated Reddit comments labeled for 27 emotion categories or Neutral. The emotion categories are admiration, amusement, anger, annoyance, approval, caring, confusion, curiosity, desire, disappointment, disapproval, disgust, embarrassment, excitement, fear, gratitude, grief, joy, love, nervousness, optimism, pride, realization, relief, remorse, sadness, surprise.
Google's query wellformedness dataset was created by crowdsourcing well-formedness annotations for 25,100 queries from the Paralex corpus. Every query was annotated by five raters each with 1/0 rating of whether or not the query is well-formed.
Strongly Generalizable Question Answering (GrailQA) is a new large-scale, high-quality dataset for question answering on knowledge bases (KBQA) on Freebase with 64,331 questions annotated with both answers and corresponding logical forms in different syntax (i.e., SPARQL, S-expression, etc.). It can be used to test three levels of generalization in KBQA: i.i.d., compositional, and zero-shot.
Hatexplain is the first benchmark hate speech dataset covering multiple aspects of the issue. Each post in the dataset is annotated from three different perspectives: the basic, commonly used 3-class classification (i.e., hate, offensive or normal), the target community (i.e., the community that has been the victim of hate speech/offensive speech in the post), and the rationales, i.e., the portions of the post on which their labelling decision (as hate, offensive or normal) is based.
Existing question answering datasets focus on dealing with homogeneous information, based either only on text or KB/Table information alone. However, as human knowledge is distributed over heterogeneous forms, using homogeneous information alone might lead to severe coverage problems. To fill in the gap, we present HybridQA, a new large-scale question-answering dataset that requires reasoning on heterogeneous information. Each question is aligned with a Wikipedia table and multiple free-form corpora linked with the entities in the table. The questions are designed to aggregate both tabular information and text information, i.e., lack of either form would render the question unanswerable.
A dataset of about 20k questions that are elicited from readers as they naturally read through a document sentence by sentence. Compared to existing datasets, INQUISITIVE questions target more towards high-level (semantic and discourse) comprehension of text. Because these questions are generated while the readers are processing the information, the questions directly communicate gaps between the reader’s and writer’s knowledge about the events described in the text, and are not necessarily answered in the document itself. This type of question reflects a real-world scenario: if one has questions during reading, some of them are answered by the text later on, the rest are not, but any of them would help further the reader’s understanding at the particular point when they asked it. This resource could enable question generation models to simulate human-like curiosity and cognitive processing, which may open up a new realm of applications.
KdConv is a Chinese multi-domain Knowledge-driven Conversionsation dataset, grounding the topics in multi-turn conversations to knowledge graphs. KdConv contains 4.5K conversations from three domains (film, music, and travel), and 86K utterances with an average turn number of 19.0. These conversations contain in-depth discussions on related topics and natural transition between multiple topics, while the corpus can also used for exploration of transfer learning and domain adaptation.\
KILT tasks training and evaluation data. - [FEVER]( | Fact Checking | fever - [AIDA CoNLL-YAGO]( | Entity Linking | aidayago2 - [WNED-WIKI]( | Entity Linking | wned - [WNED-CWEB]( | Entity Linking | cweb - [T-REx]( | Slot Filling | trex - [Zero-Shot RE]( | Slot Filling | structured_zeroshot - [Natural Questions]( | Open Domain QA | nq - [HotpotQA]( | Open Domain QA | hotpotqa - [TriviaQA]( | Open Domain QA | triviaqa - [ELI5]( | Open Domain QA | eli5 - [Wizard of Wikipedia]( | Dialogue | wow To finish linking TriviaQA questions to the IDs provided, follow the instructions [here](
Motion recognition is one of the basic cognitive capabilities of many life forms, yet identifying motion of physical entities in natural language have not been explored extensively and empirically. Literal-Motion-in-Text (LiMiT) dataset, is a large human-annotated collection of English text sentences describing physical occurrence of motion, with annotated physical entities in motion.
MC-TACO (Multiple Choice TemporAl COmmonsense) is a dataset of 13k question-answer pairs that require temporal commonsense comprehension. A system receives a sentence providing context information, a question designed to require temporal commonsense knowledge, and multiple candidate answers. More than one candidate answer can be plausible. The task is framed as binary classification: givent he context, the question, and the candidate answer, the task is to determine whether the candidate answer is plausible ("yes") or not ("no").
Machine learning models are trained to find patterns in data. NLP models can inadvertently learn socially undesirable patterns when training on gender biased text. In this work, we propose a general framework that decomposes gender bias in text along several pragmatic and semantic dimensions: bias from the gender of the person being spoken about, bias from the gender of the person being spoken to, and bias from the gender of the speaker. Using this fine-grained framework, we automatically annotate eight large scale datasets with gender information. In addition, we collect a novel, crowdsourced evaluation benchmark of utterance-level gender rewrites. Distinguishing between gender bias along multiple dimensions is important, as it enables us to train finer-grained gender bias classifiers. We show our classifiers prove valuable for a variety of important applications, such as controlling for gender bias in generative models, detecting gender bias in arbitrary text, and shed light on offensive language in terms of genderedness.
MedHop is based on research paper abstracts from PubMed, and the queries are about interactions between pairs of drugs. The correct answer has to be inferred by combining information from a chain of reactions of drugs and proteins.
MetaLWOz: A Dataset of Multi-Domain Dialogues for the Fast Adaptation of Conversation Models. We introduce the Meta-Learning Wizard of Oz (MetaLWOz) dialogue dataset for developing fast adaptation methods for conversation models. This data can be used to train task-oriented dialogue models, specifically to develop methods to quickly simulate user responses with a small amount of data. Such fast-adaptation models fall into the research areas of transfer learning and meta learning. The dataset consists of 37,884 crowdsourced dialogues recorded between two human users in a Wizard of Oz setup, in which one was instructed to behave like a bot, and the other a true human user. The users are assigned a task belonging to a particular domain, for example booking a reservation at a particular restaurant, and work together to complete the task. Our dataset spans 47 domains having 227 tasks total. Dialogues are a minimum of 10 turns long.
We introduce MKQA, an open-domain question answering evaluation set comprising 10k question-answer pairs sampled from the Google Natural Questions dataset, aligned across 26 typologically diverse languages (260k question-answer pairs in total). For each query we collected new passage-independent answers. These queries and answers were then human translated into 25 Non-English languages.
Posing reading comprehension as a generation problem provides a great deal of flexibility, allowing for open-ended questions with few restrictions on possible answers. However, progress is impeded by existing generation metrics, which rely on token overlap and are agnostic to the nuances of reading comprehension. To address this, we introduce a benchmark for training and evaluating generative reading comprehension metrics: MOdeling Correctness with Human Annotations. MOCHA contains 40K human judgement scores on model outputs from 6 diverse question answering datasets and an additional set of minimal pairs for evaluation. Using MOCHA, we train an evaluation metric: LERC, a Learned Evaluation metric for Reading Comprehension, to mimic human judgement scores.
Recent work in semantic parsing for question answering has focused on long and complicated questions, many of which would seem unnatural if asked in a normal conversation between two humans. In an effort to explore a conversational QA setting, we present a more realistic task: answering sequences of simple but inter-related questions. We created SQA by asking crowdsourced workers to decompose 2,022 questions from WikiTableQuestions (WTQ), which contains highly-compositional questions about tables from Wikipedia. We had three workers decompose each WTQ question, resulting in a dataset of 6,066 sequences that contain 17,553 questions in total. Each question is also associated with answers in the form of cell locations in the tables.
This dataset contains sentences and short paragraphs with corresponding shorter (compressed) versions. There are up to five compressions for each input text, together with quality judgements of their meaning preservation and grammaticality. The dataset is derived using source texts from the Open American National Corpus ( and crowd-sourcing.
The Third International Chinese Language Processing Bakeoff was held in Spring 2006 to assess the state of the art in two important tasks: word segmentation and named entity recognition. Twenty-nine groups submitted result sets in the two tasks across two tracks and a total of five corpora. We found strong results in both tasks as well as continuing challenges. MSRA NER is one of the provided dataset. There are three types of NE, PER (person), ORG (organization) and LOC (location). The dataset is in the BIO scheme. For more details see
Our goal is to build systems that collaborate with people by exchanging information through natural language and reasoning over structured knowledge base. In the MutualFriend task, two agents, A and B, each have a private knowledge base, which contains a list of friends with multiple attributes (e.g., name, school, major, etc.). The agents must chat with each other to find their unique mutual friend.
The NarrativeQA dataset for question answering on long documents (movie scripts, books). It includes the list of documents with Wikipedia summaries, links to full stories, and questions and answers.
NewsQA is a challenging machine comprehension dataset of over 100,000 human-generated question-answer pairs. Crowdworkers supply questions and answers based on a set of over 10,000 news articles from CNN, with answers consisting of spans of text from the corresponding articles.
This is a movie review dataset in the Korean language. Reviews were scraped from Naver movies. The dataset construction is based on the method noted in Large movie review dataset from Maas et al., 2011.
Fused Head constructions are noun phrases in which the head noun is missing and is said to be "fused" with its dependent modifier. This missing information is implicit and is important for sentence understanding.The missing heads are easily filled in by humans, but pose a challenge for computational models. For example, in the sentence: "I bought 5 apples but got only 4.", 4 is a Fused-Head, and the missing head is apples, which appear earlier in the sentence. This is a crowd-sourced dataset of 10k numerical fused head examples (1M tokens).
The researchers of OCLAR Marwan et al. (2019), they gathered Arabic costumer reviews from Google reviewsa and Zomato website ( on wide scope of domain, including restaurants, hotels, hospitals, local shops, etc.The corpus finally contains 3916 reviews in 5-rating scale. For this research purpose, the positive class considers rating stars from 5 to 3 of 3465 reviews, and the negative class is represented from values of 1 and 2 of about 451 texts.
Person SenTiment (PerSenT) is a crowd-sourced dataset that captures the sentiment of an author towards the main entity in a news article. This dataset contains annotation for 5.3k documents and 38k paragraphs covering 3.2k unique entities. The dataset consists of sentiment annotations on news articles about people. For each article, annotators judge what the author’s sentiment is towards the main (target) entity of the article. The annotations also include similar judgments on paragraphs within the article. To split the dataset, entities into 4 mutually exclusive sets. Due to the nature of news collections, some entities tend to dominate the collection. In the collection, there were four entities which were the main entity in nearly 800 articles. To avoid these entities from dominating the train or test splits, we moved them to a separate test collection. We split the remaining into a training, dev, and test sets at random. Thus our collection includes one standard test set consisting of articles drawn at random (Test Standard -- `test_random`), while the other is a test set which contains multiple articles about a small number of popular entities (Test Frequent -- `test_fixed`).
To apply eyeshadow without a brush, should I use a cotton swab or a toothpick? Questions requiring this kind of physical commonsense pose a challenge to state-of-the-art natural language understanding systems. The PIQA dataset introduces the task of physical commonsense reasoning and a corresponding benchmark dataset Physical Interaction: Question Answering or PIQA. Physical commonsense knowledge is a major challenge on the road to true AI-completeness, including robots that interact with the world and understand natural language. PIQA focuses on everyday situations with a preference for atypical solutions. The dataset is inspired by, which provides users with instructions on how to build, craft, bake, or manipulate objects using everyday materials. The underlying task is formualted as multiple choice question answering: given a question `q` and two possible solutions `s1`, `s2`, a model or a human must choose the most appropriate solution, of which exactly one is correct. The dataset is further cleaned of basic artifacts using the AFLite algorithm which is an improvement of adversarial filtering. The dataset contains 16,000 examples for training, 2,000 for development and 3,000 for testing.
This dataset is for studying computational models trained to reason about prototypical situations. Using deterministic filtering a sampling from a larger set of all transcriptions was built. It contains 9789 instances where each instance represents a survey question from Family Feud game. Each instance exactly is a question, a set of answers, and a count associated with each answer. Each line is a json dictionary, in which: 1. question - contains the question (in original and a normalized form) 2. answerstrings - contains the original answers provided by survey respondents (when available), along with the counts for each string. Because the FamilyFeud data has only cluster names rather than strings, those cluster names are included with 0 weight. 3. answer-clusters - lists clusters, with the count of each cluster and the strings included in that cluster. Each cluster is given a unique ID that can be linked to in the assessment files.
NLM produces a baseline set of MEDLINE/PubMed citation records in XML format for download on an annual basis. The annual baseline is released in December of each year. Each day, NLM produces update files that include new, revised and deleted citations. See our documentation page for more information.
Question Answering in Context 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. QuAC introduces challenges not found in existing machine comprehension datasets: its questions are often more open-ended, unanswerable, or only meaningful within the dialog context.
ReDial (Recommendation Dialogues) is an annotated dataset of dialogues, where users recommend movies to each other. The dataset was collected by a team of researchers working at Polytechnique Montréal, MILA – Quebec AI Institute, Microsoft Research Montréal, HEC Montreal, and Element AI. The dataset allows research at the intersection of goal-directed dialogue systems (such as restaurant recommendation) and free-form (also called “chit-chat”) dialogue systems.
ROPES (Reasoning Over Paragraph Effects in Situations) is a QA dataset which tests a system's ability to apply knowledge from a passage of text to a new situation. A system is presented a background passage containing a causal or qualitative relation(s) (e.g., "animal pollinators increase efficiency of fertilization in flowers"), a novel situation that uses this background, and questions that require reasoning about effects of the relationships in the background passage in the background of the situation.
scb-mt-en-th-2020: A Large English-Thai Parallel Corpus The primary objective of our work is to build a large-scale English-Thai dataset for machine translation. We construct an English-Thai machine translation dataset with over 1 million segment pairs, curated from various sources, namely news, Wikipedia articles, SMS messages, task-based dialogs, web-crawled data and government documents. Methodology for gathering data, building parallel texts and removing noisy sentence pairs are presented in a reproducible manner. We train machine translation models based on this dataset. Our models' performance are comparable to that of Google Translation API (as of May 2020) for Thai-English and outperform Google when the Open Parallel Corpus (OPUS) is included in the training data for both Thai-English and English-Thai translation. The dataset, pre-trained models, and source code to reproduce our work are available for public use.
The SemEval-2014 Task 1 focuses on Evaluation of Compositional Distributional Semantic Models on Full Sentences through Semantic Relatedness and Entailment. The task was designed to predict the degree of relatedness between two sentences and to detect the entailment relation holding between them.
ShARC is a Conversational Question Answering dataset focussing on question answering from texts containing rules. The goal is to answer questions by possibly asking follow-up questions first. It is assumed assume that the question is often underspecified, in the sense that the question does not provide enough information to be answered directly. However, an agent can use the supporting rule text to infer what needs to be asked in order to determine the final answer.
ShARC, a conversational QA task, requires a system to answer user questions based on rules expressed in natural language text. However, it is found that in the ShARC dataset there are multiple spurious patterns that could be exploited by neural models. SharcModified is a new dataset which reduces the patterns identified in the original dataset. To reduce the sensitivity of neural models, for each occurence of an instance conforming to any of the patterns, we automatically construct alternatives where we choose to either replace the current instance with an alternative instance which does not exhibit the pattern; or retain the original instance. The modified ShARC has two versions sharc-mod and history-shuffled. For morre details refer to Appendix A.3 .
The SMS Spam Collection v.1 is a public set of SMS labeled messages that have been collected for mobile phone spam research. It has one collection composed by 5,574 English, real and non-enconded messages, tagged according being legitimate (ham) or spam.
The SNLI corpus (version 1.0) is a collection of 570k human-written English sentence pairs manually labeled for balanced classification with the labels entailment, contradiction, and neutral, supporting the task of natural language inference (NLI), also known as recognizing textual entailment (RTE).
About SNOW T15: The simplified corpus for the Japanese language. The corpus has 50,000 manually simplified and aligned sentences. This corpus contains the original sentences, simplified sentences and English translation of the original sentences. It can be used for automatic text simplification as well as translating simple Japanese into English and vice-versa. The core vocabulary is restricted to 2,000 words where it is selected by accounting for several factors such as meaning preservation, variation, simplicity and the UniDic word segmentation criterion. For details, refer to the explanation page of Japanese simplification ( The original texts are from "small_parallel_enja: 50k En/Ja Parallel Corpus for Testing SMT Methods", which is a bilingual corpus for machine translation. About SNOW T23: An expansion corpus of 35,000 sentences rewritten in easy Japanese (simple Japanese vocabulary) based on SNOW T15. The original texts are from "Tanaka Corpus" (
Social Bias Frames is a new way of representing the biases and offensiveness that are implied in language. For example, these frames are meant to distill the implication that "women (candidates) are less qualified" behind the statement "we shouldn’t lower our standards to hire more women."
Here are two different adversaries, each of which uses a different procedure to pick the sentence it adds to the paragraph: AddSent: Generates up to five candidate adversarial sentences that don't answer the question, but have a lot of words in common with the question. Picks the one that most confuses the model. AddOneSent: Similar to AddSent, but just picks one of the candidate sentences at random. This adversary is does not query the model in any way.
KorQuAD 1.0 is a large-scale Korean dataset for machine reading comprehension task consisting of human generated questions for Wikipedia articles. We benchmark the data collecting process of SQuADv1.0 and crowdsourced 70,000+ question-answer pairs. 1,637 articles and 70,079 pairs of question answers were collected. 1,420 articles are used for the training set, 140 for the dev set, and 77 for the test set. 60,407 question-answer pairs are for the training set, 5,774 for the dev set, and 3,898 for the test set.
KorQuAD 2.0 is a Korean question and answering dataset consisting of a total of 100,000+ pairs. There are three major differences from KorQuAD 1.0, which is the standard Korean Q & A data. The first is that a given document is a whole Wikipedia page, not just one or two paragraphs. Second, because the document also contains tables and lists, it is necessary to understand the document structured with HTML tags. Finally, the answer can be a long text covering not only word or phrase units, but paragraphs, tables, and lists. As a baseline model, BERT Multilingual is used, released by Google as an open source. It shows 46.0% F1 score, a very low score compared to 85.7% of the human F1 score. It indicates that this data is a challenging task. Additionally, we increased the performance by no-answer data augmentation. Through the distribution of this data, we intend to extend the limit of MRC that was limited to plain text to real world tasks of various lengths and formats.
Given a partial description like "she opened the hood of the car," humans can reason about the situation and anticipate what might come next ("then, she examined the engine"). SWAG (Situations With Adversarial Generations) is a large-scale dataset for this task of grounded commonsense inference, unifying natural language inference and physically grounded reasoning. The dataset consists of 113k multiple choice questions about grounded situations (73k training, 20k validation, 20k test). 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. SWAG aims to be a benchmark for evaluating grounded commonsense NLI and for learning representations. The full data contain more information, but the regular configuration will be more interesting for modeling (note that the regular data are shuffled). The test set for leaderboard submission is under the regular configuration.
The problem of verifying whether a textual hypothesis holds the truth based on the given evidence, also known as fact verification, plays an important role in the study of natural language understanding and semantic representation. However, existing studies are restricted to dealing with unstructured textual evidence (e.g., sentences and passages, a pool of passages), while verification using structured forms of evidence, such as tables, graphs, and databases, remains unexplored. TABFACT is large scale dataset with 16k Wikipedia tables as evidence for 118k human annotated statements designed for fact verification with semi-structured evidence. The statements are labeled as either ENTAILED or REFUTED. TABFACT is challenging since it involves both soft linguistic reasoning and hard symbolic reasoning.
Taskmaster-1 is a goal-oriented conversational dataset. It includes 13,215 task-based dialogs comprising six domains. Two procedures were used to create this collection, each with unique advantages. The first involves a two-person, spoken "Wizard of Oz" (WOz) approach in which trained agents and crowdsourced workers interact to complete the task while the second is "self-dialog" in which crowdsourced workers write the entire dialog themselves.
Taskmaster is dataset for goal oriented conversations. The Taskmaster-2 dataset consists of 17,289 dialogs in the seven domains which include restaurants, food ordering, movies, hotels, flights, music and sports. Unlike Taskmaster-1, which includes both written "self-dialogs" and spoken two-person dialogs, Taskmaster-2 consists entirely of spoken two-person dialogs. In addition, while Taskmaster-1 is almost exclusively task-based, Taskmaster-2 contains a good number of search- and recommendation-oriented dialogs. All dialogs in this release were created using a Wizard of Oz (WOz) methodology in which crowdsourced workers played the role of a 'user' and trained call center operators played the role of the 'assistant'. In this way, users were led to believe they were interacting with an automated system that “spoke” using text-to-speech (TTS) even though it was in fact a human behind the scenes. As a result, users could express themselves however they chose in the context of an automated interface.
Taskmaster is dataset for goal oriented conversations. The Taskmaster-3 dataset consists of 23,757 movie ticketing dialogs. By "movie ticketing" we mean conversations where the customer's goal is to purchase tickets after deciding on theater, time, movie name, number of tickets, and date, or opt out of the transaction. This collection was created using the "self-dialog" method. This means a single, crowd-sourced worker is paid to create a conversation writing turns for both speakers, i.e. the customer and the ticketing agent.
The objective of this task is to detect hate speech in tweets. For the sake of simplicity, we say a tweet contains hate speech if it has a racist or sexist sentiment associated with it. So, the task is to classify racist or sexist tweets from other tweets. Formally, given a training sample of tweets and labels, where label ‘1’ denotes the tweet is racist/sexist and label ‘0’ denotes the tweet is not racist/sexist, your objective is to predict the labels on the given test dataset.
A translation of the word pair similarity dataset wordsim-353 to Twi. The dataset was presented in the paper Alabi et al.: Massive vs. Curated Embeddings for Low-Resourced Languages: the Case of Yorùbá and Twi (LREC 2020).
WikiAsp is a multi-domain, aspect-based summarization dataset in the encyclopedic domain. In this task, models are asked to summarize cited reference documents of a Wikipedia article into aspect-based summaries. Each of the 20 domains include 10 domain-specific pre-defined aspects.
WikiAuto provides a set of aligned sentences from English Wikipedia and Simple English Wikipedia as a resource to train sentence simplification systems. The authors first crowd-sourced a set of manual alignments between sentences in a subset of the Simple English Wikipedia and their corresponding versions in English Wikipedia (this corresponds to the `manual` config), then trained a neural CRF system to predict these alignments. The trained model was then applied to the other articles in Simple English Wikipedia with an English counterpart to create a larger corpus of aligned sentences (corresponding to the `auto`, `auto_acl`, `auto_full_no_split`, and `auto_full_with_split` configs here).
WikiHop is open-domain and based on Wikipedia articles; the goal is to recover Wikidata information by hopping through documents. The goal is to answer text understanding queries by combining multiple facts that are spread across different documents.
WikiLingua is a large-scale multilingual dataset for the evaluation of crosslingual abstractive summarization systems. The dataset includes ~770k article and summary pairs in 18 languages from WikiHow. The gold-standard article-summary alignments across languages was done by aligning the images that are used to describe each how-to step in an article.
Wizard-of-Oz (WOZ) is a dataset for training task-oriented dialogue systems. The dataset is designed around the task of finding a restaurant in the Cambridge, UK area. There are three informable slots (food, pricerange,area) that users can use to constrain the search and six requestable slots (address, phone, postcode plus the three informable slots) that the user can ask a value for once a restaurant has been offered.
XGLUE is a new benchmark dataset to evaluate the performance of cross-lingual pre-trained models with respect to cross-lingual natural language understanding and generation. The benchmark is composed of the following 11 tasks: - NER - POS Tagging (POS) - News Classification (NC) - MLQA - XNLI - PAWS-X - Query-Ad Matching (QADSM) - Web Page Ranking (WPR) - QA Matching (QAM) - Question Generation (QG) - News Title Generation (NTG) For more information, please take a look at
XOR-TyDi QA brings together for the first time information-seeking questions, open-retrieval QA, and multilingual QA to create a multilingual open-retrieval QA dataset that enables cross-lingual answer retrieval. It consists of questions written by information-seeking native speakers in 7 typologically diverse languages and answer annotations that are retrieved from multilingual document collections. There are three sub-tasks: XOR-Retrieve, XOR-EnglishSpan, and XOR-Full.
Neural abstractive summarization models are highly prone to hallucinate content that is unfaithful to the input document. The popular metric such as ROUGE fails to show the severity of the problem. The dataset consists of faithfulness and factuality annotations of abstractive summaries for the XSum dataset. We have crowdsourced 3 judgements for each of 500 x 5 document-system pairs. This will be a valuable resource to the abstractive summarization community.
The Yelp reviews dataset consists of reviews from Yelp. It is extracted from the Yelp Dataset Challenge 2015 data. The Yelp reviews full star dataset is constructed by Xiang Zhang ( from the above dataset. It is first used as a text classification benchmark in the following paper: Xiang Zhang, Junbo Zhao, Yann LeCun. Character-level Convolutional Networks for Text Classification. Advances in Neural Information Processing Systems 28 (NIPS 2015).
A translation of the word pair similarity dataset wordsim-353 to Yorùbá. The dataset was presented in the paper Alabi et al.: Massive vs. Curated Embeddings for Low-Resourced Languages: the Case of Yorùbá and Twi (LREC 2020).
ZEST tests whether NLP systems can perform unseen tasks in a zero-shot way, given a natural language description of the task. It is an instantiation of our proposed framework "learning from task descriptions". The tasks include classification, typed entity extraction and relationship extraction, and each task is paired with 20 different annotated (input, output) examples. ZEST's structure allows us to systematically test whether models can generalize in five different ways.