Back to home
All Datasets 🏷
All datasets from our datasets repository and community bucket.
Also check out the list of supported Metrics 📉.
261 results
ADE-Corpus-V2 Dataset: Adverse Drug Reaction Data. This is a dataset for Classification if a sentence is ADE-related (True) or not (False) and Relation Extraction between Adverse Drug Event and Drug. DRUG-AE.rel provides relations between drugs and adverse effects. DRUG-DOSE.rel provides relations between drugs and dosages. ADE-NEG.txt provides all sentences in the ADE corpus that DO NOT contain any drug-related adverse effects.
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.
AirDialogue, is a large dataset that contains 402,038 goal-oriented conversations. To collect this dataset, we create a contextgenerator which provides travel and flight restrictions. Then the human annotators are asked to play the role of a customer or an agent and interact with the goal of successfully booking a trip given the restrictions.
The ALT project aims to advance the state-of-the-art Asian natural language processing (NLP) techniques through the open collaboration for developing and using ALT. It was first conducted by NICT and UCSY as described in Ye Kyaw Thu, Win Pa Pa, Masao Utiyama, Andrew Finch and Eiichiro Sumita (2016). Then, it was developed under ASEAN IVO as described in this Web page. The process of building ALT began with sampling about 20,000 sentences from English Wikinews, and then these sentences were translated into the other languages. ALT now has 13 languages: Bengali, English, Filipino, Hindi, Bahasa Indonesia, Japanese, Khmer, Lao, Malay, Myanmar (Burmese), Thai, Vietnamese, Chinese (Simplified Chinese).
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.
We provide an Amazon product reviews dataset for multilingual text classification. The dataset contains reviews in English, Japanese, German, French, Chinese and Spanish, collected between November 1, 2015 and November 1, 2019. Each record in the dataset contains the review text, the review title, the star rating, an anonymized reviewer ID, an anonymized product ID and the coarse-grained product category (e.g. ‘books’, ‘appliances’, etc.) The corpus is balanced across stars, so each star rating constitutes 20% of the reviews in each language. For each language, there are 200,000, 5,000 and 5,000 reviews in the training, development and test sets respectively. The maximum number of reviews per reviewer is 20 and the maximum number of reviews per product is 20. All reviews are truncated after 2,000 characters, and all reviews are at least 20 characters long. Note that the language of a review does not necessarily match the language of its marketplace (e.g. reviews from are primarily written in German, but could also be written in English, etc.). For this reason, we applied a language detection algorithm based on the work in Bojanowski et al. (2017) to determine the language of the review text and we removed reviews that were not written in the expected language.
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.
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.
ASSET is a dataset for evaluating Sentence Simplification systems with multiple rewriting transformations, as described in "ASSET: A Dataset for Tuning and Evaluation of Sentence Simplification Models with Multiple Rewriting Transformations". The corpus is composed of 2000 validation and 359 test original sentences that were each simplified 10 times by different annotators. The corpus also contains human judgments of meaning preservation, fluency and simplicity for the outputs of several automatic text simplification systems.
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.
Multilingual information access is stipulated in the South African constitution. In practise, this is hampered by a lack of resources and capacity to perform the large volumes of translation work required to realise multilingual information access. One of the aims of the Autshumato project is to develop machine translation systems for three South African languages pairs.
Nineteen teams presented results for the Gene Mention Task at the BioCreative II Workshop. In this task participants designed systems to identify substrings in sentences corresponding to gene name mentions. A variety of different methods were used and the results varied with a highest achieved F1 score of 0.8721. Here we present brief descriptions of all the methods used and a statistical analysis of the results. We also demonstrate that, by combining the results from all submissions, an F score of 0.9066 is feasible, and furthermore that the best result makes use of the lowest scoring submissions. For more details, see: The original dataset can be downloaded from: This dataset has been converted to CoNLL format for NER using the following tool:
This is a multilingual parallel corpus created from translations of the Bible compiled by Christos Christodoulopoulos and Mark Steedman. 102 languages, 5,148 bitexts total number of files: 107 total number of tokens: 56.43M total number of sentence fragments: 2.84M
BIGPATENT, consisting of 1.3 million records of U.S. patent documents along with human written abstractive summaries. Each US patent application is filed under a Cooperative Patent Classification (CPC) code. There are nine such classification categories: A (Human Necessities), B (Performing Operations; Transporting), C (Chemistry; Metallurgy), D (Textiles; Paper), E (Fixed Constructions), F (Mechanical Engineering; Lightning; Heating; Weapons; Blasting), G (Physics), H (Electricity), and Y (General tagging of new or cross-sectional technology) There are two features: - description: detailed description of patent. - abstract: Patent abastract.
This dataset was curated from the Bing search logs (desktop users only) over the period of Jan 1st, 2020 – (Current Month - 1). Only searches that were issued many times by multiple users were included. The dataset includes queries from all over the world that had an intent related to the Coronavirus or Covid-19. In some cases this intent is explicit in the query itself (e.g., “Coronavirus updates Seattle”), in other cases it is implicit , e.g. “Shelter in place”. The implicit intent of search queries (e.g., “Toilet paper”) was extracted using random walks on the click graph as outlined in this paper by Microsoft Research. All personal data were removed.
This is the Business Scene Dialogue (BSD) dataset, a Japanese-English parallel corpus containing written conversations in various business scenarios. The dataset was constructed in 3 steps: 1) selecting business scenes, 2) writing monolingual conversation scenarios according to the selected scenes, and 3) translating the scenarios into the other language. Half of the monolingual scenarios were written in Japanese and the other half were written in English. Fields: - id: dialogue identifier - no: sentence pair number within a dialogue - en_speaker: speaker name in English - ja_speaker: speaker name in Japanese - en_sentence: sentence in English - ja_sentence: sentence in Japanese - original_language: language in which monolingual scenario was written - tag: scenario - title: scenario title
A parallel corpus of theses and dissertations abstracts in English and Portuguese were collected from the CAPES website (Coordenação de Aperfeiçoamento de Pessoal de Nível Superior) - Brazil. The corpus is sentence aligned for all language pairs. Approximately 240,000 documents were collected and aligned using the Hunalign algorithm.
This corpus is an attempt to recreate the dataset used for training XLM-R. This corpus comprises of monolingual data for 100+ languages and also includes data for romanized languages (indicated by *_rom). This was constructed using the urls and paragraph indices provided by the CC-Net repository by processing January-December 2018 Commoncrawl snapshots. Each file comprises of documents separated by double-newlines and paragraphs within the same document separated by a newline. The data is generated using the open source CC-Net repository. No claims of intellectual property are made on the work of preparation of the corpus.
ChrEn is a Cherokee-English parallel dataset to facilitate machine translation research between Cherokee and English. ChrEn is extremely low-resource contains 14k sentence pairs in total, split in ways that facilitate both in-domain and out-of-domain evaluation. ChrEn also contains 5k Cherokee monolingual data to enable semi-supervised learning.
The Circa (meaning ‘approximately’) dataset aims to help machine learning systems to solve the problem of interpreting indirect answers to polar questions. The dataset contains pairs of yes/no questions and indirect answers, together with annotations for the interpretation of the answer. The data is collected in 10 different social conversational situations (eg. food preferences of a friend). NOTE: There might be missing labels in the dataset and we have replaced them with -1. The original dataset contains no train/dev/test splits.
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.
This dataset is for evaluating the performance of intent classification systems in the presence of "out-of-scope" queries. By "out-of-scope", we mean queries that do not fall into any of the system-supported intent classes. Most datasets include only data that is "in-scope". Our dataset includes both in-scope and out-of-scope data. You might also know the term "out-of-scope" by other terms, including "out-of-domain" or "out-of-distribution".
CNN/DailyMail non-anonymized summarization dataset. There are two features: - article: text of news article, used as the document to be summarized - highlights: joined text of highlights with <s> and </s> around each highlight, which is the target summary
A dataset consisting of 502 English dialogs with 12,000 annotated utterances between a user and an assistant discussing movie preferences in natural language. It was collected using a Wizard-of-Oz methodology between two paid crowd-workers, where one worker plays the role of an 'assistant', while the other plays the role of a 'user'. The 'assistant' elicits the 'user’s' preferences about movies following a Coached Conversational Preference Elicitation (CCPE) method. The assistant asks questions designed to minimize the bias in the terminology the 'user' employs to convey his or her preferences as much as possible, and to obtain these preferences in natural language. Each dialog is annotated with entity mentions, preferences expressed about entities, descriptions of entities provided, and other statements of entities.
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.
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.
We study negotiation dialogues where two agents, a buyer and a seller, negotiate over the price of an time for sale. We collected a dataset of more than 6K negotiation dialogues over multiple categories of products scraped from Craigslist. Our goal is to develop an agent that negotiates with humans through such conversations. The challenge is to handle both the negotiation strategy and the rich language for bargaining.
Corpus of domain names scraped from Common Crawl and manually annotated to add word boundaries (e.g. "commoncrawl" to "common crawl"). Breaking domain names such as "openresearch" into component words "open" and "research" is important for applications such as Text-to-Speech synthesis and web search. Common Crawl is an open repository of web crawl data that can be accessed and analyzed by anyone. Specifically, we scraped the plaintext (WET) extracts for domain names from URLs that contained diverse letter casing (e.g. "OpenBSD"). Although in the previous example, segmentation is trivial using letter casing, this was not always the case (e.g. "NASA"), so we had to manually annotate the data. The dataset is stored as plaintext file where each line is an example of space separated segments of a domain name. The examples are stored in their original letter casing, but harder and more interesting examples can be generated by lowercasing the input first.
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:
The DBpedia ontology classification dataset is constructed by picking 14 non-overlapping classes from DBpedia 2014. They are listed in classes.txt. From each of thse 14 ontology classes, we randomly choose 40,000 training samples and 5,000 testing samples. Therefore, the total size of the training dataset is 560,000 and testing dataset 70,000. There are 3 columns in the dataset (same for train and test splits), corresponding to class index (1 to 14), title and content. The title and content are escaped using double quotes ("), and any internal double quote is escaped by 2 double quotes (""). There are no new lines in title or content.
DialogRE is the first human-annotated dialogue based relation extraction (RE) dataset aiming to support the prediction of relation(s) between two arguments that appear in a dialogue. The dataset annotates all occurrences of 36 possible relation types that exist between pairs of arguments in the 1,788 dialogues originating from the complete transcripts of Friends.
This dataset contains 30,000 messages drawn from events including an earthquake in Haiti in 2010, an earthquake in Chile in 2010, floods in Pakistan in 2010, super-storm Sandy in the U.S.A. in 2012, and news articles spanning a large number of years and 100s of different disasters. The data has been encoded with 36 different categories related to disaster response and has been stripped of messages with sensitive information in their entirety. Upon release, this is the featured dataset of a new Udacity course on Data Science and the AI4ALL summer school and is especially utile for text analytics and natural language processing (NLP) tasks and models. The input data in this job contains thousands of untranslated disaster-related messages and their English translations.
Doc2dial is dataset of goal-oriented dialogues that are grounded in the associated documents. It includes over 4500 annotated conversations with an average of 14 turns that are grounded in over 450 documents from four domains. Compared to the prior document-grounded dialogue datasets this dataset covers a variety of dialogue scenes in information-seeking conversations.
DREAM is a multiple-choice Dialogue-based REAding comprehension exaMination dataset. In contrast to existing reading comprehension datasets, DREAM is the first to focus on in-depth multi-turn multi-party dialogue understanding.
The dataset contains around 271,342 tweets. The tweets are filtered via the official Twitter API to contain tweets in Dutch language or by users who have specified their location information within Netherlands geographical boundaries. Using natural language processing we have classified the tweets for their HISCO codes. If the user has provided their location within Dutch boundaries, we have also classified them to their respective provinces The objective of this dataset is to make research data available publicly in a FAIR (Findable, Accessible, Interoperable, Reusable) way. Twitter's Terms of Service Licensed under Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) (2020-10-27)
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.
Original source: Website and documentatuion from the European Central Bank, compiled and made available by Alberto Simoes (thank you very much!) 19 languages, 170 bitexts total number of files: 340 total number of tokens: 757.37M total number of sentence fragments: 30.55M
This is a parallel corpus made out of PDF documents from the European Medicines Agency. All files are automatically converted from PDF to plain text using pdftotext with the command line arguments -layout -nopgbrk -eol unix. There are some known problems with tables and multi-column layouts - some of them are fixed in the current version. source: 22 languages, 231 bitexts total number of files: 41,957 total number of tokens: 311.65M total number of sentence fragments: 26.51M
WebNLG is a valuable resource and benchmark for the Natural Language Generation (NLG) community. However, as other NLG benchmarks, it only consists of a collection of parallel raw representations and their corresponding textual realizations. This work aimed to provide intermediate representations of the data for the development and evaluation of popular tasks in the NLG pipeline architecture (Reiter and Dale, 2000), such as Discourse Ordering, Lexicalization, Aggregation and Referring Expression Generation.
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.
In October 2012, the European Union's (EU) Directorate General for Education and Culture ( DG EAC) released a translation memory (TM), i.e. a collection of sentences and their professionally produced translations, in twenty-six languages. This resource bears the name EAC Translation Memory, short EAC-TM. EAC-TM covers up to 26 languages: 22 official languages of the EU (all except Irish) plus Icelandic, Croatian, Norwegian and Turkish. EAC-TM thus contains translations from English into the following 25 languages: Bulgarian, Czech, Danish, Dutch, Estonian, German, Greek, Finnish, French, Croatian, Hungarian, Icelandic, Italian, Latvian, Lithuanian, Maltese, Norwegian, Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish, Swedish and Turkish. All documents and sentences were originally written in English (source language is English) and then translated into the other languages. The texts were translated by staff of the National Agencies of the Lifelong Learning and Youth in Action programmes. They are typically professionals in the field of education/youth and EU programmes. They are thus not professional translators, but they are normally native speakers of the target language.
In October 2012, the European Union (EU) agency 'European Centre for Disease Prevention and Control' (ECDC) released a translation memory (TM), i.e. a collection of sentences and their professionally produced translations, in twenty-five languages. This resource bears the name EAC Translation Memory, short EAC-TM. ECDC-TM covers 25 languages: the 23 official languages of the EU plus Norwegian (Norsk) and Icelandic. ECDC-TM was created by translating from English into the following 24 languages: Bulgarian, Czech, Danish, Dutch, English, Estonian, Gaelige (Irish), German, Greek, Finnish, French, Hungarian, Icelandic, Italian, Latvian, Lithuanian, Maltese, Norwegian (NOrsk), Polish, Portuguese, Romanian, Slovak, Slovenian, Spanish and Swedish. All documents and sentences were thus originally written in English. They were then translated into the other languages by professional translators from the Translation Centre CdT in Luxembourg.
Data and code from our "Inferring Which Medical Treatments Work from Reports of Clinical Trials", NAACL 2019. This work concerns inferring the results reported in clinical trials from text. The dataset consists of biomedical articles describing randomized control trials (RCTs) that compare multiple treatments. Each of these articles will have multiple questions, or 'prompts' associated with them. These prompts will ask about the relationship between an intervention and comparator with respect to an outcome, as reported in the trial. For example, a prompt may ask about the reported effects of aspirin as compared to placebo on the duration of headaches. For the sake of this task, we assume that a particular article will report that the intervention of interest either significantly increased, significantly decreased or had significant effect on the outcome, relative to the comparator. The dataset could be used for automatic data extraction of the results of a given RCT. This would enable readers to discover the effectiveness of different treatments without needing to read the paper.
Fake news has become a major societal issue and a technical challenge for social media companies to identify. This content is difficult to identify because the term "fake news" covers intentionally false, deceptive stories as well as factual errors, satire, and sometimes, stories that a person just does not like. Addressing the problem requires clear definitions and examples. In this work, we present a dataset of fake news and satire stories that are hand coded, verified, and, in the case of fake news, include rebutting stories. We also include a thematic content analysis of the articles, identifying major themes that include hyperbolic support or condemnation of a gure, conspiracy theories, racist themes, and discrediting of reliable sources. In addition to releasing this dataset for research use, we analyze it and show results based on language that are promising for classification purposes. Overall, our contribution of a dataset and initial analysis are designed to support future work by fake news researchers.
`generated_reviews_enth` Generated product reviews dataset for machine translation quality prediction, part of [scb-mt-en-th-2020]( `generated_reviews_enth` is created as part of [scb-mt-en-th-2020]( for machine translation task. This dataset (referred to as `generated_reviews_yn` in [scb-mt-en-th-2020]( are English product reviews generated by [CTRL](, translated by Google Translate API and annotated as accepted or rejected (`correct`) based on fluency and adequacy of the translation by human annotators. This allows it to be used for English-to-Thai translation quality esitmation (binary label), machine translation, and sentiment analysis.
The GenericsKB contains 3.4M+ generic sentences about the world, i.e., sentences expressing general truths such as "Dogs bark," and "Trees remove carbon dioxide from the atmosphere." Generics are potentially useful as a knowledge source for AI systems requiring general world knowledge. The GenericsKB is the first large-scale resource containing naturally occurring generic sentences (as opposed to extracted or crowdsourced triples), and is rich in high-quality, general, semantically complete statements. Generics were primarily extracted from three large text sources, namely the Waterloo Corpus, selected parts of Simple Wikipedia, and the ARC Corpus. A filtered, high-quality subset is also available in GenericsKB-Best, containing 1,020,868 sentences. We recommend you start with GenericsKB-Best.
Giga-word corpus for French-English from WMT2010 collected by Chris Callison-Burch 2 languages, total number of files: 452 total number of tokens: 1.43G total number of sentence fragments: 47.55M
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.
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.
The dataset for the variable-misuse task, described in the ICLR 2020 paper 'Global Relational Models of Source Code' [] This is the public version of the dataset used in that paper. The original, used to produce the graphs in the paper, could not be open-sourced due to licensing issues. See the public associated code repository [] for results produced from this dataset. This dataset was generated synthetically from the corpus of Python code in the ETH Py150 Open dataset [].
This dataset is a new knowledge-base (KB) of hasPart relationships, extracted from a large corpus of generic statements. Complementary to other resources available, it is the first which is all three of: accurate (90% precision), salient (covers relationships a person may mention), and has high coverage of common terms (approximated as within a 10 year old’s vocabulary), as well as having several times more hasPart entries than in the popular ontologies ConceptNet and WordNet. In addition, it contains information about quantifiers, argument modifiers, and links the entities to appropriate concepts in Wikipedia and WordNet.
These files contain text extracted from Stormfront, a white supremacist forum. A random set of forums posts have been sampled from several subforums and split into sentences. Those sentences have been manually labelled as containing hate speech or not, according to certain annotation guidelines.
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.
HEAD-QA is a multi-choice HEAlthcare Dataset. The questions come from exams to access a specialized position in the Spanish healthcare system, and are challenging even for highly specialized humans. They are designed by the Ministerio de Sanidad, Consumo y Bienestar Social. The dataset contains questions about the following topics: medicine, nursing, psychology, chemistry, pharmacology and biology.
PUBHEALTH is a comprehensive dataset for explainable automated fact-checking of public health claims. Each instance in the PUBHEALTH dataset has an associated veracity label (true, false, unproven, mixture). Furthermore each instance in the dataset has an explanation text field. The explanation is a justification for which the claim has been assigned a particular veracity label. The dataset was created to explore fact-checking of difficult to verify claims i.e., those which require expertise from outside of the journalistics domain, in this case biomedical and public health expertise. It was also created in response to the lack of fact-checking datasets which provide gold standard natural language explanations for verdicts/labels. NOTE: There are missing labels in the dataset and we have replaced them with -1.
HindEnCorp parallel texts (sentence-aligned) come from the following sources: Tides, which contains 50K sentence pairs taken mainly from news articles. This dataset was originally col- lected for the DARPA-TIDES surprise-language con- test in 2002, later refined at IIIT Hyderabad and provided for the NLP Tools Contest at ICON 2008 (Venkatapathy, 2008). Commentaries by Daniel Pipes contain 322 articles in English written by a journalist Daniel Pipes and translated into Hindi. EMILLE. This corpus (Baker et al., 2002) consists of three components: monolingual, parallel and annotated corpora. There are fourteen monolingual sub- corpora, including both written and (for some lan- guages) spoken data for fourteen South Asian lan- guages. The EMILLE monolingual corpora contain in total 92,799,000 words (including 2,627,000 words of transcribed spoken data for Bengali, Gujarati, Hindi, Punjabi and Urdu). The parallel corpus consists of 200,000 words of text in English and its accompanying translations into Hindi and other languages. Smaller datasets as collected by Bojar et al. (2010) include the corpus used at ACL 2005 (a subcorpus of EMILLE), a corpus of named entities from Wikipedia (crawled in 2009), and Agriculture domain parallel corpus.  For the current release, we are extending the parallel corpus using these sources: Intercorp (Čermák and Rosen,2012) is a large multilingual parallel corpus of 32 languages including Hindi. The central language used for alignment is Czech. Intercorp’s core texts amount to 202 million words. These core texts are most suitable for us because their sentence alignment is manually checked and therefore very reliable. They cover predominately short sto- ries and novels. There are seven Hindi texts in Inter- corp. Unfortunately, only for three of them the English translation is available; the other four are aligned only with Czech texts. The Hindi subcorpus of Intercorp contains 118,000 words in Hindi. TED talks 3 held in various languages, primarily English, are equipped with transcripts and these are translated into 102 languages. There are 179 talks for which Hindi translation is available. The Indic multi-parallel corpus (Birch et al., 2011; Post et al., 2012) is a corpus of texts from Wikipedia translated from the respective Indian language into English by non-expert translators hired over Mechanical Turk. The quality is thus somewhat mixed in many respects starting from typesetting and punctuation over capi- talization, spelling, word choice to sentence structure. A little bit of control could be in principle obtained from the fact that every input sentence was translated 4 times. We used the 2012 release of the corpus. is a software collaboration platform that hosts many open-source projects and facilitates also collaborative localization of the tools. We downloaded all revisions of all the hosted projects and extracted the localization (.po) files. Other smaller datasets. This time, we added Wikipedia entities as crawled in 2013 (including any morphological variants of the named entitity that appears on the Hindi variant of the Wikipedia page) and words, word examples and quotes from the Shabdkosh online dictionary.
To examine the cognitive processes of remembering and imagining and their traces in language, we introduce Hippocorpus, a dataset of 6,854 English diary-like short stories about recalled and imagined events. Using a crowdsourcing framework, we first collect recalled stories and summaries from workers, then provide these summaries to other workers who write imagined stories. Finally, months later, we collect a retold version of the recalled stories from a subset of recalled authors. Our dataset comes paired with author demographics (age, gender, race), their openness to experience, as well as some variables regarding the author's relationship to the event (e.g., how personal the event is, how often they tell its story, etc.).
A Hope Speech dataset for Equality, Diversity and Inclusion (HopeEDI) containing user-generated comments from the social media platform YouTube with 28,451, 20,198 and 10,705 comments in English, Tamil and Malayalam, respectively, manually labelled as containing hope speech or not.
HoVer is an open-domain, many-hop fact extraction and claim verification dataset built upon the Wikipedia corpus. The original 2-hop claims are adapted from question-answer pairs from HotpotQA. It is collected by a team of NLP researchers at UNC Chapel Hill and Verisk Analytics.
The hrenWaC corpus version 2.0 consists of parallel Croatian-English texts crawled from the .hr top-level domain for Croatia. The corpus was built with Spidextor (, a tool that glues together the output of SpiderLing used for crawling and Bitextor used for bitext extraction. The accuracy of the extracted bitext on the segment level is around 80% and on the word level around 84%.
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.
Parallel Text Corpora for Multi-Domain Translation System created by BPPT (Indonesian Agency for the Assessment and Application of Technology) for PAN Localization Project (A Regional Initiative to Develop Local Language Computing Capacity in Asia). The dataset contains around 24K sentences divided in 4 difference topics (Economic, international, Science and Technology and Sport).
Over >25k semiautomatically generated sentence pairs illustrating well-studied pragmatic inference types. IMPPRES is an NLI dataset following the format of SNLI (Bowman et al., 2015), MultiNLI (Williams et al., 2018) and XNLI (Conneau et al., 2018), which was created to evaluate how well trained NLI models recognize several classes of presuppositions and scalar implicatures.
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.
JFLEG (JHU FLuency-Extended GUG) is an English grammatical error correction (GEC) corpus. It is a gold standard benchmark for developing and evaluating GEC systems with respect to fluency (extent to which a text is native-sounding) as well as grammaticality. For each source document, there are four human-written corrections (ref0 to ref3).
The data came from the GENIA version 3.02 corpus (Kim et al., 2003). This was formed from a controlled search on MEDLINE using the MeSH terms human, blood cells and transcription factors. From this search 2,000 abstracts were selected and hand annotated according to a small taxonomy of 48 classes based on a chemical classification. Among the classes, 36 terminal classes were used to annotate the GENIA corpus.
A parallel corpus of KDE4 localization files (v.2). 92 languages, 4,099 bitexts total number of files: 75,535 total number of tokens: 60.75M total number of sentence fragments: 8.89M
Data-To-Text Generation involves converting knowledge graph (KG) triples of the form (subject, relation, object) into a natural language sentence(s). This dataset consists of English KG data converted into paired natural language text. The generated corpus consists of ∼18M sentences spanning ∼45M triples with ∼1500 distinct relations.
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](
The LAMBADA evaluates the capabilities of computational models for text understanding by means of a word prediction task. LAMBADA is a collection of narrative passages sharing the characteristic that human subjects are able to guess their last word if they are exposed to the whole passage, but not if they only see the last sentence preceding the target word. To succeed on LAMBADA, computational models cannot simply rely on local context, but must be able to keep track of information in the broader discourse. The LAMBADA dataset is extracted from BookCorpus and consists of 10'022 passages, divided into 4'869 development and 5'153 test passages. The training data for language models to be tested on LAMBADA include the full text of 2'662 novels (disjoint from those in dev+test), comprising 203 million words.
LIAR is a dataset for fake news detection with 12.8K human labeled short statements from's API, and each statement is evaluated by a editor for its truthfulness. The distribution of labels in the LIAR dataset is relatively well-balanced: except for 1,050 pants-fire cases, the instances for all other labels range from 2,063 to 2,638. In each case, the labeler provides a lengthy analysis report to ground each judgment.
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.
A large medical text dataset (14Go) curated to 4Go for abbreviation disambiguation, designed for natural language understanding pre-training in the medical domain. For example, DHF can be disambiguated to dihydrofolate, diastolic heart failure, dengue hemorragic fever or dihydroxyfumarate
The MedDialog dataset (English) contains conversations (in English) between doctors and patients.It has 0.26 million dialogues. The data is continuously growing and more dialogues will be added. The raw dialogues are from and All copyrights of the data belong to and
MENYO-20k is a multi-domain parallel dataset with texts obtained from news articles, ted talks, movie transcripts, radio transcripts, science and technology texts, and other short articles curated from the web and professional translators. The dataset has 20,100 parallel sentences split into 10,070 training sentences, 3,397 development sentences, and 6,633 test sentences (3,419 multi-domain, 1,714 news domain, and 1,500 ted talks speech transcript domain). The development and test sets are available upon request.
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.
The dataset consists of tweets belonging to #MeToo movement on Twitter, labelled into different categories. Due to Twitter's development policies, we only provide the tweet ID's and corresponding labels, other data can be fetched via Twitter API. The data has been labelled by experts, with the majority taken into the account for deciding the final label. We provide these labels for each of the tweets. The labels provided for each data point includes -- Relevance, Directed Hate, Generalized Hate, Sarcasm, Allegation, Justification, Refutation, Support, Oppose
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.
The MRQA 2019 Shared Task focuses on generalization in question answering. An effective question answering system should do more than merely interpolate from the training set to answer test examples drawn from the same distribution: it should also be able to extrapolate to out-of-distribution examples — a significantly harder challenge. The dataset is a collection of 18 existing QA dataset (carefully selected subset of them) and converted to the same format (SQuAD format). Among these 18 datasets, six datasets were made available for training, six datasets were made available for development, and the final six for testing. The dataset is released as part of the MRQA 2019 Shared Task.
The Microsoft Terminology Collection can be used to develop localized versions of applications that integrate with Microsoft products. It can also be used to integrate Microsoft terminology into other terminology collections or serve as a base IT glossary for language development in the nearly 100 languages available. Terminology is provided in .tbx format, an industry standard for terminology exchange.
The database is derived from the NCI PID Pathway Interaction Database, and the textual mentions are extracted from cooccurring pairs of genes in PubMed abstracts, processed and annotated by Literome (Poon et al. 2014). This dataset was used in the paper “Compositional Learning of Embeddings for Relation Paths in Knowledge Bases and Text” (Toutanova, Lin, Yih, Poon, and Quirk, 2016).
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.
Translator Human Parity Data Human evaluation results and translation output for the Translator Human Parity Data release, as described in The Translator Human Parity Data release contains all human evaluation results and translations related to our paper "Achieving Human Parity on Automatic Chinese to English News Translation", published on March 14, 2018.
MultiReQA contains the sentence boundary annotation from eight publicly available QA datasets including SearchQA, TriviaQA, HotpotQA, NaturalQuestions, SQuAD, BioASQ, RelationExtraction, and TextbookQA. Five of these datasets, including SearchQA, TriviaQA, HotpotQA, NaturalQuestions, SQuAD, contain both training and test data, and three, including BioASQ, RelationExtraction, TextbookQA, contain only the test data
Multi-Domain Wizard-of-Oz dataset (MultiWOZ), a fully-labeled collection of human-human written conversations spanning over multiple domains and topics. MultiWOZ 2.1 (Eric et al., 2019) identified and fixed many erroneous annotations and user utterances in the original version, resulting in an improved version of the dataset. MultiWOZ 2.2 is a yet another improved version of this dataset, which identifies and fizes dialogue state annotation errors across 17.3% of the utterances on top of MultiWOZ 2.1 and redefines the ontology by disallowing vocabularies of slots with a large number of possible values (e.g., restaurant name, time of booking) and introducing standardized slot span annotations for these slots.
Multi-XScience, a large-scale multi-document summarization dataset created from scientific articles. Multi-XScience introduces a challenging multi-document summarization task: writing the related-work section of a paper based on its abstract and the articles it references.
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.
This paper presents the disease name and concept annotations of the NCBI disease corpus, a collection of 793 PubMed abstracts fully annotated at the mention and concept level to serve as a research resource for the biomedical natural language processing community. Each PubMed abstract was manually annotated by two annotators with disease mentions and their corresponding concepts in Medical Subject Headings (MeSH®) or Online Mendelian Inheritance in Man (OMIM®). Manual curation was performed using PubTator, which allowed the use of pre-annotations as a pre-step to manual annotations. Fourteen annotators were randomly paired and differing annotations were discussed for reaching a consensus in two annotation phases. In this setting, a high inter-annotator agreement was observed. Finally, all results were checked against annotations of the rest of the corpus to assure corpus-wide consistency. For more details, see: The original dataset can be downloaded from: This dataset has been converted to CoNLL format for NER using the following tool: Note: there is a duplicate document (PMID 8528200) in the original data, and the duplicate is recreated in the converted data.
This dataset provides version 1115 of the belief extracted by CMU's Never Ending Language Learner (NELL) and version 1110 of the candidate belief extracted by NELL. See NELL is an open information extraction system that attempts to read the Clueweb09 of 500 million web pages ( and general web searches. The dataset has 4 configurations: nell_belief, nell_candidate, nell_belief_sentences, and nell_candidate_sentences. nell_belief is certainties of belief are lower. The two sentences config extracts the CPL sentence patterns filled with the applicable 'best' literal string for the entities filled into the sentence patterns. And also provides sentences found using web searches containing the entities and relationships. There are roughly 21M entries for nell_belief_sentences, and 100M sentences for nell_candidate_sentences.
Neural-Code-Search-Evaluation-Dataset presents an evaluation dataset consisting of natural language query and code snippet pairs and a search corpus consisting of code snippets collected from the most popular Android repositories on GitHub.
A parallel corpus of News Commentaries provided by WMT for training SMT. The source is taken from CASMACAT: 12 languages, 63 bitexts total number of files: 61,928 total number of tokens: 49.66M total number of sentence fragments: 1.93M
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.
The NQ-Open task, introduced by Lee 2019, is an open domain question answering benchmark that is derived from Natural Questions. The goal is to predict an English answer string for an input English question. All questions can be answered using the contents of English Wikipedia.
NumerSense is a new numerical commonsense reasoning probing task, with a diagnostic dataset consisting of 3,145 masked-word-prediction probes. We propose to study whether numerical commonsense knowledge can be induced from pre-trained language models like BERT, and to what extent this access to knowledge robust against adversarial examples is. We hope this will be beneficial for tasks such as knowledge base completion and open-domain question answering.
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).
Offensive language identification in dravidian lanaguages dataset. The goal of this task is to identify offensive language content of the code-mixed dataset of comments/posts in Dravidian Languages ( (Tamil-English, Malayalam-English, and Kannada-English)) collected from social media.
The OHSUMED test collection is a set of 348,566 references from MEDLINE, the on-line medical information database, consisting of titles and/or abstracts from 270 medical journals over a five-year period (1987-1991). The available fields are title, abstract, MeSH indexing terms, author, source, and publication type.
The Ollie dataset includes two configs for the data used to train the Ollie informatation extraction algorithm, for 18M sentences and 3M sentences respectively. This data is for academic use only. From the authors: Ollie is a program that automatically identifies and extracts binary relationships from English sentences. Ollie is designed for Web-scale information extraction, where target relations are not specified in advance. Ollie is our second-generation information extraction system . Whereas ReVerb operates on flat sequences of tokens, Ollie works with the tree-like (graph with only small cycles) representation using Stanford's compression of the dependencies. This allows Ollie to capture expression that ReVerb misses, such as long-range relations. Ollie also captures context that modifies a binary relation. Presently Ollie handles attribution (He said/she believes) and enabling conditions (if X then). More information is available at the Ollie homepage:
This dataset is a compilation of the OneStopEnglish corpus of texts written at three reading levels into one file. Text documents are classified into three reading levels - ele, int, adv (Elementary, Intermediate and Advance). This dataset demonstrates its usefulness for through two applica-tions - automatic readability assessment and automatic text simplification. The corpus consists of 189 texts, each in three versions/reading levels (567 in total).
This is a new collection of translated movie subtitles from IMPORTANT: If you use the OpenSubtitle corpus: Please, add a link to to your website and to your reports and publications produced with the data! This is a slightly cleaner version of the subtitle collection using improved sentence alignment and better language checking. 62 languages, 1,782 bitexts total number of files: 3,735,070 total number of tokens: 22.10G total number of sentence fragments: 3.35G
OPUS-100 is English-centric, meaning that all training pairs include English on either the source or target side. The corpus covers 100 languages (including English).OPUS-100 contains approximately 55M sentence pairs. Of the 99 language pairs, 44 have 1M sentence pairs of training data, 73 have at least 100k, and 95 have at least 10k.
This is a collection of copyright free books aligned by Andras Farkas, which are available from Note that the texts are rather dated due to copyright issues and that some of them are manually reviewed (check the meta-data at the top of the corpus files in XML). The source is multilingually aligned, which is available from In OPUS, the alignment is formally bilingual but the multilingual alignment can be recovered from the XCES sentence alignment files. Note also that the alignment units from the original source may include multi-sentence paragraphs, which are split and sentence-aligned in OPUS. All texts are freely available for personal, educational and research use. Commercial use (e.g. reselling as parallel books) and mass redistribution without explicit permission are not granted. Please acknowledge the source when using the data! 16 languages, 64 bitexts total number of files: 158 total number of tokens: 19.50M total number of sentence fragments: 0.91M
A collection of translation memories provided by the JRC. Source: 25 languages, 299 bitexts total number of files: 817,410 total number of tokens: 2.13G total number of sentence fragments: 113.52M
A parallel corpus of GNOME localization files. Source: 187 languages, 12,822 bitexts total number of files: 113,344 total number of tokens: 267.27M total number of sentence fragments: 58.12M
Parallel corpora from Web Crawls collected in the ParaCrawl project 40 languages, 41 bitexts total number of files: 20,995 total number of tokens: 21.40G total number of sentence fragments: 1.12G
This is a Croatian-English parallel corpus of transcribed and translated TED talks, originally extracted from The corpus is compiled by Željko Agić and is taken from provided under the CC-BY-NC-SA license. 2 languages, total number of files: 2 total number of tokens: 2.81M total number of sentence fragments: 0.17M
A parallel corpus of Ubuntu localization files. Source: 244 languages, 23,988 bitexts total number of files: 30,959 total number of tokens: 29.84M total number of sentence fragments: 7.73M
This is a corpus of parallel sentences extracted from Wikipedia by Krzysztof Wołk and Krzysztof Marasek. Please cite the following publication if you use the data: Krzysztof Wołk and Krzysztof Marasek: Building Subject-aligned Comparable Corpora and Mining it for Truly Parallel Sentence Pairs., Procedia Technology, 18, Elsevier, p.126-132, 2014 20 languages, 36 bitexts total number of files: 114 total number of tokens: 610.13M total number of sentence fragments: 25.90M
ParaPat: The Multi-Million Sentences Parallel Corpus of Patents Abstracts This dataset contains the developed parallel corpus from the open access Google Patents dataset in 74 language pairs, comprising more than 68 million sentences and 800 million tokens. Sentences were automatically aligned using the Hunalign algorithm for the largest 22 language pairs, while the others were abstract (i.e. paragraph) aligned.
PAWS: Paraphrase Adversaries from Word Scrambling This dataset contains 108,463 human-labeled and 656k noisily labeled pairs that feature the importance of modeling structure, context, and word order information for the problem of paraphrase identification. The dataset has two subsets, one based on Wikipedia and the other one based on the Quora Question Pairs (QQP) dataset. For further details, see the accompanying paper: PAWS: Paraphrase Adversaries from Word Scrambling ( PAWS-QQP is not available due to license of QQP. It must be reconstructed by downloading the original data and then running our scripts to produce the data and attach the labels. NOTE: There might be some missing or wrong labels in the dataset and we have replaced them with -1.
PAWS-X, a multilingual version of PAWS (Paraphrase Adversaries from Word Scrambling) for six languages. This dataset contains 23,659 human translated PAWS evaluation pairs and 296,406 machine translated training pairs in six typologically distinct languages: French, Spanish, German, Chinese, Japanese, and Korean. English language is available by default. All translated pairs are sourced from examples in PAWS-Wiki. For further details, see the accompanying paper: PAWS-X: A Cross-lingual Adversarial Dataset for Paraphrase Identification ( NOTE: There might be some missing or wrong labels in the dataset and we have replaced them with -1.
PearRead is a dataset of scientific peer reviews available to help researchers study this important artifact. The dataset consists of over 14K paper drafts and the corresponding accept/reject decisions in top-tier venues including ACL, NIPS and ICLR, as well as over 10K textual peer reviews written by experts for a subset of the papers.
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`).
A parallel corpus originally extracted from The original documents are written in English and have been partly translated into 21 languages. The original manuals contain about 500,000 words. The amount of actually translated texts varies for different languages between 50,000 and 380,000 words. The corpus is rather noisy and may include parts from the English original in some of the translations. The corpus is tokenized and each language pair has been sentence aligned. 23 languages, 252 bitexts total number of files: 71,414 total number of tokens: 3.28M total number of sentence fragments: 1.38M
This new dataset is the large scale sentence aligned corpus in 11 Indian languages, viz. CVIT-PIB corpus that is the largest multilingual corpus available for Indian languages.
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.
PolEval is a SemEval-inspired evaluation campaign for natural language processing tools for Polish.Submitted solutions compete against one another within certain tasks selected by organizers, using available data and are evaluated according topre-established procedures. One of the tasks in PolEval-2019 was Machine Translation (Task-4). The task is to train as good as possible machine translation system, using any technology,with limited textual resources.The competition will be done for 2 language pairs, more popular English-Polish (into Polish direction) and pair that can be called low resourcedRussian-Polish (in both directions). Here, Polish-English is also made available to allow for training in both directions. However, the test data is ONLY available for English-Polish.
`prachathai-67k`: News Article Corpus and Multi-label Text Classificdation from The prachathai-67k dataset was scraped from the news site Prachathai. We filtered out those articles with less than 500 characters of body text, mostly images and cartoons. It contains 67,889 articles wtih 12 curated tags from August 24, 2004 to November 15, 2018. The dataset was originally scraped by @lukkiddd and cleaned by @cstorm125. You can also see preliminary exploration at
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.
This is the Penn Treebank Project: Release 2 CDROM, featuring a million words of 1989 Wall Street Journal material. This corpus has been annotated for part-of-speech (POS) information. In addition, over half of it has been annotated for skeletal syntactic structure.
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.
PubMedQA is a novel biomedical question answering (QA) dataset collected from PubMed abstracts. The task of PubMedQA is to answer research questions with yes/no/maybe (e.g.: Do preoperative statins reduce atrial fibrillation after coronary artery bypass grafting?) using the corresponding abstracts. PubMedQA has 1k expert-annotated, 61.2k unlabeled and 211.3k artificially generated QA instances. Each PubMedQA instance is composed of (1) a question which is either an existing research article title or derived from one, (2) a context which is the corresponding abstract without its conclusion, (3) a long answer, which is the conclusion of the abstract and, presumably, answers the research question, and (4) a yes/no/maybe answer which summarizes the conclusion. PubMedQA is the first QA dataset where reasoning over biomedical research texts, especially their quantitative contents, is required to answer the questions.
The dataset contains question-answer pairs to model verbal predicate-argument structure. The questions start with wh-words (Who, What, Where, What, etc.) and contain a verb predicate in the sentence; the answers are phrases in the sentence. There were 2 datsets used in the paper, newswire and wikipedia. Unfortunately the newswiredataset is built from CoNLL-2009 English training set that is covered under license Thus, we are providing only Wikipedia training set here. Please check for more details on newswire dataset. For the Wikipedia domain, randomly sampled sentences from the English Wikipedia (excluding questions and sentences with fewer than 10 or more than 60 words) were taken. This new dataset is designed to solve this great NLP task and is crafted with a lot of care.
QED, is a linguistically informed, extensible framework for explanations in question answering. A QED explanation specifies the relationship between a question and answer according to formal semantic notions such as referential equality, sentencehood, and entailment. It is an expertannotated dataset of QED explanations built upon a subset of the Google Natural Questions dataset.
The QCRI Educational Domain Corpus (formerly QCRI AMARA Corpus) is an open multilingual collection of subtitles for educational videos and lectures collaboratively transcribed and translated over the AMARA web-based platform. Developed by: Qatar Computing Research Institute, Arabic Language Technologies Group The QED Corpus is made public for RESEARCH purpose only. The corpus is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. Copyright Qatar Computing Research Institute. All rights reserved. 225 languages, 9,291 bitexts total number of files: 271,558 total number of tokens: 371.76M total number of sentence fragments: 30.93M
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.
SAMSum Corpus contains over 16k chat dialogues with manually annotated summaries. There are two features: - dialogue: text of dialogue. - summary: human written summary of the dialogue. - id: id of a example.
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 Schema-Guided Dialogue dataset (SGD) was developed for the Dialogue State Tracking task of the Eights Dialogue Systems Technology Challenge (dstc8). The SGD dataset consists of over 18k annotated multi-domain, task-oriented conversations between a human and a virtual assistant. These conversations involve interactions with services and APIs spanning 17 domains, ranging from banks and events to media, calendar, travel, and weather. For most of these domains, the SGD dataset contains multiple different APIs, many of which have overlapping functionalities but different interfaces, which reflects common real-world scenarios.
A parallel corpus of full-text scientific articles collected from Scielo database in the following languages: English, Portuguese and Spanish. The corpus is sentence aligned for all language pairs, as well as trilingual aligned for a small subset of sentences. Alignment was carried out using the Hunalign algorithm.
A new multi-target dataset of 5.4K TLDRs over 3.2K papers. SCITLDR contains both author-written and expert-derived TLDRs, where the latter are collected using a novel annotation protocol that produces high-quality summaries while minimizing annotation burden.
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.
Propagandistic news articles use specific techniques to convey their message, such as whataboutism, red Herring, and name calling, among many others. The Propaganda Techniques Corpus (PTC) allows to study automatic algorithms to detect them. We provide a permanent leaderboard to allow researchers both to advertise their progress and to be up-to-speed with the state of the art on the tasks offered (see below for a definition).
SETimes – A Parallel Corpus of English and South-East European Languages The corpus is based on the content published on the news portal. The news portal publishes “news and views from Southeast Europe” in ten languages: Bulgarian, Bosnian, Greek, English, Croatian, Macedonian, Romanian, Albanian and Serbian. This version of the corpus tries to solve the issues present in an older version of the corpus (published inside OPUS, described in the LREC 2010 paper by Francis M. Tyers and Murat Serdar Alperen). The following procedures were applied to resolve existing issues: - stricter extraction process – no HTML residues present - language identification on every non-English document – non-English online documents contain English material in case the article was not translated into that language - resolving encoding issues in Croatian and Serbian – diacritics were partially lost due to encoding errors – text was rediacritized.
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 .
SimpleQuestions is a dataset for simple QA, which consists of a total of 108,442 questions written in natural language by human English-speaking annotators each paired with a corresponding fact, formatted as (subject, relationship, object), that provides the answer but also a complete explanation. Fast have been extracted from the Knowledge Base Freebase ( We randomly shuffle these questions and use 70% of them (75910) as training set, 10% as validation set (10845), and the remaining 20% as test set.
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.
Snips' built in intents dataset was initially used to compare different voice assistants and released as a public dataset hosted at 2016-12-built-in-intents. The dataset contains 328 utterances over 10 intent classes. The related paper mentioned on the github page is and a related Medium post is .
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" (
Dataset with the text of 10% of questions and answers from the Stack Overflow programming Q&A website. This is organized as three tables: Questions contains the title, body, creation date, closed date (if applicable), score, and owner ID for all non-deleted Stack Overflow questions whose Id is a multiple of 10. Answers contains the body, creation date, score, and owner ID for each of the answers to these questions. The ParentId column links back to the Questions table. Tags contains the tags on each of these questions.
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."
The SOFC-Exp corpus consists of 45 open-access scholarly articles annotated by domain experts. A corpus and an inter-annotator agreement study demonstrate the complexity of the suggested named entity recognition and slot filling tasks as well as high annotation quality is presented in the accompanying paper.
This is a collection of parallel corpora collected by Hercules Dalianis and his research group for bilingual dictionary construction. More information in: Hercules Dalianis, Hao-chun Xing, Xin Zhang: Creating a Reusable English-Chinese Parallel Corpus for Bilingual Dictionary Construction, In Proceedings of LREC2010 (source: and Konstantinos Charitakis (2007): Using Parallel Corpora to Create a Greek-English Dictionary with UPLUG, In Proceedings of NODALIDA 2007. Afrikaans-English: Aldin Draghoender and Mattias Kanhov: Creating a reusable English – Afrikaans parallel corpora for bilingual dictionary construction 4 languages, 3 bitexts total number of files: 6 total number of tokens: 1.32M total number of sentence fragments: 0.15M
We have developed an efficient algorithm and implementation of a dictionary-based approach to named entity recognition, which we here use to identifynames of species and other taxa in text. The tool, SPECIES, is more than an order of magnitude faster and as accurate as existing tools. The precision and recall was assessed both on an existing gold-standard corpus and on a new corpus of 800 abstracts, which were manually annotated after the development of the tool. The corpus comprises abstracts from journals selected to represent many taxonomic groups, which gives insights into which types of organism names are hard to detect and which are easy. Finally, we have tagged organism names in the entire Medline database and developed a web resource, ORGANISMS, that makes the results accessible to the broad community of biologists.
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.
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 Switchboard Dialog Act Corpus (SwDA) extends the Switchboard-1 Telephone Speech Corpus, Release 2 with turn/utterance-level dialog-act tags. The tags summarize syntactic, semantic, and pragmatic information about the associated turn. The SwDA project was undertaken at UC Boulder in the late 1990s. The SwDA is not inherently linked to the Penn Treebank 3 parses of Switchboard, and it is far from straightforward to align the two resources. In addition, the SwDA is not distributed with the Switchboard's tables of metadata about the conversations and their participants.
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.
The first gold standard Tamil-English code-switched, sentiment-annotated corpus containing 15,744 comment posts from YouTube. Train: 11,335 Validation: 1,260 and Test: 3,149. This makes the largest general domain sentiment dataset for this relatively low-resource language with code-mixing phenomenon. The dataset contains all the three types of code-mixed sentences - Inter-Sentential switch, Intra-Sentential switch and Tag switching. Most comments were written in Roman script with either Tamil grammar with English lexicon or English grammar with Tamil lexicon. Some comments were written in Tamil script with English expressions in between.
This is a collection of Quran translations compiled by the Tanzil project The translations provided at this page are for non-commercial purposes only. If used otherwise, you need to obtain necessary permission from the translator or the publisher. If you are using more than three of the following translations in a website or application, we require you to put a link back to this page to make sure that subsequent users have access to the latest updates. 42 languages, 878 bitexts total number of files: 105 total number of tokens: 22.33M total number of sentence fragments: 1.01M
A freely available paraphrase corpus for 73 languages extracted from the Tatoeba database. Tatoeba is a crowdsourcing project mainly geared towards language learners. Its aim is to provide example sentences and translations for particular linguistic constructions and words. The paraphrase corpus is created by populating a graph with Tatoeba sentences and equivalence links between sentences “meaning the same thing”. This graph is then traversed to extract sets of paraphrases. Several language-independent filters and pruning steps are applied to remove uninteresting sentences. A manual evaluation performed on three languages shows that between half and three quarters of inferred paraphrases are correct and that most remaining ones are either correct but trivial, or near-paraphrases that neutralize a morphological distinction. The corpus contains a total of 1.9 million sentences, with 200 – 250 000 sentences per language. It covers a range of languages for which, to our knowledge,no other paraphrase dataset exists.
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