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.
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.
AQuaMuSe is a novel scalable approach to automatically mine dual query based multi-document summarization datasets for extractive and abstractive summaries using question answering dataset (Google Natural Questions) and large document corpora (Common Crawl)
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 (firstname.lastname@example.org) if
you have any concerns.
This dataset contains two corpora in Spanish and Catalan that consist of annotated Twitter messages for automatic stance detection. The data was collected over 12 days during February and March of 2019 from tweets posted in Barcelona, and during September of 2018 from tweets posted in the town of Terrassa, Catalonia.
Each corpus is annotated with three classes: AGAINST, FAVOR and NEUTRAL, which express the stance towards the target - independence of Catalonia.
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.
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".
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
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
omcsnet_sentences_free has 898,161 lines. omcsnet_sentences_more has
Original downloads are available here
https://github.com/commonsense/conceptnet5/wiki/Downloads. For more
information, see: https://github.com/commonsense/conceptnet5/wiki
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.
CrowS-Pairs, a challenge dataset for measuring the degree to which U.S. stereotypical biases present in the masked language models (MLMs).
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.
A large dataset of human-human negotiations on a multi-issue bargaining task, where agents who cannot observe each other’s reward functions must reach anagreement (o a deal) via natural language dialogue.
Benchmark dataset for low-resource multiclass classification, with 4,015 training, 500 testing, and 500 validation examples, each labeled as part of five classes. Each sample can be a part of multiple classes. Collected as tweets.
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.
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.
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.
The corpora comprise of files per data provider that are encoded in the IOB format (Ramshaw & Marcus, 1995). The IOB format is a simple text chunking format that divides texts into single tokens per line, and, separated by a whitespace, tags to mark named entities. The most commonly used categories for tags are PER (person), LOC (location) and ORG (organization). To mark named entities that span multiple tokens, the tags have a prefix of either B- (beginning of named entity) or I- (inside of named entity). O (outside of named entity) tags are used to mark tokens that are not a named entity.
Low-Resource Fake News Detection Corpora in Filipino. The first of its kind. Contains 3,206 expertly-labeled news samples, half of which are real and half of which are fake.
FLUE is an evaluation setup for French NLP systems similar to the popular GLUE benchmark. The goal is to enable further reproducible experiments in the future and to share models and progress on the French language.
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.
Contains 10k tweets (training set) that are labeled as hate speech or non-hate speech. Released with 4,232 validation and 4,232 testing samples. Collected during the 2016 Philippine Presidential Elections.
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.
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.
Launchpad.net 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.
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.
This new dataset is designed to assess the funniness of edited news headlines.
Existing question answering datasets focus on dealing with homogeneous information, based either only on text or KB/Table information alone. However, as human knowledge is distributed over heterogeneous forms, using homogeneous information alone might lead to severe coverage problems. To fill in the gap, we present HybridQA, a new large-scale question-answering dataset that requires reasoning on heterogeneous information. Each question is aligned with a Wikipedia table and multiple free-form corpora linked with the entities in the table. The questions are designed to aggregate both tabular information and text information, i.e., lack of either form would render the question unanswerable.
A dataset of about 20k questions that are elicited from readers as they naturally read through a document sentence by sentence. Compared to existing datasets, INQUISITIVE questions target more towards high-level (semantic and discourse) comprehension of text. Because these questions are generated while the readers are processing the information, the questions directly communicate gaps between the reader’s and writer’s knowledge about the events described in the text, and are not necessarily answered in the document itself. This type of question reflects a real-world scenario: if one has questions during reading, some of them are answered by the text later on, the rest are not, but any of them would help further the reader’s understanding at the particular point when they asked it. This resource could enable question generation models to simulate human-like curiosity and cognitive processing, which may open up a new realm of applications.
KdConv is a Chinese multi-domain Knowledge-driven Conversionsation dataset, grounding the topics in multi-turn conversations to knowledge graphs. KdConv contains 4.5K conversations from three domains (film, music, and travel), and 86K utterances with an average turn number of 19.0. These conversations contain in-depth discussions on related topics and natural transition between multiple topics, while the corpus can also used for exploration of transfer learning and domain adaptation.\
KILT tasks training and evaluation data.
- [FEVER](https://fever.ai) | Fact Checking | fever
- [AIDA CoNLL-YAGO](https://www.mpi-inf.mpg.de/departments/databases-and-information-systems/research/ambiverse-nlu/aida/downloads) | Entity Linking | aidayago2
- [WNED-WIKI](https://github.com/U-Alberta/wned) | Entity Linking | wned
- [WNED-CWEB](https://github.com/U-Alberta/wned) | Entity Linking | cweb
- [T-REx](https://hadyelsahar.github.io/t-rex) | Slot Filling | trex
- [Zero-Shot RE](http://nlp.cs.washington.edu/zeroshot) | Slot Filling | structured_zeroshot
- [Natural Questions](https://ai.google.com/research/NaturalQuestions) | Open Domain QA | nq
- [HotpotQA](https://hotpotqa.github.io) | Open Domain QA | hotpotqa
- [TriviaQA](http://nlp.cs.washington.edu/triviaqa) | Open Domain QA | triviaqa
- [ELI5](https://facebookresearch.github.io/ELI5/explore.html) | Open Domain QA | eli5
- [Wizard of Wikipedia](https://parl.ai/projects/wizard_of_wikipedia) | Dialogue | wow
To finish linking TriviaQA questions to the IDs provided, follow the instructions [here](http://github.com/huggingface/datasets/datasets/kilt_tasks/README.md).
LAMA is a dataset used to probe and analyze the factual and commonsense knowledge contained in pretrained language models. See https://github.com/facebookresearch/LAMA.
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.
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.
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.
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.
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
http://rtw.ml.cmu.edu/rtw/overview. NELL is an open information
extraction system that attempts to read the Clueweb09 of 500 million
web pages (http://boston.lti.cs.cmu.edu/Data/clueweb09/) and general
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
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.
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.
Named entities Recognition dataset for Norwegian. It is
a version of the Universal Dependency (UD) Treebank for both Bokmål and Nynorsk (UDN) where
all proper nouns have been tagged with their type according to the NER tagging scheme. UDN is a converted
version of the Norwegian Dependency Treebank into the UD scheme.
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.
The researchers of OCLAR Marwan et al. (2019), they gathered Arabic costumer reviews from Google reviewsa and Zomato
website (https://www.zomato.com/lebanon) on wide scope of domain, including restaurants, hotels, hospitals, local shops,
etc.The corpus finally contains 3916 reviews in 5-rating scale. For this research purpose, the positive class considers
rating stars from 5 to 3 of 3465 reviews, and the negative class is represented from values of 1 and 2 of about
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
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:
The “One Million Posts” corpus is an annotated data set consisting of
user comments posted to an Austrian newspaper website (in German language).
DER STANDARD is an Austrian daily broadsheet newspaper. On the newspaper’s website,
there is a discussion section below each news article where readers engage in
online discussions. The data set contains a selection of user posts from the
12 month time span from 2015-06-01 to 2016-05-31. There are 11,773 labeled and
1,000,000 unlabeled posts in the data set. The labeled posts were annotated by
professional forum moderators employed by the newspaper.
The data set contains the following data for each post:
* Post ID
* Article ID
* Headline (max. 250 characters)
* Main Body (max. 750 characters)
* User ID (the user names used by the website have been re-mapped to new numeric IDs)
* Time stamp
* Parent post (replies give rise to tree-like discussion thread structures)
* Status (online or deleted by a moderator)
* Number of positive votes by other community members
* Number of negative votes by other community members
For each article, the data set contains the following data:
* Article ID
* Publishing date
* Topic Path (e.g.: Newsroom / Sports / Motorsports / Formula 1)
Detailed descriptions of the post selection and annotation procedures are given in the paper.
## Annotated Categories
Potentially undesirable content:
* Sentiment (negative/neutral/positive)
An important goal is to detect changes in the prevalent sentiment in a discussion, e.g.,
the location within the fora and the point in time where a turn from positive/neutral
sentiment to negative sentiment takes place.
* Off-Topic (yes/no)
Posts which digress too far from the topic of the corresponding article.
* Inappropriate (yes/no)
Swearwords, suggestive and obscene language, insults, threats etc.
* Discriminating (yes/no)
Racist, sexist, misogynistic, homophobic, antisemitic and other misanthropic content.
Neutral content that requires a reaction:
* Feedback (yes/no)
Sometimes users ask questions or give feedback to the author of the article or the
newspaper in general, which may require a reply/reaction.
Potentially desirable content:
* Personal Stories (yes/no)
In certain fora, users are encouraged to share their personal stories, experiences,
anecdotes etc. regarding the respective topic.
* Arguments Used (yes/no)
It is desirable for users to back their statements with rational argumentation,
reasoning and sources.
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 instructables.com, which provides users with instructions on how to build, craft,
bake, or manipulate objects using everyday materials.
The underlying task is formualted as multiple choice question answering:
given a question `q` and two possible solutions `s1`, `s2`, a model or
a human must choose the most appropriate solution, of which exactly one is correct.
The dataset is further cleaned of basic artifacts using the AFLite algorithm which is an improvement of
adversarial filtering. The dataset contains 16,000 examples for training, 2,000 for development and 3,000 for testing.
This dataset is for studying computational models trained to reason about prototypical situations. Using deterministic filtering a sampling from a larger set of all transcriptions was built. It contains 9789 instances where each instance represents a survey question from Family Feud game. Each instance exactly is a question, a set of answers, and a count associated with each answer.
Each line is a json dictionary, in which:
1. question - contains the question (in original and a normalized form)
2. answerstrings - contains the original answers provided by survey respondents (when available), along with the counts for each string. Because the FamilyFeud data has only cluster names rather than strings, those cluster names are included with 0 weight.
3. answer-clusters - lists clusters, with the count of each cluster and the strings included in that cluster. Each cluster is given a unique ID that can be linked to in the assessment files.
NLM produces a baseline set of MEDLINE/PubMed citation records in XML format for download on an annual basis. The annual baseline is released in December of each year. Each day, NLM produces update files that include new, revised and deleted citations. See our documentation page for more information.
Question Answering in Context is a dataset for modeling, understanding,
and participating in information seeking dialog. Data instances consist
of an interactive dialog between two crowd workers: (1) a student who
poses a sequence of freeform questions to learn as much as possible
about a hidden Wikipedia text, and (2) a teacher who answers the questions
by providing short excerpts (spans) from the text. QuAC introduces
challenges not found in existing machine comprehension datasets: its
questions are often more open-ended, unanswerable, or only meaningful
within the dialog context.
ReDial (Recommendation Dialogues) is an annotated dataset of dialogues, where users
recommend movies to each other. The dataset was collected by a team of researchers working at
Polytechnique Montréal, MILA – Quebec AI Institute, Microsoft Research Montréal, HEC Montreal, and Element AI.
The dataset allows research at the intersection of goal-directed dialogue systems
(such as restaurant recommendation) and free-form (also called “chit-chat”) dialogue systems.
The Rationalized English-French Semantic Divergences (REFreSD) dataset consists of 1,039
English-French sentence-pairs annotated with sentence-level divergence judgments and token-level
rationales. For any questions, write to email@example.com.
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.
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.
ShARC is a Conversational Question Answering dataset focussing on question answering from texts containing rules. The goal is to answer questions by possibly asking follow-up questions first. It is assumed assume that the question is often underspecified, in the sense that the question does not provide enough information to be answered directly. However, an agent can use the supporting rule text to infer what needs to be asked in order to determine the final answer.
ShARC, a conversational QA task, requires a system to answer user questions based on rules expressed in natural language text. However, it is found that in the ShARC dataset there are multiple spurious patterns that could be exploited by neural models. SharcModified is a new dataset which reduces the patterns identified in the original dataset. To reduce the sensitivity of neural models, for each occurence of an instance conforming to any of the patterns, we automatically construct alternatives where we choose to either replace the current instance with an alternative instance which does not exhibit the pattern; or retain the original instance. The modified ShARC has two versions sharc-mod and history-shuffled. For morre details refer to Appendix A.3 .
The SMS Spam Collection v.1 is a public set of SMS labeled messages that have been collected for mobile phone spam research.
It has one collection composed by 5,574 English, real and non-enconded messages, tagged according being legitimate (ham) or spam.
The SNLI corpus (version 1.0) is a collection of 570k human-written English
sentence pairs manually labeled for balanced classification with the labels
entailment, contradiction, and neutral, supporting the task of natural language
inference (NLI), also known as recognizing textual entailment (RTE).
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.
Stereoset is a dataset that measures stereotype bias in language models. Stereoset consists of 17,000 sentences that
measures model preferences across gender, race, religion, and profession.
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.
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 of a 'user' and trained call center operators played the role of the 'assistant'. In this way, users were led to believe they were interacting with an automated system that “spoke” using text-to-speech (TTS) even though it was in fact a human behind the scenes. As a result, users could express themselves however they chose in the context of an automated interface.
Taskmaster is dataset for goal oriented conversations. The Taskmaster-3 dataset consists of 23,757 movie ticketing dialogs. By "movie ticketing" we mean conversations where the customer's goal is to purchase tickets after deciding on theater, time, movie name, number of tickets, and date, or opt out of the transaction. This collection was created using the "self-dialog" method. This means a single, crowd-sourced worker is paid to create a conversation writing turns for both speakers, i.e. the customer and the ticketing agent.
The core of WIT3 is the TED Talks corpus, that basically redistributes the original content published by the TED Conference website (http://www.ted.com). Since 2007,
the TED Conference, based in California, has been posting all video recordings of its talks together with subtitles in English
and their translations in more than 80 languages. Aside from its cultural and social relevance, this content, which is published under the Creative Commons BYNC-ND license, also represents a precious
language resource for the machine translation research community, thanks to its size, variety of topics, and covered languages.
This effort repurposes the original content in a way which is more convenient for machine translation researchers.
TweetQA is the first dataset for QA on social media data by leveraging news media and crowdsourcing.
The objective of this task is to detect hate speech in tweets. For the sake of simplicity, we say a tweet contains hate speech if it has a racist or sexist sentiment associated with it. So, the task is to classify racist or sexist tweets from other tweets.
Formally, given a training sample of tweets and labels, where label ‘1’ denotes the tweet is racist/sexist and label ‘0’ denotes the tweet is not racist/sexist, your objective is to predict the labels on the given test dataset.
Universal Dependencies is a project that seeks to develop cross-linguistically consistent treebank annotation for many languages, with the goal of facilitating multilingual parser development, cross-lingual learning, and parsing research from a language typology perspective. The annotation scheme is based on (universal) Stanford dependencies (de Marneffe et al., 2006, 2008, 2014), Google universal part-of-speech tags (Petrov et al., 2012), and the Interset interlingua for morphosyntactic tagsets (Zeman, 2008).
“Urdu Sentiment Corpus” (USC) shares the dat of Urdu tweets for the sentiment analysis and polarity detection.
The dataset is consisting of tweets and overall, the dataset is comprising over 17, 185 tokens
with 52% records as positive, and 48 % records as negative.
The WebNLG challenge consists in mapping data to text. The training data consists
of Data/Text pairs where the data is a set of triples extracted from DBpedia and the text is a verbalisation
of these triples. For instance, given the 3 DBpedia triples shown in (a), the aim is to generate a text such as (b).
a. (John_E_Blaha birthDate 1942_08_26) (John_E_Blaha birthPlace San_Antonio) (John_E_Blaha occupation Fighter_pilot)
b. John E Blaha, born in San Antonio on 1942-08-26, worked as a fighter pilot
As the example illustrates, the task involves specific NLG subtasks such as sentence segmentation
(how to chunk the input data into sentences), lexicalisation (of the DBpedia properties),
aggregation (how to avoid repetitions) and surface realisation
(how to build a syntactically correct and natural sounding text).
Write & Improve (Yannakoudakis et al., 2018) is an online web platform that assists non-native
English students with their writing. Specifically, students from around the world submit letters,
stories, articles and essays in response to various prompts, and the W&I system provides instant
feedback. Since W&I went live in 2014, W&I annotators have manually annotated some of these
submissions and assigned them a CEFR level.
WikiAsp is a multi-domain, aspect-based summarization dataset in the encyclopedic
domain. In this task, models are asked to summarize cited reference documents of a
Wikipedia article into aspect-based summaries. Each of the 20 domains include 10
domain-specific pre-defined aspects.
WikiLingua is a large-scale multilingual dataset for the evaluation of
crosslingual abstractive summarization systems. The dataset includes ~770k
article and summary pairs in 18 languages from WikiHow. The gold-standard
article-summary alignments across languages was done by aligning the images
that are used to describe each how-to step in an article.
The WikiMovies dataset consists of roughly 100k (templated) questions over 75k entities based on questions with answers in the open movie database (OMDb).
Arabic Version of WikiQA by automatic automatic machine translators and crowdsourced the selection of the best one to be incorporated into the corpus
The dataset extracted from Persian Wikipedia into the form of articles and highlights and cleaned the dataset into pairs of articles and highlights and reduced the articles' length (only version 1.0.0) and highlights' length to a maximum of 512 and 128, respectively, suitable for parsBERT.
WikiANN (sometimes called PAN-X) is a multilingual named entity recognition dataset consisting of Wikipedia articles annotated with LOC (location), PER (person), and ORG (organisation) tags in the IOB2 format. This version corresponds to the balanced train, dev, and test splits of Rahimi et al. (2019), which supports 176 of the 282 languages from the original WikiANN corpus.
Wizard-of-Oz (WOZ) is a dataset for training task-oriented dialogue systems. The dataset is designed around the task of finding a restaurant in the Cambridge, UK area. There are three informable slots (food, pricerange,area) that users can use to constrain the search and six requestable slots (address, phone, postcode plus the three informable slots) that the user can ask a value for once a restaurant has been offered.
XGLUE is a new benchmark dataset to evaluate the performance of cross-lingual pre-trained
models with respect to cross-lingual natural language understanding and generation.
The benchmark is composed of the following 11 tasks:
- POS Tagging (POS)
- News Classification (NC)
- Query-Ad Matching (QADSM)
- Web Page Ranking (WPR)
- QA Matching (QAM)
- Question Generation (QG)
- News Title Generation (NTG)
For more information, please take a look at https://microsoft.github.io/XGLUE/.
Neural abstractive summarization models are highly prone to hallucinate content that is unfaithful to the input
document. The popular metric such as ROUGE fails to show the severity of the problem. The dataset consists of
faithfulness and factuality annotations of abstractive summaries for the XSum dataset. We have crowdsourced 3 judgements
for each of 500 x 5 document-system pairs. This will be a valuable resource to the abstractive summarization community.
The Yelp reviews dataset consists of reviews from Yelp. It is extracted from the Yelp Dataset Challenge 2015 data.
The Yelp reviews full star dataset is constructed by Xiang Zhang (firstname.lastname@example.org) from the above dataset.
It is first used as a text classification benchmark in the following paper: Xiang Zhang, Junbo Zhao, Yann LeCun.
Character-level Convolutional Networks for Text Classification. Advances in Neural Information Processing Systems 28 (NIPS 2015).
ZEST tests whether NLP systems can perform unseen tasks in a zero-shot way, given a natural language description of
the task. It is an instantiation of our proposed framework "learning from task descriptions". The tasks include
classification, typed entity extraction and relationship extraction, and each task is paired with 20 different
annotated (input, output) examples. ZEST's structure allows us to systematically test whether models can generalize
in five different ways.