Data sets for NLG

This page lists data sets and corpora used for research in natural language generation. They are available for download over the web. If you know of a dataset which is not listed here, you can email siggen-board@aclweb.org, or just click on Edit in the upper left corner of this page and add the system yourself.

Data-to-text/Concept-to-text Generation

These datasets contain data and corresponding texts based on this data.

boxscore-data

https://github.com/harvardnlp/boxscore-data/

This dataset consists of (human-written) NBA basketball game summaries aligned with their corresponding box- and line-scores.

E2E

http://www.macs.hw.ac.uk/InteractionLab/E2E/#data

Crowdsourced restaurant descriptions with corresponding restaurant data. English.

SUMTIME

https://ehudreiter.files.wordpress.com/2016/12/sumtime.zip

Weather forecasts written by human forecasters, with corresponding forecast data, for UK North Sea marine forecasts.

WeatherGov

https://cs.stanford.edu/~pliang/data/weather-data.zip

Computer-generated weather forecasts from weather.gov (US public forecast), along with corresponding weather data.

WebNLG

https://github.com/ThiagoCF05/webnlg

Crowdsourced descriptions of semantic web entities, with corresponding RDF triples.

The Wikipedia company corpus

https://gricad-gitlab.univ-grenoble-alpes.fr/getalp/wikipediacompanycorpus

Company descriptions collected from Wikipedia. The dataset contains semantic representations, short, and long descriptions for 51K companies in English

Referring Expressions Generation

Referring expression generation is a sub-task of NLG that focuses only on the generation of referring expressions (descriptions) that identify specific entities called targets.

GRE3D3 and GRE3D7: Spatial Relations in Referring Expressions

http://jetteviethen.net/research/spatial.html

Two web-based production experiments were conducted by Jette Viethen under the supervision of Robert Dale. The resulting corpora GRE3D3 and GRE3D7 contain 720 and 4480 referring expressions, respectively. Each referring expression describes a simple object in a simple 3D scene. GRE3D3 scenes contain 3 objects and GRE3D7 scenes contain 7 objects.

RefClef, RefCOCO, RefCOCO+ and RefCOCOg

https://github.com/lichengunc/refer

Referring expressions for objects in images, and the corresponding images.

The REAL dataset

https://datastorre.stir.ac.uk/handle/11667/82

Referring expressions for real-wrold objects in images, and the corresponding images.

GeoDescriptors

https://gitlab.citius.usc.es/alejandro.ramos/geodescriptors

Geographical descriptions (eg, "Norte de Galicia") and corresponding regions on a map

TUNA Reference Corpus

https://www.abdn.ac.uk/ncs/departments/computing-science/corpus-496.php

https://www.abdn.ac.uk/ncs/documents/corpus.zip [direct download]

The TUNA Reference Corpus is a semantically and pragmatically transparent corpus of identifying references to objects in visual domains. It was constructed via an online experiment and has since been used in a number of evaluation studies on Referring Expressions Generation, as well as in two Shared Tasks: the Attribute Selection for Referring Expressions Generation task (2007), and the Referring Expression Generation task (2008). Main authors: Kees van Deemter, Albert Gatt, Ielka van der Sluis.

COCONUT Corpus

http://www.pitt.edu/~coconut/coconut-corpus.html

http://www.pitt.edu/%7Ecoconut/corpora/corpus.tar.gz [direct download]

COCONUT was a project on “Cooperative, coordinated natural language utterances”. The COCONUT corpus is a collection of computer-mediated dialogues in which two subjects collaborate on a simple task, namely buying furniture. SGML annotations were added according to the COCONUT-DRI coding scheme.

Stars2 corpus of referring expressions

A collection of 884 annotated definite descriptions produced by 56 subjects in collaborative communication involving speaker-hearer pairs in situations designed so as to challenge existing REG algorithms, with a particular focus on the issue of attribute choice in referential overspeci�fication. Link: https://drive.google.com/file/d/0B-KyU7T8S8bLZ1lEQmJRdUc1V28/view?usp=sharing Cite: https://link.springer.com/article/10.1007/s10579-016-9350-y

b5 corpus of text and referring expressions labelled with personality information

A collection of crowd sourced scene descriptions and an annotated REG corpus, both of which labelled with Big Five personality scores of their authors. Suitable for studies in personality-dependent text generation and referring expression generation. Link: https://drive.google.com/open?id=0B-KyU7T8S8bLTHpaMnh2U2NWZzQ Cite: https://www.aclweb.org/anthology/L18-1183

Dialogue

Cam4NLG

https://github.com/shawnwun/RNNLG/tree/master/data

Cam4NLG: Cam4NLG contains 4 NLG datasets for dialogue system development, each of them is in a unique domain. Each data point contains a (dialogue act, ground truth, handcrafted baseline) tuple.

CLASSiC WOZ corpus on InformationPresentation in Spoken Dialogue Systems

http://www.classic-project.org/corpora

CLASSiC is a project on Computational Learning in Adaptive Systems for Spoken Conversation. The Wizard-of-Oz corpus on Information Presentation in Spoken Dialogue Systems contains the wizards' choices on Information Presentation strategy (summary, compare, recommend , or a combination of those) and attribute selection. The domain is restaurant search in Edinburgh. Objective measures (such as dialogue length, number of database hits, number of sentences generated etc.), as well as subjective measures (the user scores) were logged.

CODA corpus Release 1.0

http://computing.open.ac.uk/coda/resources/code_form.html

This release contains approximately 700 turns of human-authored expository dialogue (by Mark Twain and George Berkeley) which has been aligned with monologue that expresses the same information as the dialogue. The monologue side is annotated with Coherence Relations (RST), and the dialogue side with Dialogue Act tags.

Summarisation

TL;DR

https://toolbox.google.com/datasetsearch/search?query=Webis-TLDR-17%20Corpus&docid=kzcwbWD9z3B4Ah3wAAAAAA%3D%3D

Dataset for abstractive summarization constructed using Reddit posts. It is the largest corpus (approximately 3 Million posts) for informal text such as Social Media text, which can be used to train neural networks for summarization technology.

Image description

Chinese

• Flickr8k-CN: http://lixirong.net/datasets/flickr8kcn

Dutch

• DIDEC: http://didec.uvt.nl
• Flickr30K https://github.com/cltl/DutchDescriptions

German

• Multi30K: http://www.statmt.org/wmt16/multimodal-task.html

Other

PIL: Patient Information Leaflet corpus

http://mcs.open.ac.uk/nlg/old_projects/pills/corpus/PIL/

http://mcs.open.ac.uk/nlg/old_projects/pills/corpus/PIL-corpus-2.0.tar.gz [direct download]

The Patient Information Leaflet (PIL) corpus] is a searchable and browsable collection of patient information leaflets available in various document formats as well as structurally annotated SGML. The PIL corpus was initially developed as part of the ICONOCLAST project at ITRI, Brighton.

Validity of BLEU Evaluation Metric

https://abdn.pure.elsevier.com/en/datasets/data-for-structured-review-of-the-validity-of-bleu

https://abdn.pure.elsevier.com/files/125166547/bleu_survey_data.zip [direct download]

Correlations between BLEU and human evaluations (for MT as well as NLG), extracted from papers in the ACL Anthology

This page was imported semi-automatically from the NLG Resources Wiki which was run by ACL SIGGEN in the years 2005–2009. Please correct conversion errors and help update its contents. Now this page is associated with the Natural Language Generation Portal.