Data sets for NLG

This page lists sets of structured data to be used as input for natural language generation tasks, or to inform research on NLG.

Focus on studying the generation target

PIL: Patient Information Leaflet corpus

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. (direct download link)

Focus on content selection, aggregation

SumTime Meteo

These data contain predictions for meteorological parameters such as precipitation, temperature, wind speed, and cloud cover at various altitudes, at regular intervals for various points in the area of interest.

The weather corpus currently exists as an Access database and, alternatively, in form of CSV (ASCII) files.

Download and Info: SumTime-Meteo

Project link: http://www.csd.abdn.ac.uk/research/sumtime/

CLASSiC WOZ corpus on InformationPresentation in Spoken Dialogue Systems

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.

Focus on generating referring expressions

Referring expression generation is a sub-task of NLG with an active research community.

COCONUT Corpus

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. (direct download link)

GRE3D3 and GRE3D7: Spatial Relations in Referring Expressions

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. The corpora and stimulus scenes are available here.

TUNA Reference Corpus

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. (direct download link)

Focus on lexicalization

...

Focus on syntax, realization

...

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.