Up-cycling Data for Natural Language Generation

Up-cycling Data for Natural Language Generation Amy Isard, Jon Oberlander, Claire Grover Language Technology Group, Institute for Language, Cognition and Communication, School of Informatics, University of Edinburgh 10 Crichton St, Edinburgh, UK {amy.isard, claire.grover}@ed.ac.uk Abstract Museums and other cultural heritage institutions have large databases of information about the objects in their collections, and existing Natural Language Generation (NLG) systems can generate fluent and adaptive texts for visitors, given appropriate input data, but there is typically a large amount of expert human effort required to bridge the gap between the available and the required data. We describe automatic processes which aim to significantly reduce the need for expert input during the conversion and up-cycling process. We detail domain-independent techniques for processing and enhancing data into a format which allows an existing NLG system to create adaptive texts. First we normalize the dates and names which occur in the data, and we link to the Semantic Web to add extra object descriptions. Then we use Semantic Web queries combined with a wide coverage grammar of English to extract relations which can be used to express the content of database fields in language accessible to a general user. As our test domain we use a database from the Edinburgh Musical Instrument Museum. Keywords: Natural Language Generation, Cultural Heritage, Semantic Web 1. Introduction There are large collections of cultural heritage data which are currently not able to be widely shared and exploited be- cause they are stored in databases whose structure and for- mat are inaccessible to the general public. We use a num- ber of Natural Language Processing techniques to bridge the gap between existing databases and Natural Language Generation systems to create varied, adaptive texts which can be tailored to particular visitors on a journey through an exhibition. A number of research projects have focussed on using NLG systems to create multilingual adaptive texts from structured cultural heritage data. ILEX (O’Donnell et al., 2001) and M-PIRO (Isard et al., 2003; Oberlander et al., 2008) worked from hand-authored resources. They used language-independent databases in a specific format con- taining up to a fifty objects and a few hundred triples describing attributes of the objects, created in collabora- tion with curators. The linguistic resources for each lan- guage available (English for ILEX, and English, Italian and Greek for M-PIRO) were also hand-created by computa- tional linguists. The generated texts could be tailored to a museum visitor’s progress through an exhibition, allow- ing for comparisons between exhibits, and preventing the repetition of background information. More recent systems have generated texts from Semantic Web ontologies; Nat- uralOWL (Galanis and Androutsopoulos, 2007; Androut- sopoulos et al., 2013) generated texts in English and Greek from OWL ontologies, and Dann´ells et al. (2013) gener- ated texts from Semantic Web data in 15 languages, but in both cases expert input was still required to create the nec- essary linguistic resources. Sun and Mellish (2007), Mellish and Pan (2008), Mellish (2010) and Androutsopoulos et al. (2013) have experi- mented with performing …

Content truncated to 500 words