Ontology extension. There is a general demand for taxonomies to organise large collections of documents into categories for browsing and exploration. In WP2 we examined four existing taxonomies that have been manually created, along with two methods for deriving taxonomies automatically from data items. We used these taxonomies to organise items from a large online cultural heritage collection automatically. We then performed two human evaluations of the taxonomies. The first measured the cohesion of the taxonomies to determine how well they group together similar items under the same concept node. The second analysed the concept relations in the taxonomies. The results show that the manual taxonomies have high quality well defined relations. However, the novel automatic method is found to generate very high cohesion. The results of the analysis are detailed in D2.2, Xxxxxxxx et al. (2012) and a summary is also presented here. The Library of Congress Subject Headings taxonomy has been manually created for the purpose of organising library collections and so might be the obvious choice to organise CH data online. However the results show that the relations within LCSH are defined less clearly than that of the Wikipedia derived taxonomies. WordNet domains performs at a similar level to the LCSH in terms of the quality of the relations. The LDA topic hierarchy gave poor results in terms of the identified topics. The topic pairs were often unrelated, and had no general-to- specific structure as would be desirable for this application. The WikiFreq hierarchy performed slightly better in this regard. Just over half the concept pairs were judged to be related. Just under a third were labelled as ‘Don’t know’ which may reflect the obscurity of the concept nodes identified. It was hoped that organising the frequency counts of the links would organise the hierarchy into a general-to-specific direction. This was not achieved, although the hierarchy does have the benefit of providing the user with an overview of the collection by immediately seeing which kind of items are most prevalent. In terms of cohesion, all the taxonomies achieved similar results except for the WikiFreq taxonomy, which achieved almost perfect cohesion. This shows how effective the Wikipedia links are in grouping together similar items. In conclusion, it was decided that a mixture of WikiFreq (for mapping items to Wikipedia categories) and the Wikipedia taxonomy to yield the taxonomy would give the best results. The...
Ontology extension. We begin by evaluating the ontology extension techniques (Section 5). Two evaluations were performed on the taxonomies. The first measured the cohesion of the item clustering, and the second gathered human judgements of the relations that were found between child- parent concept pairs in the taxonomy. For both evaluations online surveys were created using an in-house crowdsourcing interface. Links to the surveys were sent out to a mailing list comprising staff and students at a large university.
