Classifying complex topics using spatial-semantic document visualization : an evaluation of an interaction model to support open-ended search tasks
In this dissertation we propose, test and develop a novel search interaction model to address two key problems associated with conducting an open-ended search task within a classical information retrieval system: (i) the need to reformulate the query within the context of a shifting conception of the problem and (ii) the need to integrate relevant results across a number of separate results sets. In our model the user issues just one highrecall query and then performs a sequence of more focused, distinct aspect searches by browsing the static structured context of a spatial-semantic visualization of this retrieved document set. Our thesis is that unsupervised spatial-semantic visualization can automatically classify retrieved documents into a two-level hierarchy of relevance. In particular we hypothesise that the locality of any given aspect exemplar will tend to comprise a sufficient proportion of same-aspect documents to support a visually guided strategy for focused, same-aspect searching that we term the aspect cluster growing strategy. We examine spatial-semantic classification and potential aspect cluster growing performance across three scenarios derived from topics and relevance judgements from the TREC test collection. Our analyses show that the expected classification can be represented in spatial-semantic structures created from document similarities computed by a simple vector space text analysis procedure. We compare two diametrically opposed approaches to layout optimisation: a global approach that focuses on preserving the all similarities and a local approach that focuses only on the strongest similarities. We find that the local approach, based on a minimum spanning tree of similarities, produces a better classification and, as observed from strategy simulation, more efficient aspect cluster growing performance in most situations, compared to the global approach of multidimensional scaling. We show that a small but significant proportion of aspect clustering growing cases can be problematic, regardless of the layout algorithm used. We identify the characteristics of these cases and, on this basis, demonstrate a set of novel interactive tools that provide additional semantic cues to aid the user in locating same-aspect documents.