Show simple item record

dc.contributor.authorMcConky, Katie Theresa
dc.date.accessioned2016-04-05T19:13:35Z
dc.date.available2016-04-05T19:13:35Z
dc.date.issued2013
dc.identifier.isbn9781267946409
dc.identifier.other1316931412
dc.identifier.urihttp://hdl.handle.net/10477/50598
dc.description.abstractThis work covers topics in event coreference and event classification from spoken conversation. Event coreference is the process of identifying descriptions of the same event across sentences, documents, or structured databases. Existing event coreference work focuses on sentence similarity models or feature based similarity models requiring slot filling. This work shows the effectiveness of using a hybrid approach where the similarity of two events is determined by a combination of the similarity of the two event descriptions, in addition to the similarity of the event context features of location and time. A dynamic weighting approach is taken to combine the three similarity scores together. The described approach provides several benefits including improving event resolution and requiring less reliance on sophisticated natural language processing. A spatial and hierarchical based location similarity measure is developed to support the dynamic weighting event coreference approach. A complete end-to-end system is developed starting with geo-referencing toponyms from natural language text and structured documents. An experiment is conducted that shows a significant improvement in precision when using the developed spatial-hierarchical location similarity measure for an event coreference process over using location string matching techniques or Euclidean distance measures between location centroids as part of the event coreference process. Additionally, the developed spatial-hierarchical similarity measure provides robust predictable performance regardless of the parameter settings and the quality of the preceding location resolution processes. In parallel, a sentence similarity measure is developed for the event coreferencing domain based on the evaluation of location and time coupled with existing sentence similarity measures. Experiments show an increase of 30 to 64% in F-Measure performance in event coreferencing experiments when location and date are included as part of the sentence similarity measure. Seven sentence similarity measures are evaluated including semantic, word overlap and term frequency- inverse document frequency (TFIDF) based measures. TFIDF based measures are shown to have superior performance in two event coreference experiments. Finally, a novel application and implementation of conceptual spaces is applied to topic detection in military conversations to aid in battlefield event reporting. Extensions are made to existing conceptual spaces models in order to take into account contradictory information present in observations to mitigate the errors present in the noisy speech recognition data. Comparison of the conceptual spaces algorithm to a traditional approach using support vector machines indicates significant promise for this new application of conceptual spaces.
dc.languageEnglish
dc.sourceDissertations & Theses @ SUNY Buffalo,ProQuest Dissertations & Theses Global
dc.subjectCommunication and the arts
dc.subjectApplied sciences
dc.subjectConceptual spaces
dc.subjectEvent coreference
dc.subjectEvent resolution
dc.subjectSimilarity
dc.subjectTopic detection on conversation
dc.titleApplications of location similarity measures and conceptual spaces to event coreference and classification
dc.typeDissertation/Thesis


Files in this item

Thumbnail

This item appears in the following Collection(s)

Show simple item record