Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.642791
Title: Modelling aggregation motivated interactions in descriptive text generation
Author: Cheng, H.
Awarding Body: University of Edinburgh
Current Institution: University of Edinburgh
Date of Award: 2002
Availability of Full Text:
Full text unavailable from EThOS. Please contact the current institution’s library for further details.
Abstract:
This thesis aims at revealing the complex interactions between aggregation and other generation tasks, which have not been explored by previous research, and discusses what these interactions imply for generation architectures. To study these problems in detail, the thesis focuses on embedding phenomena in descriptive text. It identifies regularities in the way human authors produce complex NPs using embedding and the interactions between embedding and such processes as document structuring and referring expression generation. These findings motivate a set of preferences among coherence features, which capture the complex interactions that have been discovered. The preferences mainly include features of entity-based and relation-based coherence and aggregation. They are implemented in ILEX-TS and GA-plan, which represent two dramatically different text planning architectures (a pipeline and a non-pipeline architecture), and the behaviours of the two systems are compared. The thesis quantitatively evaluates the observed embedding rules using an annotated corpus and the output of the two generation systems using human judgement. It also makes an attempt to automatically evaluate the readability of a text. Based on the results of evaluation, the thesis is able to make a number of assertions: firstly, the effect of aggregation on the planning of entity-based and relation-based coherence demands it to be taken into account in text planning to affect the structuring of content; secondly, to produce a coherent text, it is most important to capture the interactions between generation tasks and this should ideally be done in a better way than presented in current NLG systems: and finally, it is possible to capture the preferences among coherence features in a non-sequential way. These form the main contributions of the thesis.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.642791  DOI: Not available
Share: