Designing an Anaphora Resolution Algorithm for Route Instructions

Abstract
The ambition of this dissertation is to develop an anaphora resolution algorithm for
route instructions which are taken from the Instruction-Based-Learning (IBL)-project.
Route instructions describe spatial features of the environment of the route and they
also specify temporal information about the actions (e.g. movements, turns) to be
performed. Thus, they need to be interpreted with respect to the spatial context
reference and domain-specific knowledge. References to the context are not only
created by noun phrases (referring to entities) or verb phrases (referring to
eventualities), they are also built up by anaphors for example personal or
demonstrative pronouns.
In the IBL-data a variety of anaphoric references to the spatial context occur.
Since most of the anaphoric references in the data were created by the anaphors ‘it’,
‘this’, ‘that’ and ‘This-NP’, this dissertation especially deals with these types of
anaphors. In order to resolve these anaphors, I developed the Route-Instruction-
Anaphora-Resolution-Algorithm (RIAR). Due to the fact the route instructions
specify spatial information of the environment, two other anaphors (the spatial
anaphors ‘here’ and ‘there’) frequently occur in the IBL-data. Thus, I also designed
the Spatial-Reference-Algorithm (SR). The SR-algorithm is the first algorithm which
resolves the spatial anaphors ‘here’ and ‘there’.
After analysing the given data and creating a gold standard, the conceptual
design of both algorithms is evolved with respect to recent research. The RIARalgorithm
integrates concepts of the most recent algorithm developed by Eckert &
Strube (2000) and ideas from a recent work by Byron (2002) which includes semantic
filtering in addition to the common salience calculations inspired by the Centering
Theory. In order to resolve the spatial anaphors ‘here’ and ‘there’ the SR-algorithm
takes into account that these anaphors refer to positions described by prepositional
phrases rather than to entities realized as noun phrases or clauses.
Since neither algorithm is implemented, they are only theoretically tested on a
set of unseen data in order to evaluate their performance. The RIAR-algorithm
performs quite well. It resolves 89% of all personal pronouns, 60% of all
2