Designing an Anaphora Resolution Algorithm for Route Instructions

discourse deictic anaphors Precision was 63.6% and Recall 70%. One of the most
common classification errors was that discourse deictic or vague anaphors were
classified as individuals because there was an individual antecedent available.
2.3.3 The PHORA-Algorithm
In this section, I will describe the PHORA-algorithm developed by Byron (2002)
which uses salience calculations as criteria to resolve individual anaphors but which
also takes advantage of the semantic constraints on the antecedents given by the
predicative context of the anaphor. As discourse deictic anaphors usually refer to less
salient abstract entities, semantic constraints can be used to find their referents. Byron
makes use of the observation that pronouns which do not refer to the most salient
item are typically constrained by their context and that they constrain the anaphor to
be incompatible with the more salient referent while indicating the intended referent.
Taking semantic constraints into account is useful for pronoun resolution algorithms
in order to identify the referent of discourse deictic anaphors. This task is generally
judged as being difficult and excluded from most of the pronoun resolution studies
until now.
Following Webber (1991) who suggests that each discourse unit (context) has
a ‘pseudo-DE’ that ‘stands proxy’ for its propositional context, the first action of the
algorithm is to build up the discourse entities and the discourse proxies for the actual
discourse unit (DU n ). To resolve a pronoun in the following discourse unit (DU n+1 ),
the algorithm firstly calculates the most general semantic type (T) that satisfies the
constraints of the predicative context of the pronoun. Then, the algorithm checks the
discourse entities in salience order to find a referent that matches the features of the
pronoun. Depending on the type of anaphor (either personal or demonstrative
pronoun) the algorithm uses different search orders. Finally, each discourse entity is
tested with respect to the type constraint. Every entity that matches the type constraint
of the anaphor (i.e. it either has the same semantic type as the anaphor or it is a
subtype of it) is a possible referent. In general, there is only one possible referent left
after testing the discourse entities.
18