... specified¹

Many terms have been used in the literature for the relation between anaphoric expressions and discourse entities. For example, Sidner [34] and Webber [40] argue that ``refer'' should be reserved for something people do with words, rather than something words do. Webber uses the term ``evoke'' for first references to an entity and ``access'' for subsequent references. Sidner uses the term ``specify'' for the relation between a noun phrase and a discourse entity. We primarily use Sidner's term, but use ``refer'' in a few contexts in which it seems more natural.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... Walker²

Note that Walker's model is a cache-based model for which recency is a very important but not unique criterion for determining which entities are in the cache.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... transcriptions.³

The semantic parser but not the speech recognizer was available for us to process the NMSU data. Presumably, the speech recognizer would not perform as well on the NMSU dialogs as it does on the CMU dialogs, since it was trained on the latter.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

....⁴

The anaphoric counts include the cases in which both deictic and anaphoric interpretations yield the correct result.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... state.⁵

They do not report how many multiple thread instances appear in their data.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

... (3)⁶

Recall that multiple Temporal Units specified in a single utterance are added to the focus list in order of mention and treated as separate discourse entities.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.