As mentioned earlier, our algorithm for temporal reference resolution is recency based. Specifically, the focus model is structured as a linear list of all times mentioned so far in the current dialog. The list is ordered by recency, and no entries are deleted from the list.
The candidate antecedents are as follows. For each type of anaphoric relation, the most recent Temporal Unit on the focus list that satisfies that relation, if there is one, is a candidate antecedent.
The antecedent is chosen from among the candidate antecedents based on a combined score reflecting a priori preferences for the type of anaphoric relation established, how recently the time was mentioned, and how plausible the resulting temporal interpretation would be (see Section 5). These numerical heuristics contribute to some extent to the success of the implementation, but are not critical components of the model, as shown in Section 8.3.