It should be noted that information about repeated candidates--for example, the pronominal anaphora preferences 5, 10, and 11, or the adjectival anaphora preferences 6 and 7--is usually inserted into the preference system in order to achieve knowledge about the main entities of the dialogue. However, in this experiment, information about the main topic of the dialogue has been included and so information about repeated candidates is unnecessary. Those preferences were therefore removed, improving the results.
Furthermore, we found that pronominal anaphora preferences 3 and 4 for proper nouns caused errors. This is because in the domain of the experiment there is an exaggerated use of place names where these preferences incorrectly apply. By removing them, better results were obtained.
Finally, since the usefulness of preference 9 (candidates that are not in circumstantial adjunct) has never been justified properly, it too was omitted. After removal, the precision for pronominal anaphora stayed the same.
Thus, having considered all possible applications for ordered preference management, and given that this final set of preferences represented the minimum set of preferences, we considered it to be the optimal set. We then applied this optimal set of preferences to the training corpus, obtaining a precision of 73.8% for pronominal anaphora resolution and 78.9% for adjectival anaphora.