In this subsection I will outline the realization of the item sharing method. Assume that the uniform algorithm is in the parsing mode. Then in each case a passive item is computed we automatically make available this item also for the generation mode. Thus, for example, if we are going to generate from the semantics of the parsed input we directly can return the previously computed answer during parsing as result of the generation mode (i.e., if we only consider one paraphrase). More over, if we perform generation using a different semantics as the ``parsed'' one, but which is identical with respect to some partial semantics structures (e.g., some arguments are semantically identical), then the generator also can ``reuse'' the results determined through parsing. Clearly, this kind of processing makes only sense if during parsing and generation the same grammar is used.
The restriction of sharing only passive items is plausible for the following reasons. Note that passive items have no selected element, and the value of the IN and FROM slots are the same. Assume we are in the parsing mode. Then, by means of the definition of item sets (see section 4.9), the appropriate values for the IN and FROM slot for the direction mode can directly be determined on the basis of the semantics information of the passive item. This guarantees that shared passive items produced during the parsing mode, are at their right places when they are used by the generation mode.
On the other hand, for active items, in general the chosen selected elements during parsing and generation will be different, and the essential argument of the other direction will be un-instantiated. Therefore, it would not be possible to determine the right place of an shared active item as it is the case for shared passive items.
On the basis of these observations, the structure of an item sharing approach using the uniform tabular algorithm is as follows: We assume that the uniform algorithm maintains two different agendas, one for the parsing mode and one for generation. This is no overhead, because it allows us to order the tasks of an agenda using, for instance, different preferences. Since items sets are considered as equivalence classes, that are determined on the basis of the value of the essential argument, we assume that parsing and generation have different item sets. Item sets consist of active and passive items. Now, we require that passive items are shared between the item sets determined during parsing and generation. This means, that the parser and generator each have there own private active items but can operate on the same set of passive items. Figure 4.15 illustrates the structure of the item sharing approach.
Figure 4.15: The item sharing approach using the uniform tabular
algorithm. During the different modes the uniform algorithm maintains
different agendas and private active items for the different modes.
However, passive items are shared by both directions.