Each lexicon entry sets up an option for realizing a part of the input semantics. To these we also add options set up by microtheories, which we will not discuss here due to space limitations. Two important functions of the text generation system are to determine how well each of the options fit given the input semantics, and how each of the options can potentially interact with other options. The Mikrokosmos semantic matcher performs these functions.
Figure 6: Example Semantic Representation
Figure 7: Three entries for ACQUIRE
[Beale and Viegas1996] discuss this process in more detail; here we give a simple example to highlight some of the issues. Consider the input semantic representation in Figure 6. Assume there are three lexicon entries that ``match'' the concept ACQUIRE as shown in Figure 7. The first of those entries produces a fairly generic word ``acquire'' and expects to match only an AGENT and THEME. The second entry is more specific, producing the English verb ``procure'' along with a prepositional phrase. This prepositional phrase realizes the INSTRUMENT of the event. The final entry produces the verb ``buy'' and overtly realizes the PRICE as a prepositional phrase.
Figure 8: An entry for INSTRUMENT
It is easy to see that the third entry does not fit the input semantics, as no PRICE relation is mentioned. This entry will be severely penalized for this over-generation. What about the first two entries? If we select the first entry, the INSTRUMENT relation in the input semantics will not be realized and will need to be planned separately by looking up the INSTRUMENT relation in the lexicon and retrieving something like Figure 8. On the other hand, if the second ACQUIRE entry is used, no such additional plan is needed.
In general, then, the semantic matcher keeps track of which lexicon entries cover which parts of the input semantics, which require other plans to be used with it, and which have some sort of semantic mismatch with the input semantics. The following sums up the types of mismatches that might be present, each of which receives a different penalty:
More generally, the semantic matcher sets up the basic mechanism to also deal with language gaps such as lexical divergences or semantic mismatches, [Viegas1997].