Generation in the LOLITA System: An Engineering Approach

I ~ ~ " n / ~ ~ I ( -===~ Figure 2: Simplified portion of semantic network for the realisation example Abstract transformations are low-level transformations which act on the :semantic network immediately beforetransformations are low-level transformations which act on the :semantic network immediately before realisation. Associated with each abstract transformation are theoretical issUes on why particular normal forms are chosen, rules which allow us to move away from these normal forms and' effects the transform has on the final utterance (apart from the obvious one that variations are more natural). Abstract transformations can act on and produce variations for events with, for example, antonym verbs, copula verbs, complemented verbs, verbs which have de-lexical structures, verbs which can be either specialised or generalised and events with multiple subjects. In each case the procedure is tO disconnect arcs from particular nodes and reconnect them to different ones (which might already exist or have to be constructed in the network). Figure 1 is an example network portion showing how a verb specialisation: transform can be applied. The verb in the original sentence 'Jack wounded Jill with a gun' may be specialised: into 'shot' because firearms (of which guns are specialisa~ions) usually wound by shooting. Furthermore as 'guns' are considered to be the most c o m m o n special isat ion of firearms, the instrument clause can (assuming no special Context or high precision flag is set) be dropped leaving 'Jack shot Jill'. This abstract transformation relies on a certain amount of plausible inference: if a high precision flag has been set (according to the required style of output) then the output may be modified, e.g., 'It is likely that Jack shot Jill'. As in the LOLITA system, Jacob's KING generator [4] uses a knowledge intensive semantic network to move from normal forms of representation to alternative forms. However Jacob's method requires special entries in the semantic network (e.g., a special entry representing 'hug giving' is required to produce the delexical structure 'to