Scenario comprehension

 

Since a narrative text is intended to convey the occurrence of a potential event, scenario comprehension is needed. As outlined in Figure 35, this supertask identifies the agents, objects, states, and actions at play in each scene of a story. This is a straightforward capability of the system as each concept in the memory of the reasoner is tagged with what high-level ontological category it belongs to. There is no ambiguity at this point; possible ambiguities with respect to ontological category is eliminated by the sentence processing supertask.

  

Figure 35: Scenario comprehension supertask breakdown

After the initial identification is complete, there are two primary tasks for which this supertask is responsible. First, the supertask maintains a model of each agent which is acting in the story. This acts as a complete history of what the agent has done and what description is currently attached to it. It is implemented by maintaining a distinct concept each time the agent changes in some way. These are linked together in a temporal fashion, with the agent concept being the current one. However, prior states of the agent are retrievable. Suppose, for example, an agent exists named John who begins the scenario at age twelve. A concept will be created which represents the twelve-year old John. As new information is added, the old John description is maintained. If the reasoner discovers that John is tall, for example, then a new concept is created which represents the fact that John is a tall, twelve-year old child. But, the original twelve-year old John concept still exists with a pointer between the two, as well as information concerning when the knowledge was changed. If the reasoner ever requests the John concept, then the current concept is returned; however, the reasoner can also pose questions like ``How old was John last August?'' and trace back through the conceptual representation to discover the answer.

Besides maintaining the agent information, this supertask is responsible for creating and updating the view of the actions which occur in the scenario. Each action which is discovered is analyzed in order to discover how it fits into the action sequence being maintained. Some actions will simply be added in a straightforward temporal fashion; e.g., one action is known to occur after another. Other actions will be linked via causal relations; e.g., an action is known to produce a state which enables another action to occur. The causal link between the two actions would be stored as well. Finally, some actions act as triggers for the retrieval of higher-order packets, thereby allowing the hierarchical representation described in Section 5.3.3.3 to be built.

As the supertasks are related, the information which is discovered by the scenario comprehension one is available to the others. One obvious way this occurs is the direct link between the plot of the story structure representation and the events captured by the scenario representation. Other more subtle connections exist between the agents which exist in the scenario and the characters which exist in the story structure. The story structure packets which exist in the current implementation all contain the fact that these two elements are often synonymous. Therefore, the discovery of agents by this supertask can aid the identification of characters in the story structure comprehension supertask.