In recent linguistic literature, efforts have been made to classify verbs in terms of their semantic properties (Levin 1993, St. Dizier 1995, among others). The goals of that research have been to identify semantic factors which influence and correlate with syntactic behavior. This has resulted in the identification of useful components of meaning, which are on the one hand linguistically encoded in structures such as the lexical entries of verbs or grammatical constructions, but have on the other hand great relevance for the representation of actions in the world. These components include:
In attempting to generate an unambiguous, detailed, English description from an animated simulation, the identification of such meaning components, as in the example below, provides crucial constraints that can significantly reduce the complexity of the task. Verbs can be represented in a lattice that allows semantically similar verbs, such as all motion verbs, to be closely associated with each other, and with a higher level node that captures the properties these verbs have in common. Entire clusters of verbs will correspond to the primitive actions such as MOVE and LOCOMOTE, GRASP and REACH. These generalizations combined with the unique properties of the verbs that allow them to be distinguished from each other are the very properties that are relevant to lexical choice.
A coarse-grained grouping of F-16 verbs can be made by dividing them into verbs that involve a change of state and verbs that do not. Verbs that do not necessarily change the state of an object include those requiring the Agent to check the status of the object using cognitive and sensory capabilities ( verify, inspect).
Change-of-state verbs can be subdivided into simple changes of state involving one object, such as rotate or tighten or into more complex changes of state involving symmetrical changes to more than one object, such as attach or replace. Part of the categorization of these verbs can simply include the number of arguments involved, immediately ruling out several potential choices for a particular action.
For example, the task order description for attach always involves two objects and an agent, which is the implicit you of an imperative ("Attach cap chain hook to receptacle"), and can be categorized as a CONTACT verb.
attach(L-agent, L-object1, to(L-object2)))
(agent = L-agent;
object1 = L-object1,
object2 = L-object2;
spatiotemporal =
(goal = (execution-type = establish;
relation = contact(object1, object2))))
contact(X,Y) <- locpt(X, X1),
locpt(Y, Y1),
at(X1, Y1).
The representation for connect, another CONTACT verb, is similar to that for attach except that an additional intermediary may be involved, requiring a transitive contact relationship [Pal90] ("Connect static bond cable between fuel truck and aircraft"). In generation, the inclusion of the intermediary in a CONTACT relationship would result in the choice of connect rather than attach as the lexical item.
connect(L-agent, L-object3, between(L-object1, L-object2))
(agent = L-agent;
object1 = L-object1,
object2 = L-object2,
object3 = L-object3;
spatiotemporal =
(goal = (execution-type = establish;
relation = contact(object1, object3),
contact(object3, object2))))
A more fine-grained classification of verbs can be made by focusing on the specification of the post-condition. Does the change of state involve a putting action, where an entity is directed towards a particular location, either with a specific instrument or motion as in funnel or push, or involving a specific type of entity (a liquid or mass noun) as in spill?