Lexical Choice in Generating Maintenance Activity Descriptions

We are currently engaged in a lexico-syntactic analysis of the verbs that occur in task order descriptions. This information will include the subcategorization frame or frames for each verb sense, with corpus-based frequency counts, the selectional restrictions on the associated verb arguments, and verb class membership in a specially constructed domain specific verb lattice. Having this amount of detailed information about the verbs will enable us to choose the appropriate lexical item when generating an English description from an instance of a parameterized action.

  
Figure: Representation to NL instructions control algorithm

Figure shows the algorithm to translate parameterized actions into natural language instructions. The lexical item chosen in Step 3 will be the one that most closely matches the action, in several different areas, including:

• the essence of the action, i.e.,
  • the movement or manipulation that is involved, i.e.,. grasping vs lifting,

  • the attribute/value modification involved, i.e., disengage vs open

  • cognition involved, i.e., note

  • sensing involved, i.e., inspect vs touch

• applicability conditions, i.e., in order to remove a gas cap, it must first be present

• post-conditions, i.e., the successful completion of a push action involves a change of location of the object.

• the number of arguments involved, i.e., presumably an agent as well as possibly one or more objects

• type constraints on those arguments, i.e., refueling necessarily involves an appropriate type of fuel for the device in question

The difficulty in choosing the lexical item will vary significantly from situation to situation. The most difficult choice occurs when we are given a sequence of PaT-Nets, with no clear grouping into individual actions. Trying to determine which PaT-Nets should be treated as coherent units is a combinatorily explosive search problem. However, if we can assume that the interface that created the PaT-Net sequences in the first place bracketed separate sequences appropriately as they were being generated, then our task becomes more feasible. There are still many different scenarios, depending on the type of action involved, and how specific its implementation is to the objects it is being applied to.

• Object oriented actions
• Generic action descriptions
• Motion verbs
• A detailed analysis of Grasp