Digital Signal Processing Reference
In-Depth Information
Each time the system receives input from a speaker, natural language
processing (as introduced in Sec. 2) is applied to understand the intention and
semantics of this speaker. We used a data structure,
semantic frame SF,
to
record this information, which is defined below (assuming there are
n
speakers):
For speaker
i
, represents the domain that speaker mentioned; is the
primary attribute for this domain, i.e., the purpose of the query; and is the
secondary attributes which specifies additional information needed for this
query. Note that the number of secondary attributes varies with domain. Take
the inquiry “please show me the route to the train station” as example, the
semantic frame will be:
The semantic frame of the current dialogue turn is combined with the
previous ones to determine if the goal completed. This determination is based
on whether essential information needed for a specific query is enough. For
example, if the speaker is querying the weather forecast, the essential
information would be the location (ex. city name), the weather type (ex.
temperature or rainfall density), and the time (ex. tomorrow or this afternoon).
Once a goal is completed, the system may perform database queries and
generate a proper response to the speaker. If some essential information is
missing or the speaker interactions are in conflict, further confirmation and a
repair processes should be undertaken to realize the final intention of the
speakers.
In examples 1,2, and 3, we illustrate the cases of 1) speakers who have
independent individual goals, which can be solved easily; 2) speakers with
conflicting individual goals, in which the system must resolve this problem
before further information can be relayed to the speakers; and 3) speakers
who have a common goal, which requires that they provide the necessary
information for the system in a mutually cooperative fashion.
