The Agency Pattern
Dialog-based NLP, able to also parse web pages and local files, provides a shared
front end for multiple Intelligent Agents, so we call it the Agency
design pattern.
With the "stack" of principal modules recommended below, one can start now to deploy intelligent
agent applications that function, then mature gracefully within an enterprise by expanding
later in their bolded elements:
User Interface Modules
(phased)
| MODELER's UI uses shared NLP code to generate an XTM
chart
for each input paragraph, exposing its topics
and whatever was said or asked of them. In phase I, users hand-type phrase-like WORDS scripts to build
and test semantic lexicons. In phase II, grammar rules in an English parser transliterate text
into similar scripts on the fly. In phase III, users drive
those same grammar rules by using continuous speech, not text.
|
WORDS
Expressions |
Passed as ASCII requests to a simple web app interpreter, these simple scripts can be
quickly interpreted to update or query the CONTEXT element cited below - a Topic Map
engine pre-configured with semantic lexical data that collectively forms an evolving,
domain-specific knowledge base. You can skip this whole layer and simply use Java or
C# as the scripting language, but then you have to write your
Ontology partly in that notation. Better is a distinct, high level format which hides,
specs, and sugar-coats the details of almost everything below. |
Semantic
Lexicon
Constraints |
The scripted topics in semantic lexicons, if loaded into CONTEXT before a dialog begins, can
configure MODELER to expect senses, assume default topic characteristics,
help interpret anaphora, and constrain parsing to meet semantic
checks. Such domain-specific vocabulary support is a key MODELER service that cuts
lexical ambiguity in later inputs. |
|
Contextual Topic Map
|
XTM from the UI above and reasoning modules below merges into this
topic map and
becomes persistent. Its model of discourse tracks topics of
interest, what was said or asked about them, when, and
by whom. Context is a key data base for your agent, which uses TM paradigms to
expose propositions embedded in input text. Further processing
may involve many modules, including both types below. |
|
Embedded WORDS Scripts
|
Scripts linking Context to J2EE applications
can react to each chart by triggering events or calling remote web
services, back-end enterprise applications, or their own desktop UI code. Such
reactions may help manage reasoning modules or replace them, especially in early stage Agency applications whose designs need to stay flexible.
|
Optional
Reasoning Modules |
Some J2EE applications can be (or use) RDF-OWL (or DAML, KIF, etc.) lets
an inference engine to track inputs and infer problems or answer queries by using description logic. The interface problems may be
severe outside the scope of fairly API narrow limits. High-level scripting commands can well encapsulate these, so it makes sense to publish expression
types that are designed especially for this purpose.
eithem as invetuse them
resuse the ac s . oesm
answers to any questions contained in the original chart;
or apply rules to expand it with data representing its consequences or
its rejection; or signal MODELER to load Context with new lexicons of
available senses or call predefined scripts. |
WORDS scripts
and your topic map engine provide your agents the I/O, persistent
storage, and standard interfaces demanded
by real business applications. They can also link them easily to
existing utility libraries, enterprise software, and in phase II, to
the huge base of text residing in English documents and web pages
Eager to create smart agents?
Wait no more, as what you want can be coded soon using MODELER's scripts and data inheritance, then optionally
enhanced as new lexicons and reasoning modules start becoming available.
|
