For
Thursday, March 21, 2013 dev log -
To achieve an effective natural-language AI system, it usually takes teams of academics years! Assuming we value your man hours for other things too
perhaps the best compromise is to componentise each message, where the components are based on the requirements and current state, and having a library of alternative forms for each component.
"[Predicate: Safe <Ship @ 0x10F800>]" (internally, this is the kind of thing that the AI is thinking about) and make it into a description? Well, converting it into "Escort the given ship" is easy...but what about "We need you to escort our convoy to the waypoint. We've been monitoring pirate activity in the system for some time now, and don't like the odds of a lone freighter, so we're bringing on a few wingmen in case things get hot. Keep your eye on the nearby asteroid belt, as our intelligence would seem to indicate that a hidden smuggling base therein is generating a strong pirate presence in the area."
"We need you to escort our [insert asset] to [insert location]" could be one of a number of statements, effectively saying the same thing. Also "We have [asset] that needs escorted to [location],", "We want you to escort [asset] all the way to [location]", etc. All just variations on the same theme.
Similarly "We've been monitoring [insert threat] in [insert location] for some time now" as well as "and don't like the odds of [insert package to be guarded] " and "so we're bringing on [insert escort description] in case things get hot", etc. Parameterised sentence components. If there are enough subdivided components then the variability will allow the sentences to seem fairly natural. It shows, though, that the circumstances and status would have to be available the AI so as to add the parameters convincingly. I'm not sure if the hex code has any relation to that, or if it would have to be data mined.
This also highlights the potential difficulties of adding actor voices, as sentences would be comprised of separate components and it would be hard to dovetail the tonality effectively. Not impossible, though....