A paper was recently released by a Canadian A.I. research team. The paper developed language models with the intent to create techniques and improve existing ones in order to generate engaging dialogue. It used this technology for improv theater. (At the end of the paper, there are transcripts of the conversation between professional Canadian improvisers and the A.I.—they are worth the read. Note: Canadians have been entertaining Americans with their improv talent for years. Many of the Saturday Night Live cast is from our neighbors to the north. Also, they have a strong A.I. presence, and look at this guy - a charmer.)
The Inner Guts of the A.I. Technology
How do they track conversation? They break each statement into an “utterance”. They classify each utterance into a universe. Then they quantify how likely an utterance is to appear as a category in a universe. If an utterance is more likely to appear in a specific category, it is considered to be informative. If it is unlikely to exist in a certain category, it is less informative. Then, in sequences of utterances, where one statement is made after another, a whole classification can be run on the set of utterances. An utterance may guide the overall classification of the dialogue toward a specific category, or an utterance may ambiguate the category classifications.
To clarify the dialogue classification, an utterance may be informative and entropy decreasing. To ambiguate the dialogue classification, an utterance may be uninformative (information concealing) and entropy decreasing. Thus, technically, those who cause ambiguity in conversations are creators of chaos.
That’s the basic idea. Here’s an example:
Every sentence can be classified and informs a listener, to some degree, what category the speaker is talking about. If I said, “The room is full of people.” It gets a 10% chance of talking about conference rooms and the remaining 90% is distributed among various other categories. If I say, “The room is full of people and the speaker in the front is telling everybody to take their seats, turn off their phones, and wrap up their conversations.” There is a 90% chance the sentence is about conference rooms.
The paper takes that idea and builds on it. It says that when the sentences are pieced together, statement after statement, the overall probability distribution of the system adapts at each contribution of an utterance. The collection of utterances, and its overall probability distribution, are gathered together and defined as the “Narrative Arc”. Given this larger container “Narrative Arc”, a strategy can be devised to intentionally insert sentences that increase the probability a narrative is in a certain universe or to decrease the probability a narrative is in a certain universe.
This diagram is from the paper. It is the dialogue from the first 20 lines in Shakespeare’s Romeo and Juliet. It shows how a single universe can become apparent after 20 lines of dialogue. The entropy in the system gets smaller and smaller over time.
Ambiguating the Trump Universe
Given these concepts, this is how to ambiguate the Trump universe.
There are two methods for tweaking the model. (If this were algebra, this “tweaking” of the model is similar to manipulating a linear equation to make linear or horizontal shifts):
In this ML, Boltzmann classification system, this is how we alter the model to get what we want. Change the classifications (a structural shift), alter the utterances themselves (an entropic shift).
A Structural Shift
Create a new set of classifications where the word “Trump” can be spoken, but it is not classified into the typical Trump universe. The word is used, but takes on a meaning that is totally different. Like, “Congresswoman so-and-so got Trumped!” “The Titanic went for the trump.” “So-and-so’s career took a ride on the Trump-train.” It’s more-or-less the same word, but gets classified into different sets of categories. In the new set of categories, a Trump Universe doesn’t even exist. The Trump universe has been swallowed by something larger. The word is among a set of “news” categories or “dramatic life stories” categories. A structural shift like this fits into what people widely understand as “reframing” the problem.
An Entropic Shift
Another method, is in the distribution of spoken utterances themselves. If the Trump Universe is the most likely classification, then, to minimize its prominence, an utterance must be spoken to ambiguate the Universe classification. (In other words; To increase entropy in the system, or to create a more uniform probability distribution for potential outcomes.)
If the probability distribution from subsequent utterances moves towards a more likely Trump classification (the system has an increased chance of being “Trump”), then it is necessary to respond with an utterance to ambiguate probabilities among any of the classifications. When a collection of utterances appear to specify a Trump universe (a task fulfilled by the Trump Campaign or affiliates), an equal or greater number of ambiguating utterances are required to shift the Classification from Trump to...someone else.
Are you listening, Cambridge Analyticas of the 2020 Election? You, too, reporting agencies, CNN, Fox News? (It’s not like none of these parties exist, but we can understand their operation and report accordingly.) This is the marketing strategy and the model that needs to be made to track the 2020 election (as far as speech is concerned). It is also likely that whoever is distributing large volumes of speech utterances to push the classification model in one direction or another, is a supporter of one electoral party. Their strategy can be two-fold (which it was in 2016). The two strategic moves are: i) Based on a person’s likely voting behavior, ambiguate all other candidate universes or ii) Specify a single candidate universe. Each move is presented to the appropriate audience. If the owners of all these accounts are owned by a single party, single group, possibly Russian...then we may have identified influential parties.