Data

Tools

Indexes

Applications

Experiments

Awesome

An open API service indexing awesome lists of open source software.

https://github.com/patrickkidd/ci_vectors

Experiments in calculating emotion as vectors for collective intelligence
https://github.com/patrickkidd/ci_vectors

Last synced: about 1 year ago
JSON representation

Experiments in calculating emotion as vectors for collective intelligence

Awesome Lists containing this project

README 

          
# ci_vectors :: Collective Intelligence Vectors



This is an experiment in deriving collective opinions from a list of diverse

opinions on a topic. The motivation is a key theoretical principle in the

quantitative field of Collective Intelligence, that the average of a set of

diverse viewpoints on a single topic is often more accurate to reality than a

single expert. A goal here is to take a list of english-language opinions, in

this case from a single Twitter thread, and produce some aggregated summary of

the opinions from it.



See "Theoretical Background" for more.



I don't have any experience in Language Models or machine learning, and am not

that great with data science in general. I do have a doctorate in Psychology and

advanced training in Bowen Theory, which guides some of the novel propositions

to augment findings from the field of Collective Behavior.



## Experiment 1: Twitter thread replies



A script was written to scrape twitter thread replies from a safari dev tools

network tab HAR (JSON) export of all requests. Reply text are written to a json

file. This is the raw inputs for this project. A little over 200 replies can be

produced after manually scrolling to the bottom of a twitter thread to load them

all asynchronously. Twitter appears to cap the results provided in the web

interface.



The idea is that the thread's original tweet is the "topic." All replies in that

thread are assumed to be "votes" on that topic. A goal is to start with tweets

from public figures that generate many thousands of polarized replies for every

tweet, like @jordanbpeterson or @elonmusk.



_Opinion Summary_



`test_summary.py` uses the ChatGPT API to produce an

english-language summary of all of the opinions. This requires the number of

opinions to be capped at about 175.



ChatGPT thread: https://chat.openai.com/share/78b0cde6-6374-4ef3-bb52-20f860ee8e0f



_Factor Analysis_



`test_factors.py` uses `sklearn` and `nltk` to generate a

factor analysis of terms. This script has not been tested yet. The idea here is

that instead of a single "average" of the opinons, that opions might cluster

together with different weights. That information might be useful for a person

to aggregate the factors subjectively in their head to draw a single conclusion.



ChatGPT Thread: https://chat.openai.com/share/7e9ba27d-6138-4d11-a6ee-100e2c10d3a3



_Other Ideas_: 



Here are some other attempts at getting starting points from ChatGPT:



- Emotion Vectors 1: https://chat.openai.com/share/99691cdb-9d83-4a8b-b2d5-7910a6d70f42

- Emotion Vectors 2: https://chat.openai.com/share/009f720d-2da4-4fe2-90fc-cea21cb9986d

- Straight CI: https://chat.openai.com/share/f7564a4f-e287-4518-80e2-372bd0910fdc



## Theoretical Background



A current goal of Twitter is to become the "least untrue" source of information

available. The key assumption is that if individuals are allowed to express

their opinions as freely as possible on a given topic, then it will be the best

for society's ability to make sense of that topic. It turns out that there is a

technical argument for this assumption in Collective Intelligence, which

suggests that biological systems have evolved instinctual mechanisms for making

group decisions that are far superior to a decision that any one individual

could make. These intelligience mechanisms work through local interactions

between relatively uninformed individuals that aggregate into a single and far

more sophisticated decision making system.



This principle applies to baboons making pathfinding decisions, bees making

decisions on where to move a nest, schools of fish avoiding a predator in a

manner that minimizes individual losses, and others. In the case of humans, this

pertains to the automatic verbal and non-verbal cues that occur when multiple

stakeholders engage with each other on problem. This instinctual dialog, or

*dialogos* in Latin, can hypothetically occur in debates, arguments, even wars.



A new hypothetical idea is that automatic emotional responses in humans during

dialog represent "votes" on the problem. The votes are automatically tallied

through the actual behavioral outcome of the dialog. All of this is so automatic

that we are not used to observing it while it happens.



For example, consider that persons A, B, and C are trying to figure out where

hold an office Christmas party. The dialog might look like this:



- Person A: "Let's have it at the roller rink!"

- Person B: "That place is too small."

- Person A: "But they have the best activity that everyone can do."

- Person C: "But their food is terrible."

- Person B: "How about just having it at the office?"

- Person C: "I guess that could work!" (with great positivity)

- Person A: "Hmmm, Let me think about it."

- Person B: "It would be free, and everyone knows how to get there."

- Person C: "Yeah, com on Person A, let's do it there."

- Person A: "Well, I'm not totally on board but it's two against one."

- Person B: "OK, sorry but I think it makes the most sense."

- The Christmas Party ended up at the office.



In this sequence each interaction is a vote that biases the final decision in a

particular direction. For example, "Let's have it at the roller rink!" is a

clear vote for the roller rink. "That place is too small" is clear vote against

the first vote. There are more votes that push the group's bias toward or away

the roller rink. Then a new vote for the office comes, a few more votes for and

against the office, and one with great positivity that carries a bit more weight

than the others.



Finally, the actual behavioral outcome is to have the party at the office. The

actual behavioral outcome is important because it may not always be the outcome

that was explicitly agreed upon. The main idea of CI is that the mechanism for

collective intelligence has *evolved*, which means it is instinctual/automatic,

which means that it may not be the same as what people say they will do. People

do not always do what they say they will do and *instinctual* behavior can only

be inferred through actual behavior.



Similarly, the decision was not unanimous. This is important. Agreement is not

required, and it is epected that there will be some votes that are not in line

with an explicit group decision or actual behavioral outcome. This is because

the CI process is not about *fairness*, it is about getting as close to reality

as possible. The key CI principle is that each individual provides a vote or

votes that merely biases the decions in one direction or another. The "votes"

are then tallied and then the group decision is enforced through actual

behavior.



Anecdotal evidence for this process existing at all is that each step in a

dialog series has an impact on each individual's model of the problem. People

are heavily influenced by the process of deliberation in a group. Exactly *how*

this occurs is unknown and likely very complex.



However, evidence *that* it occurs can be obtained in a single experiement. Ask

each person in an experimental group for their opinion on a topic separately

prior to being told they will participate in a group deliberation on that topic.

Then hold the group deliberation, recording each statement a person makes to the

others about the topic. The prediction is that at least some of the people will

demonstrate that their opinions have changed while the deliberation is occuring,

either verbally or non-verbally. This only occurs after seeing or hearing the

responses from others.



Most people do not really own their opinons, as demonstrated in this kind of

experiment, and more formally in this experiment where the evolution of

individual opinions through group context was demonstrated:

https://www.youtube.com/watch?v=O_m8IZzl48E&feature=youtu.be. Confederates could

flip an individual opinion on a topic merely by flipping their own. The

proposition here is that individual's opinions on a problem exist and operate

primarily in conjunction with the group of affiliated stakeholders on that

problem, and that the end result is something like an average of the individual

opinions.



A hypothesis is that this applies even when groups are highly polarized on a

topic. In fact, the proposition is that the poles *function* to balance each

other out in this instinctual CI process. In other words, in an argument the

"truth is somehwere in the middle." Providing a method for producing an opinion

"somewhere in the middle" from a list of diverse opinions would benefit

collective decision making, especially on difficult topics.