Extract emotional information from embeddings.

When working with LLMs, various embedding models capture emotional information that might be useful to work with (or without!).

An emopoint is a simplified embedding with interpretable dimensions:

1. joy vs sadness
2. anger vs fear
3. disgust vs surprise

So, for example OpenAI's text-embedding-3-small returns embeddings with 1536 dimensions. This library will convert those into 3 dimensions, losing most information except for what directly relates to emotion.

This library enables two modes:

1. Isolate emotion, converting it into 3D emopoint vectors
2. Remove emotion, stay in original dimensionality

Install

Install using your language's package manager:

JavaScript/TypeScript via NPM

npm i emopoint

and then use it

const { MODELS } = require('emopoint');

console.log(MODELS.ADA_2);

Python via PyPi

pip install emopoint

and then use it

from emopoint import MODELS

embedding = get_embeddings("James was maaaaaad")
emopoint = MODELS.ADA_3_SMALL.emb_to_emo(embedding)

Go

go get github.com/tkellogg/emopoint/go/emopoint

and then use it

import (
	emo "github.com/tkellogg/emopoint/go/emopoint"
)

func main() {
	var embeding []float32 = getEmbeddings("James was maaaaaad")
	var emopoint []float32 = emo.ADA_3_SMALL.EmbeddingToEmopoint(embedding)
}

Functions

All 3 languages have these capabilities:

• Convert embedding to emopoint — Convert an embedding (e.g. 1536 dimensions for text-embedding-3-small) to 3-dimensional space, called emopoint space that represents only emotion and nothing else.
• Remove emotion — Take an embedding and keep it in the same dimensionality, but subtract emotional information

From these operations, there's a lot more you can do:

• Get the portion of emotional information in text — Calculate the magnitude of the embedding (should be always 1.0) and subtract the magnitude of the result of remove_emotion(embedding). The result is a scalar float that represents the portion of the meaning of the text that was dedicated to emotion, as the embedding model understood it.
• Cluster on emotion — Convert to emopoint space and run a K-Means clustering algorithm
• Semantic search on emotion only — Convert to emopoint space and store in a vector database. This matches text based only on the emotional content, ignoring all factual and subjective information.
• Semantic search without emotion — Same as before, but store the result of remove_emotion(embedding). This removes noise introduced by emotion, creating closer matches and potentially enhancing the search accuracy.
• Analytics & visualizations on emotional magnitude — Calculate the magnitudes of emopoints for several texts, e.g. sections of a speech or tweets, and create visualizations on just the magnitude (portion of information dedicated to emotion).
• Analytics & visualizations on emotions — Same as before, but instead of calculating the magnitude, visualize the points in 3D emopoint space. Observe how some texts lean toward anger or joy. Analyze how emotions ebb & flow throughout a speech, and contrast that to the informational content (maybe use K-Means clustering on original content to classify the content and display those classifications as colors in a 3D scatter plot).