import { Grid, ThemeProvider, Typography } from "@mui/material"; import { buildTheme } from "../../infrastructure/theme/theme"; export default function Definition() { return ( <> This is a simple web application that creates word embeddings using the gensim model. Word embeddings are numerical representations of words that capture their semantic meanings. Each of the 50 dimensions in a GloVe model represents a latent semantic attribute, and even tough they are not directly interpretable, with som sort of reverse engineering the embedding space the latent semantic attributes can be inferred. Two main approaches can be used to achieve this:
  1. Investigating Common Category Attributes Investigating Common Category Attributes: By analyzing dimensions with the lowest variance among a group of semantically similar words.The premise is that dimensions with minimal variance may be capturing attributes that are common across the set. For instance:
    • Words: 'Spain', 'France', 'Germany', 'Japan' (all countries)
    • Low Variance Dimensions: These would theoretically indicate attributes common to all countries, potentially abstract notions like 'sovereignty', 'nationhood', or just the general category of being a 'country'.
  2. Investigating Specific Semantic Differences: By analyzing the high variance dimensions resulting from subtracting one vector from another. This subtraction aims to capture the core semantic differences between two entities. It can become more robust if more than one pair of words is selected to compare. For instance:
    • Words pairs: 'Man' and 'Woman', 'Uncle' and 'Aunt', 'Father' and 'Mother'.
    • Difference Vector (High Variance Dimensions) : These dimensions likely highlight aspects related to gender differences. The highest values in this vector suggest dimensions where the concept of 'man' and 'woman' differ most significantly, potentially capturing gender-specific traits or roles.
); }