I trained word embeddings with 300 dimensions. Now, I would like to have word embeddings with 50 dimensions: is it better to retrain the word embeddings with 50 dimensions, or can I use some dimensionality reduction method to scale the word embeddings with 300 dimensions down to 50 dimensions?
t-distributed stochastic neighbor embedding (t-SNE) is often used for dimensionality reduction in word embeddings. t-SNE maintains the relative relationships between the vectors.
Most often t-SNE is used for visualization, thus reducing the dimensions to 2 or 3. It could also reduce the dimensions down to 50.