Make EmbedText faster

[kevinmartinjos]

Right now when creating an embedding matrix random embeddings are used to OOV words. zerounk would set oov words as zero. Both are extremes.

Consequences:

1. Slow embedding creation
2. Noisly signals due to the cosine similarity between two randomly initialized oov word embeddings

Solution : For OOV words,

1. Set similarity as 1 if there's an exact match (unlike zerounk, which always set it as 0)
2. Set similarity as 0 if it's not an exact match
3. Avoid building stoi - the pymagnitude embedding already has a vocabulary. Simply add OOV terms to this vocab

[andrewyates]

Here's a SimilarityMatrix that handles this.

The idea is that OOV terms get negative indices, padding=0, and in-vocab terms get positive indices.