Are word embeddings fibered categories?
Links Kan Extension Seminar II Images
This post explores the question “What structure can we extract from a word embedding?” and posits that the answer is a fibered category, or a database!
Disclaimer: Everything I know about word embeddings come from a couple of blog posts mentioned below, so this article should not be taken too seriously. On the other hand, if this post leads to an actual paper, it would be nice if you mentioned that you first saw it here :D
You shall know a word by the company it keeps.
You shall know an object (of a category) by the company it keeps.
- Corollary of the Yoneda Lemma
Methods: word2vec and GloVe
- Clustering of closely related words
- Difference vectors encode relationship information
- Can add and subtract to form analogies!
Fibered Categories and Databases
Discrete fibrations The Grothendieck construction Database schemas (ologs)
Extracting a Fibered Category
- Compute all differences
- Cluster them, and select the largest/tightest k clusters
- Form the base graph/category from this data
- Form a discrete fibration
- Labelling arrows
- Compositionality, and overcoming “drifting”
- Sparsity of representation
- A structured database from unstructured text
- Noise, need fine tuning
- Complexity of computing difference pairs
- Collision of disparate difference pairs
- Incomplete datasets: partial functions?
- Can’t handle inclusions. More generally, objects that belong under multiple headings
- Can’t identify direction of arrows
- Can we learn the category directly?
- Clustering data points gives objects (word2vec), difference vectors gives arrows (Glove). What about higher “simplices”?
- Given a category, can we compute what it is fibered over? Any measure for what constitutes a good representation?
Written on February 18, 2018