Examine how well word embedding models can answer analogy questions of the form “A is to B as C is to [what]?” (e.g. “Athens is to Greece as Oslo is to Norway”) using vector arithmetic. Create your own embeddings or use an online site such as lamyiowce. github.io/word2viz/ or bionlp-www.utu.fi/wv_demo/. 

a. What analogies work well, and what ones fail? 

b. In handling Country:Capital analogies, is there a problem with ambiguous country names, like “Turkey”? 

c. Do analogies start to fail for rarer words? For example, “one is to 1 as two is to [what]?” will reliably retrieve “2,” but how well that hold for “ninety” or “thousand”? 

d. What research papers can help you understand word embedding analogies?

e. What else can you explore?