nlp_final_demo

I analyzed the underlying tree structures of word2vec embeddings and compared it to wordnet trees to understand how much of hypernym/hyponym information they encode

Word2Vec Trees

• Use the embeddings to find cosine distance between any 2 words
• These these distances as weigths between two nodes
• Construct a minimum spanning tree using Kruskal's algorithm

Findings

• The precision, recall and f of edges predicted are not very impressive (0.29 for ~1000 trees)
• However, the degree centrality estimations are reliable (the mean difference ~0.08 for ~1000 trees)

Meta

• The ipynb file should help you reproduce the results. You need to download Google's word2vec embeddings