Natural Language Processing - Week 4

Lecture : Machine Translation

Transforming word vectors

• Minimize the distance

• 프로베니우스 놈(Frobenius norm) : L2 Norm;Euclidean Norm을 matrix로 확장

• R is mapping matrix

K-nearest neighbors

Summary

• K-nearest neighbors, for closest matches
• Hash tables

Hash tables and hash functions

• if applying to words

• bucket of remainder

• based on where they're located in vector space

Locality sensitive hashing

• the blue plane and grey plane looks like bucket

• when dot product is
  • positive : the vector is on one side of the plane
  • zero : the vector is on the plane
  • negative : the vector is on the opposite side of the plane

Multiple Planes

• Multiple planes -> Dot products -> Hash values

• locality sensitive hashing

Approximate nearest neighbors

Searching documents

• 3 dimensions document vector