Unsupervised learning derives patterns or representations directly from data without human-assigned target labels. Word2Vec illustrates the idea with raw text: nearby words provide dataset-generated training signals, while the learned vector captures word meaning. section 6.3
Word-embedding training makes the approach concrete:
- Gather raw text from a domain such as medical journals, Wikipedia, or news articles. section 6.3.2
- Construct many local context examples with continuous bag-of-words or continuous skip-gram training. section 6.3.2
- Use prediction as a means to build vectors whose geometry captures semantic relationships. section 6.3
Word2Vec from surrounding words
A corpus mentioning Marie Curie near science-related words, the Timbers near soccer terms, and Portland near city terms supplies enough context for Word2Vec to learn useful associations without manually labeling those meanings. section 6.3
Listing 6.1 Computing nessvector
>>> from nessvec.examples.ch06.nessvectors import * #1
>>> nessvector('Marie_Curie').round(2)
placeness -0.46
peopleness 0.35 #2
animalness 0.17
conceptness -0.32
femaleness 0.26
Inspecting a learned vector for Marie Curie makes the abstract representation concrete through interpretable dimensions such as placeness, peopleness, and femaleness. section 6.3
