Word Embeddings Quiz

Answers and explanations appear as you take the quiz.

1. How does a learned word embedding represent a word for numerical comparison?
   • As a plain-text dictionary entry only
   • As a vector of learned numerical features
   • As a complete recorded pronunciation only
   • As a manually assigned grammar rule
2. How do models like word2vec learn what a word means?
   • By reading hand-written definitions from a dictionary
   • By counting letters in each word
   • By observing which other words tend to appear near it in large amounts of text
   • By asking human raters to score every word
3. What does the classic king-minus-man-plus-woman vector example illustrate?
   • Some learned representations capture relationships through vector offsets
   • Every word relation obeys an exact arithmetic law
   • Vector arithmetic verifies every factual statement
   • The vector stores a word's exact spelling only
4. What is the main difference between word2vec/GloVe embeddings and the embeddings inside modern transformer models?
   • Transformers use no numbers at all
   • GloVe only works on images
   • Modern models cannot represent meaning
   • Word2vec gives one fixed vector per word, while transformers produce context-dependent vectors
5. Which task most directly relies on comparing embedding vectors?
   • Counting the total number of words in a document
   • Converting text to uppercase
   • Semantic search that finds documents with similar meaning to a query
   • Checking a file's size on disk
6. Which scale of dimensionality is found in common released classic word-vector configurations?
   • Exactly one coordinate for every model
   • Tens or hundreds of learned coordinates per word
   • Exactly two coordinates in all word2vec models
   • At least ten million coordinates per word by definition
7. In a toy word-vector comparison, what is the cosine similarity of the nonzero vectors one-zero and zero-one?
   • One
   • Minus one
   • Two
   • Zero
8. What should the first coordinate of an arbitrary learned word embedding be assumed to mean?
   • A learned feature without a universal predetermined linguistic label
   • The word's character count in every model
   • A fixed sentiment score in every dataset
   • The document's factual reliability by definition