Distributional Hypothesis

How can machines understand that 'king' and 'monarch' are similar, or that 'bank' has multiple meanings, without human intervention? This challenge is addressed by the Distributional Hypothesis, which posits that words appearing in similar contexts tend to have similar meanings. As J.R. Firth famously stated, "You shall know a word by the company it keeps."

Consider the word "bank." If you read "The bank of the river was overgrown," you immediately understand it refers to land beside water. However, if the sentence is "I deposited money at the bank," the meaning shifts to a financial institution. The words surrounding "bank"—"river" versus "deposited money"—provide the crucial clues that differentiate its meaning. This simple observation forms the bedrock of how machines learn word semantics.

Distributional Hypothesis
The principle that words appearing in similar linguistic contexts tend to have similar semantic meanings. This implies that the meaning of a word can be inferred from the words it frequently co-occurs with.
Example: The words 'cat' and 'feline' often appear near words like 'purr,' 'whiskers,' and 'pet,' indicating their semantic similarity. Conversely, 'cat' and 'car' rarely share such contexts.

Defining Context for Words

For the Distributional Hypothesis to be useful, we need a concrete way to define a word's "context." The most common approach involves a fixed-size window around a target word. This window includes a certain number of words immediately preceding and following the target word. For instance, a window size of two would capture two words before and two words after the target. More advanced methods can also consider syntactic dependencies, analyzing grammatical relationships between words rather than just their proximity.

Extracting Context with a Fixed Window
SentenceTarget WordContext (Window Size 2)
Example 1The cat sat on the mat.cat
Example 2She bought a new car yesterday.bought
Example 3A quick brown fox jumps over the lazy dog.quick
This table illustrates how a fixed window of two words before and after the target word defines its context.
Check Your Understanding
If two words consistently appear with the same surrounding words, what does the Distributional Hypothesis suggest about their meanings?

From Context to Word Embeddings

The Distributional Hypothesis is the foundational principle behind modern word embeddings, which are dense vector representations of words. Algorithms like Word2Vec and GloVe analyze vast amounts of text to learn the contexts in which words appear. They then map each word to a point in a high-dimensional vector space, where the distance and direction between vectors capture semantic relationships. Words that share similar contexts, and thus similar meanings, will be positioned close to each other in this vector space.

Semantic Clustering in a Contextual Space
This scatter plot visualizes how words with similar meanings, derived from their shared contexts, cluster together in an abstract two-dimensional space. Words like 'king' and 'queen' are close, as are 'apple' and 'banana', reflecting their semantic relationships.
Loading chart...
Key Insight: Words with similar meanings, inferred from their shared linguistic contexts, naturally cluster together in a vector space, forming distinct semantic groups.
Key Takeaways
  • The Distributional Hypothesis states that words appearing in similar contexts tend to have similar meanings.

  • The meaning of a word is not inherent but is derived from the company it keeps within text.

  • Context is typically defined by a fixed-size window of surrounding words, though more complex methods exist.

  • This hypothesis is the fundamental principle enabling the creation of word embeddings, where semantically similar words are represented by vectors that are close in a high-dimensional space.

  • Machines can understand word similarity and polysemy (multiple meanings) by analyzing and comparing the contexts in which words appear, without explicit human labeling.

← All lessons in Word Embeddings

Ready to keep this from fading?

Bitelrn turns lessons like this into a full course — quizzes, a knowledge map, and spaced review.

Get started free