Which metric is used to measure similarity between word vectors in many embeddings?

Which metric is used to measure similarity between word vectors in many embeddings?

• A. Cosine similarity
• B. Euclidean distance
• C. Jaccard similarity
• D. Hamming distance

Answer: (A)

The main idea is that we compare word vectors by how their directions align, not how long they are. In embedding spaces, words with similar meanings tend to point in similar directions from the origin, so a measure that focuses on orientation captures semantic similarity better than one that cares about magnitude. Cosine similarity does exactly that: it looks at the angle between two vectors by taking their dot product and dividing by the product of their lengths. When vectors point in the same direction, the cosine similarity is high (close to 1); when they are orthogonal, it’s around 0; if they point in opposite directions, it’s near -1. This makes it robust to differences in vector length and often aligns well with human notions of similarity. Why other metrics don’t fit as well here: Euclidean distance combines both direction and magnitude, so two vectors that are far in length but in the same direction can be deemed quite dissimilar, which isn’t desirable for semantic similarity. Jaccard similarity is designed for comparing sets, not continuous-valued vectors. Hamming distance applies to fixed-length discrete strings, not real-valued embeddings. In practice, you often normalize word vectors to unit length, turning cosine similarity into a simple dot product, which makes the computation even more convenient and interpretable.

The main idea is that we compare word vectors by how their directions align, not how long they are. In embedding spaces, words with similar meanings tend to point in similar directions from the origin, so a measure that focuses on orientation captures semantic similarity better than one that cares about magnitude. Cosine similarity does exactly that: it looks at the angle between two vectors by taking their dot product and dividing by the product of their lengths. When vectors point in the same direction, the cosine similarity is high (close to 1); when they are orthogonal, it’s around 0; if they point in opposite directions, it’s near -1. This makes it robust to differences in vector length and often aligns well with human notions of similarity.

Why other metrics don’t fit as well here: Euclidean distance combines both direction and magnitude, so two vectors that are far in length but in the same direction can be deemed quite dissimilar, which isn’t desirable for semantic similarity. Jaccard similarity is designed for comparing sets, not continuous-valued vectors. Hamming distance applies to fixed-length discrete strings, not real-valued embeddings. In practice, you often normalize word vectors to unit length, turning cosine similarity into a simple dot product, which makes the computation even more convenient and interpretable.