Which statement illustrates a static embedding's limitation with polysemy such as 'bank'?

Which statement illustrates a static embedding's limitation with polysemy such as 'bank'?

• A. They would assign the same vector to 'bank' in 'river bank' and 'bank account'.
• B. They would distinguish the senses automatically.
• C. They generate different vectors for different senses.
• D. They require context to be disambiguated.

Answer: (A)

Static embeddings assign one fixed vector to each word form, regardless of which sense the word carries. For a polysemous word like “bank,” the same vector would represent both the river edge sense and the financial institution sense. That means the representation itself doesn’t distinguish between meanings, so the surrounding context has to carry the disambiguating signal or the downstream model has to work harder to infer which sense is intended. This is exactly what the statement illustrates: the same embedding is used in both “river bank” and “bank account,” highlighting the limitation of static embeddings in handling multiple senses. The other options describe capabilities that static embeddings don’t have—automatic sense distinction, generating different vectors for different senses, or requiring context to disambiguate—so they don’t fit as well as the described limitation.

Static embeddings assign one fixed vector to each word form, regardless of which sense the word carries. For a polysemous word like “bank,” the same vector would represent both the river edge sense and the financial institution sense. That means the representation itself doesn’t distinguish between meanings, so the surrounding context has to carry the disambiguating signal or the downstream model has to work harder to infer which sense is intended. This is exactly what the statement illustrates: the same embedding is used in both “river bank” and “bank account,” highlighting the limitation of static embeddings in handling multiple senses. The other options describe capabilities that static embeddings don’t have—automatic sense distinction, generating different vectors for different senses, or requiring context to disambiguate—so they don’t fit as well as the described limitation.