In a corpus split between British English and American English, what could cause a fully trained embedding model to assign high cosine similarity to eggplant and aubergine despite separation of texts?

In a corpus split between British English and American English, what could cause a fully trained embedding model to assign high cosine similarity to eggplant and aubergine despite separation of texts?

• A. They never co-occur in any text.
• B. They share the same spelling adjacent to each other.
• C. They would have low cosine similarity.
• D. Yes, because they appear in similar contexts.

Answer: (D)

The main idea here is distributional semantics: word meaning is captured by the contexts in which a word appears, and thus their vector representations reflect those contexts. Eggplant and aubergine are different labels for the same thing, so they tend to appear in very similar contexts—recipes, vegetables, cooking, groceries, meals—in English texts. This happens across both British and American writing, so even if the two terms are separated into dialect-specific corpora, the surrounding words that co-occur with each term are broadly alike. The embedding model then places their vectors in similar directions, leading to a high cosine similarity. Spelling tricks or adjacency of the words themselves aren’t what drives this. Cosine similarity isn’t about matching spellings but about overlapping contexts. It’s not about direct co-occurrence in the same text, but about shared contextual patterns across the corpus. So the high similarity reflects their shared contexts rather than any direct cross-dialect text pairing.

The main idea here is distributional semantics: word meaning is captured by the contexts in which a word appears, and thus their vector representations reflect those contexts. Eggplant and aubergine are different labels for the same thing, so they tend to appear in very similar contexts—recipes, vegetables, cooking, groceries, meals—in English texts. This happens across both British and American writing, so even if the two terms are separated into dialect-specific corpora, the surrounding words that co-occur with each term are broadly alike. The embedding model then places their vectors in similar directions, leading to a high cosine similarity.

Spelling tricks or adjacency of the words themselves aren’t what drives this. Cosine similarity isn’t about matching spellings but about overlapping contexts. It’s not about direct co-occurrence in the same text, but about shared contextual patterns across the corpus. So the high similarity reflects their shared contexts rather than any direct cross-dialect text pairing.