How is a co-occurrence matrix constructed?

How is a co-occurrence matrix constructed?

• A. Give each word a row and a column and count how often it occurs next to another word, filling the table.
• B. Compute the frequency of each word across the corpus only.
• C. Apply a transformer to map words to context vectors.
• D. Represent words by a one-hot vector only.

Answer: (A)

A co-occurrence matrix is built by capturing word relationships through counts of proximity. You set up a square table where every word in the vocabulary appears as both a row label and a column label. Then, as you scan the text, you look at each word and count how often it appears near another word within a defined context window (for example, the immediate neighbor or a small neighborhood). Each time word i appears near word j, you increment the cell at the intersection of i’s row and j’s column. After processing the whole corpus, that cell contains the co-occurrence count of i and j. This setup directly encodes how often words appear together, which is the essence of a co-occurrence matrix. The other options don’t fit: counting a word’s overall frequency ignores pairwise context, generating contextual vectors with a transformer describes a different process, and using only one-hot vectors describes a basic representation rather than a table of pairwise counts.

A co-occurrence matrix is built by capturing word relationships through counts of proximity. You set up a square table where every word in the vocabulary appears as both a row label and a column label. Then, as you scan the text, you look at each word and count how often it appears near another word within a defined context window (for example, the immediate neighbor or a small neighborhood). Each time word i appears near word j, you increment the cell at the intersection of i’s row and j’s column. After processing the whole corpus, that cell contains the co-occurrence count of i and j. This setup directly encodes how often words appear together, which is the essence of a co-occurrence matrix. The other options don’t fit: counting a word’s overall frequency ignores pairwise context, generating contextual vectors with a transformer describes a different process, and using only one-hot vectors describes a basic representation rather than a table of pairwise counts.