Which tool is mainly used to create custom word embeddings using n-grams?

The tool primarily utilized for creating custom word embeddings using n-grams is fastText. This is because fastText extends the concept of word embeddings by allowing for the representation of words as bags of character n-grams. This capability enables it to generate embeddings for words that may not exist in the training dataset, thus improving its performance on rare words and out-of-vocabulary terms.

By constructing embeddings based on n-grams, fastText captures subword information, which makes it particularly effective for languages with rich morphology or for domains with less frequent word occurrences. This subword modeling also helps with semantic similarity, as variations of a word can share embeddings derived from their common subword components.

The other options do have their own purposes but do not focus on n-grams in the same way. Word2vec, for instance, creates embeddings based solely on the words in a corpus without considering subword structures. Doc2vec generates embeddings for entire documents rather than individual words, and the Bag-of-words model represents text data in a way that discards the order of words, focusing only on word frequencies without capturing any contextual relationships.