Computational linguistics

What is a common tradeoff of modeling text as individual characters rather than whole words?