As a data scientist, I think losing actual words would be a loss. Words would be only used by word embeddings like word2vec, but actual words let you switch to better word embedding later.