PV-DBOW is similar to the skip-gram model, where we try to predict the context words based on the target word:

Unlike previous methods, here we do not try to predict the next words. Instead, we use a paragraph vector to classify the words in the document. But how do they work? We train the model to understand whether the word belongs to a paragraph or not. We sample some set of words and then feed it to a classifier, which tells us whether the words belong to a particular paragraph or not, and in such a way we learn the paragraph vector.