I have read on papers that Latent Dirichlet Allocation (LDA) works by identifying word cooccurances in documents. What is confusing me is since LDA uses bag-of-words approach for document representation word cooccurance information is lost. So what does it mean by LDA works by identifying word cooccurances to form topics?
Bag-of-words means that order of words is not important. Co-occurrence does not take into account the order - so there is no contradiction.