I heart some of my colleagues who are working on another latent topic model, which is different from LDA, complains that LDA like documents with similar lengths. I agree with this. But I feel that can be fixed easily. Here follows what I think.
The second term is dependent with the document length. Just consider an example document is about two topics, A, and B, and half of its words are assigned topic A, the other half are assigned topic B. So the P(z|d) distribution should have two high bins (height proportional to L/2 + alpha), and all elsewhere are short bins (height proportional to alpha). So, you see, if the document has 1000 words, alpha has trivial effect to the shape of P(z|d); but if the document contains only 2 words, alpha would have more effects on building the shape of P(z|d).
An intuitive solution to above problem is to use small alpha for short document (and vice versa). But would this break the math assumptions under LDA? No. Because this is equivalent to use different symmetric Dirichlet prior on documents with different lengths. This does not break the Dirichlet-multinomial conjugacy required by LDA’s Gibbs sampling algorithm, but just express a little more prior knowledge than using a symmetric prior for all documents. Let us set
for each document. And users need to specify parameter k as they need to specify alpha before.