OK — I fixed them. Luckily I put in alt text for this. I used to be using a LaTeX rendering service, but it’s no longer in business so the early LaTeX links are dead. Now WordPress supports LaTeX directly.

]]>(And it might be useful to me, as a practitioner, arguing with a coworker about normalizing for document length with naive bayes models).

]]>Thanks! Table 4 (page 7) of their paper has the full definition, and the one in the blog text above is wrong. If we let be the category of document , that’d be

I think there’s a remaining inconsistency with their general prior (smoothing) term and normalization; just read their defs!

]]>The equation I wrote reads: sum of all documents d whose class c’ is different from c. To make it more explicit, you may want to use y_d (the class of d) instead of c’. See Table 4 in Rennie et al.’s paper.

]]>I don’t understand where your came from — it’s not bound in the condition .

]]>compL2TfIdf(w, c) = \sum_{d: c’ \neq c} l2TfIdf(w,d)

]]>