Discrepancy from the paper

[YipingNUS]

Hi,
the eta calculation in LoadDocs.java seems to differ from the description in the paper. Specifically, in the code you assign identical eta only when sum==0. This makes the eta value of the categories whose seed word doesn't appear to be zero.

STM/STM/src/LoadDocs.java

Line 95 in ca9a7a3

if (sum == 0) {

Below is the direct translation from the paper to replace ln 95-102.

        float sumSmooth = smooth * model.numCategories;
        for (int i = 0; i < raw.length; i++)
            raw[i] = (raw[i] + smooth) / (sum + sumSmooth);

        model.eta[index] = raw;