Description
org.apache.mahout.vectorizer.collocations.llr.LLRReducer, which is part of the collocation driver, statically casts a long to an int.
private long ngramTotal;
...
int k11 = ngram.getFrequency(); /* a&b */
int k12 = gramFreq[0] - ngram.getFrequency(); /* a&!b */
int k21 = gramFreq[1] - ngram.getFrequency(); /* !b&a */
int k22 = (int) (ngramTotal - (gramFreq[0] + gramFreq[1] - ngram.getFrequency())); /* !a&!b */
These numbers are then fed into
org.apache.mahout.math.stats.LogLikelihood
specifically the function below.
public static double logLikelihoodRatio(int k11, int k12, int k21, int k22) {
// note that we have counts here, not probabilities, and that the entropy is not normalized.
double rowEntropy = entropy(k11, k12) + entropy(k21, k22);
double columnEntropy = entropy(k11, k21) + entropy(k12, k22);
double matrixEntropy = entropy(k11, k12, k21, k22);
if (rowEntropy + columnEntropy > matrixEntropy)
return 2.0 * (matrixEntropy - rowEntropy - columnEntropy);
}
In short if the long ngramTotal is larger than Integer.MAX_VALUE (which will happen in large datasets), then the driver will either crash or in the case that it casts to a negative int, will continue as usual but produce no output due to error checking.