Ratio discounting removes a fixed proportion of the probability mass in a given context, this is in contrast to absolute discounting, which removes a fixed probability mass. We have found, empirically, that for very small corpora (which is what this tool is intended for) proportional discounting works as well as or better than absolute discounting. It has also the virtue of being computationally simpler and thus allows faster model building. Intuitively, it works because the model will place less emphasis on absolute count magnitudes, which are likely to be unstable in very small corpora (for which counts will be unreliable in the first place).