Time as a Measure of Parsing Efficiency

Charniak and his colleagues have proposed implementation-independent metrics as a way of comparing the efficiency of parsing algorithms implemented on different platforms, in different languages, and with different degrees of "incidental optimization". We argue that there are easily immaginable circumstances in which their proposed metrics would mask significant differences in efficiency; we point out that their data do not, in fact, support the usability of such metrics for comparing the efficiency of different algorithms; and we analyze data for a similar metric to try to quantify the degree of variation one might expect between such metrics and actual parse time. Finally, we propose a methodology for making cross-platform comparisons through the use of reference parser implementations.