Comments on Mohri, Pereira and Riley

Throughout the history of computational linguistics the rule-based and the statistics-based approaches appeared as competing rather than complementary threads of research. Even today, many view the success of the Xerox rule-based taggers as a threat to the more statistically oriented taggers, just as a few years ago the IBM statistical approach was viewed as a threat to rule-based systems of machine translation. The historical importance of the AT&T work on weighted transducers lies in the fact that for the first time it makes possible a genuine integration of the two approaches. Under the heading of “speech-natural language integration” we usually find the shotgun marriage of two completely disjoined systems, each with its own distinct theoretical apparatus and algorithmic building blocks. In contrast, here we find surprisingly smooth integration, both in terms of underlying theory and in terms of shared algorithms. This is a very significant accomplishment, and the main goal of my comments is to situate it as a particular stage of a constant developmental trend towards greater integration. I will ask how much the good sides of the rule-based and the data-based approaches have been preserved, and what, if anything, has been lost by taking this approach. I will also ask how far the present approach can be pushed, and offer some speculative remarks on future directions. First let me take a clear and unambiguous stance on the rule-based vs. statistics-based debate: rules are better. As a simple illustration, consider Fig. 1 which shows the performance of a bank check OCR system developed by the author [2] under three conditions: using a bigram language model, a finite state grammar, and a combination of the two. As it is evident from Fig. 1, the rule-based system fares much better than the statistics-based, and in fact the latter adds very little to the performance of a system already containing the former. So the question is not so much an overarching philosophical problem of whether rules are better, but rather the more mundane practical problem of finding the rules. To the extent that the rules, constrains and representations constituting the grammar are devised by grammarians like Quirk and Greenbaum [3], computational linguists can get a free ride, and they should avail themselves of the opportunity. But to the extent that Quirk et al. represent the culmination of an extremely sophisticated descriptive tradition of a singularly deeply researched language, it appears very unlikely that more than a handful languages could be handled in the same fashion. Again as an illustration, readers are invited to consider 0 10 20 30 40 50 60 70 80