Fence Sitting We in the ELSNews community study language computationally, both with an eye to understanding language and especially language processing better, and with an eye to improving existing products and services that rely on language processing. Asked whether we're involved in science or engineering, a lot of would prefer to say "both!" and emulate illustrious predecessors in many branches of science (and engineering!) who combined work on pure science and application. (Or is it applied science and research engineering?) Einstein and Szilard, not content with leading work in theoretical physics, also obtained patents for innovative refrigeration pumps. Kepler not only systematised Tycho Brahe's observations into his famous laws (rotational period proportional to radius, equal area swept in equal time), but likewise applied his work on equations for ellipses to the problem of measuring the contents of beer barrels (in Stuttgart!). Closer to our own field, neither Turing nor von Neumann were shy about tackling practical problems. Examples of engineers with scientific credentials don't spring to mind as quickly, but maybe that's because scientists get more press to begin with, and because engineers often cultivate a "just folks" public image. Edison worked hard on his simple-man image with the remark about genius being "1% inspiration and 99% perspiration", but corresponded extensively with Maxwell, and knew all the relevant physics in areas he worked on. There's a brief, but deservedly respected tradition in which research projects ask serious scientific questions while building demonstration vehicles that the same work enables. Woods et al.'s LUNAR system developed the ATN as a processing model even while showing it off in question-answering. The blackboard architecture was developed in the HEARSAY project, which immediately put it to use in improving speech understanding. Verbmobil combined deep and shallow processing, and was likewise ambitious in supporting novel telephone translation. But it won't do to recall that good science and engineering can go hand in hand. Sometimes they don't harmonise well at all: Pat Hayes has a nice piece in which he recounts how the then president of the US Academy of Sciences, Simon Newcomb, scoffed at the prospect of human flight -- even after the Wright brothers' early successes! And even when science and engineering do co-exist in a single mind, they don't necessarily co-exist in any given piece of work springing from it. Einstein didn't make relativistic refrigeration pumps --- he was just clever in more than one way. Proposals and evaluations from computational linguistics are often treated with more then the usual refereeing savagery by funding agencies who see us waffling about science vs. engineering where we see ourselves as cleverly combining. The savagery is not entirely misplaced. To begin, projects with modest resources do not find it easy to innovate in several ways simultaneously. HEARSAY and Verbmobil are not typical. And we've all seen mediocre projects with explicit practical objectives gradually transform (by final report time) to theoretical "studies". You get the feeling that if the project had only decided definitely -- either for theory or for application, then the effort would have been more successful, and in any case clearer. And let's not forget that even practical success can be confusing in mixed-mode projects. A project several years ago aimed to support language learning with a focus on applying a particular grammar theory and formalism. They tried out a prototype and, yes, people who used the system learned faster. But they were compared to others who learned from books based on traditional grammar. One group used interactive drills, the other a book. One group was tracked automatically, the other had to be self-paced. What was the gain in scientific knowledge? And what should we apply? ... There's always the follow-up project.