FOA Home | UP: Weighting and matching against Indices


Conclusion

We've covered enormous ground in this chapter. We began by wondering just what it might mean to try to understand communicative acts, like the publication of a document or the posing of a question/query. We looked in some detail at one of the most fundamental characteristics of texts, Zipf's Law, and found it to be, in fact, quite \rikmeaning-less! But the next level of features, tokens like those produced by the machinery of Chapter 2, supported a rich analysis of the Index, as a balance between the vocabularies of the documents' authors and the searchers' subsequent queries. The vector space provides a concrete model of potential keyword vocabularies and what it might mean to match within this space. Finally, we considered an efficient implementation of nearly-complete matching. In the next chapter we consider the problem of evaluating just how well all these techniques work in practice. But there are gaps in this story that are so obvious we don't even need to measure.

Soem invovle implementation issues that can be critical, especially when faced with very large corpora [Harper92] . Parallel implementation techniques, for example those pioneered on ``massively parallel'' SIMD (single instruction/multiple data) Connection machines become important in this respect [REF589] [REF378] [REF583] . In the modern age of multiple search engines each indexing (only partially overlapping versions! [Lawrence98] ) of the WWW techniques for FUSING multiple hitlists into a single list for the same user suggests another level of parallelism in the FOA task [Voorhees95] .

But linguists, in particular, must have more serious, implmentation-independent concerns. Imagine that you are someone who has studied the range of human languages and who appreciates both their wide variety and equally remarkable commonalities. You would be appalled at the violence we have done to the syntactic structure of language. For linguists, finding out about documents by counting their words is like trying to understand Beijing by listening to a single microphone poised high over the city. You can pick up on a few basic trends (like when most people are awake) but most of the texture is missing!

{DOG BITES MAN} and {\tt MAN BITES DOG} clearly mean two different things. Word order obviously conveys meaning beyond that brought by the three words. And the problem doesn't end with word order. Look how different the meanings of these phrases are: \bit \item {\tt NEUTRALIZATION OF THE PRESENT} \item {\tt REPRESENTING NEUTRONS} \item {\tt REPRESENTATIONS, NOT NEUTRONS} \eit despite the fact that all of them (conceivably) reduce to the same set of indexable tokens! Note especially how critical the same ``noise'' words thrown away on statistical grounds (in Chapter 2) are in analyzing a sentence's syntactic structure.

The attempt to understand the phenomena of meaning -- words -- by looking for patterns in word frequency statistics alone is reminiscent of the tea leaves and entrails of this chapter's opening quote. Still, the success of many WWW search engines that use very little beyond this kind of gross analysis suggests that their is much more information in the statistics than traditional, syntactically-focussed linguists might have believed.


Top of Page | UP: Weighting and matching against Indices | ,FOA Home


FOA © R. K. Belew - 00-09-21