Encouraging words regarding applications of word sense disambiguation. Read on!
---------- Forwarded message ---------- Date: Thu, 19 Aug 2004 14:10:01 -0400 From: "Amsler, Robert" <[EMAIL PROTECTED]> To: [EMAIL PROTECTED] Subject: RE: [Senseval-discuss] more on applications of wsd I feel I should comment on this discussion of WSD since I am now involved almost completely in a significant application of such technology. At the Dept. of Energy I work on detecting what we'll call "sensitive information" in documents such that human reviewers can then decide whether the documents have to be withheld from release. The task is difficult because the volume and variety of electronic documents awaiting examination appears to be growing geometrically while the amount of sensitive information in the documents is actually not increasing. It is a typical "needle in a haystack" task of identifying a very small amount of sensitive content in a very large amount of irrelevant text. The task is performed by essentially searching for specific words and phrases whose likelihood of being related to the sensitive content is quite high, but there are many specific words which cannot readily be immediately searched for because without word sense disambiguation they return false positives on the order of 100:1 or worse. (From experience it is clear that when human readers are given very high percentages of false positives they begin to distrust the tool that provides the data and eventually will reject the term and the tool as not worth their time since it cannot discriminate between vitally important and totally irrelevant senses of a word or phrase.) While it may seem that WSD cannot achieve success until an overall score of 90% or above for general vocabulary can be realized, I do not find that a very interesting goal. I am more concerned with specific words that have extraordinarily low likelihoods of having the one precise meaning one needs to find amidst an overwhelming predominance of irrelevant meanings. It does little good in my application to speak in averages across the whole vocabulary. What would be far more useful would be explicit treatments of some highly ambiguous word or phrase from which definitive techniques for their separation might lead to the creation of general methods of attack on other specific words and phrases. I'd prefer someone to claim they had achieved 99% accuracy in distinguishing senses of some term such as "line" than in claiming they had reached 90% across the general vocabulary (with perhaps only 40% on "line" itself, but 99% on other words). I believe that WSD may achieve break-out success not from its "overall" ability, but from the ability to craft tools for specific applications to narrow vocabularies. Think of the situation with weather forecasts and MT in which being able provide French/English translations in a small domain became a successful application while general MT labors on. For an example, consider the task of distinguishing the terms for chemical elements from other words and abbreviations in text. I, Be, He, In, At and As are chemical elements, but overwhelmingly more common as English function words--how would one reliably find their chemical element senses? Likewise, Ca, Mn, Co, Ga, Al, Ne, Ar, In, La, Nd, Pa, and Nd are both chemical elements and 2-letter abbreviations for US states. Separating the two classes would be a good challenge. And the preferred solution should not depend on upper/lower case information since that is sometimes unavailable or unreliable. In general, if you talk to people doing my type of work, what they generally admit is that short acronyms are nearly impossible to disambiguate in text. They vitally need to know when something like DNA refers to the Defense Nuclear Agency, but can't afford to search for it because the now overwhelming presence of the genetic meaning will overwhelm the sense they want. Build an acronym disambiguator and the world will beat a path to your door. Robert Amsler ------------------------ Yahoo! Groups Sponsor --------------------~--> Make a clean sweep of pop-up ads. Yahoo! Companion Toolbar. Now with Pop-Up Blocker. Get it for free! http://us.click.yahoo.com/L5YrjA/eSIIAA/yQLSAA/x3XolB/TM --------------------------------------------------------------------~-> Yahoo! Groups Links <*> To visit your group on the web, go to: http://groups.yahoo.com/group/nlpatumd/ <*> To unsubscribe from this group, send an email to: [EMAIL PROTECTED] <*> Your use of Yahoo! Groups is subject to: http://docs.yahoo.com/info/terms/
