Subject: computational linguistics : the altavista " refine " option

a colleague of mine told me about the " refine " option offered by the altavista search engine ( http : / / www . altavista . digital . com / ) and how good it was . in a nutshell , the " refine " options returns a list of synonyms of , and notions related to the words in your query . indeed , it gave extremely sensible responses . perversely perhaps , i tried it in french , using " vin " ( what else ! ) as the keyword . bingo . this is what it returned : 72 % etait , etre , annees , meme , apres , etaient , derniers 59 % egalement , particulierement , differentes , possibilite 52 % qualite , vins , vin , vignoble , vigne , crus , vignes , vignerons , vigneron [ etc . . . ] far from satisfactory . " eau " and " pain " returned similar nonsense , featuring " etaient " , " etre " , " egalement " et alia in prominent positions . in fact , altavista " refine " seems decidedly adverse to foodstuff in french , " fruit " , " poisson " and " sandwich " failing equally miserably ( so did " sable " , " mer " , " lac " ) . so i was quite surprised when italian queries about " wine " returned sensible synonyms : 60 % vino , vini , vigneti , uve 40 % quantita , ettari , vitigni 39 % sapore , profumo , invecchiamento [ etc . ] " acqua " and " pane " fared equally well . so i turned my attention to spanish . spanish did quite as badly as french . this is quite puzzling for the size of the spanish data is quite large . i do n't know what inspired me , i decided to ask for an italian sandwich ( " panino " ) . bingo again ! 60 % perche , chissa , guardo , cazzo , sembrava , poiche , merda , riposto [ yes , unbelievable but true ] 54 % mangiare , specialta , birra , mangia , roba , piatti , gusti , soldi , bere 33 % avevo , scusa , aveva , stavo , rispose , facevano [ etc . ] my colleague and i scratched our collective heads , experimented some more , and came to the conclusion that the thesauri are built by a neural net ( she is heavily into neural nets ) . still , the excellent behaviour of the english thesaurus was suspect . but no , experimenting demonstrated that it could not have been a hand-crafted thesaurus . there are ways of " salting " a neural net and that is probably what digital did for english ( and perhaps for italian ) . do take a break and experiment a bit with altavista " refine " option in your favourite languages ( polish was as nonsensical as french ) . it is quite amusing . and perhaps useful : next time someone knocks at your door with a neural net for sale . . . ( i have seen queries for " kentucky fried chicken " return " chicken sexers " , " waste burners " and " singing teachers " , courtesy of a neural net ) . j . guy @ trl . telsta . com . au
