MASSACHUSEnS INSrrrrUTE OF TECHNOLOGY

AI1STI"{AC'r: .~is papcr is about an implcmentcd natural languagc intcrfacc that transl;ltcs from English into semantic net rclatioils and from sem;}l1tic net rclalions btlCk int{) ..:nglish. l.he parscr and companion gcncrator wcrc implcmentcd for two reasolls: (a) to cnablc cxperimcntal work in support of a thcory of learning by analogy; (b) to dcmonstratc tlle viability of a theory of parsing and gcncration built on commutaLivc lransfonnations. l'1le Icarning thcory was shapcd to a gretlt degrcc by cxperiments that would ha','c becn cxlraordinarily tedious to pcrf()rm witl1OUt the I~nglish intcrfacc with which the e:\pcrimcntal datAl base was prcparcd. rcvis~d, tmd rcviscd again. Inasmucl} as cufrcllt wllrk 011 Lllc 11.:(tming thcory is moving t!)ward a lcnfold incre(lsc in dt}(;l-basc size, thc l:;,nglish interface is muvinb fi.om a fjcilitatiJlg rolc to an cnt\bling one. '\l1c parsing and gcncration tllcory \1t\S t\\'o particularly important fe;ltures: (a) the samc grtlmmar is uscd l(>r both parsing and generation: (b) thc ti.ansj{)rnlations of L11C grammt\r arc commutative. The languagc gcner(Jtion proccdure convcrts a semtmtic nctwork fragmenL into kerncl fr;lmcs, chooscs t11e sct of transformations that should be performed upon each framc, executes tIlc specificd transf(>mlations, combines the altered kemcls into a sentencc, performs a pronominalization process, and finaJ1y produces tJ1C appropriatc English word string. Parsi1}g is esscntiaUy tIle revcrse of generation. The first step in the parsing process is splitting a given sentence into a s{'t of kerncl clauses along with a dcscriplion of how those clauses are hicrarchicaJ1y related to each othcr. The clauses are uscd to producc matrix and embedded kcmel fraIncs, which in turn supply arguments to rclation-crcating functions. Thc evaluatio1l of t11e rclation-creating furlctions results in the construction of the scmantic net fragmcnts.