Automatic Acquisition of Hierarchical Transduction Models for Machine Translation

Automatic Acquisition of Hierarchical Transduction Models for Machine Translation Hiyan Alshawi Srinivas Bangalore Shona Douglas XI'&T I~abs llcsear(h l S(t Park AveJme, P.O. Box 9711 l"lorham l>a.rk, NJ 07932 [.;SA Abstract \Ve describe a method for the fully automatic learning of hierarchical Iinite state translation models. The input to lhe method is transcribed sl)eech utterances and th(>ir corre.~l)on(ling hu. man translations, and the output is a set of head transducers, i.e. statistical le×ical head- outward transducers. A word-alignment func- tion and a head-ranking funcl, ion are [irst ob- tained, and then counts are generated for hy- pothesized stale transitions of head transduc- ers whose lexical translations and word order chauges are consislent with the alignment. The method has been applied to create au English- Spanish translation model for ~t speech tI'ans- lation application, with word accuracy of over 75% a,s measured by a string-dislauce compari- son to three reference tra.nslations. 1 Introduction The fully automatic constructiol, r.)f' translation lnodels offers benelits in ternls of develol)ment effort and i)otentially in robustness over met}> ods requiring hand-coding of linguistic informa- tion. Itowevei', there art disadvantages to the automatic approaches proposed so far. The var- ious methods described by P, rown el. al (1990; 1993) do not take into account tile natural strllc- turing of strings into phrases. Examph>based translation, exemplified hy tile work of Sumita and lida (1995), requires very large amounts o[ lraiiling luatcrial. The number o[' stales in a sinlple finite state model such as those used by Viler et el. (1996) becomes extremely large when faced with languages with large word order ditferences. The work reported in \'Vu (1997), which uses an inside-outside type of training algorithm to learn statistical context- free transduction, has a similar motivation to tilt, Clli'l'(?l]{ \Vol'k~ but the models \re describe here, being fully lexica.1, are more suitable for direct, statistical modelling. lu this paper, we show that both the net- work topology and parameters of a head trans- ducer translation model (Alshawi, t996b) can t)e Darned fully automatically from a bilingual corpus. It has already been S}lOWll (Alshawi eL el., i997) that a. head transducer model with hand-coded structure, can be trained to give bet- ter accuracy than a comparable transfer-based system, with snmller model size, computational requirements, and development ef['ort. \Ve have applied the learning method to el'e- ate an English-Spanish translation model for a limited domain, with word a.ccuracy of over 75% measured by a string distance comparison (as used in speech recognition) to three reference translations. The resulting translation model has been used as a comI)Ol,ent o[' an I';nglish- Spanish speech tra.nslation system. \'~)' first present the atet)s of tile 'brallbd/lc- t.io,, lraining method in Section 2. In Section 3 we describe how we obtain an alignment func- tion froHl SOllree word subsequences to target word subsequences ['or each transcribed utter- a nce and its translation. The construction of states and transitions is specitied in Section 4; the method for selecting phrase head words is described in Section 5. The …

Content truncated to 500 words