Can inversion transduction grammars generate hand alignments

Can inversion transduction grammars generate hand alignments? Anders Søgaard University of Copenhagen Njalsgade 140–2 DK-2300 Copenhagen soegaard@hum.ku.dk Abstract The adequacy of inversion transduction grammars (ITGs) has been widely de- bated, and the discussion’s crux seems to be whether the search space is inclusive enough (Zens and Ney, 2003; Wellington et al., 2006; Søgaard and Wu, 2009). Parse failure rate when parses are constrained by word alignments is one metric that has been used, but no one has studied parse failure rates of the full class of ITGs on representative hand aligned corpora. It has also been noted that ITGs in Chomsky nor- mal form induce strictly less alignments than ITGs (Søgaard and Wu, 2009). This study is the first study that directly com- pares parse failure rates for this subclass and the full class of ITGs. 1 Introduction The adequacy of grammar-based machine transla- tion formalisms is sometimes empirically evalu- ated by running all-accepting grammars on large amounts of automatically aligned text (Zens and Ney, 2003). What is studied is called alignment capacity (Søgaard and Wu, 2009) or translation equivalence modeling (Zhang et al., 2008), i.e. a formalism’s ability to generate observed align- ments or translation equivalences; and the study is closely related to the study of translation model search spaces (Zens and Ney, 2003; Dreyer et al., 2007). All-acccepting grammars are simply gram- mars that contain all possible rules that can be ex- pressed in a formalism. A grammar generates an aligned sentence pair if it can generate the two sen- tences in such a way that all aligned words are gen- c⃝ 2010 European Association for Machine Translation. erated simultaneously (Wu, 1997). What is studied is thus: Can an all-accepting grammar generate the aligned sentence pairs observed in a text? The met- ric in these studies is parse failure rate (PFR) or its inverse, i.e. the number of sentence pairs that can be generated over the total number of sentence pairs. Most alignment capacity studies use automati- cally aligned text, since hand-aligned text is hard to come by. Recently three important data sets have been released (Pad´o and Lapata, 2006; Graca et al., 2008; Buch-Kromann et al., 2010). Our ex- periments include 12 hand-aligned parallel texts of varying size. Our main contribution is evaluating the empir- ical adequacy of inversion transduction grammars (ITGs) (Wu, 1997), a popular grammar-based ma- chine translation formalism, on these data sets. It has been noted that the alignment capacity of the full class of ITGs extends that of the class of ITGs that are in Chomsky normal form (NF-ITGs) (Søgaard and Wu, 2009), i.e. while the normal form is a normal form in the sense that it does not alter the generative capacity in terms of sentence pairs, the normal form restrictions do exclude cer- tain alignment configurations. Consequently, we compare the adequacy of both ITGs and NF-ITGs. It is shown that while ITGs are more adequate than local reordering models (and in many cases also to IBM models; cf. Zens and Ney …

Content truncated to 500 words