Search Descriptions

Main Topics

Search Publications





Syntax Based Models without Linguistic Annotation


Rooted in a finite state machine approach, head automata have been developed that allow for tree representation during the translation process (Alshawi, 1996; Alshawi et al., 1997; Alshawi et al., 1998; Alshawi et al., 2000). Wang and Waibel (1998) also induce hierarchical structures in their statistical machine translation model. The use of synchronous grammars for statistical machine translation has been pioneered by Wu (1995); Wu (1996); Wu (1997); Wu and Wong (1998) with their work on inversion transduction grammars (ITG) and stochastic bracketing transduction grammars (SBTG). A different formal model are multitext grammars (Melamed, 2003; Melamed, 2004; Melamed et al., 2004) and range concatenation grammars (Søgaard, 2008). Zhang and Gildea (2005) extend the ITG formalism by lexicalizing grammar rules and apply it to a small-scale word alignment task. Armed with an efficient A* algorithm for this formalism (Zhang and Gildea, 2006), they extend the formalism with rules that are lexicalized in both input and output (Zhang and Gildea, 2006). Word alignments generated with ITG may also be used to extract phrase for phrase-based models (Sánchez and Benedí, 2006; Sánchez and Benedí, 2006b). Zhang et al. (2008) present an efficient ITG phrase alignment algorithm that uses Bayesian learning with priors.



Related Topics

New Publications