Restricting the input or the output of a grammar-induced translation to a given set of trees plays an important role in statistical machine translation. The problem for practical systems is to find a compact (and in particular, finite) representation of said restriction. For the class of synchronous treeadjoining grammars, partial solutions to this problem have been described, some being restricted to the unweighted case, some to the monolingual case. We introduce a formulation of this class of grammars which is effectively closed under input and output restrictions to regular tree languages, i.e., the restricted translations can again be represented by grammars. Moreover, we present an algorithm that constructs these grammars for input and output restriction, which is inspired by Earley’s algorithm.

Many NLP tasks make predictions that are inherently coupled to syntactic relations, but for many languages the resources required to provide such syntactic annotations are unavailable. For others it is unclear exactly how much of the syntactic annotations can be effectively leveraged with current models, and what structures in the syntactic trees are most relevant to the current task. We propose a novel method which avoids the need for any syntactically annotated data when predicting a related NLP task. Our method couples latent syntactic representations, constrained to form valid dependency graphs or constituency parses, with the prediction task via specialized factors in a Markov random field. At both training and test time we marginalize over this hidden structure, learning the optimal latent representations for the problem. Results show that this approach provides significant gains over a syntactically uninformed baseline, outperforming models that observe syntax on an English relation extraction task, and performing comparably to them in semantic role labeling. 1

Cross-language information retrieval today is dominated by techniques that rely principally on context-independent token-to-token mappings despite the fact that state-of-the-art statistical machine translation systems now have far richer translation models available in their internal representations. This paper explores combination-of-evidence techniques using three types of statistical translation models: context-independent token translation, token translation using phrase-dependent contexts, and token translation using sentence-dependent contexts. Context-independent translation is performed using statistically-aligned tokens in parallel text, phrase-dependent translation is performed using aligned statistical phrases, and sentence-dependent translation is performed using those same aligned phrases together with an n-gram language model. Experiments on retrieval of Arabic, Chinese, and French documents using English queries show that no one technique is optimal for all queries, but that statistically significant improvements in mean average precision over strong baselines can be achieved by combining translation evidence from all three techniques. The optimal combination is, however, found to be resource-dependent, indicating a need for future work on robust tuning to the characteristics of individual collections.