We present a machine translation framework that can incorporate arbitrary features of both input and output sentences. The core of the approach is a novel decoder based on lattice parsing with quasisynchronous grammar (Smith and Eisner, 2006), a syntactic formalism that does not require source and target trees to be isomorphic. Using generic approximate dynamic programming techniques, this decoder can handle “non-local ” features. Similar approximate inference techniques support efficient parameter estimation with hidden variables. We use the decoder to conduct controlled experiments on a German-to-English translation task, to compare lexical phrase, syntax, and combined models, and to measure effects of various restrictions on nonisomorphism. 1

Current SMT systems usually decode with single translation models and cannot benefit from the strengths of other models in decoding phase. We instead propose joint decoding, a method that combines multiple translation models in one decoder. Our joint decoder draws connections among multiple models by integrating the translation hypergraphs they produce individually. Therefore, one model can share translations and even derivations with other models. Comparable to the state-of-the-art system combination technique, joint decoding achieves an absolute improvement of 1.5 BLEU points over individual decoding. 1

Efficiency is a prime concern in syntactic MT decoding, yet significant developments in statistical parsing with respect to asymptotic efficiency haven’t yet been explored in MT. Recently, McDonald et al. (2005b) formalized dependency parsing as a maximum spanning tree (MST) problem, which can be solved in quadratic time relative to the length of the sentence. They show that MST parsing is almost as accurate as cubic-time dependency parsing in the case of English, and that it is more accurate with free word order languages. This paper applies MST parsing to MT, and describes how it can be integrated into a phrase-based decoder to compute dependency language model scores. Our results show that augmenting a state-ofthe-art phrase-based system with this dependency language model leads to significant improvements in TER (0.92%) and BLEU (0.45%) scores on five NIST Chinese-English evaluation test sets. 1

Hypergraphs are used in several syntaxinspired methods of machine translation to compactly encode exponentially many translation hypotheses. The hypotheses closest to given reference translations therefore cannot be found via brute force, particularly for popular measures of closeness such as BLEU. We develop a dynamic program for extracting the so called oracle-best hypothesis from a hypergraph by viewing it as the problem of finding the most likely hypothesis under an n-gram language model trained from only the reference translations. We further identify and remove massive redundancies in the dynamic program state due to the sparsity of n-grams present in the reference translations, resulting in a very efficient program. We present runtime statistics for this program, and demonstrate successful application of the hypotheses thus found as the targets for discriminative training of translation system components. 1

Recent syntactic extensions of statistical translation models work with a synchronous context-free or tree-substitution grammar extracted from an automatically parsed parallel corpus. The decoders accompanying these extensions typically exceed quadratic time complexity. This paper extends the Direct Translation Model 2 (DTM2) with syntax while maintaining linear-time decoding. We employ a linear-time parsing algorithm based on an eager, incremental interpretation of Combinatory Categorial Grammar (CCG). As every input word is processed, the local parsing decisions resolve ambiguity eagerly, by selecting a single supertag–operator pair for extending the dependency parse incrementally. Alongside translation features extracted from the derived parse tree, we explore syntactic features extracted from the incremental derivation process. Our empirical experiments show that our model significantly outperforms the state-of-the art DTM2 system. 1

Previous work has shown that high quality phrasal paraphrases can be extracted from bilingual parallel corpora. However, it is not clear whether bitexts are an appropriate resource for extracting more sophisticated sentential paraphrases, which are more obviously learnable from monolingual parallel corpora. We extend bilingual paraphrase extraction to syntactic paraphrases and demonstrate its ability to learn a variety of general paraphrastic transformations, including passivization, dative shift, and topicalization. We discuss how our model can be adapted to many text generation tasks by augmenting its feature set, development data, and parameter estimation routine. We illustrate this adaptation by using our paraphrase model for the task of sentence compression and achieve results competitive with state-of-the-art compression systems.

Cube pruning is a fast inexact method for generating the items of a beam decoder. In this paper, we show that cube pruning is essentially equivalent to A * search on a specific search space with specific heuristics. We use this insight to develop faster and exact variants of cube pruning. 1

We describe a synchronous parsing algorithm that is based on two successive monolingual parses of an input sentence pair. Although the worst-case complexity of this algorithm is and must be O(n6) for binary SCFGs, its average-case run-time is far better. We demonstrate that for a number of common synchronous parsing problems, the two-parse algorithm substantially outperforms alternative synchronous parsing strategies, making it efficient enough to be utilized without resorting to a pruned search. 1

The best systems for machine translation of natural language are based on statistical models learned from data. Conventional representation of a statistical translation model requires substantial offline computation and representation in main memory. Therefore, the principal bottlenecks to the amount of data we can exploit and the complexity of models we can use are available memory and CPU time, and current state of the art already pushes these limits. With data size and model complexity continually increasing, a scalable solution to this problem is central to future improvement. Callison-Burch et al. (2005) and Zhang and Vogel (2005) proposed a solution that we call translation by pattern matching, which we bring to fruition in this dissertation. The training data itself serves as a proxy to the model; rules and parameters are computed on demand. It achieves our desiderata of minimal offline computation and compact representation, but is dependent on fast pattern matching algorithms on text. They demonstrated its application to a common model based on the translation of contiguous substrings, but leave some open problems. Among these is a question: can this approach match the performance of conventional methods despite unavoidable differences that it induces in the model? We show how to answer this question affirmatively. The main

Many syntactic machine translation decoders, including Moses, cdec, and Joshua, implement bottom-up dynamic programming to integrate N-gram language model probabilities into hypothesis scoring. These decoders concatenate hypotheses according to grammar rules, yielding larger hypotheses and eventually complete translations. When hypotheses are concatenated, the language model score is adjusted to account for boundary-crossing n-grams. Words on the boundary of each hypothesis are encoded in state, consisting of left state (the first few words) and right state (the last few words). We speed concatenation by encoding left state using data structure pointers in lieu of vocabulary indices and by avoiding unnecessary queries. To increase the decoder’s opportunities to recombine hypothesis, we minimize the number of words encoded by left state. This has the effect of reducing search errors made by the decoder. The resulting gain in model score is smaller than for right state minimization, which we explain by observing a relationship between state minimization and language model probability. With a fixed cube pruning pop limit, we show a 3-6 % reduction in CPU time and improved model scores. Reducing the pop limit to the point where model scores tie the baseline yields a net 11 % reduction in CPU time. 1.