Ulrich Germann

Fast Decoding and Optimal Decoding for Machine Translation

A good decoding algorithm is critical to the success of any statistical machine translation system. The decoders job is to find the translation that is most likely according to set of previously learned parameters (and a formula for combining them). Since the space of possible translations is extremely large, typical decoding algorithms are only able to examine a portion of it, thus risking to miss good solutions. In this paper, we compare the speed and output quality of a traditional stack-based decoding algorithm with two new decoders: a fast greedy decoder and a slow but optimal decoder that treats decoding as an integer-programming optimization problem.

Greedy decoding for statistical machine translation in almost linear time

We present improvements to a greedy decoding algorithm for statistical machine translation that reduce its time complexity from at least cubic (O(n6) when applied naively) to practically linear time1 without sacrificing translation quality. We achieve this by integrating hypothesis evaluation into hypothesis creation, tiling improvements over the translation hypothesis at the end of each search iteration, and by imposing restrictions on the amount of word reordering during decoding.

Building a statistical machine translation system from scratch: how much bang for the buck can we expect?

We report on our experience with building a statistical MT system from scratch, including the creation of a small parallel Tamil-English corpus, and the results of a task-based pilot evaluation of statistical MT systems trained on sets of ca. 1300 and ca. 5000 parallel sentences of Tamil and English data. Our results show that even with apparently incomprehensible system output, humans without any knowledge of Tamil can achieve performance rates as high as 86% accuracy for topic identification, 93% recall for document retrieval, and 64% recall on question answering (plus an additional 14% partially correct answers).

Fast and optimal decoding for machine translation

A good decoding algorithm is critical to the success of any statistical machine translation system. The decoders job is to find the translation that is most likely according to a set of previously learned parameters (and a formula for combining them). Since the space of possible translations is extremely large, typical decoding algorithms are only able to examine a portion of it, thus risking to miss good solutions. Unfortunately, examining more of the space leads to unacceptably slow decodings.In this paper, we compare the speed and output quality of a traditional stack-based decoding algorithm with two new decoders: a fast but non-optimal greedy decoder and a slow but optimal decoder that treats decoding as an integer-programming optimization problem.

The University of Edinburgh's Neural MT Systems for WMT17

This paper describes the University of Edinburghs submissions to the WMT17 shared news translation and biomedical translation tasks. We participated in 12 translation directions for news, translating between English and Czech, German, Latvian, Russian, Turkish and Chinese. For the biomedical task we submitted systems for English to Czech, German, Polish and Romanian. Our systems are neural machine translation systems trained with Nematus, an attentional encoder-decoder. We follow our setup from last year and build BPE-based models with parallel and back-translated monolingual training data. Novelties this year include the use of deep architectures, layer normalization, and more compact models due to weight tying and improvements in BPE segmentations. We perform extensive ablative experiments, reporting on the effectivenes of layer normalization, deep architectures, and different ensembling techniques.

Cross-lingual C*ST*RD: English access to Hindi information

We present C*ST*RD, a cross-language information delivery system that supports cross-language information retrieval, information space visualization and navigation, machine translation, and text summarization of single documents and clusters of documents. C*ST*RD was assembled and trained within 1 month, in the context of DARPAs Surprise Language Exercise, that selected as source a heretofore unstudied language, Hindi. Given the brief time, we could not create deep Hindi capabilities for all the modules, but instead experimented with combining shallow Hindi capabilities, or even English-only modules, into one integrated system. Various possible configurations, with different tradeoffs in processing speed and ease of use, enable the rapid deployment of C*ST*RD to new languages under various conditions.

The Impact of Machine Translation Quality on Human Post-Editing

We investigate the effect of four different competitive machine translation systems on post-editor productivity and behaviour. The study involves four volunteers postediting automatic translations of news stories from English to German. We see significant difference in productivity due to the systems (about 20%), and even bigger variance between post-editors.

CASMACAT: A Computer-assisted Translation Workbench

CASMACAT is a modular, web-based translation workbench that offers advanced functionalities for computer-aided translation and the scientific study of human translation: automatic interaction with machine translation (MT) engines and translation memories (TM) to obtain raw translations or close TM matches for conventional post-editing; interactive translation prediction based on an MT engine’s search graph, detailed recording and replay of edit actions and translator’s gaze (the latter via eye-tracking), and the support of e-pen as an alternative input device. The system is open source sofware and interfaces with multiple MT systems.

Translation with Scarce Bilingual Resources

Machine translation of human languages is a field almost as old as computers themselves. Recent approaches to this challenging problem aim at learning translation knowledge automatically (or semi-automatically) from online text corpora, especially human-translated documents. For some language pairs, substantial translation resources exist, and these corpus-based systems can perform well. But for most language pairs, data is scarce, andcurrent techniques do not work well. To examine the gap betweenhuman and machine translators, we created an experiment in which humanbeings were asked to translate an unknown language into English on thesole basis of a very small bilingual text. Participants performed quite well,and debriefings revealed a number of valuable strategies. We discuss thesestrategies and apply some of them to a statistical translation system.

Making Semantic Interpretation Parser-Independent

We present an approach to semantic interpretation of syntactically parsed Japanese sentences that works largely parser-independent. The approach relies on a standardized parse tree format that restricts the number of syntactic configurations that the semantic interpretation rules have to anticipate. All parse trees are converted to this format prior to semantic interpretation. This setup allows us not only to apply the same set of semantic interpretation rules to output from different parsers, but also to independently develop parsers and semantic interpretation rules.