Towards Hybrid Quality-Oriented Machine Translation

We present a hybrid MT architecture, combining state-of-the-art linguistic processing with advanced stochastic techniques. Grounded in a theoretical reflection on the division of labor between rule-based and probabilistic elements in the MT task, we summarize per-component approaches to ranking, including empirical results when evaluated in isolation. Combining component-internal scores and a number of additional sources of (probabilistic) information, we explore discriminative re-ranking of n-best lists of candidate translations through an eclectic combination of knowledge sources, and provide evaluation results for various configurations. 1 Background—Motivation Machine Translation is back in fashion, with data-driven approaches and specifically Statistical MT (SMT) as the predominant paradigm— both in terms of scientific interest and evaluation results in MT competitions. But (fullyautomated) machine translation remains a hard— if not ultimately impossible—challenge. The task encompasses not only all strata of linguistic description—phonology to discourse—but in the general case requires potentially unlimited knowledge about the actual world and situated language use (Kay, 1980, 1997). Although the majority of commercial MT systems still have large sets of hand-crafted rules at their core (often using techniques first invented in the 1960s and 1970s), MT research in the once mainstream linguistic tradition has become the privilege of a small, faithful minority. Like a growing number of colleagues, we question the long-term value of purely statistical (or data-driven) approaches, both practically and scientifically. Large (parallel) training corpora remain scarce for most languages, and wordand phrase-level alignment continue to be active research topics. Assuming sufficient training material, statistical translation quality still leaves much to be desired; and probabilistic NLP experience in general suggests that one must expect ‘ceiling’ effects on system evolution. Statistical MT research has yet to find a satisfactory role for linguistic analysis; on its own, it does not further our understanding of language. Progress on combining rule-based and datadriven approaches to MT will depend on a sustained stream of state-of-the-art, MT-oriented linguistics research. The Norwegian LOGON initiative capitalizes on linguistic precision for high-quality translation and, accordingly, puts scalable, general-purpose linguistic resources—complemented with advanced stochastic components—at its core. Despite frequent cycles of overly high hopes and subsequent disillusionment, MT in our view is the type of application that may demand knowledge-heavy, ‘deep’ approaches to NLP for its ultimate, longterm success. Much like Riezler & Maxwell III (2006) and Llitjós & Vogel (2007)—being faithful minority members ourselves—we approach a hybrid MT architecture with a semantic transfer backbone as our vantage point. Plurality of approaches to grammatical description, reusability of component parts, and the interplay of linguistic and stochastic processes are among the strong points of the LOGON system. In the following, we provide a brief overview of the LOGON architecture (§ 2) and a bit of theoretical reflection on the role of probability theory