Workshop on Linguistic pre-processing for MT (Cancelled)August 30, 2009
We invite proposals for presentation at the Workshop on Linguistic pre-processing for MT, being held in conjunction with MT Summit XII.
Input for MT varies significantly in terms of spelling, terminology, word order phenomena, dialects, and sentence types, even within the same language. With user-generated content, this variability increases enormously. MT systems, and NLP systems generally, cannot cover effectively all of this variability -- usually because they are built to deal with professionally written technical or journalistic texts. Robust and reliable systems for mapping highly variable, uncontrolled writing into more consistent, tractable, "controlled" sentences will improve MT, search, and other NLP tasks. Current approaches to this problem include manually pre-editing the input texts -- as discussed for example in the series of CLAW workshops -- and/or expanding the coverage of MT systems.
One alternative approach is to pre-process or normalize the input automatically before MT. Translation of subtitles for television (Flanagan, 2006), non-fluent speech, low-quality OCR, and non-standard writing from limited-proficiency writers are only some of the application scenarios that require automatic linguistic pre-processing to improve MT output. For example, Callison-Burch (2007) showed that substitution of lexical paraphrases improved MT output. Xu & Seneff (2008) and Collins, Koehn & Kucerova (2005) re-arranged word order to improve performance of a statistical MT system. Yet another alternative approach is to produce a linguistically "enriched" input, in the form of lattices, trees, markup, etc. and allow for final interpretation later in the translation pipeline and/or with a direct feedback capability to force emergent behavior. Some approaches may even call into question the need for a strict, linear processing pipeline and may employ adaptive, iterative, or self-learning methods.
Common to all of these alternatives is the strategy of deploying significant linguistic and non-linguistic knowledge before translation itself occurs. This raises many questions about which kinds of knowledge have the biggest impact on translation, which can be automated most reliably and robustly, and which are most cost effective and scalable.
This workshop aims to compare and contrast some of the various techniques and approaches to these kinds of linguistic pre-processing for MT. The workshop will consist of a set of papers that will be selected by peer review.
We welcome submissions about the main theme of this workshop. Specific topics include but are not limited to:
Papers should not have been presented somewhere else or be under consideration for publication elsewhere, and should not identify the author(s). They should emphasize completed work rather than intended work. Each paper will be anonymously reviewed by the program committee.
Papers must be submitted in PDF format to firstname.lastname@example.org by midnight of the due date. Submissions should be in English. The papers should be attached to an email indicating contact information for the author(s) and paper’s title. Papers should not exceed 8 pages including references and tables, and should follow the formatting guidelines posted at the MT Summit web site.
For further information, contact the organizing committee at email@example.com
Mike Dillinger, Translation Optimization Partners (Primary Contact)
Last updated: Fri Apr 17 16:05:02 PDT 2009