So here's a random crazy idea I had lately. A phrase table could have multiple columns giving different scores for different probabilities from different alignments, different corpora, different domains etc. Recent work at Edinburgh, Cambridge and Sheffield has had some emphasis on adaptation of models for speech recognition purposes. I guess a similar principle could be applied to SMT. Given a text from some unknown domain the engine could perform some automated recognition test to guess which translation model best fits the text to be translated. A primitive form of automatic domain recognition and adaptation if you like.
I guess even making available multiple forms of a phrase table or a single compact version with multiple columns for scoring could even have some demand in the future. James ________________________________ From: joerg [[email protected]] Sent: 06 November 2013 13:05 To: Read, James C Cc: [email protected] Subject: Re: [Moses-support] 10 years of OPUS This is a good question. I was also wondering about this and, as I wrote, I started providing word alignments and even phrase tables for the data collected in OPUS. The idea is, of course, that people could skip running GIZA++ with standard settings over and over again on Europarl data and other standard sets. For example, you can download now word alignments for all language pairs for Europarl v7 from OPUS http://opus.lingfil.uu.se/Europarl/wordalign/ This, of course, assumes that you're happy with my tokenization and other kinds of pre-processing that may have influenced the data. In some cases you may also want to do other things like compound splitting and preordering which would require a different alignment model. I also have phrase tables in those folders as well extracted with standard settings from the grow-diag-final-and alignments. This is maybe less interesting as this assumes that you work with phrase-based SMT and that you have to work with the true cased data I provide etc. The truecaser models are, of course also available (for example for Europarl in http://opus.lingfil.uu.se/Europarl/wordalign/truecaser/). Monolingual data files with the same tokenization are also available (Europarl: http://opus.lingfil.uu.se/Europarl/mono/) As all of this takes a lot of space, I would actually like to know, if - I should keep those files on-line - I could remove some of them (for example, phrase tables that take most of the disk space) - if anyone actually is interested in using those models and alignments (and which ones) Thank you for your feedback! Jörg ********************************************************************************** Jörg Tiedemann http://stp.lingfil.uu.se/~joerg/ On Nov 6, 2013, at 8:48 AM, Read, James C wrote: I wonder if there would be a demand for ready made phrase tables generated from the data. ________________________________________ From: [email protected]<mailto:[email protected]> [[email protected]<mailto:[email protected]>] on behalf of Jorg Tiedemann [[email protected]<mailto:[email protected]>] Sent: 02 November 2013 18:54 To: [email protected]<mailto:[email protected]> Subject: [Moses-support] 10 years of OPUS After attending the 20-years-of-bitext workshop at EMNLP I suddenly realized that OPUS (http://opus.lingfil.uu.se) also has its 10-years anniversary this year (send me some champagne if you like). I will celebrate this anniversary by sending out this e-mail with some recent news and highlights. OPUS is a growing collection of parallel corpora for many languages and various domains. The collection becomes pretty big and includes a variety of data sets and tools that are not only useful for statistical machine translation. OPUS has been extended a lot since its first appearance in 2003. Actually the best birthday present would be if anyone would decide to start a mirror of OPUS. Let me know if you are interested. Here some of the highlights: - over 150 languages and language variants - over 5 billion aligned translation units - downloads in XML/XCES, plain text (Moses/SMT) and TMX - raw, tokenized and machine-annotated data - monolingual data sets (for language modeling) - search interfaces Some recent news and data sets: - EUbookshop: a large but noisy corpus (converted from PDF) - Tatoeba: a small but clean corpus with many languages - OpenSubtitles2012: an improved version of the 2011 version - coming soon: OpenSubtitles2013 - an extension of OpenSubtitles2012 - UN, MultiUN, Europarl v7: aligned for all language combinations - word alignments and phrase tables for the majority of bitexts The Web Site: http://opus.lingfil.uu.se More information: http://opus.lingfil.uu.se/trac/wiki Feedback is very welcome! And, be nice to our server! Jörg Tiedemann [email protected]<mailto:[email protected]> _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support _______________________________________________ Moses-support mailing list [email protected] http://mailman.mit.edu/mailman/listinfo/moses-support
