Context-Based Machine Translation

Context-Based Machine Translation Jaime Carbonell, Steve Klein, David Miller, Michael Steinbaum, Tomer Grassiany, Jochen Frey Meaningful Machines, LLC {jaime, steve, dave, mike, tomer, jochen}@meaningfulmachines.com Abstract Context-Based Machine Translation™ (CBMT) is a new paradigm for corpus- based translation that requires no parallel text. Instead, CBMT relies on a light- weight translation model utilizing a full- form bilingual dictionary and a sophisti- cated decoder using long-range context via long n-grams and cascaded overlap- ping. The translation process is enhanced via in-language substitution of tokens and phrases, both for source and target, when top candidates cannot be confirmed or re- solved in decoding. Substitution utilizes a synonym and near-synonym generator im- plemented as a corpus-based unsupervised learning process. Decoding requires a very large target-language-only corpus, and while substitution in target can be per- formed using that same corpus, substitu- tion in source requires a separate (and smaller) source monolingual corpus. Spanish-to-English CBMT was tested on Spanish newswire text, achieving a BLEU score of 0.6462 in June 2006, the highest BLEU reported for any language pair. Further testing also shows that quality in- creases above the reported score as the target corpus size increases and as diction- ary coverage of source words and phrases becomes more complete. 1 1 Introduction Traditional MT paradigms require either extensive transfer-rule writing by linguists and computer sci- entists or very large parallel (pre-translated) train- 1 The authors offer special acknowledgement to Eli Abir, co-founder of Meaningful Machines, who origi- nally conceived of and participated in the development of the methods described in this paper. ing corpora. The former can take person-decades to write, debug and perfect the set of rules for reason- able quality translation (e.g., at the SYSTRAN level), and acquiring parallel text for Statistical MT and Example-Based MT in sufficient quantity for comparable or better quality MT proves to be a daunting task even for major language pairs. For less-translated language pairs, accessible parallel text is simply non-existent in sufficient quantities. To address these serious challenges, CBMT is be- ing developed as a corpus-based method that re- quires neither rules nor parallel corpora. Instead, CBMT requires an extensive monolingual target- text corpus (from 50 gigabytes to 1 terabyte), a full-form bilingual dictionary, and optionally (to further improve translation quality) a smaller monolingual source-text corpus. Monolingual text can be acquired in quantities by crawling the Web, extracting pure text from HTML, and indexing such text in the manner required by the CBMT methods. The most time-consuming resource to build or acquire is the large-scale full-form bilin- gual dictionary for each language pair. CBMT exhibits two advantages over tradi- tional MT approaches: (1) higher accuracy due to the model’s ability to decode long n-grams (as evi- denced by the performance of CBMT’s still- incomplete prototype which achieved a higher BLEU score in initial testing than any other MT system to date), and (2) the ability to extend fairly rapidly to new language pairs, including those that lack sufficient parallel text. This paper describes the fundamental new techniques underlying …

Content truncated to 500 words