Capitalizing Machine Translation

Proceedings of the Human Language Technology Conference of the North American Chapter of the ACL, pages 1–8, New York, June 2006. c⃝2006 Association for Computational Linguistics Capitalizing Machine Translation Wei Wang and Kevin Knight and Daniel Marcu Language Weaver, Inc. 4640 Admiralty Way, Suite 1210 Marina del Rey, CA, 90292 {wwang, kknight, dmarcu}@languageweaver.com Abstract We present a probabilistic bilingual capi- talization model for capitalizing machine translation outputs using conditional ran- dom fields. Experiments carried out on three language pairs and a variety of ex- periment conditions show that our model significantly outperforms a strong mono- lingual capitalization model baseline, es- pecially when working with small datasets and/or European language pairs. 1 Introduction Capitalization is the process of recovering case in- formation for texts in lowercase. It is also called truecasing (Lita et al., 2003). Usually, capitalization itself tries to improve the legibility of texts. It, how- ever, can affect the word choice or order when inter- acting with other models. In natural language pro- cessing, a good capitalization model has been shown useful for tasks like name entity recognition, auto- matic content extraction, speech recognition, mod- ern word processors, and machine translation (MT). Capitalization can be viewed as a sequence la- beling process. The input to this process is a sen- tence in lowercase. For each lowercased word in the input sentence, we have several available cap- italization tags: initial capital (IU), all uppercase (AU), all lowercase (AL), mixed case (MX), and all having no case (AN). The output of capital- ization is a capitalization tag sequence. Associ- ating a tag in the output with the corresponding lowercased word in the input results in a surface form of the word. For example, we can tag the input sentence “click ok to save your changes to /home/doc.” into “click IU ok AU to AL save AL your AL changes AL to AL /home/doc MX . AN”, getting the surface form “Click OK to save your changes to /home/DOC .”. A capitalizer is a tagger that recovers the capi- talization tag for each input lowercased word, out- putting a well-capitalized sentence. Since each low- ercased word can have more than one tag, and as- sociating a tag with a lowercased word can result in more than one surface form (e.g., /home/doc MX can be either /home/DOC or /home/Doc), we need a capitalization model to solve the capitalization am- biguities. For example, Lita et al. (2003) use a tri- gram language model estimated from a corpus with case information; Chelba and Acero (2004) use a maximum entropy Markov model (MEMM) com- bining features involving words and their cases. Capitalization models presented in most previ- ous approaches are monolingual because the models are estimated only from monolingual texts. How- ever, for capitalizing machine translation outputs, using only monolingual capitalization models is not enough. For example, if the sentence “click ok to save your changes to /home/doc .” in the above example is the translation of the French sentence “CLIQUEZ SUR OK POUR ENREGISTRER VOS …

Content truncated to 500 words