Bootstrapping Translation Detection and Sentence Extraction from Comparable Corpora

Proceedings of NAACL-HLT 2016, pages 1127–1132, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics Bootstrapping Translation Detection and Sentence Extraction from Comparable Corpora Kriste Krstovski†,§ and David A. Smith‡ †Harvard-Smithsonian Center for Astrophysics, Cambridge, MA §College of Information and Computer Sciences, University of Massachusetts Amherst, Amherst, MA ‡College of Computer and Information Science, Northeastern University, Boston, MA kkrstovski@cfa.harvard.edu, dasmith@ccs.neu.edu Abstract Most work on extracting parallel text from comparable corpora depends on linguistic re- sources such as seed parallel documents or translation dictionaries. This paper presents a simple baseline approach for bootstrapping a parallel collection. It starts by observing doc- uments published on similar dates and the co- occurrence of a small number of identical to- kens across languages. It then uses fast, on- line inference for a latent variable model to represent multilingual documents in a shared topic space where it can do efficient nearest- neighbor search. Starting from the Giga- word collections in English and Spanish, we train a translation system that outperforms one trained on the WMT’11 parallel training set. 1 Introduction In statistical machine translation (SMT), the qual- ity of the translation model is highly dependent on the amount of parallel data used to build it. Paral- lel data has usually been generated through the pro- cess of human translation, which imposes signifi- cant costs when building systems for new languages and domains. To alleviate this problem, researchers have considered comparable corpora—a collection of multilingual documents that are only topically aligned but not necessary translations of each other (Fung and Cheung, 2004). While most previous ap- proaches for mining comparable corpora heavily de- pend on initializing the learning process with some translation dictionaries or parallel text, we use mul- tilingual topic models to detect document transla- tion pairs and extract parallel sentences with only minimum cross-language prior knowledge: the pub- lication dates of articles and the tendency of some vocabulary to overlap across languages. Processing only four years of Gigaword news stories in English and Spanish, we are able to outperform the WMT’11 baseline system trained on parallel News Commen- tary corpus (Table 1). 2 Prior Work on Comparable Corpora Most previous, if not all, approaches for mining comparable corpora heavily depend on bilingual re- sources, such as translation lexica, bitext, and/or a pretrained baseline MT system. This paper, in con- trast, investigates building MT systems from com- parable corpora without such resources. In a widely cited early paper, Munteanu and Marcu (2005) use a bilingual dictionary and a collection of parallel sen- tences to train IBM Model 1 and a maximum en- tropy classifier to determine whether two sentences are translations of each other. Tillmann and Xu (2009) and Smith et al. (2010) detect parallel sen- tences by training IBM Model 1 and maximum en- tropy classifiers, respectively. In later work on de- tecting sentence and phrase translation pairs, Cettolo et al. (2010) and Hoang et al. (2014) use SMT sys- tems to translate candidate documents; Quirk et al. (2007) use parallel data …

Content truncated to 500 words