Zipporah: a Fast and Scalable Data Cleaning System for Noisy Web-Crawled Parallel Corpora

Proceedings of the 2017 Conference on Empirical Methods in Natural Language Processing, pages 2945–2950 Copenhagen, Denmark, September 7–11, 2017. c⃝2017 Association for Computational Linguistics Zipporah: a Fast and Scalable Data Cleaning System for Noisy Web-Crawled Parallel Corpora Hainan Xu Philipp Koehn Department of Compute Science, Center for Language and Speech Processing, Johns Hopkins University, U.S.A., 21218 hxu31@jhu.edu phi@jhu.edu Abstract We introduce Zipporah, a fast and scal- able data cleaning system. We propose a novel type of bag-of-words translation fea- ture, and train logistic regression models to classify good data and synthetic noisy data in the proposed feature space. The trained model is used to score parallel sen- tences in the data pool for selection. As shown in experiments, Zipporah selects a high-quality parallel corpus from a large, mixed quality data pool. In particular, for one noisy dataset, Zipporah achieves a 2.1 BLEU score improvement with using 1/5 of the data over using the entire corpus. 1 Introduction Statistical machine translation (SMT) systems re- quire the use of parallel corpora for training the internal model parameters. Data quality is vital for the performance of the SMT system (Simard, 2014). To acquire a massive parallel corpus, many researchers have been using the Internet as a re- source, but the quality of data acquired from the Internet usually has no guarantee, and data clean- ing/data selection is needed before the data is used in actual systems. Usually data cleaning refers to getting rid of a small amount of very noisy data from a large data pool, and data selection refers to selecting a small subset of clean (or in-domain) data from the data pool; both have the objective of improving translation performances. For practi- cal purposes, it is highly desirable to perform data selection in a very fast and scalable manner. In this paper we introduce Zipporah1, a fast and scal- able system which can select an arbitrary size of good data from a large noisy data pool to be used in SMT model training. 1https://github.com/hainan-xv/zipporah 2 Prior Work Many researchers have studied the data clean- ing/selection problem. For data selection, there have been a lot of work on selecting a sub- set of data based on domain-matching. Duh et al. (2013) used a neural network based lan- guage model trained on a small in-domain cor- pus to select from a larger data pool. Moore and Lewis (2010) computed cross-entropy between in- domain and out-of-domain language models to se- lect data for training language models. XenC (Rousseau, 2013), an open-source tool, also se- lects data based on cross-entropy scores on lan- guage models. Axelrod et al. (2015) utilized part- of-speech tags and used a class-based n-gram lan- guage model for selecting in-domain data. There are a few works that utilize other metrics. L¨u et al. (2007) redistributed different weights for sen- tence pairs/predefined sub-models. Shah and Spe- cia (2014) described experiments on quality esti- mation which, given a source sentence, select the best translation among several options. The qe- clean system (Denkowski et …

Content truncated to 500 words