Selecting Syntactic, Non-redundant Segments in Active Learning for Machine Translation

Proceedings of NAACL-HLT 2016, pages 20–29, San Diego, California, June 12-17, 2016. c⃝2016 Association for Computational Linguistics Selecting Syntactic, Non-redundant Segments in Active Learning for Machine Translation Akiva Miura†, Graham Neubig†, Michael Paul‡, Satoshi Nakamura† † Nara Institute of Science and Technology, Japan ‡ ATR-Trek Co. Ltd., Japan miura.akiba.lr9@is.naist.jp neubig@is.naist.jp michael.paul@atr-trek.co.jp s-nakamura@is.naist.jp Abstract Active learning is a framework that makes it possible to efficiently train statistical models by selecting informative examples from a pool of unlabeled data. Previous work has found this framework effective for machine trans- lation (MT), making it possible to train bet- ter translation models with less effort, partic- ularly when annotators translate short phrases instead of full sentences. However, previous methods for phrase-based active learning in MT fail to consider whether the selected units are coherent and easy for human translators to translate, and also have problems with se- lecting redundant phrases with similar con- tent. In this paper, we tackle these problems by proposing two new methods for selecting more syntactically coherent and less redun- dant segments in active learning for MT. Ex- periments using both simulation and extensive manual translation by professional translators find the proposed method effective, achiev- ing both greater gain of BLEU score for the same number of translated words, and allow- ing translators to be more confident in their translations1. 1 Introduction In statistical machine translation (SMT) (Brown et al., 1993), large quantities of high-quality bilingual data are essential to achieve high translation accu- racy. While in many cases large corpora can be col- lected, for example by crawling the web (Resnik and 1Code to replicate the experiments can be found at https://github.com/akivajp/naacl2016 any one of the preceding claims� (a) Conventional n-gram selection method (n = 4) any one of the preceding claims� (b) Proposed maximal phrase selection method any one of the preceding claims� DT� CD� IN� DT� NNS� JJ� NP� PP� NP� NP� (c) Proposed parse subtree selection method Figure 1: Conventional and proposed data selection methods Smith, 2003), in many domains or language pairs it is still necessarily to create data by hand, either by hiring professionals or crowdsourcing (Zaidan and Callison-Burch, 2011). In these cases, active learn- ing (§2), which selects which data to annotate based on their potential benefit to the translation system, has been shown to be effective for improving SMT systems while keeping the required amount of an- notation to a minimum (Eck et al., 2005; Turchi et al., 2008; Haffari et al., 2009; Haffari and Sarkar, 2009; Ananthakrishnan et al., 2010; Bloodgood and Callison-Burch, 2010; Gonz´alez-Rubio et al., 2012; Green et al., 2014). Most work on active learning for SMT, and natu- ral language tasks in general, has focused on choos- ing which sentences to give to annotators. These 20 methods generally assign priority to sentences that contain data that is potentially useful to the MT sys- tem according to a number of criteria. For exam- ple, there are methods to select sentences that con- tain phrases that are frequent in monolingual …

Content truncated to 500 words