Identifying Semantic Divergences in Parallel Text without Annotations

Proceedings of NAACL-HLT 2018, pages 1503–1515 New Orleans, Louisiana, June 1 - 6, 2018. c⃝2018 Association for Computational Linguistics Identifying Semantic Divergences in Parallel Text without Annotations Yogarshi Vyas and Xing Niu and Marine Carpuat Department of Computer Science University of Maryland College Park, MD 20742, USA yogarshi@cs.umd.edu, xingniu@cs.umd.edu, marine@cs.umd.edu Abstract Recognizing that even correct translations are not always semantically equivalent, we auto- matically detect meaning divergences in par- allel sentence pairs with a deep neural model of bilingual semantic similarity which can be trained for any parallel corpus without any manual annotation. We show that our seman- tic model detects divergences more accurately than models based on surface features derived from word alignments, and that these diver- gences matter for neural machine translation. 1 Introduction Parallel sentence pairs are sentences that are trans- lations of each other, and are therefore often assumed to convey the same meaning in the source and target language. Occasional mis- matches between source and target have been pri- marily viewed as alignment noise (Goutte et al., 2012) due to imperfect sentence alignment tools in parallel corpora drawn from translated texts (Tiedemann, 2011; Xu and Yvon, 2016), or the noisy process of extracting parallel segments from non-parallel corpora (Fung and Cheung, 2004; Munteanu and Marcu, 2005). In contrast, we view translation as a process that inherently introduces meaning mismatches, so that even correctly aligned sentence pairs are not nec- essarily semantically equivalent. This can hap- pen for many reasons: translation lexical choice often involves selecting between near synonyms that introduce language-specific nuances (Hirst, 1995), typological divergences lead to structural mismatches (Dorr, 1994), differences in discourse organization can make it impossible to find one- to-one sentence alignments (Li et al., 2014). Cross-linguistic variations in other discourse phe- nomena such as coreference, discourse relation and modality (Lapshinova-Koltunski, 2015) com- pounded with translation effects that distinguish “translationese” from original text (Koppel and Ordan, 2011) might also lead to meaning mis- matches between source and target. In this paper, we aim to provide empirical ev- idence that semantic divergences exist in paral- lel corpora and matter for downstream applica- tions. This requires an automatic method to distin- guish semantically equivalent sentence pairs from semantically divergent pairs, so that parallel cor- pora can be used more judiciously in downstream cross-lingual NLP applications. We propose a semantic model to automatically detect whether a sentence pair is semantically divergent (Sec- tion 3). While prior work relied on surface cues to detect mis-aligments, our approach focuses on comparing the meaning of words and overlapping text spans using bilingual word embeddings (Lu- ong et al., 2015) and a deep convolutional neural network (He and Lin, 2016). Crucially, training this model requires no manual annotation. Noisy supervision is obtained automatically borrowing techniques developed for parallel sentence extrac- tion (Munteanu and Marcu, 2005). Our model can thus easily be trained to detect semantic di- vergences in any parallel corpus. We extensively evaluate our semantically- motivated models on intrinsic and extrinsic tasks: detection of divergent examples in …

Content truncated to 500 words