Exploiting Sentence Order in Document Alignment

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 5997–6007, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 5997 Exploiting Sentence Order in Document Alignment Brian Thompson Johns Hopkins University brian.thompson@jhu.edu Philipp Koehn Johns Hopkins University phi@jhu.edu Abstract We present a simple document alignment method that incorporates sentence order in- formation in both candidate generation and candidate re-scoring. Our method results in 61% relative reduction in error compared to the best previously published result on the WMT16 document alignment shared task. Our method improves downstream MT perfor- mance on web-scraped Sinhala–English docu- ments from ParaCrawl, outperforming the doc- ument alignment method used in the most re- cent ParaCrawl release. It also outperforms a comparable corpora method which uses the same multilingual embeddings, demonstrating that exploiting sentence order is beneficial even if the end goal is sentence-level bitext. 1 Introduction Document alignment is the task of finding parallel document pairs (i.e., documents that are transla- tions of each other) in a large collection of doc- uments, often crawled from the web. Aligned documents have historically been used to produce sentence-level machine translation (MT) data, but there is growing evidence that MT systems should be trained and evaluated using document-level con- text (Gong et al., 2011; L¨aubli et al., 2018; Voita et al., 2019; Junczys-Dowmunt, 2019). We exploit the simple idea that two parallel docu- ments should each contain approximately the same information, in approximately the same order. This idea can be traced back at least to the late 1990s, when STRAND (Resnik, 1998) measured how well linearized HTML tags from two documents could be aligned in order to judge whether two web pages were likely parallel. However, more recent work has primarily used unordered representations for documents, including bags of words or n-gram fea- tures and averages of sentence embeddings. Our method consists of two main parts: First, we propose a simple method for candidate generation which embeds documents into a joint semantic em- bedding space (Berry and Young, 1995; Germann, 2016), in a way that preserves some order infor- mation in each document. This enables candidate generation via fast approximate nearest neighbor search. Second, we propose re-scoring those candi- date pairs by performing sentence alignment and then scoring that alignment based on (1) the seman- tic similarity of the resulting aligned sentence pairs; (2) whether the sentence pairs are in the correct languages; and (3) the number of inserted/deleted sentences. Our re-scoring approach seeks to filter out documents pairs that contain similar informa- tion, but where the order of that information is not consistent between the two documents. Our method results in a 61% relative reduction in the false positive rate on the WMT16 document alignment shared task versus the best previously reported method. Applied to web-scraped Sinhala– English data from ParaCrawl (Ban et al., 2020), it improves MT performance by 1.2 BLEU over the document alignment method used in the latest ParaCrawl release (Buck and Koehn, 2016b), when both are used with the Vecalign sentence …

Content truncated to 500 words