Fast and highly parallelizable phrase table for statistical machine translation

Proceedings of the First Conference on Machine Translation, Volume 1: Research Papers, pages 102–109, Berlin, Germany, August 11-12, 2016. c⃝2016 Association for Computational Linguistics Fast and highly parallelizable phrase table for statistical machine translation Nikolay Bogoychev The University of Edinburgh 11 Crichton St, Edinburgh EH8 9LE United Kingdom n.bogoych@ed.ac.uk Hieu Hoang Moses Machine Translation CIC hieu@moses-mt.org Abstract Speed of access is a very important prop- erty for phrase tables in phrase based sta- tistical machine translation as they are queried many times per sentence. In this paper we present a new standalone phrase table, optimized for query speed and mem- ory locality. The phrase table is cache free and can optionally incorporate a reorder- ing table within. We are able to achieve two times faster decoding by using our phrase table in the Moses decoder in place of the current state-of-the-art phrase ta- ble solution without sacrificing translation quality. Using a new, experimental version of Moses we are able to achieve 10 times faster decoding using our novel phrase ta- ble. 1 Introduction Phrase tables are the most basic component of a statistical machine translation decoder, containing the parallel phrases necessary to perform phrase- based machine translation. Due to the noisy na- ture of phrase extraction and the large phrase vo- cabulary, phrase tables’ size can reach hundreds of gigabytes in size. Lopez (2008) describes phrase tables of size of half of terabyte. A decade ago it was prohibitively expensive for a phrase table of this size to reside in memory, even if hardware supported it: a gigabyte of RAM back in 2006 costed about a 100 USD, compared to 5 USD in 2016. Because of that for a long time Machine Translation was considered a big data problem and the engineering efforts were focused on reducing the model size. This lead to the creation of sev- eral binary phrase table implementations that tack- led the memory usage problem: Zens and Ney (2007) and Junczys-Dowmunt (2012b) developed memory mapped phrase tables which also reduce memory usage using specific datastructures. The former uses a trie (Fredkin, 1960) and the latter uses specific for the purpose phrasal rank encod- ing. Lopez (2007) and Germann (2015) developed suffix array based phrase tables, which work di- rectly with the parallel corpora in order to enable easier addition of new data, avoid long binariza- tion times and keep memory usage low, but tradi- tional precomputed phrase tables offer better per- formance. RAM prices have dropped 20 times over the past 10 years and high performance server machines have hundreds of gigabytes of memory. For those machines it is no longer needed to sac- rifice query performance in favour of compression techniques such as the one in Junczys-Dowmunt (2012a). Furthermore the machines nowadays are highly parallel and locking caches which didn’t hurt performance in the past now prevent imple- mentations from scaling. We have designed a new phrase table called ProbingPT based on lin- ear probing hash (Heafield, 2011) for storage and lock-free querying, in order …

Content truncated to 500 words