Character Mapping and Ad-hoc Adaptation: Edinburgh’s IWSLT 2020 Open Domain Translation System

Proceedings of the 17th International Conference on Spoken Language Translation (IWSLT), pages 122–129 July 9-10, 2020. c⃝2020 Association for Computational Linguistics https://doi.org/10.18653/v1/P17 122 Character Mapping and Ad-hoc Adaptation: Edinburgh’s IWSLT 2020 Open Domain Translation System Pinzhen Chen Nikolay Bogoychev Ulrich Germann School of Informatics, University of Edinburgh 10 Crichton Street, Edinburgh EH8 9AB {pinzhen.chen, n.bogoych, ulrich.germann}@ed.ac.uk Abstract This paper describes the University of Ed- inburgh’s neural machine translation systems submitted to the IWSLT 2020 open domain Japanese↔Chinese translation task. On top of commonplace techniques like tokenisation and corpus cleaning, we explore character map- ping and unsupervised decoding-time adapta- tion. Our techniques focus on leveraging the provided data, and we show the positive im- pact of each technique through the gradual im- provement of BLEU. 1 Introduction The University of Edinburgh presents its neural ma- chine translation (NMT) systems for the IWSLT 2020 open domain translation task (Ansari et al., 2020). The task requires participants to submit systems to translate between Japanese (Ja) and Chi- nese (Zh), where the sentences come from mixed domains. For training purpose, 1.96 million ex- isting sentence pairs and 59.49 million crawled sentence pairs1 are provided, making the task a high-resource one. In our experiments, we focused on three aspects: 1. Corpus cleaning which consists of hand- crafted rules and cross-entropy based meth- ods. 2. Japanese and Chinese character mapping to maximise vocabulary overlap and make em- bedding tying more intuitive. 3. Unsupervised ad-hoc adaptation during de- coding time to translate a multi-domain test set, experimented at the sentence, cluster (sub- document) and document levels. 1We mistakenly used an outdated dataset which is larger but noisier. The dataset was extracted from crawled texts with encoding issues and inconsistent handling of Japanese characters “プ” and “で”. Our techniques are mostly data-centric and each technique improves translation in terms of BLEU on our development set. In the final automatic evaluation based on 4-gram character BLEU, our systems rank 6th out of 14 for Ja→Zh and 7th out of 11 for Zh→Ja. 2 Baseline with Rule-Based Cleaning 2.1 Preprocessing We first tokenise our data at word-level, which is commonly done for the Japanese and Chinese (Bar- rault et al., 2019; Nakazawa et al., 2019). While it is unclear whether word-level or character-level models are superior (Bawden et al., 2019), word- level segmentation could resolve ambiguity and dramatically reduce sequence length. The tools we use are KyTea (Neubig et al., 2011) for Japanese and Jieba fast2 for Chinese. 2.2 Rule-based cleaning We then apply a series of rule-based cleaning oper- ations on both existing and crawled data to create baseline models. These steps are mostly inspired by submissions to the corpus filtering task at WMT 2018 (Koehn et al., 2018). The task shows that effective corpus filtering brings substantial gain in translation performance. Language identification: One way of parallel corpus filtering is to restrict source sentences to be in the source language, and target sentences to be in the target language. However, distinguishing between Japanese and Chinese, particularly short sentences, …

Content truncated to 500 words