An automatic model and Gold Standard for translation alignment of Ancient Greek

Proceedings of the 13th Conference on Language Resources and Evaluation (LREC 2022), pages 5894–5905 Marseille, 20-25 June 2022 © European Language Resources Association (ELRA), licensed under CC-BY-NC-4.0 5894 An automatic model and Gold Standard for translation alignment of Ancient Greek Tariq Yousef∗, Chiara Palladino†, Farnoosh Shamsian∗ Anise d’Orange Ferreira⋆, Michel Ferreira dos Reis⋆ ∗University of Leipzig Augustusplatz 10, 04109 Leipzig, Germany tariq.yosef@uni-leipzig.de, farnoosh.shamsian@uni-leipzig.de †Furman University 3300 Poinsett Highway, 29613, Greenville SC, USA chiara.palladino@furman.edu ⋆Universidade Estadual Paulista (UNESP) Rod. Araraquara-Ja´u Km 1 - Bairro dos Machados Machados - Araraquara/SP - CEP 14800-901, Brazil anise.ferreira@unesp.br, michelfereis@yahoo.com.br Abstract This paper illustrates a workflow for developing and evaluating automatic translation alignment models for Ancient Greek. We designed an annotation Style Guide and a gold standard for the alignment of Ancient Greek-English and Ancient Greek-Portuguese, measured inter-annotator agreement and used the resulting dataset to evaluate the performance of various translation alignment models. We proposed a fine-tuning strategy that employs unsupervised training with mono- and bilingual texts and supervised training using manually aligned sentences. The results indicate that the fine-tuned model based on XLM-Roberta is superior in performance, and it achieved good results on language pairs that were not part of the training data. Keywords: Translation Alignment, Gold Standard, Alignment Guidelines, Ancient Greek 1. Introduction Word alignment is defined as the operation of compar- ing two or more texts in order to find correspondences between their textual units. When the texts being com- pared are in different languages (also called parallel texts or parallel corpora), the task is more specifically called translation alignment. The result often takes the form of a list of pairs of items, which can be larger text chunks like documents or paragraphs, but more frequently sentences and words (Kay and R¨oscheisen, 1993; V´eronis, 2000). Translation alignment is a very important task in Natural Language Processing. Since the 90s, many models have been developed to auto- matically establish correspondences between corpora in different languages (Brown et al., 1993), and par- allel corpora are used for a variety of purposes, includ- ing neural and statistical machine translation (DeNero and Klein, 2007), automatic bilingual lexicon extrac- tion (Yousef et al., 2022), corpus linguistics (Baker, 2000), language learning (Palladino et al., 2021), and cross-lingual annotation projection (David et al., 2001; Pad´o and Lapata, 2009; M¨uller, 2017; Nicolai and Yarowsky, 2019). However, developing reliable mod- els for automatic alignment is a challenging task: cul- tural, contextual, and linguistic differences make much of the difficulty in establishing perfect correspondences across languages. Therefore, corpora of reliably man- ually aligned texts or gold standards (and guidelines to create them) are a much-desired resource, since they are often essential to train and evaluate automatic align- ment models (Dagan et al., 1999; Grac¸a et al., 2008; Lambert et al., 2005; Mareˇcek, 2008). To our best knowledge, most current studies on auto- matic translation alignment and alignment gold stan- dards are conducted on modern languages. In this pa- per, we will illustrate a workflow for evaluating trans- lation alignment models starting from Ancient …

Content truncated to 500 words