How STAR Transit NXT can help translators measure and increase their MT post-editing efficiency

Filtering and rescoring the CCMatrix corpus for Neural Machine Translation training Antoni Oliver, Sergi ´Alvarez Universitat Oberta de Catalunya (UOC) {aoliverg,salvarezvid}@uoc.edu Abstract There are several parallel corpora available for many language pairs, such as CCMa- trix, built from mass downloads of web content and automatic detection of seg- ments in one language and the transla- tion equivalent in another. These tech- niques can produce large parallel corpora, but of questionable quality. In many cases, the segments are not in the required lan- guages, or if they are, they are not transla- tion equivalents. In this article, we present an algorithm for filtering out the segments in languages other than the required ones and re-scoring the segments using SBERT. A use case on the Spanish–Asturian and Spanish–Catalan CCMatrix corpus is pre- sented. 1 Introduction 1.1 Parallel corpora crawled from the web There are several web-derived very large parallel corpora available for a high number of language pairs. Paracrawl1 (Ba˜n´on et al., 2020) is a paral- lel corpus created crawling the web searching for multilingual pages. At the moment it offers par- allel corpora from English to 38 languages and 6 additional language pairs not including English. Wikimatrix2 (Schwenk et al., 2021a) is created us- ing Wikipedia to automatically find translated sen- tences. It includes 96 languages, totalling 16,720 language pairs. CCAligned3 (El-Kishky et al., 2020) is a corpus formed by parallel or compara- ble web-document pairs in 137 languages aligned © 2023 The authors. This article is licensed under a Creative Commons 4.0 licence, no derivative works, attribution, CC- BY-ND. 1https://paracrawl.eu/ 2https://github.com/facebookresearch/LASER/ tree/main/tasks/WikiMatrix 3https://www.statmt.org/cc-aligned/ with English. From this document corpus, paral- lel segments are extracted using similarity scores of LASER4 (Artetxe and Schwenk, 2019) embed- dings from the document pairs. OSCAR5 (Abadji et al., 2022) is also a parallel corpus crawled from the web covering 166 languages. The CCMatrix6 (Schwenk et al., 2021b) corpus has the particular- ity that no document information has been used. Instead, all the segments in a given language are compared with all the segments in another lan- guage in order to detect parallel segments. To do so, they also use LASER and calculate a margin score, defined as the ratio between the cosine dis- tance between the two sentence embeddings, and the average cosine similarity of its nearest neigh- bours in both directions. This results in very large parallel corpora for 90 languages, totalling 1,197 language pairs. Some of these corpora, and CCMatrix in par- ticular, suffer from low quality, especially for lan- guage pairs with fewer resources. Two main prob- lems are easily detected by a simple visual inspec- tion: segments are not in the correct language, and source and target segments are not translation equivalents. In this paper we present a program that verifies the languages and assesses the transla- tion equivalence of the source and target segments. We evaluate the performance of the program on the CCMatrix corpus for Spanish–Asturian and Spanish–Catalan. 1.2 Automatic language detection Several language detection libraries implemented in Python …

Content truncated to 500 words