The Post-Edit Me! project

Unsupervised Feature Selection for Effective Parallel Corpus Filtering Mikko Aulamo, Ona de Gibert, Sami Virpioja, J¨org Tiedemann Department of Digital Humanities University of Helsinki, Helsinki / Finland {name.surname}@helsinki.fi Abstract This work presents an unsupervised method of selecting filters and threshold values for the OpusFilter parallel corpus cleaning toolbox. The method clusters sentence pairs into noisy and clean cate- gories and uses the features of the noisy cluster center as filtering parameters. Our approach utilizes feature importance analysis to disregard filters that do not differentiate between clean and noisy data. A randomly sampled subset of a given corpus is used for filter selection and ineffective filters are not run for the full corpus. We use a set of automatic evaluation metrics to assess the quality of translation models trained with data filtered by our method and data filtered with OpusFilter’s default parameters. The trained models cover English-German and English-Ukrainian in both directions. The proposed method outperforms the default parameters in all translation directions for almost all evaluation metrics. 1 Introduction Neural machine translation (NMT) is dependent on large parallel text corpora. Available train- ing data can often be noisy, especially if the data is retrieved by the common method of extract- ing bitexts from web crawls (Espl`a-Gomis et al., 2019; Schwenk et al., 2021; Ba˜n´on et al., 2020). Training NMT on noisy data can be detrimental to the translation models. Ensuring that the train- © 2023 The authors. This article is licensed under a Creative Commons 4.0 licence, no derivative works, attribution, CC- BY-ND. ing examples are clean sentence pairs leads to bet- ter translation quality and more efficient training (Khayrallah and Koehn, 2018). If clean paral- lel corpora are not readily available, a common practice is to refine a noisy corpus by filtering out low quality training examples. The amount and type of noise varies between different cor- pora. Selecting the kind of filters that are optimal for cleaning a specific parallel corpus can take a lot of trial and error. Several methods and tools for corpus cleaning have been proposed and de- veloped (Taghipour et al., 2011; Carpuat et al., 2017; Ram´ırez-S´anchez et al., 2020). OpusFilter (Aulamo et al., 2020) is one such toolkit. It pro- vides a selection of configurable filters, but suffers from the same issue of having to manually choose the filters and their parameters. In this work, we propose an unsupervised method of selecting ef- fective filters and filtering thresholds based on the properties of a given corpus. Our method automat- ically generates a filtering configuration file which serves as a solid starting point for finding the op- timal settings for an OpusFilter corpus cleaning pipeline. We assess the proposed method by com- paring the translation quality of models trained with data filtered with default parameters from OpusFilter and data filtered with autogenerated pa- rameters. Our implementation of the filter selec- tion method is available at https://github. com/Helsinki-NLP/OpusFilter. 2 Related work Corpus cleaning has been a part of training pipelines since the statistical machine translation …

Content truncated to 500 words