A Robust Self-Learning Method for Fully Unsupervised Cross-Lingual Mappings of Word Embeddings: Making the Method Robustly Reproducible as Well

Proceedings of the 12th Conference on Language Resources and Evaluation (LREC 2020), pages 5546–5554 Marseille, 11–16 May 2020 c⃝ European Language Resources Association (ELRA), licensed under CC-BY-NC 5546 A Robust Self-Learning Method for Fully Unsupervised Cross-Lingual Mappings of Word Embeddings: Making the Method Robustly Reproducible as Well Nicolas Garneau†, Mathieu Godbout†, David Beauchemin†, Audrey Durand, Luc Lamontagne Universit´e Laval 2325 Rue de l’Universit´e, Qu´ebec nicolas.garneau@ift.ulaval.ca, mathieu.godbout.3@ulaval.ca, david.beauchemin.5@ulaval.ca audrey.durand@ift.ulaval.ca, luc.lamontagne@ift.ulaval.ca Abstract In this paper, we reproduce the experiments of Artetxe et al. (2018b) regarding the robust self-learning method for fully unsupervised cross-lingual mappings of word embeddings. We show that the reproduction of their method is indeed feasible with some minor assump- tions. We further investigate the robustness of their model by introducing four new languages that are less similar to English than the ones proposed by the original paper. In order to assess the stability of their model, we also conduct a grid search over sensible hyperpa- rameters. We then propose key recommendations that apply to any research project in order to deliver fully reproducible research. Keywords: Machine Learning, Unsupervised Learning, Word Alignment, Cross-Lingual Word Embeddings, Reproducibility 1. Introduction The cross-lingual mapping of word embeddings is a prob- lem that has been studied more thoroughly with the rise of distributed representations induced from neural network architectures (Mikolov et al., 2013b). The goal of this task is to induce automatically a word-to-word translation dic- tionary D ∈ R|Vs|×|Vt| where D[i, j] = 1 means that the i-th word from the source vocabulary Vs is a translation of the j-th word in the target vocabulary Vt. The data used to learn the mapping from two different languages is two sets of word embeddings Xs and Xt corresponding to the vectors of the source and the target language respectively. The mapping from one language space to the other is usu- ally done with a projection matrix Wt which projects the embeddings of the target language in the same space as the source language (or vice versa), i.e. Xs ≈ XtWt. There are several methods available to achieve such a map- ping depending on the resources in hand. Given a dataset of parallel word-aligned data, supervised mapping-based ap- proaches are amongst the most popular to date (Mikolov et al., 2013a; Dinu and Baroni, 2014; Gouws and Søgaard, 2015). Several unsupervised methods (Yang et al., 2018; Conneau et al., 2017) based on Generative Adversarial Neural Networks (GANs) (Goodfellow et al., 2014) have also been proposed in the case where no seed dictionary is available. The method that we reproduce is the work of Artetxe et al. (2018b) (referenced as ”the authors”). It also falls in the unsupervised setting but is based on distances between nearest neighbors to build the initial seed dictionary. We re- fer the reader to the survey of Ruder et al. (2019) for more details, which provides an extensive overview of the differ- ent methods for learning cross-lingual mappings between two different word embedding spaces. On another hand, delivering reproducible research is too …

Content truncated to 500 words