Emerging Language Spaces Learned From Massively Multilingual Corpora

TL;DR

The paper investigates whether massively multilingual neural machine translation can learn language-independent meaning representations by treating translations as cross-lingual supervisory signals. Using a single encoder-decoder NMT with language flags trained on over 900 Bible translations, it demonstrates that a continuous language space with language embeddings emerges and that the model captures language relationships to some extent. The key findings show language clustering by family in the embedding space and the feasibility of transfer learning to unseen language pairs, albeit with limited translation adequacy due to capacity constraints. The work points toward data-driven cross-linguistic typology and suggests promising future directions for linguistic discovery using multilingual, data-rich translation signals.

Abstract

Translations capture important information about languages that can be used as implicit supervision in learning linguistic properties and semantic representations. In an information-centric view, translated texts may be considered as semantic mirrors of the original text and the significant variations that we can observe across various languages can be used to disambiguate a given expression using the linguistic signal that is grounded in translation. Parallel corpora consisting of massive amounts of human translations with a large linguistic variation can be applied to increase abstractions and we propose the use of highly multilingual machine translation models to find language-independent meaning representations. Our initial experiments show that neural machine translation models can indeed learn in such a setup and we can show that the learning algorithm picks up information about the relation between languages in order to optimize transfer leaning with shared parameters. The model creates a continuous language space that represents relationships in terms of geometric distances, which we can visualize to illustrate how languages cluster according to language families and groups. Does this open the door for new ideas of data-driven language typology with promising models and techniques in empirical cross-linguistic research?

Figures (4)

• Figure 1: Conceptual illustrations of neural machine translation and abstractions to meaning representations.
• Figure 2: Multilingual Neural MT and training data with language flags.
• Figure 3: Experimental setup: Bible translations paired with English as either source or target language are used to train one single multilingual NMT model including language flags on the source language side. Language flags are mapped onto a continuous vector space of language embeddings.
• Figure 4: Continuous language space that emerges from multilingual NMT (t-SNE plot).