Train Monolingual, Infer Bilingual

Alaeddin Selçuk Gürel, Aydın Gerek

Keywords: Cross-Lingual, NLP, Text Classification, Turkish, English, low-resource, inference, joint inference

TL;DR: We show performance increases in text classification for a monolingually fine-tuned model by jointly inferring in bilingual test set.

Abstract:

Cross-lingual transfer learning has been studied at depth. While many methods
have been developed for pretraining or fine-tuning on monolingual, multilingual
and parallel corpora with the purpose of predicting on a low-resource monolingual
test set; in this paper we investigate the feasibility of training a text classifier on a
monolingual training set and predicting on a parallel test set, jointly utilizing both
languages at inference time only.