Cross-Lingual Event Detection via Optimized Adversarial Training

Abstract

In this work, we focus on Cross-Lingual Event Detection (CLED) where a model is trained on data from a source language but its performance is evaluated on data from a second, target, language. Most recent works in this area have harnessed the language-invariant qualities displayed by pre-trained Multi-lingual Language Models (MLM). Their performance, however, reveals there is room for improvement as they mishandle delicate cross-lingual instances. We leverage the use of unlabeled data to train a Language Discriminator (LD) to discern between the source and target languages. The LD is trained in an adversarial manner so that our encoder learns to produce refined, language-invariant representations that lead to improved CLED performance. More importantly, we optimize the adversarial training by only presenting the LD with the most informative samples. We base our intuition about what makes a sample informative on two disparate metrics: sample similarity and event presence. Thus, we propose using Optimal Transport (OT) as a solution to naturally combine these two distinct information sources into the selection process. Extensive experiments on 8 different language pairs, using 4 languages from unrelated families, show the flexibility and effectiveness of our model that achieves new state-of-the-art results.

Tasks

Reproductions