Cross-Lingual Event Detection via Optimized Adversarial Training

Proceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5588 - 5599 July 10-15, 2022 ©2022 Association for Computational Linguistics Cross-Lingual Event Detection via Optimized Adversarial Training Luis F. Guzman-Nateras, Minh Van Nguyen and Thien Huu Nguyen Department of Computer and Information Science University of Oregon {lfguzman,minhnv,thien}@cs.uoregon.edu Abstract In this work, we focus on Cross-Lingual Event Detection where a model is trained on data from a source language but its performance is evaluated on data from a second, target, lan- guage. Most recent works in this area have harnessed the language-invariant qualities dis- played by pre-trained Multi-lingual Language Models. Their performance, however, reveals there is room for improvement as the cross- lingual setting entails particular challenges. We employ Adversarial Language Adaptation to train a Language Discriminator to discern be- tween the source and target languages using unlabeled data. The discriminator is trained in an adversarial manner so that the encoder learns to produce refined, language-invariant representations that lead to improved perfor- mance. More importantly, we optimize the ad- versarial training process by only presenting the discriminator with the most informative sam- ples. We base our intuition about what makes a sample informative on two disparate metrics: sample similarity and event presence. Thus, we propose leveraging Optimal Transport 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 fami- lies, show the flexibility and effectiveness of our model that achieves state-of-the-art results. 1 Introduction Event Detection (ED) is an important sub-task within the broader Information Extraction (IE) task. Event detection consists of being able to identify the words, commonly referred to as triggers, that denote the occurrence of events in a sentence, and classify them into a discrete set of event types. For example, in the sentence “Jamie bought a car yesterday.”, bought is considered the trigger of a TRANSACTION:TRANSFER-OWNERSHIP1 1Event type taken from the ACE05 dataset. event type. It is a very well studied task in which there have been lots of previous research efforts that have recently been primarily deep learning- based (Nguyen and Grishman, 2015; Chen et al., 2015; Nguyen et al., 2016a,b; Sha et al., 2018; Wad- den et al., 2019; Zhang et al., 2019a; Yang et al., 2019; Nguyen and Nguyen, 2019; Zhang et al., 2020; Liu et al., 2020). Nonetheless, ED remains quite a challenging task as the context in which a trigger occurs can change its corresponding type completely. Further- more, the same event might also be expressed by entirely different words/phrases. Additionally, the vast majority of the aforementioned efforts are lim- ited to a monolingual setting — performing ED on text belonging to a single language. Alternatively, Cross-Lingual ED (CLED) pro- poses the scenario of creating models that effec- tively perform ED on data belonging to more than one language, which brings about additional chal- lenges. For instance, trigger words present in one language might …

Content truncated to 500 words