Hybrid Knowledge Transfer for Improved Cross-Lingual Event Detection via Hierarchical Sample Selection

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 5414–5427 July 9-14, 2023 ©2023 Association for Computational Linguistics Hybrid Knowledge Transfer for Improved Cross-Lingual Event Detection via Hierarchical Sample Selection Luis F. Guzman-Nateras1, Franck Dernoncourt2, and Thien Huu Nguyen1 1 Department of Computer Science, University of Oregon, Eugene, OR, USA 2 Adobe Research, Seattle, WA, USA {lfguzman,thien}@cs.uoregon.edu, franck.dernoncourt@adobe.com Abstract In this paper, we address the Event Detection task under a zero-shot cross-lingual setting where a model is trained on a source language but evaluated on a distinct target language for which there is no labeled data available. Most recent efforts in this field follow a direct trans- fer approach in which the model is trained using language-invariant features and then di- rectly applied to the target language. However, we argue that these methods fail to take advan- tage of the benefits of the data transfer approach where a cross-lingual model is trained on target- language data and is able to learn task-specific information from syntactical features or word- label relations in the target language. As such, we propose a hybrid knowledge-transfer ap- proach that leverages a teacher-student frame- work where the teacher and student networks are trained following the direct and data transfer approaches, respectively. Our method is com- plemented by a hierarchical training-sample selection scheme designed to address the issue of noisy labels being generated by the teacher model. Our model achieves state-of-the-art re- sults on 9 morphologically-diverse target lan- guages across 3 distinct datasets, highlighting the importance of exploiting the benefits of hy- brid transfer. 1 Introduction Event Detection (ED) is a sub-task of the encom- passing Information Extraction (IE) Natural Lan- guage Processing (NLP) task. The main objective of ED is to detect and categorize the event triggers in a sentence, i.e., the words that most clearly in- dicate the occurrence of an event. Event triggers are known to be frequently related to the verb in a sentence (Majewska et al., 2021). However, they can also be other parts of speech such as nouns or adjectives. For instance, in the sentence “The ceremony was chaired by the former Secretary of State”, an ED system should recognize former as the trigger of a Personnel:End-Position event1. Generating labeled data for IE tasks such as ED can be a long and expensive endeavor. As such, most labeled ED datasets pertain to a small set of popular languages (e.g., English, Chinese, Span- ish). In turn, labeled data is scarce or non-existent for a vast majority of languages. This imbalance in annotated data availability has prompted many research efforts into zero-shot cross-lingual trans- fer learning which attempts to transfer knowledge obtained from annotated data in a high-resource source language to a low-resource target language for which no labeled data is available. There are two predominant knowledge-transfer paradigms employed by such cross-lingual methods: Data transfer and Direct transfer. Approaches that adhere to the data transfer paradigm generate pseudo-labeled data in the target language and then …

Content truncated to 500 words