Low-resource Cross-lingual Event Type Detection via Distant Supervision with Minimal Effort

Low-resource Cross-lingual Event Type Detection in Documents via Distant Supervision with Minimal Effort Aldrian Obaja Muis Naoki Otani Nidhi Vyas Ruochen Xu Yiming Yang Teruko Mitamura Eduard Hovy Language Technology Institute Carnegie Mellon University Pittsburgh, PA {amuis, notani, nkvyas, ruochenx, yiming, teruko, ehovy}@andrew.cmu.edu Abstract The use of machine learning for NLP generally requires resources for training. Tasks performed in a low-resource language usually rely on labeled data in another, typically resource-rich, lan- guage. However, there might not be enough labeled data even in a resource-rich language such as English. In such cases, one approach is to use a hand-crafted approach that utilizes only a small bilingual dictionary with minimal manual verification to create distantly supervised data. Another is to explore typical machine learning techniques, for example adversarial training of bilingual word representations. We find that in event-type detection task—the task to classify [parts of] documents into a fixed set of labels—they give about the same performance. We ex- plore ways in which the two methods can be complementary and also see how to best utilize a limited budget for manual annotation to maximize performance gain. 1 Introduction For most languages of the world, few or no language processing tools or resources exist (Baumann and Pierrehumbert, 2014). This hinders efforts to apply certain language technologies enjoyed by languages like English, in which much current research is done. To perform natural language processing tasks in resource-poor languages, one way to overcome data scarcity is to tap on resources from another resource-rich language. Assuming that there are already good resources and tools to solve the same tasks in the more resource-rich language (henceforth, aux- iliary language), the only remaining challenge is to transfer the learning process into the resource-poor language (henceforth, target language) and adapt it to the specifics of that language. One way to do this is to build a shared word representation across the two languages and train an NL engine on this shared representation, perhaps using an adversarial domain adaptation approach to handle the domain (language) shift (Chen et al., 2017). Usually, these approaches assume the availability of large labeled data in the auxiliary language, on the order of hundred thousands to millions. However, for some more complex or specialized tasks, there might not be enough available training data even in a resource-rich language such as English. A case in point is the event-type classification task over the publicly available datasets, such as ACE 20051 and TAC KBP 2 datasets, which usually contain only a few hundred to a few thousand documents. The situation frame (SF) detection task is one example of event-type classification task, where the objective is to extract from each document one or more situation frames with their corresponding arguments. A situation frame (SF) is either an issue being described in the articles, such as civil unrest or terrorism, or a need situation such as the need for water or medical aid. In our task there are 11 situation frame types, each associated with a set of …

Content truncated to 500 words