Linguistically Guided Multilingual NLP: Current Approaches, Challenges, and Future Perspectives

Abstract

The neural revolution has redefined – and many would argue, undermined – the place of traditional linguistics in natural language processing. The pace at which large unsupervised deep learning models conquer new territories of language understanding raises doubts as to the utility of rule-driven approaches and formalised linguistic knowledge in solving the challenges facing future language technology. Strikingly, the potential of leading neural architectures goes beyond solving well-defined problems and learning shortcuts to specific datasets: A growing body of evidence has revealed that they are capable of implicitly acquiring deeper linguistic knowledge directly from text. And yet, their 164success is only as certain as the availability of ever growing volumes of language data and resources, inevitable for the small elite of languages dominating the Web but far from guaranteed for thousands of others. It is clear that, to overcome data scarcity, neural models need to become smarter and more sample-efficient as language learners, strategically leveraging information available in one language to help perform a task in another. Could explicit, structured linguistic information help achieve this goal after all? From another perspective, could such human-tailored expert linguistic and symbolic knowledge provide adequate inductive biases to data-hungry neural NLP architectures, where such biases would be invaluable, especially in low-resource scenarios? In this chapter, we consider the empirical evidence for the ability of neural networks to autonomously acquire deeper linguistic understanding from data and discuss what they are still missing. To fill those gaps, we weigh the potential of typological information, intermediate dependency parsing training, linguistic knowledge transfer, and native speaker introspection to provide guidance for multilingual NLP in resource-lean scenarios.