Natural language processing and text mining technologies have experienced a revolution in the last few years, with substantial improvements in accuracy mainly due to the use of deep-learning neural networks and large pretrained models relying on huge amounts of data. Explicit representations of linguistic knowledge (such as parse trees, semantic dependencies, lexicons, linguistic rules, etc.) have lost their protagonist role in systems where neural networks perform the bulk of the task, often in an end-to-end fashion. However, it is far from guaranteed that the accuracy improvement gains from the advances in neural architectures will not plateau, as in previous occasions, highlighting the need to combine them with rich linguistic processing. Furthermore, end-to-end neural systems have limitations, especially in a context of multilingualism where low-resource languages are involved: black-box nature with limited explainability, data-induced bias, reliance on large amounts of data that may be unavailable for many of the thousands of languages existing in the world, high computational requirements, and large energy usage and contribution to global warming.