Abstract
Building models for realistic natural language tasks requires dealing withlong texts and accounting for complicated structural dependencies.Neural-symbolic representations have emerged as a way to combine the reasoningcapabilities of symbolic methods, with the expressiveness of neural networks.However, most of the existing frameworks for combining neural and symbolicrepresentations have been designed for classic relational learning tasks thatwork over a universe of symbolic entities and relations. In this paper, wepresent DRaiL, an open-source declarative framework for specifying deeprelational models, designed to support a variety of NLP scenarios. Ourframework supports easy integration with expressive language encoders, andprovides an interface to study the interactions between representation,inference and learning.