Interpreting and Grounding Pre-trained Representations for Natural Language Processing

"Interpreting and Grounding Pre-trained Representations for Natural Language Processing" is a collaboration project between Chalmers, Linköping University and Recorded Future. The project is financed by the Swedish AI-program WASP (Wallenberg AI, Autonomous Systems and Software Program), which offers a graduate school with research visits, partner universities, and visiting lecturers.

Project description

Building computers that understand human language is one of the central goals in artificial intelligence. A recent breakthrough on the way towards this goal is the development of neural models that learn deep contextualized representations of language. However, while these models have substantially advanced the state of the art in natural language processing (NLP) for a wide range of tasks, our understanding of the learned representations and our repertoire of techniques for integrating them with other knowledge representations and reasoning facilities remain severely limited. To address these gaps, we will develop new methods for the interpretation, grounding, and integration of deep contextualized representations of language, and to evaluate the usefulness of these methods in the context of threat intelligence applications together with our industrial partner, Recorded Future.

While the shift from symbolic, hand-crafted representations to neural, learned representations induced from unannotated data has led to a massive improvement in performance of NLP applications, there are at least two fundamental theoretical and practical limitations to this approach:

• Lack of interpretability. Distributed representations are difficult to interpret by humans, which makes it hard to validate them and interface them with systems that use symbolic representations. Interpretability is also important for reasons of fairness, privacy, reliability, and trust (Doshi- Velez and Kim, 2017), particularly in applications that, in addition to a prediction, also require a justification of why a certain prediction was made.
• Lack of grounding. Closely related to the first problem, most current training methods for word and sentence representations cannot take advantage of existing knowledge sources such as lexical databases and ontologies, but must rely instead on textual input exclusively. Text-external knowledge however is crucial in a large number of knowledge-heavy applications such as entity linking, relation extraction, and also the application that we target in this project, threat intelligence.

These related problems define the research agenda for this project.