Neural Models for Text: Improving Efficiency, Interpretability and Accuracy
Abstract
Modern "deep" neural models have achieved remarkably strong results in natural language processing (NLP) in recent years, achieving state-of-the-art performance on many tasks. However, a few key open problems hinder the utility of deep neural models for NLP in many real-world applications. First, such models tend to require very large volumes of training data to work well; this will not be available for many specialized domains and tasks. Second, neural models tend to be opaque Ôblack boxes . This lack of interpretability mitigates their applicability, most obviously in domains that require predictions to have a degree of transparency. This proposal seeks to address these pressing interrelated issues, in turn enabling rapid, interactive training of highly accurate and interpretable neural NLP models. This will allow wider application of cutting-edge deep learning methods for NLP. Concretely, we propose to explore methods that result in accurate neural models with minimal supervision, including active learning approaches suitable for neural architectures and approaches to incorporating domain knowledge. Furthermore, we propose novel representation learning approaches that aim to induce disentangled neural representations (embeddings) of texts and that can provide rationales for predictions. Ultimately, this work will result in interactive, interpretable and transparent neural models for NLP that can be trained with modest amounts of labeled data.
Document Details
- Document Type
- DoD Grant Award
- Publication Date
- Feb 14, 2019
- Source ID
- W911NF1810328
Entities
People
- Byron C Wallace
Organizations
- Army Contracting Command
- Northeastern University
- United States Army