Deep Reading and Learning
Abstract
Our project made significant progress on several subtasks of natural language processing (NLP) including, part of speech tagging, chunking, named entity recognition, co-reference resolution, linking, event detection, event-argument extraction, and script learning. The unifying theme is an algorithmic framework based on search-based structured prediction. Almost all tasks in NLP can be formulated as mapping a structured input, e.g., a sentence or a document, into a structured output, e.g., a knowledge base. The problem of learning this mapping from supervisory training data is called structured prediction. In search-based structured prediction, this mapping is constructed incrementally via heuristic search. We adapted several variations of heuristic search algorithms including greedy search, beam search, and limited discrepancy search to structured prediction, achieving state of the art results in multiple subtasks of NLP. We published our work in conferences such as ICML, AAAI, EMNLP and ACL and journals such as JAIR and JMLR.
Document Details
- Document Type
- Technical Report
- Publication Date
- Oct 01, 2017
- Accession Number
- AD1044902
Entities
People
- Prasad Tadepalli
- Thomas G. Dietterich
- Xiaoli Fern
Organizations
- Oregon State University