Transfer Learning for Adaptive Relation Extraction
Abstract
This project addressed the relation extraction problem in resource-poor domains. This problem is one of extracting information regarding relationships between entities, which involves extracting text spans specifying the entities involved and classifying the relationship between these entities into one of several pre-defined classes. The primary technique for solving such problems is machine learning. An obstacle to the widespread application of machine-learning methods is the necessity for relatively large amounts of annotated training data: while free text is abundantly available, it is costly to employ humans to annotate the free text with information that should be extracted from it. The focus of this project was research into technologies that could enable relation extraction systems to be quickly adapted to resource-poor domains. We applied (1) transfer learning approaches for transferring the models learned in one domain to new domains, and (2) structured learning approaches that could remove the need for powerful pre-processing modules (e.g. parsers, named entity recognizers) that may not be available in resource-poor domains.
Document Details
- Document Type
- Technical Report
- Publication Date
- Sep 13, 2011
- Accession Number
- ADA549594
Entities
People
- Hai Leong Chieu
- Jing Jiang
- Wee Sun Lee
Organizations
- Defence Science Organisation