Record Details
Learning greedy policies for the easy-first framework
ScholarsArchive at Oregon State University
| Field | Value |
|---|---|
| Title | Learning greedy policies for the easy-first framework |
| Names |
Xie, Jun
(creator) Fern, Xiaoli Z. (advisor) |
| Date Issued | 2014-03-04 (iso8601) |
| Note | Graduation date: 2014 |
| Abstract | Easy-first, a search-based structured prediction approach, has been applied to many NLP tasks including dependency parsing and coreference resolution. This approach employs a learned greedy policy (action scoring function) to make easy decisions first, which constrains the remaining decisions and makes them easier. This thesis studies the problem of learning greedy policies to ensure the success of the Easy-first approach. We propose a novel and principled online learning algorithm that solves an optimization problem to update the weights whenever a mistake happens so that 1) the greedy policy chooses a correct action, and 2) the change to the function is minimized (to avoid over-fitting). The proposed objective is non-convex and optimized via an efficient Concave-Convex Procedure (CCCP). We compare the proposed approach with existing learning approaches on both within-document coreference and cross-document joint entity and event coreference tasks. Results demonstrate that the proposed approach performs better than existing training regimes for Easy-first and is less susceptible to overfitting. |
| Genre | Thesis/Dissertation |
| Topic | Structured Prediction |
| Identifier | http://hdl.handle.net/1957/46790 |
