Difference between revisions of "Margin Infused Relaxed Algorithm"
| Line 13: | Line 13: | ||
<math>\forall t, \quad (f^{t}, \mathbf{e}^t) \in \mathcal{T} </math> and a list of <math> m </math> -best oracles <math> \mathcal{O}^t </math>: | <math>\forall t, \quad (f^{t}, \mathbf{e}^t) \in \mathcal{T} </math> and a list of <math> m </math> -best oracles <math> \mathcal{O}^t </math>: | ||
A <math>k</math> -best list of candidates is generated by <math> best_k(\cdot) </math> using the current weight vector <math>\mathbf{w}_i</math>. Each training instance <math>f_{t}</math> can have a multiple number of correct outputs or references, <math>\mathbf{e}^t</math>, in this case target translations. | A <math>k</math> -best list of candidates is generated by <math> best_k(\cdot) </math> using the current weight vector <math>\mathbf{w}_i</math>. Each training instance <math>f_{t}</math> can have a multiple number of correct outputs or references, <math>\mathbf{e}^t</math>, in this case target translations. | ||
| + | |||
| + | |||
| + | Using the <math>k</math>-best list, <math>m</math>-best oracle translations <math>\mathcal{O}^t</math> is updated by <math>oracle_{m}(\cdot)</math> in each iteration. | ||
Revision as of 18:58, 29 September 2011
This method is used by Watanabe et al., EMNLP 2007 to train an MT system a with a very large number of features of the order of millions. The training step was performed using a specific algorithm called the Margin Infused Relaxed Algorithm (MIRA) proposed by Crammer et al., 2006
Summary
MIRA is an online large-margin training algorithm which updates the weight vector according to certain margin constraints and loss function.
It is used to learn the weights of features after processing each training instance similar to structured perceptron algorithm with an additional loss function and margin constraint in its update rule.
Definition
A general definition of online training algorithms can be written down as follows:
and a list of -best oracles : A -best list of candidates is generated by using the current weight vector . Each training instance can have a multiple number of correct outputs or references, , in this case target translations.
Using the -best list, -best oracle translations is updated by in each iteration.
