1: \begin{abstract}

2: We present the \textit{Structured Weighted Violations Perceptron (SWVP)} algorithm, 

3: a new structured prediction algorithm 

4: that generalizes the Collins Structured Perceptron (CSP, \cite{CollinsPerceptron}).

5: Unlike CSP, the update rule of SWVP explicitly 

6: exploits the internal structure of the predicted labels. 

7: We prove the convergence of SWVP for linearly separable training sets, 

8: %SWVP converges to a weight 

9: %vector that separates the data, under certain conditions on the parameters of the algorithm. 

10: provide mistake and generalization bounds, 

11: %on: (a) the number of updates in the separable case; 

12: %(b) mistakes in the non-separable case; and (c) the probability to misclassify 

13: %an unseen example (generalization), 

14: and show that in the general case these bounds are tighter than those of the CSP special case.

15: In synthetic data experiments with data drawn from an HMM, various variants of SWVP 

16: substantially outperform its CSP special case. 

17: SWVP also provides encouraging initial dependency parsing results.

18: \end{abstract}

19: