Differences
This shows you the differences between two versions of the page.
Next revision | Previous revision | ||
courses:rg:non-projective-dependency-parsing-using-spanning-tree-algorithms [2011/04/18 14:22] abzianidze vytvořeno |
courses:rg:non-projective-dependency-parsing-using-spanning-tree-algorithms [2011/04/19 12:09] abzianidze |
||
---|---|---|---|
Line 14: | Line 14: | ||
* In the paper, a non-projective parser is understood as a parser allowing non-projective dependency parse trees along with projective dependency parse trees, in contrast to a projective parser, which forbids non-projective dependency parse trees. Also under the space of non-projective trees, authors mean the union space of both non-projective and projective trees. | * In the paper, a non-projective parser is understood as a parser allowing non-projective dependency parse trees along with projective dependency parse trees, in contrast to a projective parser, which forbids non-projective dependency parse trees. Also under the space of non-projective trees, authors mean the union space of both non-projective and projective trees. | ||
* The score of dependency trees are commonly represented as the sum of the scores of all edges in the tree, and the score of an edge is a dot product weight vector and feature vector (containing information about nodes - words). | * The score of dependency trees are commonly represented as the sum of the scores of all edges in the tree, and the score of an edge is a dot product weight vector and feature vector (containing information about nodes - words). | ||
- | * Chu-Liu-Edmonds algorithm for finding maximum spanning trees takes in general < | + | * Chu-Liu-Edmonds algorithm for finding maximum spanning trees takes in general < |
- | * In the training phase, two modified versions of the Margin Infused relaxed Algorithm (MIRA) | + | * In the training phase, two modified versions of the Margin Infused relaxed Algorithm (MIRA) |
* Experiments are done on the Czech PDT. In particular, on entire PDT (Czech-A) and on 23% portion of PDT including only non-projective dependency trees (Czech-B). The introduced algorithm is competing to other 3 dependency parsers (2 projective and 1 pseudo-projective). | * Experiments are done on the Czech PDT. In particular, on entire PDT (Czech-A) and on 23% portion of PDT including only non-projective dependency trees (Czech-B). The introduced algorithm is competing to other 3 dependency parsers (2 projective and 1 pseudo-projective). | ||
* Chu-Liu-Edmonds MST algorithm with Factored MIRA shows the best results on both Czech-A and Czech-B, and slightly lower results than McD2005 projective parser on English projective dependency trees. | * Chu-Liu-Edmonds MST algorithm with Factored MIRA shows the best results on both Czech-A and Czech-B, and slightly lower results than McD2005 projective parser on English projective dependency trees. |