Arabic syntactic word vectors
=============================
The following word vectors have been used in our paper on PP attachment, following the methods proposed by: 
Bansal et al.: http://www.aclweb.org/anthology/P14-2131.pdf 
Levy and Goldberg: http://www.aclweb.org/anthology/P14-2131.pdf 
The vectors provided here are of 100 dimensions representing word lemmas. Get in touch for vectors of other dimensions and for vectors representing word tokens.

Vectors a la Bansal et al.: words; all (including grandparent and label vectors)
Vectors a la Levy and Goldberg: words; contexts

The vectors were trained from parsed texts in the following manner. First, we converted the morphologically-processed arTenTen corpus to CoNLL format with the SPMRL shared-task scripts. Then we parsed the corpus with a second-order MST parser, trained on the CATiB dependency treebank. Finally, we created syntactic word vectors using the word2vec tool as described in the original papers, with the following differences: contrary to Levy and Goldberg, we do not collapse prepositional relations and we set a count threshold of 10. For the method of Bansal et al., we used a window size of 2 because we found it to be better in our case than their window size of 1.

If you use our Arabic syntacic vectors, please cite the following paper:

Yonatan Belinkov, Tao Lei, Regina Barzilay, and Amir Globerson. Accepted. Exploring Compositional Architectures and Word Vector Representations for Prepositional Phrase Attachment. Transactions of the Association for Computational Linguistics (TACL). 

as well as the appropriate paper that introduced the original English vectors (Bansal et al. or Levy and Goldberg).