==============================
introduction
==============================
In this package we release the data for our new sense similarity wmfvec (100 dimensions) proposed in our ACL 2012 short paper, so that everyone can use our sense similarity in the future.  If you publish materials based on this data, please refer to the source as follows to help others to reproduce your experiments:

@INPROCEEDINGS {Guo:12,
AUTHOR		= {Weiwei Guo and Mona Diab},
TITLE		= {Learning the Latent Semantics of a Concept from its Definition},
BOOKTITLE	= {Proceedings of the 50th Annual Meeting of the Association for Computational Linguistics},
YEAR		= {2012},
}


==============================
files
==============================
There are 2 data files.  The two files are line aligned, which means the information in the same line of the 2 files belongs to the same WordNet sense.
1. wmf_vec: The file that contains WMF output. Each line contains the 100 dimension vector for a WN sense. Note that the vectors are not combined with neighbor senses' vectors.
2. neighbor.osp: In WordNet, offset+pos uniquely identifies a sense (offset means the offset of the sense in WordNet/dict/data.POS file). Hence we represent each sense as offset+pos.
   		 neighbor.osp contains the neighbor information. For example, the first line in neighbor.osp is "00001740#a: 00002098#a 05200169#n 05616246#n". It means the first sense is 00001740#a; its latent vector is stored in the first line of file "wmf_vec"; and it has three neighbors: 00002098#a 05200169#n 05616246#n.



==============================
test script
==============================
test.pl: This script returns a similarity value used in the paper. It will:
1. read the 2 files mentioned above, 
2. get the sum of the vector and its neighbor senses' vectors, as described in section 2.3, 
3. return the inner product of the resulting two vectors of input senses.

The command to run the script: perl test.pl wmf_vec neighbors.osp 1 02330245#n 05436752#n

It has 5 parameters:
the files of wmf_vec and neighbors, the weight of neighbors senses, and the two senses in offset+pos format.
The weight of neighbor senses (wnb) is used to compute the final vector of a sense (in the paper, we use wnb = 1):
Q^{new}_{\cdot,j} = Q_{\cdot,j} + \sum_k^{k\in N(j)} wnb*Q_{\cdot,k}

test case:
02330245#n(mouse#n#1) 05436752#n(gene#n#1) should return 26.999537332678.  (elesk return 68)
03793489#n(computer_mouse#n#1) 05436752#n(gene#n#1) should return 7.13896257925662.  (elesk return 80)
