Joint Word Representation Learning using a Corpus and a Semantic Lexicon
This software implements the joint word representation learning method proposed in the AAAI-2016 submission titled "Joint Word Representation Learning using a Corpus and a Semantic Lexicon".
-
The source code is in ./src/reps.cc and is written in C++. You will need C++0x supported compiler to compile the source.
-
The code uses the Eigen library. (http://eigen.tuxfamily.org/index.php?title=Main_Page) If you have your path variables properly configured in the C++ compiler, it should be able to read the library from the current location, without having to install the Eigen library. If you do not wish to install Eigen, then simply copy the source code of Eigen into a directory named 'eigen' at the same level as 'src'.
The code is implemented and tested only on Ubtuntu Linux. It might not compile/run on other platforms without further configurations.
Go to ./src and type make
A small co-occurrence matrix containing 100k elements is provided in ./work/edges.sample. A Makefile is provided that trains word representations using synonym relation. To run this type
make run which will internall call
./reps --dim=300 --epohs=20 --model=../work/model --alpha=0.01 --lmda=10000 --edges=../work/edges.sample --pairs=../work/synonyms
Here, --dim is the dimensionality of the word vectors,
--epohs is the maximum number of iterations,
--alpha is the initial learning rate in AdaGrad,
--lmda is the regularization coefficient in the proposed method,
--edges is the co-occurrence matrix, and
--pairs is the lexicon file (8 relation types provided in ./work)
The learnt word representations will be written to ./work/model, specified by --model.