README.md

This implementation is based on the cnn library for this software to function. The paper is "Shift-Reduce Constituent Parsing with Neural Lookahead Features".

Building

mkdir build
cd build
cmake .. -DEIGEN3_INCLUDE_DIR=/path/to/eigen
make    

Preprocessing

You can get the constituent hierarchy by scripts/conhier_s.py for s-type constituent hierarchy and scripts/conhier_e.py for e-type constituent hierachy

./scripts/conhier_s.py [training data in bracketed format] > [s-type training data]
./scripts/conhier_e.py [training data in bracketed format] > [e-type training data]

./scripts/conhier_s.py [development data in bracketed format] > [s-type development data]
./scripts/conhier_e.py [development data in bracketed format] > [e-type development data]

./scripts/conhier_s.py [test data in bracketed format] > [s-type test data]
./scripts/conhier_e.py [test data in bracketed format] > [e-type test data]

The directory data contains the related data.

Training

./build/impl/s-hierarchy-trainer train.s dev.s 
./build/impl/e-hierarchy-trainer train.e dev.e

Decoding

./build/impl/s-hierarchy-trainer train.s test.s [s_model]
./build/impl/e-hierarchy-decoder train.e test.e [e_model]

It will automatically generate the output file test.sOUT and test.eOUT, respectively, which then can be used as extra features on the lookahead implementation of ZPar

Usage of ZPar

./scripts/combine.py [.sOUT] [.eOUT] > [extra feature]

You can follow the ZPar instruction to complie constituent parser with implementation of "jiangming".

./conparser [in] [out] zpar_model [extra feature]

Citation

@article{TACL927,
    author = {Liu, Jiangming  and Zhang, Yue },
    title = {Shift-Reduce Constituent Parsing with Neural Lookahead Features},
    journal = {Transactions of the Association for Computational Linguistics},
    volume = {5},
    year = {2017},
    issn = {2307-387X},
    url = {https://transacl.org/ojs/index.php/tacl/article/view/927},
    pages = {45--58}
}