Concept dictionary using Eilenberg machines
Branch: master
Clone or download
Adam Kovacs
Adam Kovacs Proper tab fix
Latest commit b97275b Jan 25, 2019

README.md

4lang

This repository provides

  • the 4lang concept dictionary, which contains manually written concept definitions. (Learn more about the filelds of the tsv file here)
  • the text_to_4lang module, which creates concept graph representations from running text
  • the dict_to_4lang module, which builds more of these definitions from human-readable dictionaries

Dependencies

pymachine

Our tools require an installation of the pymachine implementation of Eilenberg-machines.

hunmorph

For lemmatization, 4lang uses the hunmorph tool, on most UNIX-based systems you can use these pre-compiled executables and models (just extract them in your 4lang directory). If they don't work on your system, you may have to download and recompile hunmorph and/or the model it uses following the instructions here. This process is quite error-prone, but please reach out to us and we'll be happy to help you!

NOTE: All remaining dependencies are required only for building 4lang graphs, so in case you only want to use the graphs we provide (e.g. for the machine similarity component of our Semeval STS system), you can skip the rest of this section and continue to download pre-compiled graphs.

Stanford Parser, CoreNLP, jython

For parsing dictionary definitions, 4lang requires the Stanford Dependency Parser. Additionally, text_to_4lang.py requires the Stanford CoreNLP toolkit for parsing and coreference resolution, while the dict_to_4lang tool requires jython for customized parsing via the Stanford Parser API. Both tools require a copy of the RNN-based parser model for English, which is distributed alongside the Stanford Parser.

Currently, text_to_4lang requires the installation of the corenlp-server package. Just download the repository and follow the instructions in its README to build the package and start the server (mvn package; mvn exec:java -D server), the text_to_4lang module will then be able to connect.

After downloading and installing these tools, all you need to do is edit the stanford and corenlp sections of the default configuration file conf/default.cfg so that the relevant fields point to your installations of each tool and your copy of the englishRNN.ser.gz model (more on config files below).

Downloading pre-compiled graphs

We provide serialized machine graphs built from 4lang definitions as well as from the English Wiktionary (using the dict_to_4lang module). Unpacking this archive in your 4lang directory will place them in the data/machines directory, which is the default location for compiled machine graphs.

Environment variables

The location of your installations of the above third-party tools, as well as 4lang must be specified via environment variables. These variables must always be set, there are no fallback values to avoid strange bugs. Here's an example of a bashrc file setting all required variables:

export FOURLANGPATH=/home/recski/projects/4lang
export JYTHONPATH=/home/recski/projects/jython/jython/bin/jython
export STANFORDPATH=/home/recski/projects/stanford_dp
export MAGYARLANCPATH=/home/recski/projects/4lang/magyarlanc
export HUNTOOLSBINPATH=/home/recski/sandbox/huntools_binaries

Note that the JYTHONPATH variable must point to the jython binary directly (and not a directory), since various jython installations may have different directory structures.

Usage

Semeval STS

To use 4lang from our Semeval STS system you just need to edit the 4langpath and hunmorph_path attributes in your semeval config file so that they point to your 4lang directory and the downloaded hunmorph binaries, respectively.

Dict_to_4lang and Text_to_4lang

To run each module on small test datasets, simply run

python src/dict_to_4lang.py
python src/text_to_4lang.py

Both tools can be configured by editing a copy of conf/default.cfg and running

python src/dict_to_4lang.py MY_CONFIG_FILE

to build 4lang-style definitions from a monolingual dictionary such as Wiktionary or Longman

cat INPUT_FILE | python src/text_to_4lang.py MY_CONFIG_FILE

to create concept graphs from running English text

The config file

Contact

This repository is maintained by Gábor Recski. Questions, suggestions, bug reports, etc. are very welcome and can be sent by email to recski at aut bme hu.

Publications

If you use the 4lang module, please cite:

@inproceedings{Kornai:2015a,
    author    = {Kornai, Andr\'as  and  \'{A}cs, Judit  and  Makrai, M\'{a}rton  and  Nemeskey, D\'{a}vid M\'{a}rk  and  Pajkossy, Katalin  and  Recski, G\'{a}bor},
    title     = {Competence in lexical semantics},
    booktitle = {Proceedings of the Fourth Joint Conference on Lexical and Computational Semantics},
    month     = {June},
    year      = {2015},
    address   = {Denver, Colorado},
    publisher = {Association for Computational Linguistics},
    pages     = {165--175},
    url       = {ht