graph-based processing of multi-level annotated corpora

corpus, linguistics, nlp, graph, networkx, annotation, conversion, converter, natural-language-processing, python
pip install discoursegraphs==0.4.14



Latest version BSD License Build status Test coverage Code Issues Docker build status

This library enables you to process linguistic corpora with multiple levels of annotations by:

  1. converting the different annotation formats into separate graphs and
  2. merging these graphs into a single multidigraph (based on the common tokenization of the annotation layers)
  3. exporting your (merged) graphs into several output formats
  4. visualizing linguistic graphs directly in an IPython notebook

Import formats

So far, the following formats can be imported and merged:

  • TigerXML (a format for representing tree-like syntax graphs with secondary edges)
  • NeGra Export Format (a format used i.a. for the TüBa-D/Z Treebank)
  • Penn Treebank format (an s-expressions/lisp/brackets format for representing syntax trees)
  • a number of formats for Rhetorical Structure Theory:
    • RS3 (a format used by RSTTool to annotate documents with Rhetorical Structure Theory)
    • the .dis "LISP" format used by the RST-DT corpus
    • URML (a format for underspecified rhetorical structure trees)
  • MMAX2 (a format / GUI tool for annotating spans and connections between them (e.g. coreferences)
  • CoNLL 2009 and CoNLL 2010 formats (used for annotating i.a. dependency parses and coreference links)
  • ConanoXML (a format for annotating connectives, used by Conano)
  • Decour (an XML format used by a corpus of DEceptive statements in Italian COURts)
  • EXMARaLDA, a format for annotating spans in spoken or written language
  • an ad-hoc plain text format for annotating expletives (you're probably not interested in)

Export formats

discoursegraphs can export graphs into the following formats / for the following tools:

  • dot format, which is used by the open source graph visualization software graphviz
  • geoff format, used by the neo4j graph database
  • GEXF and GraphML (common interchange formats for graphs used by various tools such as Gephi and Cytoscape)
  • PAULA XML 1.1, an exchange format for linguistic data (exporter is still buggy)
  • EXMARaLDA, a tool for annotating spans in spoken or written language
  • CoNLL 2009 (so far, only tokens, sentence boundaries and coreferences are exported)


This should work on both Linux and Mac OSX using Python 2.7 and either pip or easy_install.

Install from PyPI

pip install discoursegraphs # prepend 'sudo' if needed

or, if you're oldschool:

easy_install discoursegraphs # prepend 'sudo' if needed

Install from source

sudo apt-get install python-dev libxml2-dev libxslt-dev pkg-config graphviz-dev libgraphviz-dev -y
sudo easy_install -U setuptools
git clone https://github.com/arne-cl/discoursegraphs.git
cd discoursegraphs
sudo python setup.py install


The command line interface of DiscourseGraphs allows you to merge syntax, rhetorical structure, connectives and expletives annotation files into one graph and to store this graph in one of several output formats (e.g. the geoff format used by the neo4j graph database or the dot format used by the graphviz plotting tool).

discoursegraphs -t syntax/maz-13915.xml -r rst/maz-13915.rs3 -c connectors/maz-13915.xml -a anaphora/tosik/das/maz-13915.txt -o dot
dot -Tpdf doc.dot > discoursegraph.pdf # generates a PDF from the dot file

If you're interested in working with just one of those layers, you'll have to call the code directly:

import discoursegraphs as dg
tiger_docgraph = dg.read_tiger('syntax/doc.xml')
rst_docgraph = dg.read_rs3('rst/doc.rs3')
expletives_docgraph = dg.read_anaphoricity('expletives/doc.txt')

All the document graphs generated in this example are derived from the networkx.MultiDiGraph class, so you should be able to use all of its methods.


Source code documentation is available here, but you can always get an up-to-date local copy using Sphinx.

You can generate an HTML or PDF version by running these commands in the docs directory:

make latexpdf

to produce a PDF (docs/_build/latex/discoursegraphs.pdf) and

make html

to produce a set of HTML files (docs/_build/html/index.html).


If you'd like to visualize your graphs, you will also need:

License and Citation

This software is released under a 3-Clause BSD license. If you use discoursegraphs in your academic work, please cite the following paper:

Neumann, A. 2015. discoursegraphs: A graph-based merging tool and converter for multilayer annotated corpora. In Proceedings of the 20th Nordic Conference of Computational Linguistics (NODALIDA 2015), pp. 309-312.

  title={discoursegraphs: A graph-based merging tool and converter for multilayer annotated corpora},
  author={Neumann, Arne},
  booktitle={Proceedings of the 20th Nordic Conference of Computational Linguistics (NODALIDA 2015)},


Arne Neumann

People who downloaded this also like

  • SaltNPepper: a converter framework for various linguistic data formats
  • educe: a library for handling discourse-annotated corpora (SDRT, RST and PDTB)
  • treetools: a library for converting treebanks and grammar extraction (supports i.a. TigerXML and Negra/Tüba-Export formats)
  • TCFnetworks: library for creating graphs from annotated text corpora (based on TCF).