Visualization of Linguistic Graphs with discoursegraphs

each format that can be parsed by discoursegraphs (RST, MMAX2, TigerXML, Conano ...) is converted into a networkx multidigraph
networkx graphs can be converted into dot/graphviz format for visulization

WARNING: dot/graphviz doesn't sort nodes

we would have to add additional (invisible) edges between the token nodes to make them appear in the right order
the token order can be extracted from the .tokens list of a DiscourseDocumentGraph (or any of its subclasses)



In [1]:

    
%matplotlib inline

# install the gvmagic extension by uncommenting/running it once: 
# %install_ext https://raw.github.com/cjdrake/ipython-magic/master/gvmagic.py

%load_ext gvmagic

import os
import discoursegraphs as dg



In [2]:

    
CORPUS_DIR = os.path.expanduser('~/corpora/potsdam-commentary-corpus-2.0.0/')



In [3]:

    
rstgraph = dg.read_rs3(os.path.join(CORPUS_DIR, 'rst', 'maz-1423.rs3'))
rstgraph_untokenized = dg.read_rs3(os.path.join(CORPUS_DIR, 'rst', 'maz-1423.rs3'), tokenize=False)



In [4]:

    
%dotstr dg.print_dot(rstgraph_untokenized)

Let's look at some statistics about the created document graph



In [5]:

    
dg.info(rstgraph_untokenized)









    



Name: maz-1423.rs3
Type: RSTGraph
Number of nodes: 33
Number of edges: 30
Average in degree:   0.9091
Average out degree:   0.9091 

Node statistics
===============

number of nodes with layers
	rst - 31
	rst:segment - 17
	rst:group - 14
	rst:root - 1
	discoursegraph - 1

number of nodes with attributes
	layers - 33
	rst:segment_type - 32
	label - 32
	rst:rel_name - 23
	rst:text - 17
	rst:group_type - 15
	metadata - 2

Edge statistics
===============

number of edges with layers
	rst - 30

number of edges with attributes
	layers - 30
	rst:rel_type - 30
	edge_type - 30
	rst:rel_name - 30
	label - 30

most common source edges
	rst:29 - 3
	rst:30 - 2
	rst:33 - 2
	rst:19 - 2
	rst:23 - 2

most common target edges
	rst:30 - 1
	rst:32 - 1
	rst:33 - 1
	rst:12 - 1
	rst:13 - 1

We can also plot attribute value statistics (e.g. POS distributions, RST relation counts)



In [17]:

    
dg.util.plot_attribute_distribution(rstgraph, rstgraph.edges(), 'rst:rel_name')



In [9]:

    
cdg = dg.read_conano(os.path.join(CORPUS_DIR, 'connectors', 'maz-1423.xml'))
cdg_untokenized = dg.read_conano(os.path.join(CORPUS_DIR, 'connectors', 'maz-1423.xml'), tokenize=False)



In [10]:

    
%dotstr dg.print_dot(cdg_untokenized)



In [11]:

    
%dotstr dg.print_dot(cdg)



In [12]:

    
mdg = dg.read_mmax2(os.path.join(CORPUS_DIR, 'coreference', 'maz-1423.mmax'))



In [13]:

    
%dotstr dg.print_dot(mdg)



In [14]:

    
tdg = dg.read_tiger(os.path.join(CORPUS_DIR, 'syntax', 'maz-1423.xml'))
%dotstr dg.print_dot(tdg)

Let's merge all these graphs to make them visually more appealing!1!!

we can only merge tokenized graphs (with identical tokenization)

Warning: original graphs are manipulated while merging

if a graph is merged into an existing graphs, its node IDs will be changed to match those of the existing graph



In [15]:

    
merged_graph = dg.DiscourseDocumentGraph()
merged_graph.merge_graphs(rstgraph)
merged_graph.merge_graphs(cdg)
merged_graph.merge_graphs(mdg)
merged_graph.merge_graphs(tdg)



In [16]:

    
%dotstr dg.print_dot(merged_graph)

Beautiful, isn't it?



In [ ]: