What is Alearbres ?

Alearbres¹ is a collection of scripts that is intended to facilitate the generation of random dependency trees.

You can used Alearbres to generate artificial treebanks with various constraints :

• random structures
• random (projective and non-projective) linearisations
• random linearisations that optimize for Dependency Length Minimization

Main functionalities

Creating random non-ordered dependency trees :

Main script : random_structure.py

# create a random tree with 5 nodes
>>> t = DependencyTree(5)
Tree
	Edges: [(1, 0), (2, 1), (3, 0), (4, 2)]

# transform it to conllu format
>>> tree = t.toTree()
# write it to a file
conll3.trees2conllFile([tree], "sample_random-trees.conllu")

You can visualize the tree as a stemma using the draw_forest() function from the script² visualize_dependencies_udpipe.py

import visualize_dependencies_udpipe as visualize
visualize.draw_forest("sample_random-forest.conllu") 

Now if we want to build several trees at the same time, for example 10 trees of size 3 and 20 trees of size 20 :

specs = {3:10, 4:20}
forest = build_random_forest(specs)
conll3.trees2conllFile(forest, "sample_random-forest.conllu")

Randomly reordering nodes inside a dependency tree

Main script : random_linearisation.py

>>> conll = """1	le	le	_	_	_	3	det	_	_
2	petit	petit	_	_	_	3	amod	_	_
3	chat	chat	_	_	_	4	subj	_	_
4	dort	dort	_	_	_	0	root	_	_
5	très	très	_	_	_	6	advmod	_	_
6	bien	bien	_	_	_	4	advmod	_	_
"""
>>> tree = conll3.conll2tree(conll) # transform the conll string into a tree format

# create random linearisations
>>> random_tree = create_random_pj_linearisation(tree) # projective
r_nonproj = >>> create_random_nonpj_linearisation(tree) # non-projective

# look at the projective tree
>>> tree
1	dort	dort	_	_	_	0	root	_	_
2	le	le	_	_	_	4	det	_	_
3	petit	petit	_	_	_	4	amod	_	_
4	chat	chat	_	_	_	1	subj	_	_
5	très	très	_	_	_	6	advmod	_	_
6	bien	bien	_	_	_	1	advmod	_	_

Which once translated into graphical format gives :

# write the trees to a conll file
trees = [random_tree, r_nonproj]
conll3.trees2conllFile(trees, "sample_randomly-linearised-trees.conllu") 

Optimally reordering nodes in a dependency tree with respect to Dependency Length Minimization

Main script : dlm.py

conll = """1	le	le	_	_	_	3	det	_	_
2	petit	petit	_	_	_	3	amod	_	_
3	chat	chat	_	_	_	4	subj	_	_
4	dort	dort	_	_	_	0	root	_	_
5	très	très	_	_	_	6	advmod	_	_
6	bien	bien	_	_	_	4	advmod	_	_
"""
# store the dependency tree in a dict
>>> tree = conll3.conll2tree(conll)

# find the optimal sequence of nodes
>>> linearization = optimal_linearization(tree)
>>> print(linearization)
[1,3,2,4,6,5]
	
# rewrite the conll according to the new linearisation
>>> new_tree = random_linearisation.rewrite_tree(tree, linearization)
	
# see the conll
>>> print(new_tree.conllu())
1	le	le	_	_	_	2	det	_	_
2	chat	chat	_	_	_	4	subj	_	_
3	petit	petit	_	_	_	2	amod	_	_
4	dort	dort	_	_	_	0	root	_	_
5	bien	bien	_	_	_	4	advmod	_	_
6	très	très	_	_	_	5	advmod	_	_

Which gives us the nice tree below :

You can also of course use it on random trees to generate trees such as this one:

There's some weird trees I would really like to generate but Alearbres does not cover them...

Great ! We're always interested in expanding the types of constraints that can be added during the generation process. If there is some specific constraint that you would be interested in, please submit a feature-request and we will look into it.

Get the source

git clone https://github.com/marinecourtin/Alearbres.git

Changelog

18/09/2019 : first release

---

1 : Alearbres is a mashup of aléatoire and arbres (random trees in French)

2 : This script was kindly provided by Peter Uhrig.