This library contains a modified implementation of the Maximum Spanning Tree (MST) algorithm proposed by Tarjan (1977) and Camerini et al. (1979). The MST algorithm is a popular decoding algorithm for graph-based depenedency parsing. However, dependency trees often have a constraint that only one edge may emanate from the root. Gabow and Tarjan (1984) suggest an efficient extension to the algorithm for this which is included in this library. The implementation of this library runs in the expected O(N^2) and is based on the implementation of Stanojević and Cohen. soon.
This library also contains a simplified version of the K-Best MST algorithm of Camerini et al. (1980) that runs in O(K N^2). Additionally, the library contains a new algorithm for finding the K-Best dependency trees subject to a root constraint.
A detailed description of these algorithms including proofs of correctness can be found in "Mind the Root: Decoding Arborescences for Dependency Parsing" and "On Finding the K-best Non-projective Dependency Trees" .
A runtime experiment on random weights is available to run in notes/runtime_experiment.py
This code is for the papers Please Mind the Root: Decoding Arborescences for Dependency Parsing and On Finding the K-best Non-projective Dependency Trees featured in EMNLP 2020 and ACL 2021 respectively. Please cite as:
@inproceedings{zmigrod-etal-2020-please,
title = "Please Mind the Root: {D}ecoding Arborescences for Dependency Parsing",
author = "Zmigrod, Ran and
Vieira, Tim and
Cotterell, Ryan",
booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)",
month = nov,
year = "2020",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://www.aclweb.org/anthology/2020.emnlp-main.390",
doi = "10.18653/v1/2020.emnlp-main.390",
pages = "4809--4819",
}
@inproceedings{zmigrod-etal-2021-finding,
title = "On Finding the K-best Non-projective Dependency Trees",
author = "Zmigrod, Ran and
Vieira, Tim and
Cotterell, Ryan",
booktitle = "Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)",
month = aug,
year = "2021",
address = "Online",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2021.acl-long.106",
doi = "10.18653/v1/2021.acl-long.106",
pages = "1324--1337",
}- Python version >= 3.6
Installation:
git clone https://github.com/rycolab/spanningtrees
cd spanningtrees
pip install -e .We support graph creations using numpy arrays.
We consider 0 to be the root node of the graph.
Note that the root node has no incoming edges and we do not have any self loops.
import numpy as np
from spanningtrees.mst import MST
from spanningtrees.kbest import KBest
graph = np.array(([
[ 0., 71., 52., 67.],
[ 0., 0., 32., 46.],
[ 0., 2., 0., 79.],
[ 0., 65., 16., 0.]
]), dtype=float)
mst = MST(graph)
kbest = KBest(graph)The MST and root-constrained MST can then be calculated by:
print(f"Unconstrained MST: {mst.mst()}")
print(f"Constrained MST: {mst.mst(True)}")Output:
Unconstrained MST: [-1 0 0 2]
Constrained MST: [-1 3 0 2]
The following can be instead used to find the K-Best trees
print("Unconstrained K-best")
for tree in kbest.kbest():
print(kbest.weight(tree), tree)
# Root constraint
print("Constrained K-best")
for tree in kbest.kbest(True):
print(kbest.weight(tree), tree)Output:
Unconstrained K-best
202.0 [-1 0 0 2]
196.0 [-1 3 0 2]
190.0 [-1 0 0 0]
184.0 [-1 3 0 0]
182.0 [-1 0 1 2]
170.0 [-1 0 1 0]
169.0 [-1 0 0 1]
164.0 [-1 3 1 0]
154.0 [-1 0 3 0]
149.0 [-1 0 1 1]
148.0 [-1 3 3 0]
133.0 [-1 0 3 1]
133.0 [-1 2 0 2]
121.0 [-1 2 0 0]
100.0 [-1 2 0 1]
85.0 [-1 2 3 0]
Constrained K-best
196.0 [-1 3 0 2]
182.0 [-1 0 1 2]
164.0 [-1 3 1 0]
149.0 [-1 0 1 1]
148.0 [-1 3 3 0]
133.0 [-1 2 0 2]
133.0 [-1 0 3 1]
100.0 [-1 2 0 1]
85.0 [-1 2 3 0]
This code repository focuses on decoding MSTs. A useful library to use during training and learning of edges weights can be found here.
Other libraries for performing MST computations are Stanojević and Cohen, networkx, and stanza.