This project provides a reference implementation of the inference algorithms for globally normalized distributions over directed spanning trees. In natural language processing, such models are used to develop nonprojective dependency parsers. The current implementation shows the matrix-tree theorem in action. The implementation is carefully tested against brute-force algorithms that explicitly marginalizes over directed trees.
Citation: If you found this useful, please cite it as
@software{vieira-spanningtree,
author = {Tim Vieira},
title = {A reference implementation of algorithms for distributions over spanning trees},
url = {https://github.com/timvieira/spanning_tree}
}
References: In 2007, three different groups published similar methods for inference in discriminative nonprojective dependency parsing.
-
Koo, Globerson, Carreras and Collins (EMNLP'07) Structured Prediction Models via the Matrix-Tree Theorem
-
Smith and Smith (EMNLP'07) Probabilistic Models of Nonprojective Dependency Trees
-
McDonald & Satta (IWPT'07) On the Complexity of Non-Projective Data-Driven Dependency Parsing