On the Transitivity of Hypernym-Hyponym Relations
in Data-Driven Lexical Taxonomies
Jiaqing Liang, Yi Zhang, Yanghua Xiao
∗
Shanghai Key Laboratory of Data Science, School of Computer Science, Fudan University
l.j.q.light@gmail.com, {z
yi11, shawyh}@fudan.edu.cn
Haixun Wang
Facebook, USA
haixun@gmail.com
Wei Wang
School of Computer Science,
Fudan University
weiwang1@fudan.edu.cn
Pinpin Zhu
Xiaoi Research, Shanghai Xiaoi
Robot Technology Co. LTD., China.
pp@xiaoi.com
Abstract
Taxonomy is indispensable in understanding natural lan-
guage. A variety of large scale, usage-based, data-driven
lexical taxonomies have been constructed in recent years.
Hypernym-hyponym relationship, which is considered as the
backbone of lexical taxonomies can not only be used to cat-
egorize the data but also enables generalization. In particu-
lar, we focus on one of the most prominent properties of the
hypernym-hyponym relationship, namely, transitivity, which
has a significant implication for many applications. We show
that, unlike human crafted ontologies and taxonomies, transi-
tivity does not always hold in data-driven lexical taxonomies.
We introduce a supervised approach to detect whether tran-
sitivity holds for any given pair of hypernym-hyponym rela-
tionships. Besides solving the inferencing problem, we also
use the transitivity to derive new hypernym-hyponym re-
lationships for data-driven lexical taxonomies. We conduct
extensive experiments to show the effectiveness of our ap-
proach.
Introduction
Knowledge bases are playing an increasingly important
role in many applications. Most knowledge bases, including
WordNet (Miller 1995), Cyc (Lenat and Guha 1989), and
Freebase (Bollacker et al. 2008), are manually crafted by
human experts or community efforts. The coverage of man-
ual knowledge bases, such as WordNet, is far from being
complete (Sang 2007). For example, the concepts and in-
stances below Animals and People in WordNet is quite lim-
ited (Pantel and Pennacchiotti 2006; Hovy, Kozareva, and
Riloff 2009).
Much attention thus has been paid on deriving knowledge
bases by automatic extraction from big corpora. The data-
driven approaches produce many knowledge bases such as
KnowItAll (Etzioni et al. 2004), NELL (Mitchell et al.
∗
Correspondence author. This paper was supported by National
Key Basic Research Program of China under No.2015CB358800,
by the National NSFC (No.61472085, U1509213), by Shanghai
Municipal Science and Technology Commission foundation key
project under No.15JC1400900, by Shanghai Municipal Science
and Technology project under No.16511102102.
Copyright
c
2017, Association for the Advancement of Artificial
Intelligence (www.aaai.org). All rights reserved.
2015), and Probase (Wu et al. 2012). Data-driven knowledge
bases in general are larger than manual knowledge bases,
covering more entities, concepts as well as their relation-
ships. For example, Freebase has thousands of types, while
Probase has millions of concepts. With a larger coverage,
data-driven knowledge bases are better at supporting large
scale text understanding and many other tasks.
Data-driven Lexical Taxonomy
In this paper, we concentrate on a particular knowledge
base: lexical taxonomy built by data-driven approaches. A
lexical taxonomy consists of the hypernym-hyponym rela-
tions between terms. One term A is a hypernym of another
term B if A’s meaning covers the meaning of B or much
broader (Sang 2007). For example, furniture is a hy-
pernym of chair. The opposite term for hypernym is hy-
ponym. So chair is a hyponym of furniture.Weuse
the expression hyponym(A, B) for a hypernym-hyponym re-
lationship, which means A is a hyponym of B.
Hypernym-hyponym relations are backbones of text un-
derstanding. The reason hypernym-hyponym relationships
hold such significance is that they enable generalization,
which lies at the core of human cognition as well as at the
core of machine inferencing for text understanding. To see
this, hyponym(iphone, smart phone) enables machine
to understand the search intent of iphone (i.e. smart
phone). hyponym(galaxy s4, smart phone) further
allows to recommend the related keyword galaxy s4.
Many automatically harvested lexical taxonomies such as
Probase, YAGO (Suchanek, Kasneci, and Weikum 2007),
WikiTaxonomy (Ponzetto and Strube 2008), are extracted
from web corpora or Wikipedia by certain syntactic pat-
terns (such as Hearst patterns (Hearst 1992)) or heuristic
rules. For example, a sentence “...famous basketball play-
ers such as Michael Jordan ...” is considered an evidence
for the claim that term Michael Jordan is a hyponym
of term famous basketball player, while this sen-
tence follows one Hearst pattern.
Problem Statement
In this paper, we focus on one of the most important proper-
ties of the hypernym-hyponym relationship: transitivity.For
Proceedings of the Thirty-First AAAI Conference on Artificial Intelligence (AAAI-17)