Construction of semantic bootstrapping models for relation extraction
Chunyun Zhang
a,
⇑
, Weiran Xu
a
, Zhanyu Ma
a
, Sheng Gao
a
, Qun Li
b
, Jun Guo
a
a
Pattern Recognition and Intelligent System Laboratory, Beijing University of Posts and Telecommunications, Beijing, China
b
School of Computer Science and Technology, Nanjing University of Posts and Telecommunications, Nanjing, China
article info
Article history:
Received 14 July 2014
Received in revised form 16 March 2015
Accepted 17 March 2015
Available online 25 March 2015
Keywords:
Relation extraction
Bootstrapping
Trigger word
Kernel
Pattern learning
abstract
Traditionally, pattern-based relation extraction methods are usually based on iterative bootstrapping
model which generally implies semantic drift or low recall problem. In this paper, we present a novel
semantic bootstrapping framework that uses semantic information of patterns and flexible match
method to address such problem. We introduce formalization for this class of bootstrapping models,
which allows semantic constraint to guide learning iterations and use flexible bottom-up kernel to com-
pare patterns. To obtain the insights of reliability and applicability of our framework, we applied it to the
English Slot Filling (ESF) task of Knowledge Based Population (KBP) at Text Analysis Conference (TAC).
Experimental results show that our framework obtains performance superior to the state of the art.
Ó 2015 Elsevier B.V. All rights reserved.
1. Introduction
Relation extraction (RE) is an important but unsolved problem
in information extraction (IE). It focuses on extracting structured
relations from unstructured sources such as documents or webs,
which can potentially benefit a wide range of natural language pro-
cessing (NLP) tasks such as question answering, ontology learning,
and summarization [1].
To solve the RE problem, a number of machine learning
approaches have been recently applied. One common paradigm
is the usage of bootstrapping [2] to learn relation patterns. The
popularity of this framework lies in its ability to learn sufficient
patterns and instances simply by iterations starting from a small
number of seeds. Its central assumption is the pattern-relation
duality principle [3] that good seed samples lead to good patterns,
while good patterns help to extract good instances. Here, good pat-
terns are usually referred to patterns that have high coverage (high
recall) and low error rate (high precision), and good instances are
instances that are realized by good patterns. Systems such as
DIPRE [3], Snowball [4], and ExDisco [5] took a small set of
domain-specific examples as seeds and an unannotated corpus as
input. The seed examples can be either target relation instances
or sample linguistic patterns in which the linguistic arguments
correspond to the target relation arguments. New instances or
new patterns will be found in the documents where the seed is
located. The new instances or patterns will be used as new seed
for the next iteration. However, Komachi’ analysis in [6] showed
that semantic drift is an inherent property of iterative bootstrap-
ping algorithms and, therefore, poses a fundamental problem.
Hence, these systems without semantic constraint are greatly trou-
bled by the problem of semantic drift.
Relation patterns are defined as the structured features of the
context of the entity and its attribute value (e.g. Bill Gates and
Microsoft of the relation org:founded_by of organization entity) in
a target relation mentioning [7]. Consequently, how well the sys-
tem performs largely depends on how well patterns are repre-
sented. However, most existing patterns are with inflexible
representation or without semantic constraint. Patterns in [3,4,
7–9] using shallow syntactic features have poor performances in
the extraction of the relations that are ambiguous or lexically dis-
tant in their expression. Dependency patterns [10–15] have been
shown to perform better, since they are more informative for rela-
tion extraction. The shortest dependency pattern (SDP) and the
subject–verb–object (SVO) pattern, among other dependency pat-
terns, are two commonly used patterns [10,1,12,13]. However,
due to less semantic constraint, they gain the generality at the cost
of lacking specific information and thus may produce semantic
drift in bootstrapping iterations.
Similarity method, a measure which determines whether a pat-
tern or instance derived from a new sentence is relation oriented or
not, is another important key method for bootstrapping model.
Unfortunately, the existing similarity methods are rigid or unsuit-
able for extracting relations expressed in complex structure pat-
terns, since they cannot weigh the relative importance of
different features of patterns only by using exact match method
[3,7,8] or cosine-like method [12,4]. Kernel methods
[16,10,11,17,15] have been proven to be effective in measuring
http://dx.doi.org/10.1016/j.knosys.2015.03.017
0950-7051/Ó 2015 Elsevier B.V. All rights reserved.
⇑
Corresponding author.
E-mail address: zhangchunyun1009@126.com (C. Zhang).
Knowledge-Based Systems 83 (2015) 128–137
Contents lists available at ScienceDirect
Knowledge-Based Systems
journal homepage: www.elsevier.com/locate/knosys