Kernel Methods for Relation Extraction

Graph Data Management Lab, School of Computer ScienceGDM@FUDAN http://gdm.fudan.edu.cn

Kernel Methods for Relation Extraction

杨振东Journal of Machine Learning Research 3 (2003) 1083-11062003 Dmitry Zelenko, Chinatsu Aone and Anthony Richardella

Graph Data Management Lab, School of Computer ScienceGDM@FUDAN http://gdm.fudan.edu.cn

•person-affiliation(organization) John Smith is the chief scientist of the Hardcom Corporation

•organization-location The IBM is an American multinational technology and consulting corporation, with headquarters in New York, United States.

Extract What Relation

Graph Data Management Lab, School of Computer ScienceGDM@FUDAN http://gdm.fudan.edu.cn

•Prerequisiteshallow parsingMachine Learning (Andrew NG)

Graph Data Management Lab, School of Computer ScienceGDM@FUDAN http://gdm.fudan.edu.cn

•We believe that shallow parsing (Abney, 1990) is an important prerequisite for information extraction.

“John Smith is the chief scientist of the Hardcom Corporation”

The types “PNP”, “Det”, “Adj”, and “Prep” denote “Personal Noun Phrase”, “Determiner(限定词 )”, “Adjective”, and “Preposition”, respectively.

shallow parsing

Graph Data Management Lab, School of Computer ScienceGDM@FUDAN http://gdm.fudan.edu.cn

The person and organization under consideration will receive the member and affiliation roles, respectively. The rest of the nodes will receive none roles reflecting that they do not participate in the relation.

shallow parsing contd.

Graph Data Management Lab, School of Computer ScienceGDM@FUDAN http://gdm.fudan.edu.cn

•Kernel function defines the similarity of the object X and the object Y, denoting K(X,Y).

Definition

John Smith is the chief scientist of the Hardcom Corporation

James Brown was a scientist at the University of Illinois

Graph Data Management Lab, School of Computer ScienceGDM@FUDAN http://gdm.fudan.edu.cn

Definition 1: Each node may have a different number of attributes. The attributes are named, and each node necessarily has attributes with names “Type” and “Role”.

Definition contd.

Graph Data Management Lab, School of Computer ScienceGDM@FUDAN http://gdm.fudan.edu.cn

Definition contd.• Definition 2: An (unlabeled)

relation example is defined inductively as follows:

• Let p be a node, then the pair P = (p, []) is a relation example, where by [] we denote an empty sequence.

• Let p be a node, and [P1,P2 ,…,Pl ] be a sequence of relation examples. Then, the pair P = (p, [P1,P2 ,…,Pl ]) is a relation example.

Graph Data Management Lab, School of Computer ScienceGDM@FUDAN http://gdm.fudan.edu.cn

•We denote by P.p the first element of the example pair, by P.c the second element of the example pair..

Definition contd.

{0,1}),t( ),( k

RolepPRolepPandTypepPTypepPifotherwisepPpPt ........,1,021

2121).,.( 　

…….

……. …….

•We first define a matching function and a similarity function on nodes.

Graph Data Management Lab, School of Computer ScienceGDM@FUDAN http://gdm.fudan.edu.cn

t()=1

John Smith is the chief scientist of the Hardcom Corporation

James Brown was a scientist at the University of Illinois

Graph Data Management Lab, School of Computer ScienceGDM@FUDAN http://gdm.fudan.edu.cn

k()=1

)..p..(....,1,021

2121).,.( headPheadpPtextpPtextpPifotherwisepPpPk

Graph Data Management Lab, School of Computer ScienceGDM@FUDAN http://gdm.fudan.edu.cn

Definition contd.•Then, for two relation examples P1,P2, we define the similarity function K(P1,P2) in terms of similarity function of the parent nodes and the similarity function Kc of the children.

(1)

…….

•We now give a general definition of Kc in terms of similarities of children subsequences. We first introduce some helpful notation.

Graph Data Management Lab, School of Computer ScienceGDM@FUDAN http://gdm.fudan.edu.cn

The formula enumerates all subsequences of relation example children with matching parents, accumulates the similarity for each subsequence by adding the corresponding child examples’ similarities, and decreases the similarity by the factor of reflecting how spread out the subsequences within children sequences. Finally, the similarity of two children sequences is the sum all matching subsequences similarities.

Definition contd.

(2)

A1 A2 A3 A4 B1 B2 B3

l(*) = 2 (A1,A2-> B1,B2) (A1,A3-> B1,B2) (A2,A4-> B1,B3) ……

Graph Data Management Lab, School of Computer ScienceGDM@FUDAN http://gdm.fudan.edu.cn

Contiguous Subtree Kernels

A1 A2 A3 A4 B1 B2 B3

l(*) = 2 (A1,A2-> B1,B2) (A2,A3-> B1,B2) (A2,A3-> B2,B3) ……

Sparse Subtree Kernels

Graph Data Management Lab, School of Computer ScienceGDM@FUDAN http://gdm.fudan.edu.cn

assume that = 0.5

Graph Data Management Lab, School of Computer ScienceGDM@FUDAN http://gdm.fudan.edu.cn

)..p..(....,1,021

2121).,.( headPheadpPtextpPtextpPifotherwisepPpPk

RolepPRolepPandTypepPTypepPifotherwisepPpPt ........,1,021

2121).,.( 　

Graph Data Management Lab, School of Computer ScienceGDM@FUDAN http://gdm.fudan.edu.cn

=k(P1.Sentence.p,P2.Sentence.p)+Kc([P1.Person,P1.Verb,P1.PNP],[P2.Person,P2.Verb,P2.PNP])

= 0 + 0.5( K(P1.Person,P2.Person) + K(P1.Verb,P2.Verb) + K(P1.PNP,P2.PNP) ) +0.52( K(P1.Person,P2.Person)+K(P1.Verb,P2.Verb)+K(P1.Verb,P2.Verb) + K(P1.PNP,P2.PNP) )+0.53( K(P1.Person,P2.Person)+K(P1.Verb,P2.Verb)+K(P1.PNP,P2.PNP) )= 0.5( k(P1.Person,P2.Person) + k(P1.Verb,P2.Verb) + K(P1.PNP,P2.PNP) ) + 0.52( k(P1.Person,P2.Person) + 2k(P1.Verb,P2.Verb) + K(P1.PNP,P2.PNP) ) + 0.53( k(P1.Person,P2.Person) + k(P1.Verb,P2.Verb) + K(P1.PNP,P2.PNP) )

…….

= 2.765625

K(P1,P2)

Graph Data Management Lab, School of Computer ScienceGDM@FUDAN http://gdm.fudan.edu.cn

Experiments•The (text) corpus for our experiments comprises 200 news articles from different news agencies and publications (Associated Press,Wall Street Journal, Washington Post, Los Angeles Times, Philadelphia Inquirer).

Learning curves (of F-measure) for the person-affiliation relation (on the left) and org-location relation (on the right), comparing feature-based learning algorithms with kernel-based learning algorithms.

Graph Data Management Lab, School of Computer ScienceGDM@FUDAN http://gdm.fudan.edu.cn

Experiments contd.

Learning curve (of F-measure) for the person-affiliation relation (on the left) and org-location relation (on the right), comparing kernel-based learning algorithms with different kernels.

Graph Data Management Lab, School of Computer ScienceGDM@FUDAN http://gdm.fudan.edu.cn