1

Click here to load reader

An Alignment-based Pattern Representation Model for Information Extraction

Embed Size (px)

Citation preview

Page 1: An Alignment-based Pattern Representation Model for Information Extraction

An Alignment-based Pattern Representation Modelfor Information Extraction

Seokhwan Kim, Minwoo Jeong, Gary Geunbae Lee{megaup, stardust, gblee}@postech.ac.kr

Abstract - In this paper, we propose an alternative pattern representation model and the effective method of utilizing it. While the previous pattern representation models completely depend on the result of dependency analysis, our approach is basically based on the lexical alignment and considers the result of dependency analysis only as a meaningful feature of the alignment process. In this way, we can cope with the errors of incomplete dependency analysis. An evaluation of a scenario template task shows that our proposed model outperforms the previous syntax-dependent models.

Pattern Representation Model for Information Extraction Information Extraction Extracting the defined number of relevant arguments from natural language documents Subtasks

# of arguments subtask1 named-entity recognition2 binary relation extraction

more than 2 relation/event extraction

Approach Automatic Pattern Learning Pattern Representation Model Pattern Learning Algorithm

Pattern Representation Model for IE Problem Definition Ex)

Related Works Lexical Sequence Pattern Models A set of lexical sequences

Syntax-dependent Pattern Models A set of subtrees (from D-tree)

Method

Experiment

Our Approach Pattern Model Lexical Sequence Pattern + Term Weight (from Dependency Analysis)

Soft Pattern Matching Sequence Alignment

Pattern Sequences Extraction1) Searching the sentences containing all arguments of each tuple in source documents2) Segmenting out subpart of the sentence based on clausal boundaries3) Replacing the parts of arguments in the sub-sentence with argument labels Computing Term Weights

Ex)

Pattern Matching Sequence Alignment Based on a Dynamic Programming Alignment Matrix

Matrix Computation

About 50 peasants have been kidnapped by terrorists of the FMNL

incident type kidnappingprep_ind terroristsprep_org FMNLhum_tgt peasants

Extraction Pattern

?

<hum_tgt> be kidnappedbe kidnapped by <prep_ind>

kidnapped

peasants terrorists

FMNL

nsubjpass agent

prep_of

(kidnapped ({HUM_TGT}-nsubjpass)(kidnapped ({PREP_IND}-agent))(kidnapped ({PREP_IND}-agent ({PREP_ORG}-prep_of)))

<HUM_TGT> of [NP] have been kidnapped by <PRED_IND>

1 0.33 0.33 4 4 4 1.5 1.5

<HUM_TGT> of [NP] have been kidnapped by <PREP_IND>

about 50 peasants have been kidnapped by terrorists

wi = (ri + c) / (di + c)wi : weight of i-th termri : number of relevant terms within a subtree, ti as rootdi : distance from root nodec : for smoothing (default:1)

kidnapped

<HUM_TGT> <PREP_IND>

nsubjpass agent

prep_of

<PREP_ORG>[NP]

prep_of

(3+1)/(0+1) = 4

(2+1)/(1+1) = 1.5

(1+1)/(2+1) =0.67

(1+1)/(1+1) = 1

(0+1)/(2+1) = 0.33

peasants have been kidnapped by terrorists<HUM_TGT> 1 0 0 0 0 0

of 0 1 0 0 0 0[NP] 0 0.66 1 0 0 0have 0 4 3 2 1 0been 0 0 8 7 6 5

kidnapped 0 0 4 12 11 10by 0 0 0 8 13.5 12.5

<PRED_IND> 1.5 0.5 0 0 0 15

Mi-1,j-1 + sim i-1,j-1 * wi-1

Mi-1,j + gp * wi-1

Mi,j-1 + gp * wi

0

Mi,j = max

Experimental Setup Data MUC-3/4 Data About the Terrorism Events Simpler template structure with 4 slots perp_ind, perp_org, phys_tgt, hum_tgt Dev-set (training), Test-set (evaluation) Preprocessing Dependency Parsing and NP-chunking Stanford Parser Extracting Pattern Candidates Selecting all pattern candidates for test Without pattern filtering To compare not the pattern filtering method, but the representative performance among pattern models

Experimental Result Pattern Models SVO Model (Yangarber ‘00) Linked-Chain Model (Greenwood ‘06) Subtree Model (Sudo ‘03) Our Model Result

Our proposed model achieved much higher recall than the other models with similar precision

Model Precision Recall F-measure

SVO 21.74 20.62 21.16

Linked-Chain 20.04 26.55 22.84

Subtree 23.34 32.73 27.25

Alignment 23.35 45.62 30.89