Presentation is loading. Please wait.

Presentation is loading. Please wait.

Multi-Task Transfer Learning for Weakly- Supervised Relation Extraction Jing Jiang Singapore Management University ACL-IJCNLP 2009.

Similar presentations


Presentation on theme: "Multi-Task Transfer Learning for Weakly- Supervised Relation Extraction Jing Jiang Singapore Management University ACL-IJCNLP 2009."— Presentation transcript:

1 Multi-Task Transfer Learning for Weakly- Supervised Relation Extraction Jing Jiang Singapore Management University ACL-IJCNLP 2009

2 Aug 5, 2009ACL-IJCNLP 20092 Relation Extraction Task definition: to label the semantic relation between a pair of entities in a sentence (fragment) …[leader arg-1 ] of a minority [government arg-2 ]… PHYSPER-SOCEMP-ORGNIL PHYS: Physical PER-SOC: Personal / Social EMP-ORG: Employment / Membership / Subsidiary

3 Aug 5, 2009ACL-IJCNLP 20093 Supervised Learning Current solution: supervised machine learning (e.g. [Zhou et al. 2005], [Bunescu & Mooney 2005], [Zhang et al. 2006]) Training data is needed for each relation type …[leader arg-1 ] of a minority [government arg-2 ]… arg-1 word: leaderarg-2 type: ORG dependency: arg-1  of  arg-2 EMP-ORGPHYSPER-SOCNIL

4 Aug 5, 2009ACL-IJCNLP 20094 Challenge in Practice New relation type (in a new domain): no training data or a few seed instances In this work, we study weakly-supervised relation extraction –A few seed instances of the target relation type –Many instances of other auxiliary relation types –Additional human knowledge about the target relation type Main idea: Auxiliary relation types can help!

5 Aug 5, 2009ACL-IJCNLP 20095 Syntactic Similarity across Relation Types …[leader arg-1 ] of a minority [government arg-2 ]… arg-1 word: leaderarg-2 type: ORG dependency: arg-1  of  arg-2 the youngest [son arg-1 ] of ex-director [Suharto arg-2 ] the [Socialist People’s Party arg-1 ] of [Montenegro arg-2 ] EMP-ORG PER-SOC GPE-AFF

6 Aug 5, 2009ACL-IJCNLP 20096 Syntactic Similarity Syntactic Pattern Relation InstanceRelation Type (Subtype) arg-2 arg-1Arab leadersOTHER-AFF (Ethnic) his fatherPER-SOC (Family) South Jakarta Prosecution Office GPE-AFF (Based-in) arg-1 [verb] arg-2Yemen [sent] planes to Baghdad ART (User-or- Owner) His wife [had] three young children PER-SOC (Family) Jody Scheckter [paced] Farrari to both victories EMP-ORG (Employ- Staff)

7 Aug 5, 2009ACL-IJCNLP 20097 Problem Formulation based on Transfer Learning Domain adaptation and transfer learning (e.g. [Blitzer et al. 2006], [Hal Daume III 2007]) our goal: PER-SOCEMP-ORG We apply our previous framework ([Jiang & Zhai 2007b]) –Similar in spirit to [Evgeniou & Pontil 2004] and [Daume III, 2007]

8 Aug 5, 2009ACL-IJCNLP 20098 Review of Relation Extraction Basics Linear classifier …[leader arg-1 ] of a minority [government arg-2 ]… 10..1..10..1.. arg-2 type: ORG arg-2 type: PER dependency: arg-1  of  arg-2 4.5 0.3. 6.7. arg-2 type: ORG feature vectorweight vector in linear classifier dependency: arg-1  of  arg-2 EMP-ORG

9 Aug 5, 2009ACL-IJCNLP 20099 General vs. Specific Features Assumption: some features are commonly useful for different relation types, while other features are specific for individual relation types : weight vector for target type : weight vector for k’th auxiliary type common weight vector in a lower H dimensional space

10 Aug 5, 2009ACL-IJCNLP 200910 Learning Framework loss function on the target seed instances loss function on the auxiliary training instances 10 4 1

11 Aug 5, 2009ACL-IJCNLP 200911 General Features Which subset of features should be captured by ? common weight vector in a lower H dimensional space

12 Aug 5, 2009ACL-IJCNLP 200912 Feature Separation Automatic separation within the learning framework (see [Jiang & Zhai 2007b]) Human guidance –Argument word features: features that contain head word of an argument E.g. arg-1 word: sister –Entity type features: features that contain the entity type (subtype) of an argument E.g. arg-2 type: ORG Combined

13 Aug 5, 2009ACL-IJCNLP 200913 Imposing Entity Type Constraint Fix the possible entity types for the arguments for the target relation type Filter out the relation instances that do not satisfy the constraint in the end

14 Aug 5, 2009ACL-IJCNLP 200914 Experiment Setup ACE 2004, 7 relation types –6 types  auxiliary types 1 type  target type 5-fold cross validation # seed instances: 10

15 Aug 5, 2009ACL-IJCNLP 200915 Methods Compared BL: train on seed instances only BL-A: train on seed and auxiliary training instances together w/o feature separation TL-auto: transfer learning w/ automatic feature separation TL-guide: transfer learning w/ human-guided feature separation TL-comb: automatic feature separation combined with human guidance TL-NE: TL-comb + entity type constraint

16 Aug 5, 2009ACL-IJCNLP 200916 Comparison Target TypeBLBL-ATL-autoTL- guide TL- comb TL-NE PhysicalP0.0000.16920.29200.29340.33250.5056 R0.0000.08480.16960.17220.23830.2316 F0.0000.11300.21460.21700.27770.3176 Personal/SocialP1.0000.08040.10050.30690.32140.6412 R0.03860.17080.15980.72450.76860.7631 F0.07430.10930.12340.43110.45330.6969 Employment /Membership /Subsidiary P0.92310.35610.52300.54280.59730.7145 R0.00750.18500.26170.26480.36320.3601 F0.01480.24350.34880.35590.45180.4789 AverageP0.81240.14750.24120.27030.29920.4231 R0.02120.24320.38320.47640.55090.5464 F0.04060.15320.25320.29580.34230.4132

17 Aug 5, 2009ACL-IJCNLP 200917 Effect of λ λμTλμT 100100010000 P0.62650.31620.2992 R0.11700.39590.5509 F0.18470.29830.3423 Performance of TL-comb. λ μ k = 10 4, λ ν = 1.

18 Aug 5, 2009ACL-IJCNLP 200918 Number of Seed Instances

19 Aug 5, 2009ACL-IJCNLP 200919 Sensitivity of H

20 Aug 5, 2009ACL-IJCNLP 200920 Conclusions We proposed to apply a multi-task transfer learning framework to the weakly-supervised relation extraction problem. We defined two kinds of type-specific features. Our experiments show that automatic feature separation combined with human guidance and entity type constraint can significantly outperform the baselines.

21 Aug 5, 2009ACL-IJCNLP 200921 Thank You! Questions?

22 Aug 5, 2009ACL-IJCNLP 200922 Related Work [Zhou et al. 2008]: Different way of modeling commonality among relation types. [Banko & Etzioni, 2008]: Open-domain relation extraction. No target relation type. [Xu et al. 2008]: Rule-based adaptation. Same type.

23 Aug 5, 2009ACL-IJCNLP 200923 Hypothesized Type-Specific Features


Download ppt "Multi-Task Transfer Learning for Weakly- Supervised Relation Extraction Jing Jiang Singapore Management University ACL-IJCNLP 2009."

Similar presentations


Ads by Google