際際滷

際際滷Share a Scribd company logo

Building a Graph of Names and Contextual Patterns for Named Entity ClassificationEcir09 poster

Grupo HULAT
Grupo HULAT

Authors: C辿sar de Pablo S叩nchez, Paloma Mart鱈nez ECIR 2009: Proceedings of the 31th European Conference on IR Research on Advances in Information Retrieval, Tolouse, France (April 6-9 2009)

1 of 1
Building a Graph of Names and Contextual Patterns for Named Entity Classi鍖cation C卒esar de Pablo S卒anchez and Paloma Mart卒脹nez LABDA, Computer Science Dept., Universidad Carlos III de Madrid {cdepablo,pmf}@inf.uc3m.es Objectives NERC for multilingual applications Bootstrap a name list and indicative patterns Large document collection Few example seeds for every class Nseeds < 40 Language independence (as an aim) Initial assumptions Dual bootstrapping One sense per entity type (name) Indelibility of class assignments Counter-training: learn several classes at once Query based exploration of the indexed collection. PERSON(x) Left patterns Right patterns Num Name Num Text Num Text 15 Fernando Arrabal 0 Gobierno del presidente 6 , ### esta tarde 64 Teodoro Obiang 1 Gobierno del ### 9 , vencedor 68 Salvador Allende 12 gobierno del presidente 21 y el ex 128 Peres 13 presidente del pa卒脹s , 26 , viajar卒a 156 Edouard Balladur 29 actual presidente 34 , y su colega 332 Grachov 47 palabras de 42 , visitar卒a 423 Calder卒on 50 cuyo ### , 49 , y el l卒脹der 450 Colom 60 presidente , 63 y el presidente 522 Joaqu卒脹n Almunia 61 reuni卒on con 65 se entrevist卒o Direct Evaluation: Name Lists (AvgPrec) Model PER LOC ORG M / T Mean PLO 94.8 52.7 67.1 71.5 PLOM 93.0 44.8 79.3 75.0 73.0 PLOT 94.8 87.4 81.1 40.9 76.0 Name Classi鍖cation Model P R F Acc baseline CONLL 26.27 56.48 35.86 ORG 39.34 entities PLO 77.33 54.34 63.83 64.04 PLOM 78.85 51.53 62.36 66.24 PLOT 78.72 41.58 54.42 62.18 entities+patterns PLO 66.12 57.97 61.78 63.17 PLOM 73.65 61.73 67.17 71.29 PLOT 66.35 56.62 61.10 62.50 Algorithm Pattern selection and evaluation 1. Rank by Support, 鍖lter min-support, select top-k 2. Evaluate min-Acc: Acc(p) = Pos Pos+Neg 3. Evaluate min-Conf: Conf(p) == PosNeg Pos+Neg+Unk Entity selection and evaluation 1. Rank by Support, 鍖lter min-support, select top-k 2. Evaluate min-Conf: Confslot(a) = 1 i (1 Confpattern(pi)) , ConfNE(a) = Confleft(a) Confright(a) Conclusions Ef鍖cient bootstrapping from large indexed collections with less seeds Already useful for NERC F-measure is lower than supervised machine learning More classes improves precision, not always recall Future work Other languages and domains Complex semantic models Language independence and NE Recognition Seed selection and improve effectiveness Acknowledgements: This work has been supported by the Regional Government of Madrid under the Research Network MAVIR (S-0505/TIC-0267) and by the Spanish Ministry of Education under the project BRAVO (TIN2007-67407-C03-01).

More Related Content

Building a Graph of Names and Contextual Patterns for Named Entity ClassificationEcir09 poster

  • 1. Building a Graph of Names and Contextual Patterns for Named Entity Classi鍖cation C卒esar de Pablo S卒anchez and Paloma Mart卒脹nez LABDA, Computer Science Dept., Universidad Carlos III de Madrid {cdepablo,pmf}@inf.uc3m.es Objectives NERC for multilingual applications Bootstrap a name list and indicative patterns Large document collection Few example seeds for every class Nseeds < 40 Language independence (as an aim) Initial assumptions Dual bootstrapping One sense per entity type (name) Indelibility of class assignments Counter-training: learn several classes at once Query based exploration of the indexed collection. PERSON(x) Left patterns Right patterns Num Name Num Text Num Text 15 Fernando Arrabal 0 Gobierno del presidente 6 , ### esta tarde 64 Teodoro Obiang 1 Gobierno del ### 9 , vencedor 68 Salvador Allende 12 gobierno del presidente 21 y el ex 128 Peres 13 presidente del pa卒脹s , 26 , viajar卒a 156 Edouard Balladur 29 actual presidente 34 , y su colega 332 Grachov 47 palabras de 42 , visitar卒a 423 Calder卒on 50 cuyo ### , 49 , y el l卒脹der 450 Colom 60 presidente , 63 y el presidente 522 Joaqu卒脹n Almunia 61 reuni卒on con 65 se entrevist卒o Direct Evaluation: Name Lists (AvgPrec) Model PER LOC ORG M / T Mean PLO 94.8 52.7 67.1 71.5 PLOM 93.0 44.8 79.3 75.0 73.0 PLOT 94.8 87.4 81.1 40.9 76.0 Name Classi鍖cation Model P R F Acc baseline CONLL 26.27 56.48 35.86 ORG 39.34 entities PLO 77.33 54.34 63.83 64.04 PLOM 78.85 51.53 62.36 66.24 PLOT 78.72 41.58 54.42 62.18 entities+patterns PLO 66.12 57.97 61.78 63.17 PLOM 73.65 61.73 67.17 71.29 PLOT 66.35 56.62 61.10 62.50 Algorithm Pattern selection and evaluation 1. Rank by Support, 鍖lter min-support, select top-k 2. Evaluate min-Acc: Acc(p) = Pos Pos+Neg 3. Evaluate min-Conf: Conf(p) == PosNeg Pos+Neg+Unk Entity selection and evaluation 1. Rank by Support, 鍖lter min-support, select top-k 2. Evaluate min-Conf: Confslot(a) = 1 i (1 Confpattern(pi)) , ConfNE(a) = Confleft(a) Confright(a) Conclusions Ef鍖cient bootstrapping from large indexed collections with less seeds Already useful for NERC F-measure is lower than supervised machine learning More classes improves precision, not always recall Future work Other languages and domains Complex semantic models Language independence and NE Recognition Seed selection and improve effectiveness Acknowledgements: This work has been supported by the Regional Government of Madrid under the Research Network MAVIR (S-0505/TIC-0267) and by the Spanish Ministry of Education under the project BRAVO (TIN2007-67407-C03-01).