List your company
ResearchFormerly WizWand

Part-of-Speech Tagging on Penn TreeBank (test)

97.96AccuracyMeta-BiLSTM
-2.91615223.27284949.4618575.650851Jun 24, 2015Dec 18, 2015Jun 12, 2016Dec 6, 2016Jun 1, 2017Nov 25, 2017May 21, 2018
Updated 8mo ago

Evaluation results

MethodLinks
Meta-BiLSTMmodel type=sentence-based character model2018.0597.96
C2W(tanh)+word (sskip)hand-crafted features=no, additional data=yes2015.0897.78
Ling et al. (2015)2016.0397.78
Ling et al. (2015)2017.0397.78
Ling et al. (2015)2016.1197.78
Choi2018.0597.64
C2W+featureshand-crafted features=yes, additional data=no2015.0897.57
DMNEnsemble=top 4 models2015.0697.56
Kumar et al. (2016)2016.1197.56
Huang et al.2018.0597.55
Bi-directional LSTM-CNNs-CRFneural network based=true2016.0397.55
Huang et al. (2015)2017.0397.55
Ma & Hovy (2016)2017.0397.55
Transfer Learning ModelTransfer setting=w/o transfer2017.0397.55
Transfer Learning ModelTransfer setting=w/ transfer2017.0397.55
JMTallTasks=Joint all tasks2016.1197.55
Ma and Hovy (2016)2016.1197.55
C2W+word (sskip)hand-crafted features=no, additional data=yes2015.0897.54
Søgaard2018.0597.5
SCNN2015.0697.5
SCCNhand-crafted features=yes, additional data=yes2015.0897.5
Søgaard (2011)2016.0397.5
Søgaard (2011)2016.1197.5
Our Local (B=8)Normalization Strategy=Local, Beam Size (B)=82016.0397.45
Andor et al.2018.0597.44
Spoustova et al.2015.0697.44
Morčehand-crafted features=yes, additional data=yes2015.0897.44
Our Local (B=1)Normalization Strategy=Local, Beam Size (B)=12016.0397.44
Our Global (B=8)Normalization Strategy=Global, Beam Size (B)=82016.0397.44
word (sskip)hand-crafted features=no, additional data=yes2015.0897.42
Dozat et al.2018.0597.41
Suzuki et al.2015.0697.4
C2Wrepresentation=compositional character-to-word2015.0897.36
C2Whand-crafted features=no, additional data=no2015.0897.36
structReghand-crafted features=yes, additional data=no2015.0897.36
Sun (2014)2016.0397.36
word+featureshand-crafted features=yes, additional data=no2015.0897.34
Shen et al. (2007)2016.0397.33
Stanfordfeatures=default set of features2015.0897.32
Stanford 2.0hand-crafted features=yes, additional data=no2015.0897.32
CNNhand-crafted features=no, additional data=yes2015.0897.32
Santos and Zadrozny (2014)neural network based=true2016.0397.32
Collobert et al. (2011)neural network based=true2016.0397.29
Collobert et al. (2011)2017.0397.29
Collobert et al. (2011)2016.1197.29
Manning (2011)2016.0397.28
Tsuruoka et al. (2011)2016.1197.28
Sogaard2015.0697.27
Toutanova et al. (2003)2016.0397.27
Toutanova et al. (2003)2016.1197.27
Linear CRF2016.0397.17
Giménez and Màrquez (2004)2016.0397.16
SVMTool2015.0697.15
Wordrepresentation=word lookup tables2015.0896.97
wordhand-crafted features=no, additional data=no2015.0896.7
BI-LSTM-CRF (Senna)extra data=Yes2015.080.9755
Semi-supervised condensed nearest neighborextra data=Yes2015.080.975
BI-LSTM-CRFextra data=No2015.080.9743
CRFs with structure regularizationextra data=No2015.080.9736
Bidirectional perceptron learningextra data=No2015.080.9733
Conv network tagger (senna)extra data=Yes2015.080.9729
Maximum entropy cyclic dependency networkextra data=No2015.080.9724
SVM based taggerextra data=No2015.080.9716
Conv network taggerextra data=No2015.080.9637