Introduction to Machine Learning NPFL 054 http://ufal.mff.cuni.cz/course/npfl054 Barbora Hladká Martin Holub {Hladka | Holub} @ufal.mff.cuni.cz Charles University in Prague, Faculty of Mathematics and Physics, Institute of Formal and Applied Linguistics NPFL054, 2015 Hladká & Holub & Lukšová Demo 1, page 1/12
Demo 1 Verb Patterns Classification Purpose of the demo task = to show several things related to gold standard data for a supervised machine learning task, especially • Manual annotation and basic data analysis • Gold Standard data distribution • Inter-annotator agreement • Confusion matrices • Error analysis NPFL054, 2015 Hladká & Holub & Lukšová Demo 1, page 2/12
Demo 1 Verb Patterns Classification Annotation experiment 2015 NPFL054, 2015 Hladká & Holub & Lukšová Demo 1, page 3/12
Gold Standard – data distributions Cry Enlarge GS histogram GS histogram 250 140 120 200 100 150 80 60 100 40 50 20 0 0 u x 1 4 7 u 1 2 3 4 NPFL054, 2015 Hladká & Holub & Lukšová Demo 1, page 4/12
Manual annotation • Annotated data • 6 x 10 sentences with CRY • 6 x 10 sentences with ENLARGE • 4 groups A, B, C, D • the same data set annotated by each group • We analyse • which group is closer to the Gold Standard • the inter-annotator agreement between groups NPFL054, 2015 Hladká & Holub & Lukšová Demo 1, page 5/12
A, B, C, D distributions - CRY Cry A Cry B 30 30 20 10 10 0 0 u x 1 4 7 u x 1 4 7 Cry C Cry D 30 25 20 10 10 0 0 u x 1 4 7 u x 1 4 7 NPFL054, 2015 Hladká & Holub & Lukšová Demo 1, page 6/12
A, B, C, D distributions - ENLARGE Enlarge A Enlarge B 30 30 20 20 10 10 0 0 u 1 2 3 4 u 1 2 3 4 Enlarge C Enlarge D 30 20 10 10 0 0 u 1 2 3 4 u 1 2 3 4 NPFL054, 2015 Hladká & Holub & Lukšová Demo 1, page 7/12
Confusion matrices - CRY group agreement (%) disagreement (%) A 41 (68,3) 19 (31,7) B 40 (66,7) 20 (33,3) C 40 (66,7) 20 (33,3) D 45 (75,0) 15 (25,0) NPFL054, 2015 Hladká & Holub & Lukšová Demo 1, page 8/12
Confusion matrices - ENLARGE group agreement (%) disagreement (%) A 38 (63,3) 22 (36,7) B 28 (46,7) 32 (53,3) C 28 (46,7) 32 (53,3) D 36 (60,0) 24 (40,0) NPFL054, 2015 Hladká & Holub & Lukšová Demo 1, page 9/12
Inter-annotator agreement (IAA) groups verb Cohen’s Kappa Fleiss’s Kappa A-B cry 0.355 A-C cry 0.276 A-D cry 0.406 B-C cry 0.366 B-D cry 0.407 C-D cry 0.327 A-B-C-D cry – 0.353 A-B enlarge 0.306 A-C enlarge 0.413 A-D enlarge 0.296 B-C enlarge 0.222 B-D enlarge 0.324 C-D enlarge 0.365 A-B-C-D enlarge – 0.319 NPFL054, 2015 Hladká & Holub & Lukšová Demo 1, page 10/12
A + B + C + D vs. GS – CRY • Number of agreements: 171 (71.3 %) • Number of disagreements: 69 (28.7 %) NPFL054, 2015 Hladká & Holub & Lukšová Demo 1, page 11/12
A + B + C + D vs. GS – ENLARGE • Number of agreements: 138 (57.5 %) • Number of disagreements: 102 (42.5 %) NPFL054, 2015 Hladká & Holub & Lukšová Demo 1, page 12/12
Recommend
More recommend