Skip to content

Latest commit

 

History

History
149 lines (120 loc) · 11.4 KB

Experiment.md

File metadata and controls

149 lines (120 loc) · 11.4 KB
Raw

Experiment

Baseline

  • 10 fold cross-validation
  • 7:3 train:test
  • training and testing use original data
  • embedding dimension
    • cw2vec: 100
    • PiPiDai: 300

Valid Data: the performance of cross-validation on the last epoch

Model Complexity Dataset Embedding Word Seg lr Valid Acc Valid F1 Test Acc Test F1
ERCNN 2722885 Ant cw2vec char 0.001 81.85% 45.01% 81.57% 44.93%
ERCNN 3825985 Ant cw2vec word 0.001 81.85% 45.01% 81.57% 44.93%
ERCNN 2657085 CCKS cw2vec char 0.001 70.91% 70.90% 70.28% 70.25%
ERCNN 3088685 CCKS cw2vec word 0.001 77.18% 77.18% 74.76% 74.74%
ERCNN 3649685 PiPiDai PiPiDai char 0.001 82.43% 82.25% 81.21% 81.04%
ERCNN 6341885 PiPiDai PiPiDai word 0.001 86.16% 86.11% 83.72% 83.67%
Model Complexity Dataset Embedding Word Seg lr Valid Acc Valid F1 Test Acc Test F1
Transformer 913493 Ant cw2vec char 0.001 84.02% 60.68% 82.91% 58.95%
Transformer 2016593 Ant cw2vec word 0.001 83.66% 57.37% 82.58% 55.36%
Transformer 847693 CCKS cw2vec char 0.001 75.51% 75.46% 74.39% 74.34%
Transformer 1279293 CCKS cw2vec word 0.001 79.53% 79.50% 78.06% 78.04%
Transformer 3473093 PiPiDai PiPiDai char 0.001 51.99% 34.21% 51.72% 34.09%
Transformer 6165293 PiPiDai PiPiDai word 0.001 51.99% 34.21% 51.72% 34.09%

Deprecated

Fixed embedding

Complexity: trainable/all parameters

Model Complexity Dataset Embedding Word Seg Valid Acc Valid F1 Test Acc Test F1
SiameseCNN 38601/251501 Ant cw2vec (fixed) char 67.24% - 50.54% -
SiameseCNN 38601/1354601 Ant cw2vec (fixed) word 66.90% - 49.64% -
SiameseCNN 38601/185701 CCKS cw2vec (fixed) char 68.55% - 64.21% -
SiameseCNN 38601/617301 CCKS cw2vec (fixed) word 67.15% - 64.59% -
SiameseCNN 38601/717501 PiPiDai PiPiDai (fixed) char 62.46% - 60.69% -
SiameseCNN 38601/3409701 PiPiDai PiPiDai (fixed) word 60.79% - 62.55% -

2019/9/22,23 Use Balanced Data

  • training using balance data, test using original data
Model Complexity Dataset Embedding Word Seg Valid Acc Valid F1 Test Acc Test F1
SiameseCNN 251501 Ant cw2vec char 72.85% 73.57% 57.62% 50.76%
SiameseCNN 1354601 Ant cw2vec word 74.00% 74.75% 56.67% 49.38%
SiameseCNN 185701 CCKS cw2vec char 76.96% 77.61% 75.57% 75.50%
SiameseCNN 617301 CCKS cw2vec word 77.31% 77.94% 77.43% 77.32%
SiameseCNN 717501 PiPiDai PiPiDai char 71.06% 71.15% 73.75% 73.67%
SiameseCNN 3409701 PiPiDai PiPiDai word 72.19% 71.85% 74.50% 74.38%
Model Complexity Dataset Embedding Word Seg Valid Acc Valid F1 Test Acc Test F1
ERCNN 2722885 Ant cw2vec char 95.12% 96.75% 61.94% 48.74%
ERCNN 3825985 Ant cw2vec word 97.58% 99.26% 41.46% 39.42%
ERCNN 2657085 CCKS cw2vec char 49.39% 33.33% 50.02% 33.34%
ERCNN 3088685 CCKS cw2vec word 95.07% 96.25% 51.65% 48.79%
ERCNN 3649685 PiPiDai PiPiDai char 95.34% 95.77% 52.87% 49.33%
ERCNN 6341885 PiPiDai PiPiDai word 83.31% 83.25% 55.94% 55.59%
Model Complexity Dataset Embedding Word Seg Valid Acc Valid F1 Test Acc Test F1
Transformer 913493 Ant cw2vec char 69.20% 68.11% 75.70% 62.77%
Transformer 2016593 Ant cw2vec word 74.65% 74.64% 73.25% 62.90%
Transformer 847693 CCKS cw2vec char 77.61% 78.10% 72.78% 72.63%
Transformer 1279293 CCKS cw2vec word 78.95% 79.89% 72.32% 71.95%
Transformer 3473093 PiPiDai PiPiDai char 49.77% 33.33% 51.72% 34.09%
Transformer 6165293 PiPiDai PiPiDai word 49.77% 33.33% 51.72% 34.09%
  1. Basically Transformer predict everything to be positive in PiPiDai
  2. When the model doesn't work, its loss will stick on 13.815511

2019/9/24 Old Siamese Structure (The last dense layer)

Model Complexity Dataset Embedding Word Seg lr Valid Acc Valid F1 Test Acc Test F1
SiameseCNN 4533869 Ant cw2vec char 0.00001 73.46% 51.62% 73.37% 52.09%
SiameseCNN 5636969 Ant cw2vec word 0.00001 73.61% 52.21% 73.65% 51.86%
SiameseCNN 4468069 CCKS cw2vec char 0.001 78.03% 77.67% 73.81% 73.38%
SiameseCNN 4899669 CCKS cw2vec word 0.001 77.87% 77.28% 72.04% 71.24%
SiameseCNN 7097021 PiPiDai PiPiDai char 0.0001 81.17% 81.15% 77.63% 77.59%
SiameseCNN 9789221 PiPiDai PiPiDai word 0.0001 81.17% 81.15% 77.63% 77.59%
  • this model predict everything to be positive on PiPiDai dataset...

2019/9/23 Old SiameseCNN (using TextCNN structure)

Model Complexity Dataset Embedding Word Seg lr Valid Acc Valid F1 Test Acc Test F1
SiameseCNN 251501 Ant cw2vec char 0.001 81.88% 45.17% 81.58% 45.03%
SiameseCNN 1354601 Ant cw2vec word 0.001 83.35% 53.69% 81.45% 45.96%
SiameseCNN 185701 CCKS cw2vec char 0.001 84.28% 84.25% 75.07% 74.96%
SiameseCNN 617301 CCKS cw2vec word 0.001 87.78% 87.78% 77.43% 77.38%
SiameseCNN 717501 PiPiDai PiPiDai char 0.001 73.30% 73.18% 68.93% 68.81%
SiameseCNN 3409701 PiPiDai PiPiDai word 0.001 84.17% 84.09% 74.68% 74.55%
Model Complexity Dataset Embedding Word Seg lr Valid Acc Valid F1 Test Acc Test F1
SiameseRNN 288365 Ant cw2vec char 0.001 81.85% 45.01% 81.57% 44.93%
SiameseRNN 1391465 Ant cw2vec word 0.001 81.85% 45.01% 81.57% 44.93%
SiameseRNN 222565 CCKS cw2vec char 0.001 64.98% 64.96% 63.15% 63.13%
SiameseRNN 654165 CCKS cw2vec word 0.001 70.18% 70.17% 66.23% 66.19%
SiameseRNN 779965 PiPiDai PiPiDai char 0.001 69.99% 69.87% 68.10% 68.00%
SiameseRNN 3472165 PiPiDai PiPiDai word 0.001 73.36% 73.21% 70.76% 70.64%
Model Complexity Dataset Embedding Word Seg lr Valid Acc Valid F1 Test Acc Test F1
SiameseLSTM 475757 Ant cw2vec char 0.001 81.85% 45.01% 81.57% 44.93%
SiameseLSTM 1578857 Ant cw2vec word 0.001 81.85% 45.01% 81.57% 44.93%
SiameseLSTM 409957 CCKS cw2vec char 0.001 64.49% 64.42% 62.98% 62.92%
SiameseLSTM 841557 CCKS cw2vec word 0.001 69.45% 69.44% 65.71% 65.70%
SiameseLSTM 1044157 PiPiDai PiPiDai char 0.001 67.58% 67.46% 66.51% 66.40%
SiameseLSTM 3736357 PiPiDai PiPiDai word 0.001 70.90% 70.86% 68.56% 68.53%

Ant Financial

(during training)

Model Word Segment Embedding Batch Preprocessing Epoch Average Loss Accuracy Remark
Random - - - - - 82% -
Paper (rejected) char cw2vec (100d) - - - 76.89% not sure how dev set been generated
original Keras word cw2vec (100d) none - - 83% learned nothing???
ERCN word cw2vec (100d) none 10 0.4611 82% learned nothing
ERCN char cw2vec (100d) none 10 0.4611 82% learned nothing
ERCNN-Transformer char cw2vec (100d) none 10 0.4128 83% learned nothing

Quora

(during training)

Model Embedding Batch Preprocessing Epoch Average Loss Accuracy Remark
Random GloVe (300d) - - - 63% -
Paper (rejected) GloVe (300d) - - - 88.15% not sure how dev set been generated
ERCNN GloVe (300d) none 10 0.4094 80% using the same model... but this learned something
ERCNN-Transformer GloVe (300d) none 5 10.2011 63% learned nothing (to be improved)