wwmmqq/DeepTextClassifier

Method

1-EMNLP2014 Convolutional Neural Networks for Sentence Classification

2-AAAI2015 Recurrent Convolutional Neural Networks for Text Classification

3-NAACL2016 Hierarchical Attention Networks for Document Classification

4-ICLR2017 A Structured Self-attentive Sentence Embedding

Experiments

2D - viewer

can be find in folder view

Question Classification (QC 2002)

http://cogcomp.org/Data/QA/QC/

type	objective	average length	class	train/test	Lan
sentence	question types	10	6	15452/500	English

AVG (average word embedding) training

Epoch 0: train cost: 0.0157731532153, acc: 0.856604866744
Epoch 0: test acc: 0.914
Epoch 1: train cost: 0.011120560211, acc: 0.90527230591
Epoch 1: test acc: 0.926
Epoch 2: train cost: 0.0110226789984, acc: 0.899188876014
Epoch 2: test acc: 0.93
Epoch 3: train cost: 0.0108223813667, acc: 0.898899188876
Epoch 3: test acc: 0.93
Epoch 4: train cost: 0.0105730641033, acc: 0.910486674392
Epoch 4: test acc: 0.93
Epoch 5: train cost: 0.0107601115206, acc: 0.901796060255
Epoch 5: test acc: 0.93
Epoch 6: train cost: 0.0109058949116, acc: 0.905561993048
Epoch 6: test acc: 0.93
Epoch 7: train cost: 0.010843385788, acc: 0.897740440324
Epoch 7: test acc: 0.93
Epoch 8: train cost: 0.0106531062346, acc: 0.91077636153
Epoch 8: test acc: 0.93
Epoch 9: train cost: 0.0101846210305, acc: 0.912224797219
Epoch 9: test acc: 0.93

AVGTFIDF (average word embedding tfidf) training

Epoch 0: train cost: 0.0227936226724, acc: 0.743337195829
Epoch 0: test acc: 0.858
Epoch 1: train cost: 0.0163344583568, acc: 0.823001158749
Epoch 1: test acc: 0.888
Epoch 2: train cost: 0.015335006076, acc: 0.832560834299
Epoch 2: test acc: 0.892
...
Epoch 9: train cost: 0.0152274934685, acc: 0.83516801854
Epoch 9: test acc: 0.892

kim cnn training

Epoch 2:
    cost: 0.00132063299802
    train acc: 0.989860950174
    dev acc: 0.974

GRU training

Epoch 1:
    cost: 0.000869356350728
    train acc: 0.992178447277
    dev acc: 0.984

GRU-CNN training

Epoch 3:
    train cost: 0.000631529912567
    train acc: 0.993626882966
    dev acc: 0.984

RCNN training

Epoch 1:
    train cost: 0.00179642121795
    train acc: 0.980011587486
    dev acc: 0.98
Epoch 2:
    train cost: 0.000787947041509
    train acc: 0.992757821553
    dev acc: 0.982
Epoch 3:
    train cost: 0.000631529912567
    train acc: 0.993626882966
    dev acc: 0.984

RNN-ATT training

Epoch 1:
    train cost: 0.00179642121795
    train acc: 0.980011587486
    dev acc: 0.98
Epoch 2:
    train cost: 0.000787947041509
    train acc: 0.992757821553
    dev acc: 0.982
Epoch 3:
    train cost: 0.000631529912567
    train acc: 0.993626882966
    dev acc: 0.984

Structured Self-attentive training

Epoch 1:
    train cost: 0.179252121705, acc: 0.960892236385
    dev acc: 0.968
Epoch 2:
    train cost: 0.174375187066, acc: 0.985225955968
    dev acc: 0.972
....
....
Epoch 8:
    train cost: 0.172080848612, acc: 0.985515643105
    dev acc: 0.972

GRU with Center Loss

Epoch 0: train cost: 0.00327359643783, acc: 0.98348783314
Epoch 0: test acc: 0.974
Epoch 1: train cost: 0.00208708796292, acc: 0.991888760139
Epoch 1: test acc: 0.98
Epoch 2: train cost: 0.00197653938085, acc: 0.99449594438
Epoch 2: test acc: 0.976
Epoch 3: train cost: 0.00187900724474, acc: 0.996234067207
Epoch 3: test acc: 0.978

Center Divide

Epoch 0: train cost: 0.00184123501496, acc: 0.982908458864
Epoch 0: test acc: 0.974
Epoch 1: train cost: 0.000830719525279, acc: 0.991888760139
Epoch 1: test acc: 0.974
Epoch 2: train cost: 0.000683305005285, acc: 0.993047508691
Epoch 2: test acc: 0.978
Epoch 3: train cost: 0.000309743383767, acc: 0.997392815759
Epoch 3: test acc: 0.982
Epoch 4: train cost: 0.000337060144333, acc: 0.995654692932
Epoch 4: test acc: 0.986

NLPCC 2017 News headline Classification

https://github.com/FudanNLP/nlpcc2017_news_headline_categorization

type	objective	average length	class	train/dev/test	Lan
sentence	news headline	10	6	0/0/0	Chinese