python 2.7、java 8、jieba 0.39、 pandas 0.22.0、sklearn 0.19.1、 gensim 3.4.0、xgboost 0.71、lightgbm 2.1.1 tensorflow 1.5、pytorch 0.4.0、keras 2.1.6、

model epoch/step batch train_acc test_acc p r f1 tune_word bn ccks kerasqqp 20 100 0.780 0.584 no kerasqqp 20 100 0.780 0.584 no kerasqqp 40 100 0.770 0.611 no kerasqqp 26353 100 0.88 0.707 0.625 yes kerasqqp 26870 100 0.88 0.653 0.343 0.669 0.454 yes kerasqqp 26870 100 0.839 0.754 0.429 0.421 0.425 no sialstm 26870 100 0.798 0.720 0.628 no manhattan sialstm 6870 100 0.783 0.781 0 0 0 yes no no manhattan sialstm 6870 100 0.90 0.742 0.418 0.453 0.435 yes no no max, log sialstm 6870 100 0.84 0.752 0.413 0.309 0.354 yes no no basic att sialstm 21467 32 0.90 0.751 0.438 0.473 45.4/64.7 yes no no max, log

siacnn 26870 0.714 0.749 0.484 no yes siacnn 26870 0.682 0.768 0.491 no no esim 21078 100 0.932 0.793 0.516 0.597 0.554 no yes yes esim 83967 32 0.85 0.781 0.5 67.5 57.5/72.1 no no yes 5.47h esim 29167 32 0.90 0.562 0.30 0.759 43.2/61.2 no no only 2.85h esim 29167 32 0.92 0.802 52 60 56.3/71.5 no no pretrain 1.49h esim 29167 32 0.762 53 51 52/69 yes no no 0.97h esim 21467 32 0.86 0.79 51.9 68.3 59/73.1 no no no 0.9h

bimpm 19389 0.96 0.794 0.529 0.554 0.541 yes no no 0.8714 rnet 6870 100 81 81 57 52 54.7/71.4 no no no 0.9154

model snli quora BiMPM 86.9 88.69 ESIM 88.0 Datt 86.8 DR-BiLSTM 88.5 DIIN 88.0 KIM 88.6 CAFE 88.5 DRCN 88.9 DMAN 88.8

model epoch batch train test p r f1 tune_w bn l2 ccks rnet 20 100 89 81.9 49.7 55.4 52.4/70.7 no no dense no 1.17h rnet 20 100 86 81.5 48.8 57.6 52.9/70.9 no no all no 1.14h 59.98 esim 20 100 87 80.8 47.6 61.7 53.7/71.3 no no all no 0.89h 61.84 esim 20 100 86.4 81 48 63.9 54.8/71.9 no no all no sent vec abs(-) * esim 20 100 85 81.7 49.4 62.9 55.3/72.3 no no all no 63.12 lr decay + highway esim 20 100 85 78.3 43.8 72.4 54.6/72 no no all no 3.26h translate esim 55.0/72.1 3.76h esim ema 52.8/70.7 esim focal loss 53.0/71

bimpm 20 100 85 81.1 47.9 55 51.2/69.9 yes no all no 1.10h
bimpm 20 100 87 80 45.7 58.4 51.3/69.7 no no all no 1.04h
qanet 20 100 82 82 51.2 57.4 54.1/71.7 no no all no 0.47h 60.89 lr decay no fitting qanet 20 100 81 82.5 51.3 53.1 52.2/70.7 no no all no no ema tflm 5+5 64 83.6 57 33 42.7/67.9 yes yes

https://github.com/PaddlePaddle/models/tree/develop/fluid/language_model

• A Structured Self-Attentive Sentence Embedding
• stacking
• easy, hard

ontology, BabelNet intent

同义词 反义词 上位词 下位词 Universal Sentence Encoder 相同字符数 莱文斯坦距离(编辑距离), 欧式距离，余弦距离 SimHash LDA/LSA BLUE, 共同单词 pointwise mutual information PMI

短文本相似度计算 https://www.zhihu.com/question/49424474 深度语义模型 https://zhuanlan.zhihu.com/p/33537217 http://www.sohu.com/a/222501203_717210 https://www.kaggle.com/c/quora-question-pairs/discussion/34325 https://www.kaggle.com/c/multinli-matched-evaluation/leaderboard

拍拍贷 https://www.ppdai.ai/mirror/goToMirrorDetail?mirrorId=1&tabindex=1 天池 https://tianchi.aliyun.com/competition/introduction.htm?spm=5176.100066.0.0.798f33afftdSM9&raceId=231661

还款期 数 【 】