AI_LAW

all kinds of baseline models for long text classificaiton( text categorization)

Update: Joint Model for law cases prediction is released.

run python HAN_train.py to train the model for predict accusation, relevant articles and term of imprisonment.

challenges of this task:

  1. the description of law case(called facts) is quite long, the average words is around 400. it is not only long, but contain lots of information,

    even for a human being, you need to pay lots of attention before you can tell what's it about. for machine it need to have ability to handle long distance dependency,

    and pay attention on most important information.

  2. multiple sub tasks are included in this task. you are not only to predict accusations, but also need to predict relevant articles and term of imprisonment.

  3. this is a multi-label classification problem. given a fact, it may exists one or more than one accusations and serveral relevant articles.

  4. this also a imbalanced classification problem, while some labels have many data, other labels only have few data. to get best performance, you are not only

    to balanced precision and recall for a single label, but also need to balanced importance among different labels based on requirement or your evaluation matrix.

if you have any suggestion or problem, a better idea on how to slove this problem,find a bug, or want to find a team member, you are welcomed to can contact with me:

[email protected]. you can also commit your code to this repository.

1.Desc

this repository contain models that learn from law cases to make a prediction:

crime(accusation): which kinds of crime the bad guy did according to law system

relevant articles: specific law that is used for this case

term of imprisonment: whether it is death penalty or life imprisonment, how many years in prison(imprisonment)

find more about task, data or even start smart AI competition by check here:

http://cail.cipsc.org.cn/index.html

you can download sample data here, password is:d6e9

2.Data Processing: Data Distribution,Pre-processing, Generate Training & Validation set, Data Mining Features

2.1 Data Distribution

  1. total examples, crime,relevant articles(stage one):

    data_train: 155k

    data_valid(data_valid.json): 17k

    data_test(data_test.json): 33k

    number of total [accusation]: 202

    number of total [relevant_article]:183

    total examples, crime,relevant articles(stage two):

    data set(cail2018_big.json): around 1.7 million, we had to generate training/validation/test set ourself.

    number of labels is same as stage one.

  2. length of inputs(facts of the law case, after word tokenize)

    average length: 279

    max length: 9

    mini length: 19447

    length percentage:[(50, 0.01), (75, 0.04), (100, 0.07), (150, 0.19), (200, 0.19), (300, 0.25), (400, 0.11), (500, 0.05), (600, 0.03), (700, 0.06)]

    as we can see the majority of length is between 100 and 500. percentage of length less than 400 is 87%, less then 500 is 0.92.

    notice: here (50, 0.01) means length less than 50 is 1%, (75, 0.04) means length greater than 50 and less than 75 is 4%.

    that's take length as 500. pad for length less it, and truncate words exceed this number(truncate from start words, not end).

  3. label distribution(accusation, relevant article):

    how many accusations for a case?

     average length:1.2; max_length:6;mini_length:1; around 78% cases only has one accusation. most cases's accusation is less and equal than 3.
    
     length of accusation distribution:[(1, 0.78), (2, 0.0), (3, 0.2), (4, 0.019), (5, 0.002), (6, 0.001), (7, 0.0001)]
    

    how many relevant articles for a case?

     average length:1.45; max_length:9; mini_length:1; around 70% cases only has one relevant articles;
    
     length of relevant articles:[(1, 0.71), (2, 0.0), (3, 0.18), (4, 0.07), (5, 0.03), (6, 0.01), (7, 0.004)].
    
  4. Top20 accusation and its frequency( as you can see that data imbalance problem is not a big problem here; however in big data set this is a serious problem)

    Theft: 10051
    
    Smuggling, trafficking, transporting, and making drugs: 8872
    
    Intentional injury: 6377
    
    Robbery: 5020
    
    Fraud: 3536
    
    Bribery: 3496
    
    Provocation: 3290
    
    Dangerous driving: 2758
    
    Organization, coercion, seduction, shelter, and introduction of prostitution: 2647
    
    Manufacture, trafficking, and Disseminating Obscene Articles: 2617
    

2.2 Pre-processing

  1. preprocess value for imprisonment.

    range of imprisonment from 0 to 300(=12*25), it is raw value too big, not may lead to less efficient for model to learn.

    we will normalize imprisonment using following format:

      imprisonment= (imprisonment- imprisonment_mean)/imprisonment_std
    

    where imprisonment_mean stand for mean of imprisonment, imprisonment_std stand for stand deviation of imprisonment.

    we can easily to get its mean:, 26.2, and std:33.5 from training data. during test we will re-scale value back:

     imprisonment_test=(imprisonment_test+imprisonment_mean)*imprisonment_std
    

    notice: this preprocess is remove. we usually do normalize on features, not target.

  2. normalize value of money:

    there are lots of int and float in this data related to money. such as 12343 or 3446.56, to make it easy for the model to learn it, we normalize

    int and float. for example to 10000 or 3000. check method replace_money_value in data_util.py

2.3 Data Mining Features for handling Few-Shot Labels

  1. there are some labels only associate with very few samples, like 40,30 or 20. it is very difficult for machine learning models

    to learn it.

    we believe for each law fact, there are some sailent words that can help us to decide what accusation or related law it is.

    based on this insight, we list these sailent words for few-shot labels, count how many times matched between this words and inputs,

    normalize them, and get a vector as our additional feature.

    you can find this features under data_mining/data_mining_features.py

3.Evaluation: F1 score(Micro,Macro)

for task1(predict accusation) and task2(predict relevant article), there are multi-label classification problem. and as we already seen from

previous description, label distribution may be skewed or imbalanced. so we use f1 score which computed from precision and recall.

rembember that:

   True Postive=TP=[predict=1,truth=1]

   False Postive=FP=[predict=1,truth=0]

   False Negative=FN=[predict=0,truth=1]

   precison=TP/(TP+FP). precision is among all labels that you predict as postive, how many are real postive

   recall=TP/(TP+FN).   recall is among all true labels, how many you predict as postive.

   f1=(2*precision*recall)/(precision+recall). it's value is among:[0,1].

f1_micro_macro

   finally we compute:

        score=((f1_macro+f1_micro)/2.0)*100

Question1: here we have multi-label, it is not binary classifcaiton problem, how do we define TP,FP,FN and f1 score?

   we will first get confusing matrix(TP,FP,FN), then compute f1 score.

   for your better understanding, i will give you an example. suppose target_label=[2,12,88], predict_label=[2,12,13,10]

   as you can see, unique_label=[2,12,13,10,88]. and two labels exists in both side, label_in_common=[2,12]. and we use a dict to store TP,FP,FN for each class.

   dict_count[class]=(num_TP,num_FP,num_FN)

   1) we will go though this unique_label, and count TP,FP,FN for each class:

        for the first element 2, it exists in predict_label(you predict it is true), and also exists in target_label(actually it is true),

             so it is True Positive(TP), ===> dict[2]=(1,0,0)

        for the second element 12, it is similiar like first element, also True Positive(TP). ===> dict[12]=(1,0,0)

        for the third element 13, it not exists in predict_label(predict=0), but exists in target_label(truth=1),

             so it is False Negative(FN),===> dict[13]=(0,0,1)

        for the fourth element 10, predict=1,truth=0 ===>False Postive(FP) ===> dict[10]=(0,1,0)

        for the last element 88,   predict=0,truth=1 ===>False Negative(FN) ===>dict[88]=(0,0,0)

   2) secondly, we compute total TP,FP,FN:

      TP=2,FP=1,FN=2

   3) finally, we compute P,R and F1 score:

      P_micro=TP/(TP+FP)=2/3=0.6777

      R_micro=TP/(TP+FN)=2/4=0.5

      F1_micro=(2*P_micro*R_micro)/(P_micro+R_micro)=0.575

   for detail, check compute_confuse_matrix() and compute_micro_macro() at evaluation_matrix.py

Question2: the above steps is for only one input. but suppose after we go through several inputs, and got:

      dict_count[2]=(20,5,7)

      where we define format in above: dict_count[class]=(num_TP,num_FP,num_FN)

      what's the f1 score for class 2?

      this is similiar we saw in the last step(#3), we will compute Precsion,Recall, then f1 score:

         P_label2=TP/(TP+FP)=20/(20+5)=0.80

         R_label2=TP/(TP+FN)=20/(20+7)=0.74

         f1_label2=(2*P*R)/(P+R)=0.76

       notice: if you want to compute f1_macro, you can go through each class just same as label2, then compute average value among all classes as f1_macro.

Question3: how many labels should we retrieve once we compute the logits for a input?

      remember we are in mulit-label classification scenario. given a input, there may exist multi-label at same time. we can't just get the max possible

      label as our target, which is usually implement by softmax function. that is saying that we can think this is a multi-binary classification problem,

      for each class, we just consider whether it is exists or not. by applying sigmoid function for each class, we will make sure that possibility will be

      between 0 and 1. if the possibility is greater(or equal) than 0.5, we think it is exists(predict=true).

      in a word, we look each class seperately, and get all classes that possibilities we predict greater than a threshold.


 for detail above evaluation matrix, check evaluation_matrix.py

4.Imbalance Classification for Skew Data

4.1 Over-sampling

since some labels associate with many data, others may only contain few data. we will try to use sampling to handle this problem.

over sampling is used:

we will first get frequency of each label in accusation and relevant article. given a input, we will get label of accusation and relevant article, and there

frequency, then compute average value. we also set a threshold(e.g. 1500), add compute how many copy should we have in our training data(num_copy=threshold/average).

e.g. below is what we did:

input1:('freq_accusation:', 1230, 'freq_article:', 2973, ';num_copy:', 1)

input2:('freq_accusation:', 167, 'freq_article:', 3525, ';num_copy:', 1)

input3:('freq_accusation:', 282, 'freq_article:', 304, ';num_copy:', 5)

input4:('freq_accusation:', 225, 'freq_article:', 22, ';num_copy:', 12)

input5:('freq_accusation:', 489, 'freq_article:', 487, ';num_copy:', 3)

input6:('freq_accusation:', 1134, 'freq_article:', 1148, ';num_copy:', 1)

why we take average value for frequency of accusation and relevant article?

  as accsuation and relevant articles are related in many cases, we will use average

check transform_data_to_index() under data_util.py

4.2 Generating Training and Validation set

training data is used all data from training set with sampling.

 based on frequency of each labels, we also do over-sampling for those labels that are too frequent. 
 
 check down sampling: data_mining/data_construct_training_set.py, 
      
 frequency of accusation: cache_text_cnn/accusation_freq.txt
 
 frequency of law: cache_text_cnn/article_freq.txt

validation data is based on validation set released on the early stage, but we remove those data that is

 overlap with training data.
 
 for distribution of validation data, you can find it in cache_text_cnn/accusation_freq_valid.txt

5.Transfer Learning & Pretrained Word Embedding

download pretrained word embedding from https://github.com/Embedding/Chinese-Word-Vectors and enable flag 'use_embedding' during training.

or download from https://pan.baidu.com/s/1o7MWrnc, password: wzqv , choose '.bin' file, embedding size is 64.

command to import word embedding, especially for words contain chinese:

import gensim

from gensim.models import KeyedVectors

word2vec_model = KeyedVectors.load_word2vec_format(word2vec_model_path, binary=True, unicode_errors='ignore')

6.Models

  1. HAN: hierarchical attention network(completed)

    embedding-->word level bi-lstm encoder-->word level attention-->sentence level bi-lstm encoder-->sentence level attention

  2. TextCNN(multi-layers)

    embedding-->CNN1(BN-->Relu)--->CNN2(BN-->Relu)--->Max-pooling

  3. DPCNN: deep pyramid cnn for text categorization

    text region embedding-->CNN1.1-->CNN2.1-->(Pooling/2-->CONV-->CONV)*N-->Pooling

  4. c-gru: CNN followed by GRU

    embedding--->CNN(BN-->Relu)--->bi-GRU

  5. gru-c: GRU followed by CNN

    embedding-->bi-GRU--->CNN(BN--->Relu)

  6. simple_pooling

a) embedding-->max over each dimension of word embedding

b) embedding-->mean over each dimension of word embedding

c) embedding-->concat of max pooling and mean pooling

d) embedding-->mean over each context(context is define as words around itself)-->max over context for each dimension.
  1. self-attention(transformer) TODO

  2. Convolutional Sequence to Sequence Learning TODO

7.Performance

overall

performance on validation dataset(seperate from training data) train for first 1000 steps(around one epoch):

('1.Accasation Score:', 70.8571253641409, ';2.Article Score:', 71.28389708554336, ';3.Penalty Score:', 69.78947368421059, ';Score ALL:', 211.93049613389485)

han_1000steps

Performance on test env(small data, 155k training data),online:

Model Accasation Score Relevant Score Penalty Score
HAN 77.63 75.29 52.65
TextCNN(multiple layers) 79.91 76.87 53.62
c-gru
gru-c

Performance on test env(big data, 1.5 million training data),online:

Model Accasation Score Relevant Score Penalty Score Total Score
TextCNN-multiple layers(online) 84.8 82.9 68.20 235.9
Deep Pyramid CNN(local) 89.0 86.4 78.6 254
Hierarchical Attention Network(local) 85.1 84.0 79.2 248.3

Notice:

1.offline(local) score is lower than online(official) score for about 4.0. for example, dpcnn's online performance may be 85 or less.

  1. we used data_valid.json from data as our valdation data, which is provided in the first stage of competition.
 but there may exists cases where sample overlap from this validation set and our big 1.7 million training set.
 
 so for better evaluation perpose, you may remove those overlapped data from training data, or use other validation set.

89.03954996862663, ';2.Article Score:', 86.38077500531911, ';3.Penalty Score:', 78.64466689362311, ';Score ALL:', 254.06499186756886)

F1 score for each accusation on validation set:

accusation: f1 score(validation set)
经济犯:0.0
打击报复证人:0.0
非法制造、买卖、运输、储存危险物质:0.0
高利转贷:0.0
倒卖车票、船票:0.0
走私:0.0
贷款诈骗:0.536580606819
非法持有毒品:0.578942996339
非法制造、销售非法制造的注册商标标识:0.599994600029
过失投放危险物质:0.66665777783
协助组织卖淫:0.666661527813
.......
.......
.......
拐骗儿童:0.999993750027
走私珍贵动物、珍贵动物制品:0.999993750027
引诱、教唆、欺骗他人吸毒:0.999994000026
编造、故意传播虚假恐怖信息:0.999994000026
对单位行贿:0.999994090935
聚众扰乱公共场所秩序、交通秩序:0.999994090935
帮助犯罪分子逃避处罚:0.999994166692
非法获取公民个人信息:0.999994230795
非法买卖制毒物品:0.999994375025
脱逃:0.999994375025
盗掘古文化遗址、古墓葬:0.99999444447

8.Modeling Dependencies of Sub Tasks and Labels

Modeling Dependencies of Sub Tasks:

given a fact of law case, and a accusation(or charge) of this law case, which relevant law(s) should it apply?

as a matter of fact, relevant law and accusation are highly relevanted. for example, if charge of a law case is fraud,

relevant laws(s) usually will be limited to some laws that are connected to fraud; another example is if someone was charged

of murder, the punishment will be always very serious, like sentence to death or life imprisonment.

from this insights, we try to model denpendency of sub tasks(accusation,relevant laws,punishment):

  1. basic, we use joint model to solve these three task together in a single model where we sharing a common features extracted from
 fact of law case. the feature used for a task can also be used for other tasks. and we want to let the model to learn a 

 sharing feature so well that all three tasks can be benefited from it.
  1. secondly, as these tasks are dependent between each other in some extent, and get some idea from densely connected models
 from recently development, we design our model as follow:

let common shared feature h , accusation feature as h_accu, relevant law feature as h_law, and punishement feature as h_puni.

 a. we will pass accusation feature h_acc together with original common shared feature h as a new feature to compute 

 relevant law feature as h_law.

 b.we will also pass accusation feature h_acc,original common shared feature h, toegether with relevant law feature h_law

 to compute punishment feature.

we are doing this, so that the decision we made about accusation, will also affect the decision of predict

relevant law. the decision we made about accusation, relevant punishment will also influence punishment.

Modeling Dependencies of Labels of accusation:

 when doing error analysis, we found that there are some kinds of dependency between labels of accusation given a fact.
 
 for example, when some one was selling fake branded product, he or she will be charged of 'selling good of counterfeit registered trademarks',
 
 meanwhile he or she will also be charged as 'economic criminal'. 
 
 previously we assumed that labels are independent from each other, as we see from computer vision where you can find a cat
 
 and there is also a person or not. but now we believe we this is not true, and we better model this task, we should use a better
 
 loss function.

9.Error Analysis

during each validation, f1 score for each label in accusation will be printed in debug_accusation_accuracy.txt file,

those f1 scores are lower will be in list first, so you can pay more attention on to improve accuracy of these labels in accusation task.

10.Usage

train:

 python train.py

it will report macro f1 score and micro f1 score when doing validation, save f1 score for each label on validation 
 
 during each epoch, and save checkpoint to predictor/checkpoint/

optional parameters:

--model: the name of model you will use. {text_cnn,han,dp_cnn,c_gru,c_gru2,gru,pooling} [text_cnn]

--use_pretrained_embedding: whether use pretrained embedding or not. download it as discussed on section #5, otherwise set it to False. {True,False} [True]

--embed_size: embedding size

--hidden_size: hidden size

--tokenize_style: how to tokenize your input(sentences/document).if you set this to char, you are training a char model. {word, char} [word]

predict(or test):

 python main.py

zip your model so that you can upload for testing purpose, run:

 zip -r ai_law_predictor.zip predictor  ===>it will zip all resources in directory predictor/ as a zip file.

11.Environment

python 2 + tensorflow 1.8

for people use python3, you can just remove below lines:

  import sys

  reload(sys)

  sys.setdefaultencoding('utf-8')

12.Model Details

1).Hierarchical Attention Network:

Implementation of Hierarchical Attention Networks for Document Classification

Structure:

embedding

Word Encoder: word level bi-directional GRU to get rich representation of words

Word Attention:word level attention to get important information in a sentence

Sentence Encoder: sentence level bi-directional GRU to get rich representation of sentences

Sentence Attention: sentence level attention to get important sentence among sentences

One Layer MLP for transform document representation to each sub task's features.

FC+Softmax

HAN

check inference_han method from HAN_model.py under directory of predictor

2).TextCNN(Multiple Layers):

Implementation of <a href="http://www.aclweb.org/anthology/D14-1181"> Convolutional Neural Networks for Sentence Classification </a>

Structure: embedding-->CNN1(BN-->Relu)--->CNN2(BN-->Relu)--->Max-pooling-->concat features--->Fully Connected Layer

In order to get very good result with TextCNN, you also need to read carefully about this paper <a href="https://arxiv.org/abs/1510.03820">A Sensitivity Analysis of (and Practitioners' Guide to) Convolutional Neural Networks for Sentence Classification</a>: it give you some insights of things that can affect performance. although you need to  change some settings according to your specific task.

Convolutional Neural Network is main building box for solve problems of computer vision. Now we will show how CNN can be used for NLP, in in particular, text classification. Sentence length will be different from one to another. So we will use pad to get fixed length, n. For each token in the sentence, we will use word embedding to get a fixed dimension vector, d. So our input is a 2-dimension matrix:(n,d). This is similar with image for CNN.

Firstly, we will do convolutional operation to our input. It is a element-wise multiply between filter and part of input. We use k number of filters, each filter size is a 2-dimension matrix (f,d). Now the output will be k number of lists. Each list has a length of n-f+1. each element is a scalar. Notice that the second dimension will be always the dimension of word embedding. We are using different size of filters to get rich features from text inputs. And this is something similar with n-gram features.

Secondly, we will do max pooling for the output of convolutional operation. For k number of lists, we will get k number of scalars.

Thirdly, we will concatenate scalars to form final features. It is a fixed-size vector. And it is independent from the size of filters we use.

Finally, we will use linear layer to project these features to per-defined labels.

text_cnn_multiplelayers

model_law

check inference_text_cnn method from HAN_model.py under directory of predictor

3)DPCNN: Deep Pyramid CNN

Deep Pyramid Convolutional Neural Networks for Text Categorization

text region embedding-->CNN1.1-->CNN2.1-->(Pooling/2-->CONV-->CONV)*N-->Pooling

this model is used for text categorization, you can think it is a text classification model with text length is quite long.

basicly, it is a deep convolutional neural networks with repeat of building block: max-pooling and multiple layers of CNN.

to make it easy for train this deep model, it also used skip connection as ResNet does. different from other deep models,

conventionally when you reduce space size(input size for each layer), we will also increase depth of channels, so that computation resource in each layer is fixed.

in deep pyramid CNN, it keep depth of channels, and gradually recuce space size(input size).

main features of the model:

a.downsampling with the number of feature maps fixed ===> reduce computation. as total computation time is twice

the computation time of a single block.

b. shortcut connections with pre-activation ====> so that it can train very deep neural network, similiar with ResNet.

c. no need for dimension matching ===> although skip connection is used, but dimension is fixed(e.g. 250).

d. text region embedding: embedding of a region of text convering one or more words.

deep_pyramid_compare

DPCNN

China law research cup judicial artificial intelligence challenge:

Task 1 crime prediction: predict the accused's conviction according to the description and facts of the case in the criminal legal documents;

Task 2 Recommendation of the relevant law: to predict the relevant laws in this case according to the description and facts of the case in the

criminal legal documents;

Task 3 time of imprisonment:  predict the defendant's sentence length.

Data set including 2.68 million criminal law legal documents, criminal 183, relevant article 202 of law, sentence length including

 0-25 years, life imprisonment, sentence to death.

13.Data Mining Features

data mining feature is complete, will add desc here as soon as possible

14.Conclusion

  1. it is possible to solve the problem in a way of joint model. each sub task shared same input and representation.

  2. a single evaluation matrix(here micro and macro of f1 score) is important for evaluate how well you done on your task,

and to know how well it is when you changed something.

  1. add more here.

GitHub