Implementation of the paper 2019 IMP conferrence(https://arxiv.org/abs/2204.12093) in Kaohsiung, Taiwan - Chen, C. L., Huang, P. Y., Lin J., and Huang, Y. T. 2019 Approach to Predicting News ─ A Precise Multi-LSTM Network With BERT.

While predicting news from LTN website during 2019 July 25th to August 5th, the accuracy may up to 99%. The accuracy of experiment is related to the corpus, that is, if one news including many new words, this program may do bad.

Author: 
	Lana Chen(k66inthesky@gmail.com)
	Amanda Huang(amanda10702@gmail.com)
Advisor:
	Meng-Chang Chen(mcc@iis.sinica.edu.tw)
Update: Nov.11th,2019
Target: To predict an unknown news/article from the eight categories(Technology, Finance, Politics, Entertainment, International, Sports, Health, Fashion)
File description:
	To use this project, you should download the whole package of files and the bert_model.ckpt.data-00000-of-00001,
	and follow the following install instructions.
	You only need to execute two programs--news2E.ipynb and predict.py.
	It's optional for replacinging the context in 'myinput.csv' if you wish to predict your own news.

Installation:

(Request)
-Python(3.6.2)
-Tensorflow(1.15)
-Numpy
-json
-Pandas
-Keras
-re

(Optional)
-collections
-copy
-math
-six
-shutil
-unicodedata

(MUST DOWNLOAD FILE)
"bert_model.ckpt.data-00000-of-00001"
(It's one of the file inside this folder "chinese_L-12_H-768_A-12"
from this website: https://github.com/google-research/bert)
AND PUT IT IN THE FOLDER:
"Predict_News/chinese_L-12_H-768_A-12/"

OR YOU CAN JUST DOWNLOAD THE WHOLE FOLDER "chinese_L-12_H-768_A-12"
from this website: https://github.com/google-research/bert)

Notice:

*input:
	
	CHINESE ONLY!!!
	
	Should be like the example'myinput.csv':
		two rows: 
			-context
			-your news // It can be any size, but 30*20 words per paragraph may be better.
			
*MUST CREATE AN EMPTY FOLDER, "embedding" BY YOURSELF 

File:

|-chinese_L-12_H-768_A-12
|-embedding
	|-myoutput.npy
|-news2E.ipynb #turn your input news(myinput.csv) into mytmpfile.jsonl and myoutput.npy(located at ./embedding/)
|-news2E_optional #only need to choose one of the two--news2E.ipynb or news2E_optional
|-predict.py #predict the last output to one of the eight categories 
|-extract_features.py
|-modeling.py
|-myinput.csv #it's optional for replacinging the context
|-mytmpfile.jsonl
|-optimization.py
|-our_model.h5 # it had already been trained with 28,768 corpus
|-our_model_weight.h5 # it had already been trained with 28,768 corpus
|-tokenization.py

Input(input will be a table/dataframe with only one row):

Output(You can find the answer on the bottom):

If you have any questions, please feel free to ask;)