[N.K. provocation] Results

LDA Topic Model output (20 Topics):


Topic #13 and #17 can be interpreted as “armed provocation” and “nuclear provocation”, respectively. Each of #14 and #18 can be interpreted as “South-North Dialogue” and “international talks”.

The numbers of articles that belong to each category are shown in the graph below


Independent variable: the inverse degree of support for the unification of the people (1: necessary / 5: unnecessary)


Training NER[0] – Training custom Word Vectors from 10-K/Q filings

NER with word vectors.

Word vectors are particularly useful for terms which aren’t well represented in your labelled training data. For instance, if you’re doing named entity recognition, there will always be lots of names that you don’t have examples of. For instance, imagine your training data happens to contain some examples of the term “Microsoft”, but it doesn’t contain any examples of the term “Symantec”. In your raw text sample, there are plenty of examples of both terms, and they’re used in similar contexts. The word vectors make that fact available to the entity recognition model. It still won’t see examples of “Symantec” labelled as a company. However, it’ll see that “Symantec” has a word vector that usually corresponds to company terms, so it can make the inference.


INPUT data : All of 10-K, 10-Q filings available on SEC, with basic preprocessing steps except for lemmatization, N grams.



As my RAM was restricted to 16 GB, I found the Data Streaming technique is especially useful in my case. (Helpful introductions on training Word Vectors using Gensim : https://rare-technologies.com/word2vec-tutorial/ , https://rare-technologies.com/data-streaming-in-python-generators-iterators-iterables/)

import os
import sys
import re
from gensim.models import Word2Vec

import logging
format='%(asctime)s : %(levelname)s : %(message)s',
level=logging.INFO) #Display the progress of training word vectors

dirname = 'D:/10_k' # directory where 10-K/Q filings are downloaded
class MySentences(object):
    def __init__(self, dirname):
        self.dirname = dirname            
    def __iter__(self):
        for f_dir in os.listdir(self.dirname):
            for qname in os.listdir(os.path.join(self.dirname, f_dir, 'EDGAR\\10-X_C', f_dir[-4:])):
                #print(os.path.join(dirname, f_dir, 'EDGAR\\10-X_C', f_dir[-4:]))
                for fname in os.listdir(os.path.join(self.dirname, f_dir, 'EDGAR\\10-X_C', f_dir[-4:], qname)):
                    if fname.endswith(".txt"):
                        for line in open(os.path.join(self.dirname, f_dir, 'EDGAR\\10-X_C', f_dir[-4:], qname, fname), 'r', encoding = 'utf-8'):
                            yield line.split()
sentences_ = MySentences(dirname)
model = Word2Vec(min_count = 10, size = 200)
model.build_vocab(sentences_) # to test multiple parameters later, it is much convenient to first build vocabulary and save it
#model = Word2Vec.load('D:/mltool/word2vec/LMLM_word2vec_MIN_10')

sentences_ = MySentences(dirname)
model.window = 5
model.workers = 5
model.size = 200
model.sg = 0 # allows faster training
model.train(sentences__, total_examples = model.corpus_count, epochs = 10)
model.wv.save_word2vec_format('D:/mltool/kv_LMLM_dim_200_MIN_10', binary = False) # save keyed vector from the model

Now, finally build spacy model using keyed vector constructed right before.

Type into CMD window:

python -m spacy init-model en your_spacy_model_name --vectors-loc keyed vector location




[NK Provocation Index][1] Identification 1 – LDA model


import re
import os
import sys
import pandas as pd
import numpy as np
import pandas as pd
from pprint import pprint
import random
import gensim
import gensim.corpora as corpora
from konlpy.tag import Twitter
from operator import itemgetter
import datetime as dt 
import logging
logging.basicConfig(format='%(asctime)s : %(levelname)s : %(message)s', level=logging.INFO)
import warnings

data = []
dirname = 'D://nk//data'
header = {}
k = 0
for f_dir in os.listdir(dirname):
    for fname in os.listdir(os.path.join(dirname, f_dir)):
        k += 1
        if k % 10000 == 0:
        f = open(os.path.join(dirname, f_dir, fname), 'r', encoding = 'utf-8')
        data.append([f_dir, f.read()])

#Document specific Preprocessing

a1 = re.compile('등록\s*\:\s*(\d{4}\s*\-\s*\d{2}\s*\-\s*\d{2})')
a2 = re.compile('입력\s*(\d{4}\s*[\.\-]\s*\d{2}\s*[\.\-]\s*\d{2})')
a2_2 = re.compile('(\d{4}\s*[\.\-]\s*\d{2}\s*[\.\-]\s*\d{2})')
a2_3 = re.compile('등록\s*\:\(\d{4}\-\d{2}\-\d{2})')
a3 = re.compile('((?:19|20)\d{6})')
a4 = re.compile('(\d{4}\s*\-\s*\d{2}\s*\-\s*\d{2})')
for item in data:
    k = 0
    for a in [a1, a2, a2_2, a2_3, a4, a3]:
        if a.search(item[1]):
            date = re.sub('[\s\-\.]', '', a.search(item[1])[1])
            k = 1
            data_date.append([item[0], date, item[1]])
            if len(str(date)) != 8:
    #if k == 0 :

twitter = Twitter()

def sent_to_words(sentences):
    return twitter.morphs(sentences)  # deacc=True removes punctuations

def remove_stopwords(texts):
    return [[word for word in preprocess(str(doc)) if word not in stopwords] for doc in texts]

def make_bigrams(texts):
    return [bigram_mod[doc] for doc in texts]

def make_trigrams(texts):
    return [trigram_mod[bigram_mod[doc]] for doc in texts]

def preprocess(doc):
    doc = re.sub('\s+', ' ', doc)
    doc = re.sub('[A-Za-z]+[0-9]+', '', doc)
    doc = re.sub('[a-zA-Z]+', ' ', doc)
    doc = re.sub('\s+', ' ', doc)
    return doc

#Remove all the one character words except for the name of country

country_list = ['미', '북', '러', '중', '일', '한', '군', '핵', '당', '말', '남']
data = [[re.sub('[^가-힣\s\_]', '', word) for word in item] for item in data]
data = [[word for word in item if (len(word) > 1) or (word in country_list)] for item in data]

data_words = list(zip(header, data))

#Remove document specific stopwords

stop_words = ['아티클', '중앙일보', '조선일보', '동아일보', '한겨레', '구독', '관련기사', '아티', '클관련', '추가', '지면보기',
'종합', '뉴스', '사진', '밝혔', '이라고', '등록', '라고', '라며', '내용', '보다', '경우', '지역', '위해', '이라는', '그런', '처럼', '이나', '같은', '보다', '는데', '다면', '그것', '이제',
'때문', '다시', '많은', '정도', '일이', '없었', '되었', '인가', '않는', '베스트추천', '기자', '수정']
data_words = [[item[0], item[1], [word for word in item[2] if word not in stop_words]] for item in data_words]

def get_topic(txt):
    corpus = id2word.doc2bow(txt)
    topic = list(lda_model.get_document_topics(corpus))
    return sorted(topic, key=itemgetter(1))[-1][0]

#return topics and the corresponding words and weights.

#Drawing Graph

result = [[item[0], item[1], get_topic(item[2])] for item in data_words]
df = pd.DataFrame(result)
df.columns = ['news', 'date', 'topic']
df.date = df.date.astype(int)
df = df[df.date >= 199501]

#Remove errors in date information

for i in range(len(df)):
    if int(str(df['date'].iloc[i])[4:6]) > 12 or int(str(df['date'].iloc[i])[4:6]) == 0:
        print(i, df['date'].iloc[i])
for i in range(len(df)):
        x = df['date'].iloc[i]
        dt.date(int(str(x)[0:4]), int(str(x)[4:6]), int(str(x)[6:]))
        print(i, df['date'].iloc[i])

df.drop(df.index[[42981, 43438]], inplace = True)
df.drop(df.index[[65986, 74283]], inplace = True)

#Process date information  

df['date'] = df['date'].apply(lambda x: dt.date(int(str(x)[0:4]), int(str(x)[4:6]), int(str(x)[6:])))
df['date'] = pd.to_datetime(df['date'])

#Adjust sample ratio by putting different weight.


count_d = df.groupby(['date', 'topic']).size().reset_index(name = 'count')
count_d['adj'] = count_d['count']
mask1 = (count_d['date'] < dt.date(2005, 1, 1))
mask2 = (dt.date(2005, 1, 1) < count_d['date']) & (count_d['date']< dt.date(2009, 10, 17))
count_d.loc[mask1, 'adj'] = count_d.loc[mask1, 'count'] * 2
count_d.loc[mask2, 'adj'] = count_d.loc[mask2, 'count'] * (4/3)


df_ = df.copy()
df_['date'] = pd.to_datetime(df_['date'])
df_.set_index('date', inplace = True)
df_ = df_.to_period('M').to_timestamp('M')
count_m = df_.groupby(['date', 'topic']).size().reset_index(name = 'count')
count_m['adj'] = count_m['count']
mask1 = (count_m['date'] < dt.date(2005, 1, 1))
mask2 = (dt.date(2005, 1, 1) &lt; count_m[&#039;date&#039;]) &amp; (count_m[&#039;date&#039;]<dt>= dt.date(2000,1,1)) &amp; (count_d.date = dt.date(2000,1,1)) &amp; (count_m.date = dt.date(2000,1,1)) &amp; (count_y.date = dt.date(2018, 1, 1))
            count_y.loc[mask3, 'adj'] = count_y.loc[mask3, 'count'] * 6/5
            return None
        #color = '#00BFFF'
        x = xy['date']
        if adj:
            y = xy['adj']
            y = xy['count']
        #ax.plot(x, y, alpha = 0.8, c = color, linewidth=1.3)
        #print(label, topic)
        ax.plot(x, y, alpha = 0.7, linewidth=1.3, label = 'topic # : %s(%s)'%(label, str(topic)))
        ax.set_xlim(dt.date(2000, 1,1), dt.date(2018, 12, 31))
        #ax.set_xlim(min(x), max(x))
    #ax.set_title("Spread and GDP", fontsize = 20)
    ax.set_xlabel('year', fontsize = 20)
    ax.set_ylabel('Number of Ariticles' , fontsize= 24)
    ax.grid(color='grey', linestyle='-', linewidth=0, alpha = 1)
    #ax.set_xticks([1995, 1997, 1999, 2001, 2003, 2005, 2007, 2009, 2011, 2013, 2015, 2017])
    ax.tick_params(axis = 'both', labelsize = 17)
    #if period == 'Y':
    #    plt.xticks(np.arange(dt.date(2000, 12, 31), dt.date(2018, 12, 31), dt.timedelta(731)))
    plt.xticks(np.arange(dt.date(2000, 1,1), dt.date(2018, 12, 31), dt.timedelta(731)))

#Draw Graphs (Day, Month, Year frequency on provocation and peace topics respectively) 
dp([12, 16], ['provocation', 'nuclear'], 'D', adj = True)
dp([13, 14], ['South-North', 'Global'], 'D', adj = True)
dp([12, 16], ['provocation', 'nuclear'], 'M', adj = True)
dp([13, 14], ['South-North', 'Global'], 'M', adj = True)
dp([12, 16], ['provocatoin', 'nuclear'], 'Y', adj = True)
dp([13, 14], ['South-North', 'US-North'], 'Y', adj = True)

#Number of Total Articles

fig, ax = plt.subplots(figsize = (9.5, 6.5), dpi = 100)
xy = df.groupby(['date']).size().reset_index(name = 'count')
xy = xy[xy['date'] &gt; dt.date(1995, 1, 1)]
#color = '#00BFFF'
x = xy['date']
y = xy['count']
#ax.plot(x, y, alpha = 0.8, c = color, linewidth=1.3<code>)
#print(label, topic)
ax.plot(x, y, alpha = 0.8, linewidth=1.3, label = 'total')
ax.set_xlim(min(x), max(x))
#ax.set_title("Spread and GDP", fontsize = 20)
ax.set_xlabel('year', fontsize = 20)
ax.set_ylabel('Number of Ariticles' , fontsize= 24)
ax.grid(color='grey', linestyle='-', linewidth=0, alpha = 1)
#ax.set_xticks([1995, 1997, 1999, 2001, 2003, 2005, 2007, 2009, 2011, 2013, 2015, 2017])
#ax.tick_params(axis = 'both', labelsize = 17)
plt.xticks(np.arange(min(x), max(x), dt.timedelta(730)))

[NK Provocation Index][0] Intro

  • North Korean’s military provocations and nuclear threats are likely to hamper Korean Economic Growth
  • Possible Channel : Increased risk lead to Investment, Saving to decrease
  • X(N.K. Provocation)  (–> X'(Investment, saving(consumption) rate)  –> Y(Economic Growth)
  • Identification 1 : Measuring the degree of N.K. Provocation by number of articles belong to ‘Provocation/Nuclear threats’ topic (LDA topic model)
  • Identification 2 : Causality? VAR may be helpful

[NKPR 0] Building Caffe on Window (for anaconda environment)

I recall that installing Caffe on Window was one of the hardest steps on this project.


  • However,  some (small) problems arises depending on the different environment one has.  For me, installing VS 2015 raised error ;  a setup package is either missing or damaged, but no perfect help for this problem exists on the web. (Spend two days repeating shredding the whole VS 2015/reinstalling)


  • In addition, building PyCaffe requires python 3.5, while I have been using python 3.6 (anaconda) for my previous works. Since I do not want to change my working environment, I tried to install PyCaffe using anaconda environment setting(python 3.5). There are some settings that should be modified before installing.


  1. Create new environment for python 3.5. (e.g. conda create -n py35 python = 3.5.0 anaconda)
  2. Before using cmd, call the anaconda environment (e.g. conda activate py35)
  3. When modifying caffe\caffe\scripts\build_win.cmd according to the video above, set CONDA_ROOT variable as location to the python 3.5. environment conda
  4. Now follow the video!
  5. Done!