[ ML 文章收集 ] Topic Modeling and Latent Dirichlet Allocation (LDA) in Python

Source From Here 
Preface 
Topic modeling is a type of statistical modeling for discovering the abstract “topics” that occur in a collection of documents. Latent Dirichlet Allocation (LDA) is an example of topic model and is used to classify text in a document to a particular topic. It builds a topic per document model and words per topic model, modeled as Dirichlet distributions. 

Here we are going to apply LDA to a set of documents and split them into topics. Let’s get started! 

The Data 
The data set we’ll use is a list of over one million news headlines published over a period of 15 years and can be downloaded from Kaggle: (abcnews-date-text.csv.zip) 

view plaincopy to clipboardprint?

1. import pandas as pd  
2.   
3. data = pd.read_csv('abcnews-date-text.csv', error_bad_lines=False);  
4. data_text = data[['headline_text']]  
5. data_text['index'] = data_text.index  
6. documents = data_text  

Take a peek of the data: 

view plaincopy to clipboardprint?

1. print(len(documents))  
2. documents.head()  

Data Pre-processing 
We will perform the following steps: 

* Tokenization: Split the text into sentences and the sentences into words. Lowercase the words and remove punctuation.
* Words that have fewer than 3 characters are removed.
* All stopwords are removed.
* Words are lemmatized — words in third person are changed to first person and verbs in past and future tenses are changed into present. (lemmatisation)
* Words are stemmed — words are reduced to their root form. (stemming)

Loading gensim and nltk libraries: 

view plaincopy to clipboardprint?

1. import gensim  
2. from gensim.utils import simple_preprocess  
3. from gensim.parsing.preprocessing import STOPWORDS  
4. from nltk.stem import WordNetLemmatizer, SnowballStemmer  
5. from nltk.stem.porter import *  
6. import numpy as np  
7. np.random.seed(2018)  
8. import nltk  
9. nltk.download('wordnet')  
10.   
11. stemmer = PorterStemmer()  

Write a function to perform lemmatize and stem preprocessing steps on the data set: 

view plaincopy to clipboardprint?

1. def lemmatize_stemming(text):  
2.     return stemmer.stem(WordNetLemmatizer().lemmatize(text, pos='v'))  
3.   
4. def preprocess(text):  
5.     result = []  
6.     for token in gensim.utils.simple_preprocess(text):  
7.         if token not in gensim.parsing.preprocessing.STOPWORDS and len(token) > 3:  
8.             result.append(lemmatize_stemming(token))  
9.     return result  

Select a document to preview after preprocessing: 

view plaincopy to clipboardprint?

1. doc_sample = documents[documents['index'] == 4310]['headline_text'].iloc[0]  
2.   
3. print('original document: ')  
4. words = []  
5. for word in doc_sample.split(' '):  
6.     words.append(word)  
7. print(words)  
8. print('\n\n tokenized and lemmatized document: ')  
9. print(preprocess(doc_sample))  

Output: 

original document:
['rain', 'helps', 'dampen', 'bushfires']


tokenized and lemmatized document:
['rain', 'help', 'dampen', 'bushfir']

Preprocess the headline text, saving the results as ‘processed_docs’: 

view plaincopy to clipboardprint?

1. processed_docs = documents['headline_text'].map(preprocess)  
2. processed_docs[:10]  

Bag of Words on the Data set 
Create a dictionary from ‘processed_docs’ containing the number of times a word appears in the training set: 

view plaincopy to clipboardprint?

1. dictionary = gensim.corpora.Dictionary(processed_docs)  
2.   
3. count = 0  
4. for k, v in dictionary.iteritems():  
5.     print(k, v)  
6.     count += 1  
7.     if count > 10:  
8.         break  

Output: 

0 broadcast
1 commun
2 decid
3 licenc
4 awar
5 defam
6 wit
7 call
8 infrastructur
9 protect
10 summit

Filter out tokens that appear in: 

* less than 15 documents (absolute number) or
* more than 0.5 documents (fraction of total corpus size, not absolute number).
* after the above two steps, keep only the first 100000 most frequent tokens.

view plaincopy to clipboardprint?

1. dictionary.filter_extremes(no_below=15, no_above=0.5, keep_n=100000)  

Gensim doc2bow 
For each document we create a dictionary reporting how many words and how many times those words appear. Save this to ‘bow_corpus’, then check our selected document earlier: 

view plaincopy to clipboardprint?

1. bow_corpus = [dictionary.doc2bow(doc) for doc in processed_docs]  
2. bow_corpus[4310]  

Output: 

[(76, 1), (112, 1), (483, 1), (4021, 1)]

Preview Bag Of Words for our sample preprocessed document: 

view plaincopy to clipboardprint?

1. bow_doc_4310 = bow_corpus[4310]  
2. for i in range(len(bow_doc_4310)):  
3.     print("Word {} (\"{}\") appears {} time.".format(bow_doc_4310[i][0],   
4.                                                      dictionary[bow_doc_4310[i][0]],   
5.                                                      bow_doc_4310[i][1]))  

Output: 

Word 76 ("bushfir") appears 1 time.
Word 112 ("help") appears 1 time.
Word 483 ("rain") appears 1 time.
Word 4021 ("dampen") appears 1 time.

TF-IDF 
Create tf-idf model object using models.TfidfModel on ‘bow_corpus’ and save it to ‘tfidf’, then apply transformation to the entire corpus and call it ‘corpus_tfidf’. Finally we preview TF-IDF scores for our first document: 

view plaincopy to clipboardprint?

1. from gensim import corpora, models  
2. tfidf = models.TfidfModel(bow_corpus)  
3. corpus_tfidf = tfidf[bow_corpus]  
4. from pprint import pprint  
5. for doc in corpus_tfidf:  
6.     pprint(doc)  
7.     break  

Output: 

[(0, 0.5903603121911333),
(1, 0.3852450692300274),
(2, 0.4974556050119205),
(3, 0.505567858418396)]

Train our lda model using gensim.models.LdaMulticore and save it to ‘lda_model’: 

view plaincopy to clipboardprint?

1. lda_model = gensim.models.LdaMulticore(bow_corpus, num_topics=10, id2word=dictionary, passes=2, workers=2)  

For each topic, we will explore the words occurring in that topic and its relative weight: 

view plaincopy to clipboardprint?

1. for idx, topic in lda_model.print_topics(-1):  
2.     print('Topic: {} \nWords: {}'.format(idx, topic))  

Can you distinguish different topics using the words in each topic and their corresponding weights? 

Running LDA using TF-IDF 

view plaincopy to clipboardprint?

1. lda_model_tfidf = gensim.models.LdaMulticore(corpus_tfidf, num_topics=10, id2word=dictionary, passes=2, workers=4)  
2. for idx, topic in lda_model_tfidf.print_topics(-1):  
3.     print('Topic: {} Word: {}'.format(idx, topic))  

Again, can you distinguish different topics using the words in each topic and their corresponding weights? 

Performance evaluation by classifying sample document using LDA Bag of Words model 
We will check where our test document would be classified: 

view plaincopy to clipboardprint?

1. processed_docs[4310]  

Output: 

['rain', 'help', 'dampen', 'bushfir']

Show out the top 5 topics according to score decreasingly: 

view plaincopy to clipboardprint?

1. for index, score in sorted(lda_model[bow_corpus[4310]], key=lambda tup: -1*tup[1])[:5]:  
2.     print("\nScore: {}\t \nTopic: {}".format(score, lda_model.print_topic(index, 10)))  

Our test document has the highest probability to be part of the topic that our model assigned, which is the accurate classification. 

Performance evaluation by classifying sample document using LDA TF-IDF model. 

view plaincopy to clipboardprint?

1. for index, score in sorted(lda_model_tfidf[bow_corpus[4310]], key=lambda tup: -1*tup[1])[:5]:  
2.     print("\nScore: {}\t \nTopic: {}".format(score, lda_model_tfidf.print_topic(index, 10)))  

Our test document has the highest probability to be part of the topic that our model assigned, which is the accurate classification. 

Testing model on unseen document 

view plaincopy to clipboardprint?

1. unseen_document = 'How a Pentagon deal became an identity crisis for Google'  
2. bow_vector = dictionary.doc2bow(preprocess(unseen_document))  
3. for index, score in sorted(lda_model[bow_vector], key=lambda tup: -1*tup[1]):  
4.     print("Score: {}\t Topic: {}".format(score, lda_model.print_topic(index, 4)))  

Source code can be found on Github (Or lda_example.ipynb). I look forward to hearing any feedback or questions. 

Supplement 
* Youtube - (ML 7.7.A1) Dirichlet distribution