In [ ]:
# Common imports
%matplotlib inline
import matplotlib.pyplot as plt
import pylab
import numpy as np
# import pandas as pd
# import os
from os.path import isfile, join
# import scipy.io as sio
# import scipy
import zipfile as zp
# import shutil
# import difflib
import gensim
In this block we will work with collections of text documents. The objectives will be:
We will work with a collection of research projects funded by the US National Science Foundation, that you can find under the ./data directory. These files are publicly available from the NSF website.
NSF project information is provided in XML files. Projects are yearly grouped in .zip files, and each project is saved in a different XML file. To explore the structure of such files, we will use the file 160057.xml. Parsing XML files in python is rather easy using the ElementTree module.
In [ ]:
xmlfile = '../data/1600057.xml'
with open(xmlfile,'r') as fin:
print(fin.read())
XML is an inherently hierarchical data format, and the most natural way to represent it is with a tree. The ElementTree module has two classes for this purpose:
ElementTree represents the whole XML document as a treeElement represents a single node in this treeWe can import XML data by reading an XML file:
In [ ]:
import xml.etree.ElementTree as ET
root = ET.fromstring(open(xmlfile,'r').read())
The code below implements a function that parses the XML files and provides as its output a dictionary with fields:
project_code (string)
title (string)
abstract (string)
budget (float)
year (string)
institution (tuple with elements: name, zipcode, and statecode)
In [ ]:
def parse_xmlproject(xml_string):
"""This function processess the specified XML field,
and outputs a dictionary with the desired project information
:xml_string: String with XML content
:Returns: Dictionary with indicated files
"""
root = ET.fromstring(xml_string)
dictio = {}
for child in root[0]:
if child.tag.lower() == 'awardtitle':
dictio['title'] = child.text
elif child.tag.lower() == 'awardeffectivedate':
dictio['year'] = str(child.text[-4:])
elif child.tag.lower() == 'awardamount':
dictio['budget'] = float(child.text)
elif child.tag.lower() == 'abstractnarration':
dictio['abstract'] = child.text
elif child.tag.lower() == 'awardid':
dictio['project_code'] = child.text
elif child.tag.lower() == 'institution':
#For the institution we have to access the children elements
#and search for the name, zipcode, and statecode only
name = ''
zipcode = ''
statecode = ''
for child2 in child:
if child2.tag.lower() == 'name':
name = child2.text
elif child2.tag.lower() == 'zipcode':
zipcode = child2.text
elif child2.tag.lower() == 'statecode':
statecode = child2.text
dictio['institution'] = (name, zipcode, statecode)
return dictio
parse_xmlproject(open(xmlfile,'r').read())
Now, we will use the function you just implemented, to create a database that we will use throughout this module.
For simplicity, and given that the dataset is not too large, we will keep all projects in the RAM. The dataset will consist of a list containing the dictionaries associated to each of the considered projects in a time interval.
In [ ]:
# Construct an iterator (or a list) for the years you want to work with
years = [2015, 2016]
datafiles_path = '../data/'
NSF_data = []
for year in years:
zpobj = zp.ZipFile(join(datafiles_path, str(year)+'.zip'))
for fileinzip in zpobj.namelist():
if fileinzip.endswith('xml'):
#Some files seem to be incorrectly parsed
try:
project_dictio = parse_xmlproject(zpobj.read(fileinzip))
if project_dictio['abstract']:
NSF_data.append(project_dictio)
except:
pass
We will extract some characteristics of the constructed dataset:
In [ ]:
print('Number of projects in dataset:', len(NSF_data))
# Budget
budget_data = list(map(lambda x: x['budget'], NSF_data))
print('Average budget of projects in dataset:', np.mean(budget_data))
# Institutions
insti_data = list(map(lambda x: x['institution'], NSF_data))
print('Number of unique institutions in dataset:', len(set(insti_data)))
# Counts per year
counts = dict()
for project in NSF_data:
counts[project['year']] = counts.get(project['year'],0) + 1
print('Breakdown of projects by starting year:')
for el in counts:
print(el, ':', counts[el])
For the rest of this notebook, we will work with the abstracts only. The list of all abstract will be the corpus we will work with.
In [ ]:
corpus_raw = list(map(lambda x: x['abstract'], NSF_data))
abstractlen_data = list(map(lambda x: len(x), corpus_raw))
print('Average length of projects abstracts (in characters):', np.mean(abstractlen_data))
Topic modelling algorithms process vectorized data. In order to apply them, we need to transform the raw text input data into a vector representation. To do so, we will remove irrelevant information from the text data and preserve as much relevant information as possible to capture the semantic content in the document collection.
Thus, we will proceed with the following steps:
For the first steps, we will use some of the powerful methods available from the Natural Language Toolkit. In order to use the word_tokenize method from nltk, you might need to get the appropriate libraries using nltk.download(). You must select option "d) Download", and identifier "punkt"
In [ ]:
from nltk import download
# You should comment this code fragment if the package is already available.
# download('punkt')
# download('stopwords')
We will create a list that contains just the abstracts in the dataset. As the order of the elements in a list is fixed, it will be later straightforward to match the processed abstracts to metadata associated to their corresponding projects.
Exercise 1: Generate a corpus of processed documents. For each document in corpus_raw complete the following steps:
In [ ]:
from nltk.tokenize import word_tokenize
from nltk.stem import SnowballStemmer, WordNetLemmatizer
from nltk.corpus import stopwords
# Maybe you can try the stemmer too.
# stemmer = SnowballStemmer('english')
wnl = WordNetLemmatizer()
stopwords_en = stopwords.words('english')
# Initialize ouput corpus
corpus_clean = []
ndocs = len(corpus_raw)
for n, text in enumerate(corpus_raw):
if not n%100:
print('\rTokenizing document', n, 'out of', ndocs, end='', flush=True)
# Tokenize each text entry.
# tokens = <FILL IN>
# tokens_filtered = <FILL IN>
# tokens_lemmatized = <FILL IN>
# tokens_clean = <FILL IN>
# Add the new token list as a new element to corpus_clean (that will be a list of lists)
# corpus_clean.<FILL IN>
print('\n\n The corpus has been tokenized. Check the result for the first abstract:')
print(corpus_raw[0])
print(corpus_clean[0])
Up to this point, we have transformed the raw text collection in a list of documents, where each documen is a collection of the words that are most relevant for semantic analysis. Now, we need to convert these data (a list of token lists) into a numerical representation (a list of vectors, or a matrix). To do so, we will start using the tools provided by the gensim library.
As a first step, we create a dictionary containing all tokens in our text corpus, and assigning an integer identifier to each one of them.
In [ ]:
# Create dictionary of tokens
D = gensim.corpora.Dictionary(corpus_clean)
n_tokens = len(D)
print('The dictionary contains', n_tokens, 'terms')
print('First terms in the dictionary:')
for n in range(10):
print(str(n), ':', D[n])
We can also filter out terms that appear in too few or too many of the documents in the dataset:
In [ ]:
no_below = 5 #Minimum number of documents to keep a term in the dictionary
no_above = .75 #Maximum proportion of documents in which a term can appear to be kept in the dictionary
D.filter_extremes(no_below=no_below, no_above=no_above, keep_n=25000)
n_tokens = len(D)
print('The dictionary contains', n_tokens, 'terms')
print('First terms in the dictionary:')
for n in range(10):
print(str(n), ':', D[n])
In the second step, let us create a numerical version of our corpus using the doc2bow method. In general, D.doc2bow(token_list) transforms any list of tokens into a list of tuples (token_id, n), one per each token in token_list, where token_id is the token identifier (according to dictionary D) and n is the number of occurrences of such token in token_list.
Exercise 2: Apply the doc2bow method from gensim dictionary D, to all tokens in every document in clean_abstracts. The result must be a new list named corpus_bow where each element is a list of tuples (token_id, number_of_occurrences).
In [ ]:
# corpus_bow = <FILL IN>
At this point, it is good to make sure to understand what has happened. In clean_abstracts we had a list of token lists. With it, we have constructed a Dictionary, D, which assigns an integer identifier to each token in the corpus.
After that, we have transformed each article (in clean_abstracts) in a list tuples (id, n).
In [ ]:
print('Original document (after cleaning):')
print(corpus_clean[0])
print('Sparse vector representation (first 10 components):')
print(corpus_bow[0][:10])
print('Word counts for the first project (first 10 components):')
print(list(map(lambda x: (D[x[0]], x[1]), corpus_bow[0][:10])))
Note that we can interpret each element of corpus_bow as a sparse_vector. For example, a list of tuples
[(0, 1), (3, 3), (5,2)]
for a dictionary of 10 elements can be represented as a vector, where any tuple (id, n) states that position id must take value n. The rest of positions must be zero.
[1, 0, 0, 3, 0, 2, 0, 0, 0, 0]
These sparse vectors will be the inputs to the topic modeling algorithms.
As a summary, the following variables will be relevant for the next chapters:
D: A gensim dictionary. Term strings can be accessed using the numeric identifiers. For instance, D[0] contains the string corresponding to the first position in the BoW representation.corpus_bow: BoW corpus. A list containing an entry per project in the dataset, and consisting of the (sparse) BoW representation for the abstract of that project.NSF_data: A list containing an entry per project in the dataset, and consisting of metadata for the projects in the datasetThe way we have constructed the corpus_bow variable guarantees that the order is preserved, so that the projects are listed in the same order in the lists corpus_bow and NSF_data.
Before starting with the semantic analyisis, it is interesting to observe the token distribution for the given corpus.
In [ ]:
# SORTED TOKEN FREQUENCIES (I):
# Create a "flat" corpus with all tuples in a single list
corpus_bow_flat = [item for sublist in corpus_bow for item in sublist]
# Initialize a numpy array that we will use to count tokens.
# token_count[n] should store the number of ocurrences of the n-th token, D[n]
token_count = np.zeros(n_tokens)
# Count the number of occurrences of each token.
for x in corpus_bow_flat:
# Update the proper element in token_count
# scode: <FILL IN>
# Sort by decreasing number of occurences
ids_sorted = np.argsort(- token_count)
tf_sorted = token_count[ids_sorted]
ids_sorted is a list of all token ids, sorted by decreasing number of occurrences in the whole corpus. For instance, the most frequent term is
In [ ]:
print(D[ids_sorted[0]])
which appears
In [ ]:
print("{0} times in the whole corpus".format(tf_sorted[0]))
In the following we plot the most frequent terms in the corpus.
In [ ]:
# SORTED TOKEN FREQUENCIES (II):
plt.rcdefaults()
# Example data
n_art = len(NSF_data)
n_bins = 25
hot_tokens = [D[i] for i in ids_sorted[n_bins-1::-1]]
y_pos = np.arange(len(hot_tokens))
z = tf_sorted[n_bins-1::-1]/n_art
plt.figure()
plt.barh(y_pos, z, align='center', alpha=0.4)
plt.yticks(y_pos, hot_tokens)
plt.xlabel('Average number of occurrences per article')
plt.title('Token distribution')
plt.show()
In [ ]:
# SORTED TOKEN FREQUENCIES:
# Example data
plt.figure()
plt.semilogy(tf_sorted)
plt.ylabel('Total number of occurrences')
plt.xlabel('Token rank')
plt.title('Token occurrences')
plt.show()
In [ ]:
# cold_tokens = <FILL IN>
print("There are {0} cold tokens, which represent {1}% of the total number of tokens in the dictionary".format(
len(cold_tokens), float(len(cold_tokens))/n_tokens*100))
In [ ]:
all_counts = [(D[el], D.dfs[el]) for el in D.dfs]
all_counts = sorted(all_counts, key=lambda x: x[1])
There are several implementations of the LDA topic model in python:
lda.gensim.models.ldamodel.LdaModelsklearn.decompositionIn the following sections we explore the use of gensim
Since we already have computed the dictionary and documents BoW representation using Gensim, computing the topic model is straightforward using the LdaModel() function. Please, refer to Gensim API documentation for more information on the different parameters accepted by the function:
Exercise 3: Create an LDA model with the 50 topics using corpus_bow and the dictionary, D.
In [ ]:
num_topics = 50
# This might take some time...
# ldag = gensim.models.ldamodel.LdaModel(<FILL IN>)
In [ ]:
ldag.print_topics(num_topics=-1, num_words=10)
A more useful visualization is provided by the python LDA visualization library, pyLDAvis.
Before executing the next code fragment you might need to install pyLDAvis:
>> pip install (--user) pyLDAvis
In [ ]:
import pyLDAvis.gensim as gensimvis
import pyLDAvis
vis_data = gensimvis.prepare(ldag, corpus_bow, D)
pyLDAvis.display(vis_data)
In addition to visualization purposes, topic models are useful to obtain a semantic representation of documents that can later be used with some other purpose:
Essentially, the idea is that the topic model provides a (semantic) vector representation of documents, and use probability divergences to measure document similarity. The following functions of the LdaModel class will be useful in this context:
get_topic_terms(topic_id): Gets vector of the probability distribution among words for the indicated topicget_document_topics(bow_vector): Gets (sparse) vector with the probability distribution among topics for the provided documentExercise 4: Show the probability distribution over words for topic 0.
In [ ]:
# <SOL>
# </SOL>
Exercise 5: Show the probability distribution over topics for document 0.
In [ ]:
print(ldag.get_document_topics(corpus_bow[0]))
An alternative to the use of the get_document_topics() function is to directly transform a dataset using the ldag object as follows. You can apply this transformation to several documents at once, but then the result is an iterator from which you can build the corresponding list if necessary
In [ ]:
print(ldag[corpus_bow[0]])
print('When applied to a dataset it will provide an iterator')
print(ldag[corpus_bow[:3]])
print('We can rebuild the list from the iterator with a one liner')
print([el for el in ldag[corpus_bow[:3]]])
Finally, Gensim provides some useful functions to convert between formats, and to simplify interaction with numpy and scipy. The following code fragment converts a corpus in sparse format to a full numpy matrix
In [ ]:
reduced_corpus = [el for el in ldag[corpus_bow[:3]]]
reduced_corpus = gensim.matutils.corpus2dense(reduced_corpus, num_topics).T
print(reduced_corpus)
Exercise 6: Build a function that returns the most relevant projects for a given topic
In [ ]:
def most_relevant_projects(ldag, topicid, corpus_bow, nprojects=10):
"""This function returns the most relevant projects in corpus_bow
: ldag: The trained topic model object provided by gensim
: topicid: The topic for which we want to find the most relevant documents
: corpus_bow: The BoW representation of documents in Gensim format
: nprojects: Number of most relevant projects to identify
: Returns: A list with the identifiers of the most relevant projects
"""
print('Computing most relevant projects for Topic', topicid)
print('Topic composition is:')
print(ldag.show_topic(topicid))
#<SOL>
#</SOL>
#To test the function we will find the most relevant projects for a subset of the NSF dataset
project_id = most_relevant_projects(ldag, 17, corpus_bow[:10000])
#Print titles of selected projects
for idproject in project_id:
print(NSF_data[idproject]['title'])
Exercise 7: Build a function that computes the semantic distance between two documents. For this, you can use the functions (or code fragments) provided in the library dist_utils.py.
In [ ]:
def pairwase_dist(doc1, doc2):
"""This function returns the Jensen-Shannon
distance between the corresponding vectors of the documents
: doc1: Semantic representation for the doc1 (a vector of length ntopics)
: doc2: Semantic representation for the doc2 (a vector of length ntopics)
: Returns: The JS distance between doc1 and doc2 (a number)
"""
#<SOL>
#</SOL>
In [ ]:
Exercise 8: Explore the influence of the concentration parameters, $alpha$ and $eta$. In particular observe how do topic and document distributions change as these parameters increase.
In [ ]:
Exercise 9: Note that we have not used the terms in the article titles, though the can be expected to contain relevant words for the topic modeling. Include the title words in the analyisis. In order to give them a special relevante, insert them in the corpus several times, so as to make their words more significant.
In [ ]:
In [ ]:
#print(NSF_data[0].keys())
#print(NSF_data[0]['institution'])
def strNone(str_to_convert):
if str_to_convert is None:
return ''
else:
return str_to_convert
with open('NSF_nodes.csv','w') as fout:
fout.write('Id;Title;Year;Budget;UnivName;UnivZIP;State\n')
for project in NSF_data:
fout.write(project['project_code']+';'+project['title']+';')
fout.write(project['year']+';'+str(project['budget'])+';')
fout.write(project['institution'][0]+';')
fout.write(strNone(project['institution'][1])+';')
fout.write(strNone(project['institution'][2])+'\n')
In [ ]: