In [1]:
# Common imports
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
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.
(As a side note, there are many other available text collections to work with. In particular, the NLTK library has many examples, that you can explore using the nltk.download()
tool.
import nltk
nltk.download()
for instance, you can take the gutemberg dataset
Mycorpus = nltk.corpus.gutenberg
text_name = Mycorpus.fileids()[0]
raw = Mycorpus.raw(text_name)
Words = Mycorpus.words(text_name)
Also, tools like Gensim or Sci-kit learn include text databases to work with).
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.
To introduce some common functions to work with XML files we will follow this tutorial.
In [2]:
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 XLM data by reading an XML file:
In [3]:
import xml.etree.ElementTree as ET
tree = ET.parse(xmlfile)
root = tree.getroot
or directly reading a string:
In [4]:
root = ET.fromstring(open(xmlfile,'r').read())
fromstring()
parses XML from a string directly into an Element
, which is the root element of the parsed tree. Other parsing functions may create an ElementTree
, but we will not cover them here.
As an Element
, root has a tag and a dictionary of attributes:
In [5]:
print(root.tag)
print(root.attrib)
It also has children nodes over which we can iterate:
In [6]:
for child in root:
print(child.tag, child.attrib)
Children are nested, and we can access specific child nodes by index. We can also access the text of specified elements. For instance:
In [7]:
for child in root[0]:
print(child.tag, child.attrib, child.text)
The presented classes and functions are all you need to solve the following exercise. However, there are many other interesting functions that can probably make it easier for you to work with XML files. For more information, please refer to the ElementTree API.
In [8]:
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
"""
#<SOL>
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
#</SOL>
parse_xmlproject(open(xmlfile,'r').read())
Out[8]:
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 [9]:
# Construct an iterator (or a list) for the years you want to work with
years = range(2015,2017)
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 [10]:
print('Number of projects in dataset:', len(NSF_data))
####
budget_data = list(map(lambda x: x['budget'], NSF_data))
print('Average budget of projects in dataset:', np.mean(budget_data))
####
insti_data = list(map(lambda x: x['institution'], NSF_data))
print('Number of unique institutions in dataset:', len(set(insti_data)))
####
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])
In [11]:
#<SOL>
abstractlen_data = list(map(lambda x: len(x['abstract']), NSF_data))
print('Average length of projects abstracts (in characters):', np.mean(abstractlen_data))
#</SOL>
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 [12]:
import nltk
# You should comment this code fragment if the package is already available.
# Select option "d) Download", and identifier "punkt"
# nltk.download()
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.
In [13]:
from nltk.tokenize import word_tokenize
NSF_abstracts = list(map(lambda x: x['abstract'], NSF_data))
tokenized_abstracts = []
nprojects = len(NSF_abstracts)
for n, abstract in enumerate(NSF_abstracts):
if not n%100:
print('\rTokenizing abstract', n, 'out of', nprojects, end='', flush=True)
tokenized_abstracts.append(word_tokenize(abstract))
print('\n\n The corpus has been tokenized. Check the result for the first abstract:')
print(NSF_abstracts[0])
print(tokenized_abstracts[0])
By looking at the tokenized corpus you may verify that there are many tokens that correspond to punktuation signs and other symbols that are not relevant to analyze the semantic content. They can be removed using the stemming or lemmatization tools from nltk
.
The homogeneization process will consist of:
In [14]:
filtered_abstracts = []
for n, abstract in enumerate(tokenized_abstracts):
if not n%100:
print('\rFiltering abstract', n, 'out of', nprojects, end='', flush=True)
#<SOL>
filtered_abstracts.append([el.lower() for el in abstract if el.isalnum()])
#</SOL>
print('\n',filtered_abstracts[0])
At this point, we can choose between applying a simple stemming or ussing lemmatization. We will try both to test their differences.
The lemmatizer from NLTK is based on WordNet. If you have not used wordnet before, you will likely need to download it from nltk (use the nltk.download() command)
In [15]:
stemmer = nltk.stem.SnowballStemmer('english')
from nltk.stem import WordNetLemmatizer
wnl = WordNetLemmatizer()
print('Result for the first abstract in dataset applying stemming')
print([stemmer.stem(el) for el in filtered_abstracts[0]])
print('Result for the first abstract in the dataset applying lemmatization')
print([wnl.lemmatize(el) for el in filtered_abstracts[0]])
One of the advantages of the lemmatizer method is that the result of lemmmatization is still a true word, which is more advisable for the presentation of text processing results and lemmatization.
However, without using contextual information, lemmatize() does not remove grammatical differences. This is the reason why "is" or "are" are preserved and not replaced by infinitive "be".
As an alternative, we can apply .lemmatize(word, pos), where 'pos' is a string code specifying the part-of-speech (pos), i.e. the grammatical role of the words in its sentence. For instance, you can check the difference between wnl.lemmatize('is')
and wnl.lemmatize('is, pos='v')
.
In [16]:
lemmatized_abstracts = []
for n, abstract in enumerate(filtered_abstracts):
if not n%100:
print('\rLemmatizing abstract', n, 'out of', nprojects, end='', flush=True)
#<SOL>
lemmatized_abstracts.append([wnl.lemmatize(el) for el in abstract])
#</SOL>
print('Result for the first abstract in the dataset applying lemmatization')
print('\n',lemmatized_abstracts[0])
In [17]:
from nltk.corpus import stopwords
stopwords_en = stopwords.words('english')
clean_abstracts = []
for n, abstract in enumerate(lemmatized_abstracts):
if not n%100:
print('\rCleaning abstract', n, 'out of', nprojects, end='', flush=True)
# Remove all tokens in the stopwords list and append the result to clean_abstracts
# <SOL>
clean_tokens = [token for token in abstract if token not in stopwords_en]
# </SOL>
clean_abstracts.append(clean_tokens)
print('\n Let us check tokens after cleaning:')
print(clean_abstracts[0])
Up to this point, we have transformed the raw text collection of articles in a list of articles, where each article is a collection of the word roots 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 [18]:
import gensim
# Create dictionary of tokens
D = gensim.corpora.Dictionary(clean_abstracts)
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 [19]:
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
.
In [20]:
corpus_bow = [D.doc2bow(doc) for doc in clean_abstracts]
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 [21]:
print('Original article (after cleaning):')
print(clean_abstracts[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
.
In [22]:
all_counts = [(D[el], D.dfs[el]) for el in D.dfs]
all_counts = sorted(all_counts, key=lambda x: x[1])
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:
In [32]:
import gensim
num_topics = 50
ldag = gensim.models.ldamodel.LdaModel(corpus=corpus_bow, id2word=D, num_topics=num_topics)
In [33]:
ldag.print_topics(num_topics=-1, num_words=10)
Out[33]:
A more useful visualization is provided by the python LDA visualization library, pyLDAvis.
Before executing the next code fragment you need to install pyLDAvis:
>> pip install (--user) pyLDAvis
In [34]:
import pyLDAvis.gensim as gensimvis
import pyLDAvis
vis_data = gensimvis.prepare(ldag, corpus_bow, D)
pyLDAvis.display(vis_data)
Out[34]:
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 document
In [39]:
ldag.get_topic_terms(topicid=0)
Out[39]:
In [40]:
ldag.get_document_topics(corpus_bow[0])
Out[40]:
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 [43]:
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 [59]:
reduced_corpus = [el for el in ldag[corpus_bow[:3]]]
reduced_corpus = gensim.matutils.corpus2dense(reduced_corpus, num_topics).T
print(reduced_corpus)
In [80]:
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>
document_topic = [el for el in ldag[corpus_bow]]
document_topic = gensim.matutils.corpus2dense(document_topic, ldag.num_topics).T
return np.argsort(document_topic[:,topicid])[::-1][:nprojects].tolist()
#</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'])
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 [19]:
#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')