Managing textual data using pandas
Contents
Managing textual data using pandas¶
This section introduces how to prepare and manage textual data for analysis using pandas, a Python library for working with tabular data.
After reading this section, you should know:
how to import data into a pandas DataFrame
how to explore data stored in a pandas DataFrame
how to append data to a pandas DataFrame
how to save the data in a pandas DataFrame
Importing data to pandas¶
Let’s start by importing the pandas library.
Note that we can control the name of the imported module using the as addition to the import command. pandas is commonly abbreviated pd.
This allows us to use the variable pd to refer to the pandas library.
import pandas as pd
Importing data from a single file¶
You must often load and prepare the data yourself, either from a single file or from multiple files.
Typical formats for distributing corpora include CSV files, which stands for Comma-separated Values, and JSON, which stands for JavaScript Object Notation or simple plain text files.
pandas provides plenty of functions for reading data in various formats. You can even try importing Excel sheets!
The following example shows how to load a corpus from a CSV file for processing in Python using the SFU Opinion and Comments Corpus (SOCC) (Kolhatkar et al. 2020).
Let’s load a part of the SFU Opinion and Comments Corpus, which contains the opinion articles from The Globe and Mail, a Canadian newspaper.
We can use the read_csv() function from pandas to read files with comma-separated values, such as the SOCC corpus.
The read_csv() function takes a string object as input, which defines a path to the input file.
# Read the CSV file and assign the output to the variable 'socc'
socc = pd.read_csv('data/socc_gnm_articles.csv')
pandas does all the heavy lifting and returns the contents of the CSV file in a pandas DataFrame, which is data structure native to pandas.
# Examine the type of the object stored under the variable 'socc'
type(socc)
pandas.core.frame.DataFrame
Let’s use the head() method of a DataFrame to check out the first five rows of the DataFrame.
# Print the first five rows of the DataFrame
socc.head(5)
| article_id | title | article_url | author | published_date | ncomments | ntop_level_comments | article_text | |
|---|---|---|---|---|---|---|---|---|
| 0 | 26842506 | The Tories deserve another mandate - Stephen H... | http://www.theglobeandmail.com/opinion/editori... | GLOBE EDITORIAL | 2015-10-16 EDT | 2187.0 | 1378.0 | <p>All elections are choices among imperfect a... |
| 1 | 26055892 | Harper hysteria a sign of closed liberal minds | http://www.theglobeandmail.com/opinion/harper-... | Konrad Yakabuski | 2015-08-24 EDT | 1103.0 | 455.0 | <p>If even a fraction of the darkness that his... |
| 2 | 6929035 | Too many first nations people live in a dream ... | http://www.theglobeandmail.com/opinion/too-man... | Jeffrey Simpson | 2013-01-05 EST | 1164.0 | 433.0 | <p>Large elements of aboriginal Canada live in... |
| 3 | 19047636 | The Globe's editorial board endorses Tim Hudak... | http://www.theglobeandmail.com/opinion/editori... | GLOBE EDITORIAL | 2014-06-06 EDT | 905.0 | 432.0 | <p>Over four days, The Globe editorial board l... |
| 4 | 11672346 | Disgruntled Arab states look to strip Canada o... | http://www.theglobeandmail.com/news/world/disg... | Campbell Clark | 2013-05-02 EDT | 1129.0 | 411.0 | <p>Growing discontent among Arab nations over ... |
As you can see, the DataFrame has a tabular form.
The DataFrame contains several columns such as article_id, title and article_text, accompanied by an index for each row (0, 1, 2, 3, 4).
The .at[] accessor can be used to inspect a single item in the DataFrame.
Let’s examine the value in the column title at index 123.
socc.at[123, 'title']
"How Toronto got a 'world-class,' gold-plated, half-billion-dollar empty train"
Importing data from multiple files¶
Another common scenario is that you have multiple files with text data, which you want to load into pandas.
Let’s first collect the files that we want to load.
# Import the patch library
from pathlib import Path
# Create a Path object that points to the directory with data
corpus_dir = Path('data')
# Get all .txt files in the corpus directory
corpus_files = list(corpus_dir.glob('*.txt'))
# Check the corpus files
corpus_files
[PosixPath('data/WP_1990-08-10-25A.txt'),
PosixPath('data/NYT_1991-01-16-A15.txt'),
PosixPath('data/WP_1991-01-17-A1B.txt')]
To accommodate our data, let’s create an empty pandas DataFrame and specify its shape in advance, that is, the number of rows (index) and the names of the columns columns.
We can determine the number of rows needed using Python’s range() function. This function generates a list of numbers that fall within certain range, which we can use for the index of the DataFrame.
In this case, we define a range() between 0 and the number of text files in the directory, which are stored under the variable corpus_files. We retrieve their number using the len() function, which returns the length of Python objects, if applicable.
For the columns of the DataFrame, we simply create columns for filenames and their textual content by providing a list of strings to the columns argument.
# Create a DataFrame and assign the result to the variable 'df'
df = pd.DataFrame(index=range(0, len(corpus_files)), columns=['filename', 'text'])
# Call the variable to inspect the output
df
| filename | text | |
|---|---|---|
| 0 | NaN | NaN |
| 1 | NaN | NaN |
| 2 | NaN | NaN |
Now that we have an empty data with rows for each file in the corpus, we can loop over the Path objects under corpus_files, read their contents and add them to the DataFrame.
# Loop over the corpus files and count each loop using enumerate()
for i, f in enumerate(corpus_files):
# Read the file contents
text = f.read_text(encoding='utf-8')
# Get the filename from the Path object
filename = f.name
# Assign the text from the file to index 'i' at column 'text'
# using the .at accessor – note that this modifies the DataFrame
# "in place" – you don't need to assign the result into a variable
df.at[i, 'text'] = text
# We then do the same to the filename
df.at[i, 'filename'] = filename
Let’s check the result by calling the variable df.
# Call the variable to check the output
df
| filename | text | |
|---|---|---|
| 0 | WP_1990-08-10-25A.txt | *We Don’t Stand for Bullies': Diverse Voices ... |
| 1 | NYT_1991-01-16-A15.txt | U.S. TAKING STEPS TO CURB TERRORISM: F.B.I. I... |
| 2 | WP_1991-01-17-A1B.txt | U.S., Allies Launch Massive Air War Against T... |
As you can see, the DataFrame has been populated with filenames and text.
Now that we know how to load data into DataFrames, we can turn towards accessing and manipulating the data that they store.
Examining DataFrames¶
pandas DataFrames can hold a lot of information, which is often organised into columns.
The columns present in a DataFrame are accessible through the attribute columns.
# Retrieve the columns and their names
socc.columns
Index(['article_id', 'title', 'article_url', 'author', 'published_date',
'ncomments', 'ntop_level_comments', 'article_text'],
dtype='object')
pandas provides various methods for examining the contents of entire columns, which can be accessed just like the keys and values of a Python dictionary.
The brackets [] can be used to access entire columns by placing the column name within the brackets as a string.
Let’s retrieve the contents of the column author, which contains author information.
# Retrieve the contents of the column 'author' in the DataFrame 'socc'
socc['author']
0 GLOBE EDITORIAL
1 Konrad Yakabuski
2 Jeffrey Simpson
3 GLOBE EDITORIAL
4 Campbell Clark
...
10334 GLOBE EDITORIAL
10335 GLOBE EDITORIAL
10336 GLOBE EDITORIAL
10337 Adam Radwanski
10338 GLOBE EDITORIAL
Name: author, Length: 10339, dtype: object
As you can see, the column author contains 10399 objects, as indicated by the Length and dtype properties. The numbers on the left-hand side give the index, that is, the row numbers.
The columns of a pandas DataFrame consist of another object type, namely pandas Series. You can think of the DataFrame as an entire table, whose columns consist of Series.
We can verify this by examining their type using Python’s type() function.
# Check the type of 'socc' and 'socc['author']'
type(socc), type(socc['author'])
(pandas.core.frame.DataFrame, pandas.core.series.Series)
When printing out the contents of a DataFrame or Series, pandas omits everything between the first and last five rows by default. This is convenient when working with thousands of rows.
This also applies to the output for methods such as value_counts(), which allows counting the number of unique values in a Series.
# Count unique values in the column 'author'
socc['author'].value_counts()
GLOBE EDITORIAL 2712
Jeffrey Simpson 649
Margaret Wente 547
Konrad Yakabuski 404
Gary Mason 365
...
MARK MATTSON AND KRYSTYN TULLY 1
KEVIN LACEY 1
Adrian Morrow 1
Jessica Scott-Reid 1
Kenneth Oppel 1
Name: author, Length: 1896, dtype: int64
Not surprisingly, the editorial team at the The Globe and Mail is responsible for most of the editorials!
Let’s take another look at the data by visualising the result by calling the .plot() method for the author information column. This method calls an external library named matplotlib, which can be used to produce all kinds of plots and visualisations.
More specifically, we instruct the plot() method to draw a bar chart by providing the string bar to the kind argument.
We also use the brackets [:10] to limit the output to the ten most profilic authors.
# This is some Jupyter magic that allows us to render matplotlib plots in the notebooks!
# You only need to enter this command once.
%matplotlib inline
# Count the values in the column 'author' and clip the result to top-10 before plotting.
socc['author'].value_counts()[:10].plot(kind='bar')
<AxesSubplot:>
For columns with numerical values, we can also use the describe() method to get basic descriptive statistics on the data.
# Get basic descriptive statistics for the column 'ntop_level_comments'
socc['ntop_level_comments'].describe()
count 10339.000000
mean 26.384273
std 39.786923
min 0.000000
25% 1.000000
50% 14.000000
75% 35.000000
max 1378.000000
Name: ntop_level_comments, dtype: float64
As we can see, the column ntop_level_comments has a total of 10339 rows.
The average number of comments received by an editorial is approximately 26, but this number fluctuates, as the standard deviation from the average is nearly 40.
Some editorials do not have any comments at all, as indicated by the minimum value of 0.
The lowest quartile shows that 25% of the data has only one comment or less (none).
The second quartile (50%), which is also known as the median, indicates that half of the data has less than 14 comments and half has more than 14 comments.
The third quartile shows that 75% of the data has 35 comments or less.
The most commented editorial has 1378 comments.
What if we would like to find the articles with zero comments?
We can use the DataFrame accessor .loc to access specific rows based on their values.
The number of comments is stored in the column ntop_level_comments, but we also need to specify that the DataFrame stored under the variable socc contains the column that we wish to examine.
This causes the somewhat repetitive reference to the socc DataFrame, which is nevertheless necessary for being explicit.
We also need to provide a command for “is equal to”. Since the single equal sign = is reserved for assigning variables in Python, two equal signs == are used for comparison.
Finally, we place the value we want to evaluate against on the right-hand side of the double equal sign ==, that is, zero for no comments.
# Get rows with no top level comments
socc.loc[socc['ntop_level_comments'] == 0]
| article_id | title | article_url | author | published_date | ncomments | ntop_level_comments | article_text | |
|---|---|---|---|---|---|---|---|---|
| 7797 | 33441604 | Joseph Boyden, where are you from? | http://www.theglobeandmail.com/opinion/joseph-... | Hayden King | 2016-12-28 EST | 0.0 | 0.0 | <p>Hayden King teaches in the School of Public... |
| 7798 | 33316285 | Globe editorial: Rejoice! Congress just gave t... | http://www.theglobeandmail.com/opinion/editori... | GLOBE EDITORIAL | 2016-12-13 EST | 0.0 | 0.0 | <p>The United States may have just elected a p... |
| 7799 | 33009790 | Police and La Presse: Warrants not warranted | http://www.theglobeandmail.com/opinion/editori... | GLOBE EDITORIAL | 2016-11-23 EST | 0.0 | 0.0 | <p>The discovery that the Montreal Police obta... |
| 7800 | 32970624 | The Galloway affair: Salem comes to UBC | http://www.theglobeandmail.com/opinion/the-gal... | Margaret Wente | 2016-11-22 EST | 0.0 | 0.0 | <p>I have a question about the Steven Galloway... |
| 7801 | 32927142 | Justice delayed: the law of unintended consequ... | http://www.theglobeandmail.com/opinion/unreaso... | BENJAMIN PERRIN | 2016-11-19 EST | 0.0 | 0.0 | <p>Benjamin Perrin is a law professor at the U... |
| ... | ... | ... | ... | ... | ... | ... | ... | ... |
| 10334 | 533784 | WTO action on China's rare-earth quotas makes ... | http://www.theglobeandmail.com/opinion/editori... | GLOBE EDITORIAL | 2012-03-14 EDT | 0.0 | 0.0 | <p>The confusingly named substances known as '... |
| 10335 | 533594 | A customer-friendly Finance Department | http://www.theglobeandmail.com/opinion/editori... | GLOBE EDITORIAL | 2012-03-13 EDT | 0.0 | 0.0 | <p>Of the many things that frustrate the retai... |
| 10336 | 533508 | Video raises questions about Nik Zoricic's 'fr... | http://www.theglobeandmail.com/opinion/editori... | GLOBE EDITORIAL | 2012-03-12 EDT | 0.0 | 0.0 | <p>Officials and fans are mourning the death o... |
| 10337 | 533504 | McGuinty can't afford misgivings about gaming | http://www.theglobeandmail.com/news/politics/m... | Adam Radwanski | 2012-03-12 EDT | 0.0 | 0.0 | <p>Unlike so many of the other measures that m... |
| 10338 | 533471 | In Russia, Canada should look for investment, ... | http://www.theglobeandmail.com/opinion/editori... | GLOBE EDITORIAL | 2012-03-12 EDT | 0.0 | 0.0 | <p>As if in swift response to Prime Minister V... |
2542 rows × 8 columns
This returns a total of 2542 rows where the value of the column ntop_level_comments is zero.
For more complex views of the data, we can also combine multiple criteria using the & symbol, which is the Python operator for “AND”.
Note that individual criteria must be placed in parentheses () to perform the operation.
Let’s check if the first author in our result, Hayden King, wrote any other articles with zero comments.
# Get number of top level comments for author Hayden King
socc.loc[(socc['ntop_level_comments'] == 0) & (socc['author'] == 'Hayden King')]
| article_id | title | article_url | author | published_date | ncomments | ntop_level_comments | article_text | |
|---|---|---|---|---|---|---|---|---|
| 7797 | 33441604 | Joseph Boyden, where are you from? | http://www.theglobeandmail.com/opinion/joseph-... | Hayden King | 2016-12-28 EST | 0.0 | 0.0 | <p>Hayden King teaches in the School of Public... |
Extending DataFrames¶
You can easily add information to pandas DataFrames.
One common scenario could involve loading some data from an external file (such as a CSV or JSON file), performing some analyses and storing the results to the same DataFrame.
We can easily add an empty column to the DataFrame. This is achieved using the column accessor [] and the Python datatype None.
Let’s add a new column named comments_ratio to the DataFrame socc.
# Add a new column named 'comments_ratio' to the DataFrame
socc['comments_ratio'] = None
# Print out the first five rows of the DataFrame
socc.head(5)
| article_id | title | article_url | author | published_date | ncomments | ntop_level_comments | article_text | comments_ratio | |
|---|---|---|---|---|---|---|---|---|---|
| 0 | 26842506 | The Tories deserve another mandate - Stephen H... | http://www.theglobeandmail.com/opinion/editori... | GLOBE EDITORIAL | 2015-10-16 EDT | 2187.0 | 1378.0 | <p>All elections are choices among imperfect a... | None |
| 1 | 26055892 | Harper hysteria a sign of closed liberal minds | http://www.theglobeandmail.com/opinion/harper-... | Konrad Yakabuski | 2015-08-24 EDT | 1103.0 | 455.0 | <p>If even a fraction of the darkness that his... | None |
| 2 | 6929035 | Too many first nations people live in a dream ... | http://www.theglobeandmail.com/opinion/too-man... | Jeffrey Simpson | 2013-01-05 EST | 1164.0 | 433.0 | <p>Large elements of aboriginal Canada live in... | None |
| 3 | 19047636 | The Globe's editorial board endorses Tim Hudak... | http://www.theglobeandmail.com/opinion/editori... | GLOBE EDITORIAL | 2014-06-06 EDT | 905.0 | 432.0 | <p>Over four days, The Globe editorial board l... | None |
| 4 | 11672346 | Disgruntled Arab states look to strip Canada o... | http://www.theglobeandmail.com/news/world/disg... | Campbell Clark | 2013-05-02 EDT | 1129.0 | 411.0 | <p>Growing discontent among Arab nations over ... | None |
Let’s populate the column with some data by calculating which percentage of the comments are top-level comments, assuming that a high percentage of top-level comments indicates comments about the article, whereas a lower percentage indicates more discussion about the comments posted.
To get the proportion of top-level comments out of all comments, we must divide the number of top-level comments by the number of all comments.
# Populate the 'comments_ratio' column by calculating the ratio of top-level comments and comments
socc['comments_ratio'] = socc['ntop_level_comments'] / socc['ncomments']
Column accessors can be used very flexibly to access and manipulate data stored in the DataFrame, as exemplified by the division.
# Print out the first five rows of the DataFrame
socc.head(5)
| article_id | title | article_url | author | published_date | ncomments | ntop_level_comments | article_text | comments_ratio | |
|---|---|---|---|---|---|---|---|---|---|
| 0 | 26842506 | The Tories deserve another mandate - Stephen H... | http://www.theglobeandmail.com/opinion/editori... | GLOBE EDITORIAL | 2015-10-16 EDT | 2187.0 | 1378.0 | <p>All elections are choices among imperfect a... | 0.630087 |
| 1 | 26055892 | Harper hysteria a sign of closed liberal minds | http://www.theglobeandmail.com/opinion/harper-... | Konrad Yakabuski | 2015-08-24 EDT | 1103.0 | 455.0 | <p>If even a fraction of the darkness that his... | 0.412511 |
| 2 | 6929035 | Too many first nations people live in a dream ... | http://www.theglobeandmail.com/opinion/too-man... | Jeffrey Simpson | 2013-01-05 EST | 1164.0 | 433.0 | <p>Large elements of aboriginal Canada live in... | 0.371993 |
| 3 | 19047636 | The Globe's editorial board endorses Tim Hudak... | http://www.theglobeandmail.com/opinion/editori... | GLOBE EDITORIAL | 2014-06-06 EDT | 905.0 | 432.0 | <p>Over four days, The Globe editorial board l... | 0.477348 |
| 4 | 11672346 | Disgruntled Arab states look to strip Canada o... | http://www.theglobeandmail.com/news/world/disg... | Campbell Clark | 2013-05-02 EDT | 1129.0 | 411.0 | <p>Growing discontent among Arab nations over ... | 0.364039 |
As you can see, the column comments_ratio now stores the result of our calculation!
However, we should also keep in mind that some articles did not receive any comments at all: thus we would have divided zero by zero.
Let’s examine these cases again by retrieving articles without comments, and use the head() method to limit the output to the first five rows.
# Print out the first five comments with no top-level comments
socc.loc[socc['ntop_level_comments'] == 0].head(5)
| article_id | title | article_url | author | published_date | ncomments | ntop_level_comments | article_text | comments_ratio | |
|---|---|---|---|---|---|---|---|---|---|
| 7797 | 33441604 | Joseph Boyden, where are you from? | http://www.theglobeandmail.com/opinion/joseph-... | Hayden King | 2016-12-28 EST | 0.0 | 0.0 | <p>Hayden King teaches in the School of Public... | NaN |
| 7798 | 33316285 | Globe editorial: Rejoice! Congress just gave t... | http://www.theglobeandmail.com/opinion/editori... | GLOBE EDITORIAL | 2016-12-13 EST | 0.0 | 0.0 | <p>The United States may have just elected a p... | NaN |
| 7799 | 33009790 | Police and La Presse: Warrants not warranted | http://www.theglobeandmail.com/opinion/editori... | GLOBE EDITORIAL | 2016-11-23 EST | 0.0 | 0.0 | <p>The discovery that the Montreal Police obta... | NaN |
| 7800 | 32970624 | The Galloway affair: Salem comes to UBC | http://www.theglobeandmail.com/opinion/the-gal... | Margaret Wente | 2016-11-22 EST | 0.0 | 0.0 | <p>I have a question about the Steven Galloway... | NaN |
| 7801 | 32927142 | Justice delayed: the law of unintended consequ... | http://www.theglobeandmail.com/opinion/unreaso... | BENJAMIN PERRIN | 2016-11-19 EST | 0.0 | 0.0 | <p>Benjamin Perrin is a law professor at the U... | NaN |
For these rows, the comments_ratio column contains values marked as NaN or “not a number”.
This indicates that the division was performed on these cells as well, but the result was not a number.
pandas automatically ignores NaN values when performing calculations, as show by the .describe() method.
# Get descriptive statistics for the column 'comments_ratio'
socc['comments_ratio'].describe()
count 7797.000000
mean 0.537057
std 0.205398
min 0.083333
25% 0.384615
50% 0.485714
75% 0.647059
max 1.000000
Name: comments_ratio, dtype: float64
Note the difference in the result for the count. Only 7797 items out of 10399 were included in the calculation.
What if we would like to do some natural language processing and store the results in the DataFrame?
Let’s select articles that fall within the first quartile in terms of the ratio of original comments to all comments made (comments_ratio) and have received more than 200 comments (ncomments).
# Filter the DataFrame for highly commented articles and assign the result to the variable 'talk'
talk = socc.loc[(socc['comments_ratio'] <= 0.384) & (socc['ncomments'] >= 200)]
# Call the variable to examine the output
talk.head(5)
| article_id | title | article_url | author | published_date | ncomments | ntop_level_comments | article_text | comments_ratio | |
|---|---|---|---|---|---|---|---|---|---|
| 2 | 6929035 | Too many first nations people live in a dream ... | http://www.theglobeandmail.com/opinion/too-man... | Jeffrey Simpson | 2013-01-05 EST | 1164.0 | 433.0 | <p>Large elements of aboriginal Canada live in... | 0.371993 |
| 4 | 11672346 | Disgruntled Arab states look to strip Canada o... | http://www.theglobeandmail.com/news/world/disg... | Campbell Clark | 2013-05-02 EDT | 1129.0 | 411.0 | <p>Growing discontent among Arab nations over ... | 0.364039 |
| 5 | 26691065 | Fifty years in Canada, and now I feel like a s... | http://www.theglobeandmail.com/opinion/fifty-y... | SHEEMA KHAN | 2015-10-07 EDT | 1142.0 | 376.0 | <p>'Too broken to write,' I told my editor, af... | 0.329247 |
| 6 | 25731634 | I'm Canadian - and I should have a right to vote | http://www.theglobeandmail.com/opinion/im-cana... | Donald Sutherland | 2015-07-28 EDT | 1021.0 | 348.0 | <p>My name is Donald Sutherland. My wife's nam... | 0.340842 |
| 8 | 13647608 | A nation of $100,000 firefighters | http://www.theglobeandmail.com/opinion/a-natio... | Margaret Wente | 2013-08-08 EDT | 1102.0 | 338.0 | <p>Everyone loves firefighters. They save live... | 0.306715 |
Let’s import spaCy, load a medium-sized language model for English and assign this model to the variable nlp.
# Import the spaCy library
import spacy
# Note that we now load a medium-sized language model!
nlp = spacy.load('en_core_web_md')
Let’s limit processing to article titles and create a placeholder column to the DataFrame named processed_title.
# Create a new column named 'processed_title'
talk['processed_title'] = None
pandas warns about performing this command, because talk is only a slice or a view into the DataFrame.
Assigning a new column to only a part of the DataFrame would cause problems by breaking the tabular structure.
We can fix the situation by creating a deep copy of the slice using Python’s .copy() method.
# Create a deep copy of the DataFrame
talk = talk.copy()
Let’s try creating an empty column again.
# Create a new column named 'processed_title'
talk['processed_title'] = None
# Print out the first five rows of the DataFrame
talk.head(5)
| article_id | title | article_url | author | published_date | ncomments | ntop_level_comments | article_text | comments_ratio | processed_title | |
|---|---|---|---|---|---|---|---|---|---|---|
| 2 | 6929035 | Too many first nations people live in a dream ... | http://www.theglobeandmail.com/opinion/too-man... | Jeffrey Simpson | 2013-01-05 EST | 1164.0 | 433.0 | <p>Large elements of aboriginal Canada live in... | 0.371993 | None |
| 4 | 11672346 | Disgruntled Arab states look to strip Canada o... | http://www.theglobeandmail.com/news/world/disg... | Campbell Clark | 2013-05-02 EDT | 1129.0 | 411.0 | <p>Growing discontent among Arab nations over ... | 0.364039 | None |
| 5 | 26691065 | Fifty years in Canada, and now I feel like a s... | http://www.theglobeandmail.com/opinion/fifty-y... | SHEEMA KHAN | 2015-10-07 EDT | 1142.0 | 376.0 | <p>'Too broken to write,' I told my editor, af... | 0.329247 | None |
| 6 | 25731634 | I'm Canadian - and I should have a right to vote | http://www.theglobeandmail.com/opinion/im-cana... | Donald Sutherland | 2015-07-28 EDT | 1021.0 | 348.0 | <p>My name is Donald Sutherland. My wife's nam... | 0.340842 | None |
| 8 | 13647608 | A nation of $100,000 firefighters | http://www.theglobeandmail.com/opinion/a-natio... | Margaret Wente | 2013-08-08 EDT | 1102.0 | 338.0 | <p>Everyone loves firefighters. They save live... | 0.306715 | None |
To retrieve the title for each article from the column title, feed it to the language model under nlp for processing and store the output into the column processed_title, we need to use the apply() method of a DataFrame.
As the name suggests, the apply() method applies whatever is provided as input to the method to each row in the column.
In this case, we pass the language model nlp to the apply() method, essentially retrieving the titles stored as string objects in the column title and “applying” the language model nlp to them.
We assign the output to the DataFrame column named processed_title.
# Apply the language model under 'nlp' to the contents of the DataFrame column 'title'
talk['processed_title'] = talk['title'].apply(nlp)
# Call the variable to check the output
talk
| article_id | title | article_url | author | published_date | ncomments | ntop_level_comments | article_text | comments_ratio | processed_title | |
|---|---|---|---|---|---|---|---|---|---|---|
| 2 | 6929035 | Too many first nations people live in a dream ... | http://www.theglobeandmail.com/opinion/too-man... | Jeffrey Simpson | 2013-01-05 EST | 1164.0 | 433.0 | <p>Large elements of aboriginal Canada live in... | 0.371993 | (Too, many, first, nations, people, live, in, ... |
| 4 | 11672346 | Disgruntled Arab states look to strip Canada o... | http://www.theglobeandmail.com/news/world/disg... | Campbell Clark | 2013-05-02 EDT | 1129.0 | 411.0 | <p>Growing discontent among Arab nations over ... | 0.364039 | (Disgruntled, Arab, states, look, to, strip, C... |
| 5 | 26691065 | Fifty years in Canada, and now I feel like a s... | http://www.theglobeandmail.com/opinion/fifty-y... | SHEEMA KHAN | 2015-10-07 EDT | 1142.0 | 376.0 | <p>'Too broken to write,' I told my editor, af... | 0.329247 | (Fifty, years, in, Canada, ,, and, now, I, fee... |
| 6 | 25731634 | I'm Canadian - and I should have a right to vote | http://www.theglobeandmail.com/opinion/im-cana... | Donald Sutherland | 2015-07-28 EDT | 1021.0 | 348.0 | <p>My name is Donald Sutherland. My wife's nam... | 0.340842 | (I, 'm, Canadian, -, and, I, should, have, a, ... |
| 8 | 13647608 | A nation of $100,000 firefighters | http://www.theglobeandmail.com/opinion/a-natio... | Margaret Wente | 2013-08-08 EDT | 1102.0 | 338.0 | <p>Everyone loves firefighters. They save live... | 0.306715 | (A, nation, of, $, 100,000, firefighters) |
| ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... |
| 1694 | 30474884 | A dangerous moment in history: Can the politic... | http://www.theglobeandmail.com/opinion/can-the... | Konrad Yakabuski | 2016-06-16 EDT | 239.0 | 50.0 | <p>As anyone trying to maintain perspective wh... | 0.209205 | (A, dangerous, moment, in, history, :, Can, th... |
| 1735 | 32088785 | Clinton shines in first debate, and not just i... | http://www.theglobeandmail.com/opinion/editori... | GLOBE EDITORIAL | 2016-09-27 EDT | 232.0 | 49.0 | <p>For those who wondered whether Hillary Clin... | 0.211207 | (Clinton, shines, in, first, debate, ,, and, n... |
| 2213 | 30508530 | U.S. gun control: Don't look for logic after O... | http://www.theglobeandmail.com/opinion/us-gun-... | Konrad Yakabuski | 2016-06-20 EDT | 243.0 | 40.0 | <p>The script is by now tediously formulaic. A... | 0.164609 | (U.S., gun, control, :, Do, n't, look, for, lo... |
| 2301 | 31605288 | Let's make sure Ontario's sex-ed curriculum is... | http://www.theglobeandmail.com/opinion/lets-ma... | DEBRA SOH | 2016-08-30 EDT | 239.0 | 39.0 | <p>Debra W. Soh is a sex writer and sexual neu... | 0.163180 | (Let, 's, make, sure, Ontario, 's, sex, -, ed,... |
| 2302 | 24363093 | Dad rules when sex ed collides with religion | http://www.theglobeandmail.com/opinion/dad-rul... | MICHAEL ADAMS | 2015-05-11 EDT | 222.0 | 39.0 | <p>Michael Adams is founder and president of t... | 0.175676 | (Dad, rules, when, sex, ed, collides, with, re... |
519 rows × 10 columns
We now have the processed titles in a separate column named processed_title!
Let’s examine the first row in the DataFrame talk, whose index is 2.
# Get the value in the column 'processed_title' at row with index 2
talk.at[2, 'processed_title']
Too many first nations people live in a dream palace
# Check the type of the contained object
type(talk.at[2, 'processed_title'])
spacy.tokens.doc.Doc
As you can see, the cell contains a spaCy Doc object.
Let’s now define our own Python function to fetch lemmas for each noun in the title.
Python functions are defined using the command def, which is followed by the name of the function, in this case get_nouns.
The input to the function is given in parentheses that follow the name of the function.
In this case, we name a variable for the input called nlp_text. This is an arbitrary variable, which is needed for referring to whatever is being provided as input to the function. To put it simply, you can think of this variable as referring to any input that will be eventually provided to the function.
# Define a function named 'get_nouns' that takes a single object as input.
# We refer to this input using the variable name 'nlp_text'.
def get_nouns(nlp_text):
# First we make sure that the input is of correct type
# by using the assert command to check the input type
assert type(nlp_text) == spacy.tokens.doc.Doc
# Let's set up a placeholder list for our lemmas
lemmas = []
# We begin then begin looping over the Doc object
for token in nlp_text:
# If the fine-grained POS tag for the token is a noun (NN)
if token.tag_ == 'NN':
# Append the token lemma to the list of lemmas
lemmas.append(token.lemma_)
# When the loop is complete, return the list of lemmas
return lemmas
Now that we have defined our function, we can use the function with the .apply() method to collect all nouns to the column nouns.
# Apply the 'get_nouns' function to the column 'processed_title'
talk['nouns'] = talk['processed_title'].apply(get_nouns)
# Call the variable to examine the output
talk
| article_id | title | article_url | author | published_date | ncomments | ntop_level_comments | article_text | comments_ratio | processed_title | nouns | |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 2 | 6929035 | Too many first nations people live in a dream ... | http://www.theglobeandmail.com/opinion/too-man... | Jeffrey Simpson | 2013-01-05 EST | 1164.0 | 433.0 | <p>Large elements of aboriginal Canada live in... | 0.371993 | (Too, many, first, nations, people, live, in, ... | [dream, palace] |
| 4 | 11672346 | Disgruntled Arab states look to strip Canada o... | http://www.theglobeandmail.com/news/world/disg... | Campbell Clark | 2013-05-02 EDT | 1129.0 | 411.0 | <p>Growing discontent among Arab nations over ... | 0.364039 | (Disgruntled, Arab, states, look, to, strip, C... | [agency] |
| 5 | 26691065 | Fifty years in Canada, and now I feel like a s... | http://www.theglobeandmail.com/opinion/fifty-y... | SHEEMA KHAN | 2015-10-07 EDT | 1142.0 | 376.0 | <p>'Too broken to write,' I told my editor, af... | 0.329247 | (Fifty, years, in, Canada, ,, and, now, I, fee... | [class, citizen] |
| 6 | 25731634 | I'm Canadian - and I should have a right to vote | http://www.theglobeandmail.com/opinion/im-cana... | Donald Sutherland | 2015-07-28 EDT | 1021.0 | 348.0 | <p>My name is Donald Sutherland. My wife's nam... | 0.340842 | (I, 'm, Canadian, -, and, I, should, have, a, ... | [right] |
| 8 | 13647608 | A nation of $100,000 firefighters | http://www.theglobeandmail.com/opinion/a-natio... | Margaret Wente | 2013-08-08 EDT | 1102.0 | 338.0 | <p>Everyone loves firefighters. They save live... | 0.306715 | (A, nation, of, $, 100,000, firefighters) | [nation] |
| ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... | ... |
| 1694 | 30474884 | A dangerous moment in history: Can the politic... | http://www.theglobeandmail.com/opinion/can-the... | Konrad Yakabuski | 2016-06-16 EDT | 239.0 | 50.0 | <p>As anyone trying to maintain perspective wh... | 0.209205 | (A, dangerous, moment, in, history, :, Can, th... | [moment, history, centre, hold] |
| 1735 | 32088785 | Clinton shines in first debate, and not just i... | http://www.theglobeandmail.com/opinion/editori... | GLOBE EDITORIAL | 2016-09-27 EDT | 232.0 | 49.0 | <p>For those who wondered whether Hillary Clin... | 0.211207 | (Clinton, shines, in, first, debate, ,, and, n... | [debate, comparison] |
| 2213 | 30508530 | U.S. gun control: Don't look for logic after O... | http://www.theglobeandmail.com/opinion/us-gun-... | Konrad Yakabuski | 2016-06-20 EDT | 243.0 | 40.0 | <p>The script is by now tediously formulaic. A... | 0.164609 | (U.S., gun, control, :, Do, n't, look, for, lo... | [gun, control, logic] |
| 2301 | 31605288 | Let's make sure Ontario's sex-ed curriculum is... | http://www.theglobeandmail.com/opinion/lets-ma... | DEBRA SOH | 2016-08-30 EDT | 239.0 | 39.0 | <p>Debra W. Soh is a sex writer and sexual neu... | 0.163180 | (Let, 's, make, sure, Ontario, 's, sex, -, ed,... | [sex, ed, curriculum] |
| 2302 | 24363093 | Dad rules when sex ed collides with religion | http://www.theglobeandmail.com/opinion/dad-rul... | MICHAEL ADAMS | 2015-05-11 EDT | 222.0 | 39.0 | <p>Michael Adams is founder and president of t... | 0.175676 | (Dad, rules, when, sex, ed, collides, with, re... | [sex, religion] |
519 rows × 11 columns
As you can see, an empty DataFrame column is actually not required for adding new data, because pandas creates a new column automatically through assignment, as exemplified by talk['nouns'].
We can also easily extract information from DataFrames into Python’s native data structures.
The tolist() method, for instance, can be used to extract the contents of a pandas Series into a list.
# Cast pandas Series to a list
noun_list = talk['nouns'].tolist()
# Call the variable to check the output
noun_list[:10]
[['dream', 'palace'],
['agency'],
['class', 'citizen'],
['right'],
['nation'],
[],
['reform'],
['leader', 'parade'],
['pm'],
['government', 'monopoly']]
What we have now under noun_list is a list of lists, because each row in the nouns column contains a list.
Let’s loop over the list and collect the items into a single list named final_list using the extend() method of a Python list.
# Set up the placeholder list
final_list = []
# Loop over each list in the list of lists
for nlist in noun_list:
# Extend the final list with the current list
final_list.extend(nlist)
Let’s briefly examine the first ten items in final list and then count the number of items in the list.
# Print the first ten items in the list
final_list[:10]
['dream',
'palace',
'agency',
'class',
'citizen',
'right',
'nation',
'reform',
'leader',
'parade']
# Check the length of the list
len(final_list)
884
To plot the 10 most frequent nouns, we can cast the final_list into a pandas Series, count the occurrences of each lemma using value_counts() and plot the result using the plot() method.
# Convert the list into a pandas Series, count unique nouns
# using the value_counts() method, get the 10 most frequent
# items [:10] and plot the result into a bar chart using the
# plot() method and its attribute 'kind'.
pd.Series(final_list).value_counts()[:10].plot(kind='bar')
<AxesSubplot:>
Saving DataFrames¶
pandas DataFrames can be easily saved as pickled objects using the to_pickle() method.
The to_pickle() method takes a string as input, which defines a path to the file in which the DataFrame should be written.
Let’s pickle the DataFrame with the three articles stored under df into a file named pickled_df.pkl into the directory data.
# Write the DataFrame to disk using pickle
df.to_pickle('data/pickled_df.pkl')
We can easily check if the data has been saved successfully by reading the file contents using the read_pickle() method.
# Read the pickled DataFrame and assign the result to 'df_2'
df_2 = pd.read_pickle('data/pickled_df.pkl')
# Call the variable to examine the output
df_2
| filename | text | |
|---|---|---|
| 0 | WP_1990-08-10-25A.txt | *We Don’t Stand for Bullies': Diverse Voices ... |
| 1 | NYT_1991-01-16-A15.txt | U.S. TAKING STEPS TO CURB TERRORISM: F.B.I. I... |
| 2 | WP_1991-01-17-A1B.txt | U.S., Allies Launch Massive Air War Against T... |
Let’s compare the DataFrames, which returns a Boolean value (True/False) for each cell.
# Compare DataFrames 'df' and 'df_2'
df == df_2
| filename | text | |
|---|---|---|
| 0 | True | True |
| 1 | True | True |
| 2 | True | True |
This section should have given you a basic idea of the pandas library and how DataFrames can be used to store and manipulate textual data.
In Part III, you will learn more about natural language processing techniques and their applications.
