Ccmmutty logo
Commutty IT
23 min read

03.10 Working With Strings
One strength of Python is its relative ease in handling and manipulating string data. Pandas builds on this and provides a comprehensive set of vectorized string operations that become an essential piece of the type of munging required when working with (read: cleaning up) real-world data. In this section, we'll walk through some of the Pandas string operations, and then take a look at using them to partially clean up a very messy dataset of recipes collected from the Internet.

Introducing Pandas String Operations

We saw in previous sections how tools like NumPy and Pandas generalize arithmetic operations so that we can easily and quickly perform the same operation on many array elements. For example:
import numpy as np
x = np.array([2, 3, 5, 7, 11, 13])
x * 2
array([ 4, 6, 10, 14, 22, 26])
This vectorization of operations simplifies the syntax of operating on arrays of data: we no longer have to worry about the size or shape of the array, but just about what operation we want done. For arrays of strings, NumPy does not provide such simple access, and thus you're stuck using a more verbose loop syntax:
data = ['peter', 'Paul', 'MARY', 'gUIDO']
[s.capitalize() for s in data]
['Peter', 'Paul', 'Mary', 'Guido']
This is perhaps sufficient to work with some data, but it will break if there are any missing values. For example:
data = ['peter', 'Paul', None, 'MARY', 'gUIDO']
[s.capitalize() for s in data]

AttributeError                            Traceback (most recent call last)

<ipython-input-3-fc1d891ab539> in <module>()
      1 data = ['peter', 'Paul', None, 'MARY', 'gUIDO']
----> 2 [s.capitalize() for s in data]

<ipython-input-3-fc1d891ab539> in <listcomp>(.0)
      1 data = ['peter', 'Paul', None, 'MARY', 'gUIDO']
----> 2 [s.capitalize() for s in data]

AttributeError: 'NoneType' object has no attribute 'capitalize'
Pandas includes features to address both this need for vectorized string operations and for correctly handling missing data via the str attribute of Pandas Series and Index objects containing strings. So, for example, suppose we create a Pandas Series with this data:
import pandas as pd
names = pd.Series(data)
0 peter 1 Paul 2 None 3 MARY 4 gUIDO dtype: object
We can now call a single method that will capitalize all the entries, while skipping over any missing values:
0 Peter 1 Paul 2 None 3 Mary 4 Guido dtype: object
Using tab completion on this str attribute will list all the vectorized string methods available to Pandas.

Tables of Pandas String Methods

If you have a good understanding of string manipulation in Python, most of Pandas string syntax is intuitive enough that it's probably sufficient to just list a table of available methods; we will start with that here, before diving deeper into a few of the subtleties. The examples in this section use the following series of names:
monte = pd.Series(['Graham Chapman', 'John Cleese', 'Terry Gilliam',
                   'Eric Idle', 'Terry Jones', 'Michael Palin'])

Methods similar to Python string methods

Nearly all Python's built-in string methods are mirrored by a Pandas vectorized string method. Here is a list of Pandas str methods that mirror Python string methods:
Notice that these have various return values. Some, like lower(), return a series of strings:
0 graham chapman 1 john cleese 2 terry gilliam 3 eric idle 4 terry jones 5 michael palin dtype: object
But some others return numbers:
0 14 1 11 2 13 3 9 4 11 5 13 dtype: int64
Or Boolean values:
0 False 1 False 2 True 3 False 4 True 5 False dtype: bool
Still others return lists or other compound values for each element:
0 [Graham, Chapman] 1 [John, Cleese] 2 [Terry, Gilliam] 3 [Eric, Idle] 4 [Terry, Jones] 5 [Michael, Palin] dtype: object
We'll see further manipulations of this kind of series-of-lists object as we continue our discussion.

Methods using regular expressions

In addition, there are several methods that accept regular expressions to examine the content of each string element, and follow some of the API conventions of Python's built-in re module:
match()Call re.match() on each element, returning a boolean.
extract()Call re.match() on each element, returning matched groups as strings.
findall()Call re.findall() on each element
replace()Replace occurrences of pattern with some other string
contains()Call on each element, returning a boolean
count()Count occurrences of pattern
split()Equivalent to str.split(), but accepts regexps
rsplit()Equivalent to str.rsplit(), but accepts regexps
With these, you can do a wide range of interesting operations. For example, we can extract the first name from each by asking for a contiguous group of characters at the beginning of each element:
monte.str.extract('([A-Za-z]+)', expand=False)
0 Graham 1 John 2 Terry 3 Eric 4 Terry 5 Michael dtype: object
Or we can do something more complicated, like finding all names that start and end with a consonant, making use of the start-of-string (^) and end-of-string ($) regular expression characters:
0 [Graham Chapman] 1 [] 2 [Terry Gilliam] 3 [] 4 [Terry Jones] 5 [Michael Palin] dtype: object
The ability to concisely apply regular expressions across Series or Dataframe entries opens up many possibilities for analysis and cleaning of data.

Miscellaneous methods

Finally, there are some miscellaneous methods that enable other convenient operations:
get()Index each element
slice()Slice each element
slice_replace()Replace slice in each element with passed value
cat()Concatenate strings
repeat()Repeat values
normalize()Return Unicode form of string
pad()Add whitespace to left, right, or both sides of strings
wrap()Split long strings into lines with length less than a given width
join()Join strings in each element of the Series with passed separator
get_dummies()extract dummy variables as a dataframe
Vectorized item access and slicing
The get() and slice() operations, in particular, enable vectorized element access from each array. For example, we can get a slice of the first three characters of each array using str.slice(0, 3). Note that this behavior is also available through Python's normal indexing syntax–for example, df.str.slice(0, 3) is equivalent to df.str[0:3]:
0 Gra 1 Joh 2 Ter 3 Eri 4 Ter 5 Mic dtype: object
Indexing via df.str.get(i) and df.str[i] is likewise similar.
These get() and slice() methods also let you access elements of arrays returned by split(). For example, to extract the last name of each entry, we can combine split() and get():
0 Chapman 1 Cleese 2 Gilliam 3 Idle 4 Jones 5 Palin dtype: object
Indicator variables
Another method that requires a bit of extra explanation is the get_dummies() method. This is useful when your data has a column containing some sort of coded indicator. For example, we might have a dataset that contains information in the form of codes, such as A="born in America," B="born in the United Kingdom," C="likes cheese," D="likes spam":
full_monte = pd.DataFrame({'name': monte,
                           'info': ['B|C|D', 'B|D', 'A|C',
                                    'B|D', 'B|C', 'B|C|D']})
The get_dummies() routine lets you quickly split-out these indicator variables into a DataFrame:
With these operations as building blocks, you can construct an endless range of string processing procedures when cleaning your data.
We won't dive further into these methods here, but I encourage you to read through "Working with Text Data" in the Pandas online documentation, or to refer to the resources listed in Further Resources.

Example: Recipe Database

These vectorized string operations become most useful in the process of cleaning up messy, real-world data. Here I'll walk through an example of that, using an open recipe database compiled from various sources on the Web. Our goal will be to parse the recipe data into ingredient lists, so we can quickly find a recipe based on some ingredients we have on hand.
The scripts used to compile this can be found at, and the link to the current version of the database is found there as well.
As of Spring 2016, this database is about 30 MB, and can be downloaded and unzipped with these commands:
# !curl -O
# !gunzip recipeitems-latest.json.gz
The database is in JSON format, so we will try pd.read_json to read it:
    recipes = pd.read_json('recipeitems-latest.json')
except ValueError as e:
    print("ValueError:", e)
ValueError: Trailing data
Oops! We get a ValueError mentioning that there is "trailing data." Searching for the text of this error on the Internet, it seems that it's due to using a file in which each line is itself a valid JSON, but the full file is not. Let's check if this interpretation is true:
with open('recipeitems-latest.json') as f:
    line = f.readline()
(2, 12)
Yes, apparently each line is a valid JSON, so we'll need to string them together. One way we can do this is to actually construct a string representation containing all these JSON entries, and then load the whole thing with pd.read_json:
# read the entire file into a Python array
with open('recipeitems-latest.json', 'r') as f:
    # Extract each line
    data = (line.strip() for line in f)
    # Reformat so each line is the element of a list
    data_json = "[{0}]".format(','.join(data))
# read the result as a JSON
recipes = pd.read_json(data_json)
(173278, 17)
We see there are nearly 200,000 recipes, and 17 columns. Let's take a look at one row to see what we have:
_id {'oid': '5160756b96cc62079cc2db15'} cookTime PT30M creator NaN dateModified NaN datePublished 2013-03-11 description Late Saturday afternoon, after Marlboro Man ha... image ingredients Biscuits\n3 cups All-purpose Flour\n2 Tablespo... name Drop Biscuits and Sausage Gravy prepTime PT10M recipeCategory NaN recipeInstructions NaN recipeYield 12 source thepioneerwoman totalTime NaN ts {'date': 1365276011104} url Name: 0, dtype: object
There is a lot of information there, but much of it is in a very messy form, as is typical of data scraped from the Web. In particular, the ingredient list is in string format; we're going to have to carefully extract the information we're interested in. Let's start by taking a closer look at the ingredients:
count 173278.000000 mean 244.617926 std 146.705285 min 0.000000 25% 147.000000 50% 221.000000 75% 314.000000 max 9067.000000 Name: ingredients, dtype: float64
The ingredient lists average 250 characters long, with a minimum of 0 and a maximum of nearly 10,000 characters!
Just out of curiousity, let's see which recipe has the longest ingredient list:
'Carrot Pineapple Spice & Brownie Layer Cake with Whipped Cream & Cream Cheese Frosting and Marzipan Carrots'
That certainly looks like an involved recipe.
We can do other aggregate explorations; for example, let's see how many of the recipes are for breakfast food:
Or how many of the recipes list cinnamon as an ingredient:
We could even look to see whether any recipes misspell the ingredient as "cinamon":
This is the type of essential data exploration that is possible with Pandas string tools. It is data munging like this that Python really excels at.

A simple recipe recommender

Let's go a bit further, and start working on a simple recipe recommendation system: given a list of ingredients, find a recipe that uses all those ingredients. While conceptually straightforward, the task is complicated by the heterogeneity of the data: there is no easy operation, for example, to extract a clean list of ingredients from each row. So we will cheat a bit: we'll start with a list of common ingredients, and simply search to see whether they are in each recipe's ingredient list. For simplicity, let's just stick with herbs and spices for the time being:
spice_list = ['salt', 'pepper', 'oregano', 'sage', 'parsley',
              'rosemary', 'tarragon', 'thyme', 'paprika', 'cumin']
We can then build a Boolean DataFrame consisting of True and False values, indicating whether this ingredient appears in the list:
import re
spice_df = pd.DataFrame(dict((spice, recipes.ingredients.str.contains(spice, re.IGNORECASE))
                             for spice in spice_list))
Now, as an example, let's say we'd like to find a recipe that uses parsley, paprika, and tarragon. We can compute this very quickly using the query() method of DataFrames, discussed in High-Performance Pandas: eval() and query():
selection = spice_df.query('parsley & paprika & tarragon')
We find only 10 recipes with this combination; let's use the index returned by this selection to discover the names of the recipes that have this combination: