Text Wrangling and Regex¶

Adapted from Lisa Yan, Will Fithian, Joseph Gonzalez, Deborah Nolan, Sam Lau

Updated by Bella Crouch

Working with text: applying string methods and regular expressions

In [1]:
import matplotlib.pyplot as plt
import numpy as np
import pandas as pd
import seaborn as sns
import zipfile
import pandas as pd

Demo: Canonicalizing County Names¶

In [2]:
states = pd.read_csv("data/county_and_state.csv")
populations = pd.read_csv("data/county_and_population.csv")

# display allows us to view a DataFrame without returning it as an object
display(states)

display(populations)
County State
0 De Witt County IL
1 Lac qui Parle County MN
2 Lewis and Clark County MT
3 St John the Baptist Parish LS
County Population
0 DeWitt 16798
1 Lac Qui Parle 8067
2 Lewis & Clark 55716
3 St. John the Baptist 43044

Both of these DataFrames share a "County" column. Unfortunately, formatting differences mean that we can't directly merge the two DataFrames using the "County"s.

In [3]:
states.merge(populations, left_on="County", right_on="County")
Out[3]:
County State Population

To address this, we can canonicalize the "County" string data to apply a common formatting.

In [4]:
def canonicalize_county(county_series):
    return (county_series.str.lower()                           # lowercase
            .str.replace(' ', '', regex=True)                   # remove space
            .str.replace('&', 'and', regex=True)                # replace &
            .str.replace('.', '', regex=True)                   # remove dot
            .str.replace('county', '', regex=True)              # remove "county"
            .str.replace('parish', '', regex=True)              # remove "parish" 
            )
In [5]:
states["County"] = canonicalize_county(states["County"])
populations["County"] = canonicalize_county(populations["County"])

display(states)

display(populations)
County State
0 dewitt IL
1 lacquiparle MN
2 lewisandclark MT
3 stjohnthebaptist LS
County Population
0 dewitt 16798
1 lacquiparle 8067
2 lewisandclark 55716
3 stjohnthebaptist 43044

Now, the merge works as expected!

In [6]:
states.merge(populations, left_on="County", right_on="County")
Out[6]:
County State Population
0 dewitt IL 16798
1 lacquiparle MN 8067
2 lewisandclark MT 55716
3 stjohnthebaptist LS 43044

Demo: Extracting Log Data¶

In [7]:
log_fname = 'data/log.txt'
with open(log_fname, 'r') as f:
    log_lines = f.readlines()
log_lines
Out[7]:
['169.237.46.168 - - [26/Jan/2014:10:47:58 -0800] "GET /stat141/Winter04/ HTTP/1.1" 200 2585 "http://anson.ucdavis.edu/courses/"\n',
 '193.205.203.3 - - [2/Feb/2005:17:23:6 -0800] "GET /stat141/Notes/dim.html HTTP/1.0" 404 302 "http://eeyore.ucdavis.edu/stat141/Notes/session.html"\n',
 '169.237.46.240 - "" [3/Feb/2006:10:18:37 -0800] "GET /stat141/homework/Solutions/hw1Sol.pdf HTTP/1.1"\n']

Suppose we want to extract the day, month, year, hour, minutes, seconds, and timezone. Looking at the data, we see that these items are not in a fixed position relative to the beginning of the string. That is, slicing by some fixed offset isn't going to work.

In [8]:
log_lines[0][20:31]
Out[8]:
'26/Jan/2014'
In [9]:
log_lines[1][20:31]
Out[9]:
'/Feb/2005:1'

Instead, we'll need to use some more sophisticated thinking. Let's focus on only the first line of the file.

In [10]:
first = log_lines[0]
first
Out[10]:
'169.237.46.168 - - [26/Jan/2014:10:47:58 -0800] "GET /stat141/Winter04/ HTTP/1.1" 200 2585 "http://anson.ucdavis.edu/courses/"\n'
In [11]:
pertinent = first.split("[")[1].split(']')[0] # find the text enclosed in square brackets
day, month, rest = pertinent.split('/')       # split up the date/month/year 
year, hour, minute, rest = rest.split(':')    # split up the hour:minute:second
seconds, time_zone = rest.split(' ')          # split the timezone after the blank space
day, month, year, hour, minute, seconds, time_zone
Out[11]:
('26', 'Jan', '2014', '10', '47', '58', '-0800')

This worked, but felt fairly "hacky" – the code above isn't particularly elegant. A much more sophisticated but common approach is to extract the information we need using a regular expression.

Regular Expressions¶

String Extraction with Regex¶

Python re.findall returns a list of all extracted matches:

In [12]:
import re

text = "My social security number is 123-45-6789 bro, or actually maybe it’s 321-45-6789.";
pattern = r"[0-9]{3}-[0-9]{2}-[0-9]{4}"
re.findall(pattern, text)
Out[12]:
['123-45-6789', '321-45-6789']


Now, let's see vectorized extraction in pandas:

.str.findall returns a Series of lists of all matches in each record.

In [13]:
df_ssn = pd.DataFrame(
    ['987-65-4321',
     'forty',
     '123-45-6789 bro or 321-45-6789',
     '999-99-9999'],
    columns=['SSN'])
df_ssn
Out[13]:
SSN
0 987-65-4321
1 forty
2 123-45-6789 bro or 321-45-6789
3 999-99-9999
In [14]:
# -> Series of lists
pattern = r"[0-9]{3}-[0-9]{2}-[0-9]{4}"
df_ssn['SSN'].str.findall(pattern)
Out[14]:
0                 [987-65-4321]
1                            []
2    [123-45-6789, 321-45-6789]
3                 [999-99-9999]
Name: SSN, dtype: object

Extraction Using Regex Capture Groups¶

The Python function re.findall, in combination with parentheses returns specific substrings (i.e., capture groups) within each matched string, or match.

In [15]:
text = """I will meet you at 08:30:00 pm tomorrow"""       
pattern = ".*(\d\d):(\d\d):(\d\d).*"
matches = re.findall(pattern, text)
matches
Out[15]:
[('08', '30', '00')]
In [16]:
# the three capture groups in the first matched string
hour, minute, second = matches[0]


In pandas, we can use .str.extract to extract each capture group of only the first match of each record into separate columns.

In [17]:
# back to SSNs
df_ssn
Out[17]:
SSN
0 987-65-4321
1 forty
2 123-45-6789 bro or 321-45-6789
3 999-99-9999
In [18]:
# Will extract the first match of all groups
pattern_group_mult = r"([0-9]{3})-([0-9]{2})-([0-9]{4})" # 3 groups
df_ssn['SSN'].str.extract(pattern_group_mult)
Out[18]:
0 1 2
0 987 65 4321
1 NaN NaN NaN
2 123 45 6789
3 999 99 9999

Alternatively, .str.extractall extracts all matches of each record into separate columns. Rows are then MultiIndexed by original record index and match index.

In [19]:
# -> DataFrame, one row per match
df_ssn['SSN'].str.extractall(pattern_group_mult)
Out[19]:
0 1 2
match
0 0 987 65 4321
2 0 123 45 6789
1 321 45 6789
3 0 999 99 9999

Canonicalization with Regex¶

In regular Python, canonicalize with re.sub (standing for "substitute"):

In [20]:
text = '<div><td valign="top">Moo</td></div>'
pattern = r"<[^>]+>"
re.sub(pattern, '', text)
Out[20]:
'Moo'


In pandas, canonicalize with Series.str.replace.

In [21]:
# example dataframe of strings
df_html = pd.DataFrame(['<div><td valign="top">Moo</td></div>',
                   '<a href="http://ds100.org">Link</a>',
                   '<b>Bold text</b>'], columns=['Html'])
df_html
Out[21]:
Html
0 <div><td valign="top">Moo</td></div>
1 <a href="http://ds100.org">Link</a>
2 <b>Bold text</b>
In [22]:
# Series -> Series
df_html["Html"].str.replace(pattern, '', regex=True).to_frame()
Out[22]:
Html
0 Moo
1 Link
2 Bold text

Revisiting Text Log Processing using Regex¶

Python re version¶

In [23]:
line = log_lines[0]
display(line)

pattern = r'\[(\d+)\/(\w+)\/(\d+):(\d+):(\d+):(\d+) (.+)\]'
day, month, year, hour, minute, second, time_zone = re.findall(pattern, line)[0] # get first match
day, month, year, hour, minute, second, time_zone

'169.237.46.168 - - [26/Jan/2014:10:47:58 -0800] "GET /stat141/Winter04/ HTTP/1.1" 200 2585 "http://anson.ucdavis.edu/courses/"\n'
Out[23]:
('26', 'Jan', '2014', '10', '47', '58', '-0800')

pandas version¶

In [24]:
df = pd.DataFrame(log_lines, columns=['Log'])
df
Out[24]:
Log
0 169.237.46.168 - - [26/Jan/2014:10:47:58 -0800...
1 193.205.203.3 - - [2/Feb/2005:17:23:6 -0800] "...
2 169.237.46.240 - "" [3/Feb/2006:10:18:37 -0800...

Option 1: Series.str.findall

In [25]:
pattern = r'\[(\d+)\/(\w+)\/(\d+):(\d+):(\d+):(\d+) (.+)\]'
df['Log'].str.findall(pattern)
Out[25]:
0    [(26, Jan, 2014, 10, 47, 58, -0800)]
1      [(2, Feb, 2005, 17, 23, 6, -0800)]
2     [(3, Feb, 2006, 10, 18, 37, -0800)]
Name: Log, dtype: object


Option 2: Series.str.extractall

In [26]:
df['Log'].str.extractall(pattern)
Out[26]:
0 1 2 3 4 5 6
match
0 0 26 Jan 2014 10 47 58 -0800
1 0 2 Feb 2005 17 23 6 -0800
2 0 3 Feb 2006 10 18 37 -0800

Wrangling either of these two DataFrames into a nice format (like below) is left as an exercise for you! You will do a related problem on the homework.

Day Month Year Hour Minute Second Time Zone
0 26 Jan 2014 10 47 58 -0800
1 2 Feb 2005 17 23 6 -0800
2 3 Feb 2006 10 18 37 -0800
In [27]:
# your code here
...




Real World Case Study: Restaurant Data¶

In this example, we will show how regexes can allow us to track quantitative data across categories defined by the appearance of various text fields.

In this example we'll see how the presence of certain keywords can affect quantitative data:

How do restaurant health scores vary as a function of the number of violations that mention a particular keyword?
(e.g., unclean surfaces, vermin, permits, etc.)

In [28]:
vio = pd.read_csv('data/violations.csv', header=0, names=['bid', 'date', 'desc'])
desc = vio['desc']
vio.head()
Out[28]:
bid date desc
0 19 20171211 Inadequate food safety knowledge or lack of ce...
1 19 20171211 Unapproved or unmaintained equipment or utensils
2 19 20160513 Unapproved or unmaintained equipment or utensi...
3 19 20160513 Unclean or degraded floors walls or ceilings ...
4 19 20160513 Food safety certificate or food handler card n...
In [29]:
counts = desc.value_counts()
counts.shape
Out[29]:
(14253,)

That's a lot of different descriptions!! Can we canonicalize at all? Let's explore two sets of 10 rows.

In [30]:
counts[:10]
Out[30]:
Unclean or degraded floors walls or ceilings                          999
Unapproved or unmaintained equipment or utensils                      659
Inadequately cleaned or sanitized food contact surfaces               493
Improper food storage                                                 476
Inadequate and inaccessible handwashing facilities                    467
Moderate risk food holding temperature                                452
Wiping cloths not clean or properly stored or inadequate sanitizer    418
Moderate risk vermin infestation                                      374
Unclean nonfood contact surfaces                                      369
Food safety certificate or food handler card not available            353
Name: desc, dtype: int64
In [31]:
# Hmmm...
counts[50:60]
Out[31]:
Unclean or degraded floors walls or ceilings  [ date violation corrected: 11/29/2017 ]              16
Unclean or degraded floors walls or ceilings  [ date violation corrected: 9/19/2017 ]               16
Inadequate HACCP plan record keeping                                                                16
Unclean or degraded floors walls or ceilings  [ date violation corrected: 11/27/2017 ]              15
Unclean or degraded floors walls or ceilings  [ date violation corrected: 12/7/2017 ]               15
Inadequately cleaned or sanitized food contact surfaces  [ date violation corrected: 9/26/2017 ]    14
Unclean or degraded floors walls or ceilings  [ date violation corrected: 11/28/2017 ]              14
Unclean or degraded floors walls or ceilings  [ date violation corrected: 9/6/2017 ]                14
Unapproved or unmaintained equipment or utensils  [ date violation corrected: 9/19/2017 ]           14
Unapproved  living quarters in food facility                                                        13
Name: desc, dtype: int64
In [32]:
# Use regular expressions to cut out the extra info in square braces.
vio['clean_desc'] = (vio['desc']
             .str.replace(r'\s*\[.*\]$', '', regex=True)
             .str.strip()       # removes leading/trailing whitespace
             .str.lower())
vio.head()
Out[32]:
bid date desc clean_desc
0 19 20171211 Inadequate food safety knowledge or lack of ce... inadequate food safety knowledge or lack of ce...
1 19 20171211 Unapproved or unmaintained equipment or utensils unapproved or unmaintained equipment or utensils
2 19 20160513 Unapproved or unmaintained equipment or utensi... unapproved or unmaintained equipment or utensils
3 19 20160513 Unclean or degraded floors walls or ceilings ... unclean or degraded floors walls or ceilings
4 19 20160513 Food safety certificate or food handler card n... food safety certificate or food handler card n...
In [33]:
# canonicalizing definitely helped
vio['clean_desc'].value_counts().shape
Out[33]:
(68,)
In [34]:
vio['clean_desc'].value_counts().head()
Out[34]:
unclean or degraded floors walls or ceilings               3507
moderate risk food holding temperature                     2542
inadequate and inaccessible handwashing facilities         2529
unapproved or unmaintained equipment or utensils           2382
inadequately cleaned or sanitized food contact surfaces    2301
Name: clean_desc, dtype: int64

Remember our research question:

How do restaurant health scores vary as a function of the number of violations that mention a particular keyword?
(e.g., unclean surfaces, vermin, permits, etc.)


Below, we use regular expressions and df.assign() (documentation) to method chain our creation of new boolean features, one per keyword.

In [35]:
# use regular expressions to assign new features for the presence of various keywords
# regex metacharacter | 
with_features = (vio
 .assign(is_unclean     = vio['clean_desc'].str.contains('clean|sanit'))
 .assign(is_high_risk = vio['clean_desc'].str.contains('high risk'))
 .assign(is_vermin    = vio['clean_desc'].str.contains('vermin'))
 .assign(is_surface   = vio['clean_desc'].str.contains('wall|ceiling|floor|surface'))
 .assign(is_human     = vio['clean_desc'].str.contains('hand|glove|hair|nail'))
 .assign(is_permit    = vio['clean_desc'].str.contains('permit|certif'))
)
with_features.head()
Out[35]:
bid date desc clean_desc is_unclean is_high_risk is_vermin is_surface is_human is_permit
0 19 20171211 Inadequate food safety knowledge or lack of ce... inadequate food safety knowledge or lack of ce... False False False False False True
1 19 20171211 Unapproved or unmaintained equipment or utensils unapproved or unmaintained equipment or utensils False False False False False False
2 19 20160513 Unapproved or unmaintained equipment or utensi... unapproved or unmaintained equipment or utensils False False False False False False
3 19 20160513 Unclean or degraded floors walls or ceilings ... unclean or degraded floors walls or ceilings True False False True False False
4 19 20160513 Food safety certificate or food handler card n... food safety certificate or food handler card n... False False False False True True



EDA¶

That's the end of our text wrangling. Now let's do some more analysis to analyze restaurant health as a function of the number of violation keywords.

To do so we'll first group so that our granularity is one inspection for a business on particular date. This effectively counts the number of violations by keyword for a given inspection.

In [36]:
count_features = (with_features
 .groupby(['bid', 'date'])
 .sum()
 .reset_index()
)
count_features.iloc[255:260, :]
Out[36]:
bid date is_unclean is_high_risk is_vermin is_surface is_human is_permit
255 489 20150728 5 0 2 3 0 0
256 489 20150807 1 0 0 1 0 0
257 489 20160308 2 2 1 0 1 0
258 489 20160721 2 1 1 1 0 1
259 489 20161220 3 0 1 2 0 0

Check out our new dataframe in action:

In [37]:
count_features[count_features['is_vermin'] > 1].head(5)
Out[37]:
bid date is_unclean is_high_risk is_vermin is_surface is_human is_permit
255 489 20150728 5 0 2 3 0 0
291 527 20170821 1 1 2 1 1 1
1508 2622 20160526 4 2 2 3 0 0
1573 2721 20150422 2 1 2 1 0 0
1746 2945 20150921 2 1 2 2 2 1

Now we'll reshape this "wide" table into a "tidy" table using a pandas feature called pd.melt (documentation) which we won't describe in any detail, other than that it's effectively the inverse of pd.pivot_table.

Our granularity is now a violation type for a given inspection (for a business on a particular date).

In [38]:
violation_type_df = pd.melt(count_features, id_vars=['bid', 'date'],
            var_name='feature', value_name='num_vios')

# show a particular inspection's results
violation_type_df[(violation_type_df['bid'] == 489) & (violation_type_df['date'] == 20150728)]
Out[38]:
bid date feature num_vios
255 489 20150728 is_unclean 5
12517 489 20150728 is_high_risk 0
24779 489 20150728 is_vermin 2
37041 489 20150728 is_surface 3
49303 489 20150728 is_human 0
61565 489 20150728 is_permit 0

Remember our research question:

How do restaurant health scores vary as a function of the number of violations that mention a particular keyword?
(e.g., unclean surfaces, vermin, permits, etc.)


We have the second half of this question! Now let's join our table with the inspection scores, located in inspections.csv.

In [39]:
# read in the scores
inspection_df = pd.read_csv('data/inspections.csv',
                  header=0,
                  usecols=[0, 1, 2],
                  names=['bid', 'score', 'date'])
inspection_df.head()
Out[39]:
bid score date
0 19 94 20160513
1 19 94 20171211
2 24 98 20171101
3 24 98 20161005
4 24 96 20160311

While the inspection scores were stored in a separate file from the violation descriptions, we notice that the primary key in inspections is (bid, date)! So we can reference this key in our join.

In [40]:
# join scores with the table broken down by violation type
violation_type_and_scores = (
    violation_type_df
    .merge(inspection_df, on=['bid', 'date'])
)
violation_type_and_scores.head(12)
Out[40]:
bid date feature num_vios score
0 19 20160513 is_unclean 1 94
1 19 20160513 is_high_risk 0 94
2 19 20160513 is_vermin 0 94
3 19 20160513 is_surface 1 94
4 19 20160513 is_human 1 94
5 19 20160513 is_permit 1 94
6 19 20171211 is_unclean 0 94
7 19 20171211 is_high_risk 0 94
8 19 20171211 is_vermin 0 94
9 19 20171211 is_surface 0 94
10 19 20171211 is_human 0 94
11 19 20171211 is_permit 1 94



Let's plot the distribution of scores, broken down by violation counts, for each inspection feature (is_clean, is_high_risk, is_vermin, is_surface).

In [41]:
# you will learn this syntax next week. Focus on interpreting for now.
sns.catplot(x='num_vios', y='score',
               col='feature', col_wrap=2,
               kind='box',
               data=violation_type_and_scores);

Above we can observe:

  • The inspection score generally goes down with increasing numbers of violations, as expected.
  • Depending on the violation keyword, inspections scores on average go down at slightly different rates.
  • For example, that if a restaurant inspection involved 2 violations with the keyword "vermin", the average score for that inspection would be a little bit below 80.