Working with text data — pandas 3.0.0.dev0+2095.g2e141aaf99 documentation (original) (raw)

Text data types#

There are two ways to store text data in pandas:

1. object dtype NumPy array.
2. StringDtype extension type.

We recommend using StringDtype to store text data.

Prior to pandas 1.0, object dtype was the only option. This was unfortunate for many reasons:

1. You can accidentally store a mixture of strings and non-strings in anobject dtype array. It’s better to have a dedicated dtype.
2. object dtype breaks dtype-specific operations like DataFrame.select_dtypes(). There isn’t a clear way to select just text while excluding non-text but still object-dtype columns.
3. When reading code, the contents of an object dtype array is less clear than 'string'.

Currently, the performance of object dtype arrays of strings andarrays.StringArray are about the same. We expect future enhancements to significantly increase the performance and lower the memory overhead ofStringArray.

Warning

StringArray is currently considered experimental. The implementation and parts of the API may change without warning.

For backwards-compatibility, object dtype remains the default type we infer a list of strings to:

In [1]: pd.Series(["a", "b", "c"]) Out[1]: 0 a 1 b 2 c dtype: object

To explicitly request string dtype, specify the dtype:

In [2]: pd.Series(["a", "b", "c"], dtype="string") Out[2]: 0 a 1 b 2 c dtype: string

In [3]: pd.Series(["a", "b", "c"], dtype=pd.StringDtype()) Out[3]: 0 a 1 b 2 c dtype: string

Or astype after the Series or DataFrame is created:

In [4]: s = pd.Series(["a", "b", "c"])

In [5]: s Out[5]: 0 a 1 b 2 c dtype: object

In [6]: s.astype("string") Out[6]: 0 a 1 b 2 c dtype: string

You can also use StringDtype/"string" as the dtype on non-string data and it will be converted to string dtype:

In [7]: s = pd.Series(["a", 2, np.nan], dtype="string")

In [8]: s Out[8]: 0 a 1 2 2 dtype: string

In [9]: type(s[1]) Out[9]: str

or convert from existing pandas data:

In [10]: s1 = pd.Series([1, 2, np.nan], dtype="Int64")

In [11]: s1 Out[11]: 0 1 1 2 2 dtype: Int64

In [12]: s2 = s1.astype("string")

In [13]: s2 Out[13]: 0 1 1 2 2 dtype: string

In [14]: type(s2[0]) Out[14]: str

Behavior differences#

These are places where the behavior of StringDtype objects differ fromobject dtype:

1. For StringDtype, string accessor methodsthat return numeric output will always return a nullable integer dtype, rather than either int or float dtype, depending on the presence of NA values. Methods returning boolean output will return a nullable boolean dtype.
   In [15]: s = pd.Series(["a", None, "b"], dtype="string")
   In [16]: s
   Out[16]:
   0 a

1
2 b
dtype: string
In [17]: s.str.count("a")
Out[17]:
0 1
1
2 0
dtype: Int64
In [18]: s.dropna().str.count("a")
Out[18]:
0 1
2 0
dtype: Int64
Both outputs are Int64 dtype. Compare that with object-dtype:
In [19]: s2 = pd.Series(["a", None, "b"], dtype="object")
In [20]: s2.str.count("a")
Out[20]:
0 1.0
1 NaN
2 0.0
dtype: float64
In [21]: s2.dropna().str.count("a")
Out[21]:
0 1
2 0
dtype: int64
When NA values are present, the output dtype is float64. Similarly for methods returning boolean values.
In [22]: s.str.isdigit()
Out[22]:
0 False
1
2 False
dtype: boolean
In [23]: s.str.match("a")
Out[23]:
0 True
1
2 False
dtype: boolean 2. Some string methods, like Series.str.decode() are not available on StringArray because StringArray only holds strings, not bytes. 3. In comparison operations, arrays.StringArray and Series backed by a StringArray will return an object with BooleanDtype, rather than a bool dtype object. Missing values in a StringArraywill propagate in comparison operations, rather than always comparing unequal like numpy.nan.

Everything else that follows in the rest of this document applies equally tostring and object dtype.

String methods#

Series and Index are equipped with a set of string processing methods that make it easy to operate on each element of the array. Perhaps most importantly, these methods exclude missing/NA values automatically. These are accessed via the str attribute and generally have names matching the equivalent (scalar) built-in string methods:

In [24]: s = pd.Series( ....: ["A", "B", "C", "Aaba", "Baca", np.nan, "CABA", "dog", "cat"], dtype="string" ....: ) ....:

In [25]: s.str.lower() Out[25]: 0 a 1 b 2 c 3 aaba 4 baca 5 6 caba 7 dog 8 cat dtype: string

In [26]: s.str.upper() Out[26]: 0 A 1 B 2 C 3 AABA 4 BACA 5 6 CABA 7 DOG 8 CAT dtype: string

In [27]: s.str.len() Out[27]: 0 1 1 1 2 1 3 4 4 4 5 6 4 7 3 8 3 dtype: Int64

In [28]: idx = pd.Index([" jack", "jill ", " jesse ", "frank"])

In [29]: idx.str.strip() Out[29]: Index(['jack', 'jill', 'jesse', 'frank'], dtype='object')

In [30]: idx.str.lstrip() Out[30]: Index(['jack', 'jill ', 'jesse ', 'frank'], dtype='object')

In [31]: idx.str.rstrip() Out[31]: Index([' jack', 'jill', ' jesse', 'frank'], dtype='object')

The string methods on Index are especially useful for cleaning up or transforming DataFrame columns. For instance, you may have columns with leading or trailing whitespace:

In [32]: df = pd.DataFrame( ....: np.random.randn(3, 2), columns=[" Column A ", " Column B "], index=range(3) ....: ) ....:

In [33]: df Out[33]: Column A Column B 0 0.469112 -0.282863 1 -1.509059 -1.135632 2 1.212112 -0.173215

Since df.columns is an Index object, we can use the .str accessor

In [34]: df.columns.str.strip() Out[34]: Index(['Column A', 'Column B'], dtype='object')

In [35]: df.columns.str.lower() Out[35]: Index([' column a ', ' column b '], dtype='object')

These string methods can then be used to clean up the columns as needed. Here we are removing leading and trailing whitespaces, lower casing all names, and replacing any remaining whitespaces with underscores:

In [36]: df.columns = df.columns.str.strip().str.lower().str.replace(" ", "_")

In [37]: df Out[37]: column_a column_b 0 0.469112 -0.282863 1 -1.509059 -1.135632 2 1.212112 -0.173215

Note

If you have a Series where lots of elements are repeated (i.e. the number of unique elements in the Series is a lot smaller than the length of theSeries), it can be faster to convert the original Series to one of typecategory and then use .str.<method> or .dt.<property> on that. The performance difference comes from the fact that, for Series of type category, the string operations are done on the .categories and not on each element of theSeries.

Please note that a Series of type category with string .categories has some limitations in comparison to Series of type string (e.g. you can’t add strings to each other: s + " " + s won’t work if s is a Series of type category). Also,.str methods which operate on elements of type list are not available on such aSeries.

Warning

The type of the Series is inferred and is one among the allowed types (i.e. strings).

Generally speaking, the .str accessor is intended to work only on strings. With very few exceptions, other uses are not supported, and may be disabled at a later point.

Splitting and replacing strings#

Methods like split return a Series of lists:

In [38]: s2 = pd.Series(["a_b_c", "c_d_e", np.nan, "f_g_h"], dtype="string")

In [39]: s2.str.split("_") Out[39]: 0 [a, b, c] 1 [c, d, e] 2 3 [f, g, h] dtype: object

Elements in the split lists can be accessed using get or [] notation:

In [40]: s2.str.split("_").str.get(1) Out[40]: 0 b 1 d 2 3 g dtype: object

In [41]: s2.str.split("_").str[1] Out[41]: 0 b 1 d 2 3 g dtype: object

It is easy to expand this to return a DataFrame using expand.

In [42]: s2.str.split("_", expand=True) Out[42]: 0 1 2 0 a b c 1 c d e 2 3 f g h

When original Series has StringDtype, the output columns will all be StringDtype as well.

It is also possible to limit the number of splits:

In [43]: s2.str.split("_", expand=True, n=1) Out[43]: 0 1 0 a b_c 1 c d_e 2 3 f g_h

rsplit is similar to split except it works in the reverse direction, i.e., from the end of the string to the beginning of the string:

In [44]: s2.str.rsplit("_", expand=True, n=1) Out[44]: 0 1 0 a_b c 1 c_d e 2 3 f_g h

replace optionally uses regular expressions:

In [45]: s3 = pd.Series( ....: ["A", "B", "C", "Aaba", "Baca", "", np.nan, "CABA", "dog", "cat"], ....: dtype="string", ....: ) ....:

In [46]: s3 Out[46]: 0 A 1 B 2 C 3 Aaba 4 Baca 5
6 7 CABA 8 dog 9 cat dtype: string

In [47]: s3.str.replace("^.a|dog", "XX-XX ", case=False, regex=True) Out[47]: 0 A 1 B 2 C 3 XX-XX ba 4 XX-XX ca 5
6 7 XX-XX BA 8 XX-XX 9 XX-XX t dtype: string

Changed in version 2.0.

Single character pattern with regex=True will also be treated as regular expressions:

In [48]: s4 = pd.Series(["a.b", ".", "b", np.nan, ""], dtype="string")

In [49]: s4 Out[49]: 0 a.b 1 . 2 b 3 4
dtype: string

In [50]: s4.str.replace(".", "a", regex=True) Out[50]: 0 aaa 1 a 2 a 3 4
dtype: string

If you want literal replacement of a string (equivalent to str.replace()), you can set the optional regex parameter to False, rather than escaping each character. In this case both pat and repl must be strings:

In [51]: dollars = pd.Series(["12", "-$10", "$10,000"], dtype="string")

These lines are equivalent

In [52]: dollars.str.replace(r"-$", "-", regex=True) Out[52]: 0 12 1 -10 2 $10,000 dtype: string

In [53]: dollars.str.replace("-$", "-", regex=False) Out[53]: 0 12 1 -10 2 $10,000 dtype: string

The replace method can also take a callable as replacement. It is called on every pat using re.sub(). The callable should expect one positional argument (a regex object) and return a string.

Reverse every lowercase alphabetic word

In [54]: pat = r"[a-z]+"

In [55]: def repl(m): ....: return m.group(0)[::-1] ....:

In [56]: pd.Series(["foo 123", "bar baz", np.nan], dtype="string").str.replace( ....: pat, repl, regex=True ....: ) ....: Out[56]: 0 oof 123 1 rab zab 2 dtype: string

Using regex groups

In [57]: pat = r"(?P\w+) (?P\w+) (?P\w+)"

In [58]: def repl(m): ....: return m.group("two").swapcase() ....:

In [59]: pd.Series(["Foo Bar Baz", np.nan], dtype="string").str.replace( ....: pat, repl, regex=True ....: ) ....: Out[59]: 0 bAR 1 dtype: string

The replace method also accepts a compiled regular expression object from re.compile() as a pattern. All flags should be included in the compiled regular expression object.

In [60]: import re

In [61]: regex_pat = re.compile(r"^.a|dog", flags=re.IGNORECASE)

In [62]: s3.str.replace(regex_pat, "XX-XX ", regex=True) Out[62]: 0 A 1 B 2 C 3 XX-XX ba 4 XX-XX ca 5
6 7 XX-XX BA 8 XX-XX 9 XX-XX t dtype: string

Including a flags argument when calling replace with a compiled regular expression object will raise a ValueError.

In [63]: s3.str.replace(regex_pat, 'XX-XX ', flags=re.IGNORECASE)

ValueError: case and flags cannot be set when pat is a compiled regex

removeprefix and removesuffix have the same effect as str.removeprefix and str.removesuffix added inPython 3.9:

Added in version 1.4.0.

In [64]: s = pd.Series(["str_foo", "str_bar", "no_prefix"])

In [65]: s.str.removeprefix("str_") Out[65]: 0 foo 1 bar 2 no_prefix dtype: object

In [66]: s = pd.Series(["foo_str", "bar_str", "no_suffix"])

In [67]: s.str.removesuffix("_str") Out[67]: 0 foo 1 bar 2 no_suffix dtype: object

Concatenation#

There are several ways to concatenate a Series or Index, either with itself or others, all based on cat(), resp. Index.str.cat.

Concatenating a single Series into a string#

The content of a Series (or Index) can be concatenated:

In [68]: s = pd.Series(["a", "b", "c", "d"], dtype="string")

In [69]: s.str.cat(sep=",") Out[69]: 'a,b,c,d'

If not specified, the keyword sep for the separator defaults to the empty string, sep='':

In [70]: s.str.cat() Out[70]: 'abcd'

By default, missing values are ignored. Using na_rep, they can be given a representation:

In [71]: t = pd.Series(["a", "b", np.nan, "d"], dtype="string")

In [72]: t.str.cat(sep=",") Out[72]: 'a,b,d'

In [73]: t.str.cat(sep=",", na_rep="-") Out[73]: 'a,b,-,d'

Concatenating a Series and something list-like into a Series#

The first argument to cat() can be a list-like object, provided that it matches the length of the calling Series (or Index).

In [74]: s.str.cat(["A", "B", "C", "D"]) Out[74]: 0 aA 1 bB 2 cC 3 dD dtype: string

Missing values on either side will result in missing values in the result as well, unless na_rep is specified:

In [75]: s.str.cat(t) Out[75]: 0 aa 1 bb 2 3 dd dtype: string

In [76]: s.str.cat(t, na_rep="-") Out[76]: 0 aa 1 bb 2 c- 3 dd dtype: string

Concatenating a Series and something array-like into a Series#

The parameter others can also be two-dimensional. In this case, the number or rows must match the lengths of the calling Series (or Index).

In [77]: d = pd.concat([t, s], axis=1)

In [78]: s Out[78]: 0 a 1 b 2 c 3 d dtype: string

In [79]: d Out[79]: 0 1 0 a a 1 b b 2 c 3 d d

In [80]: s.str.cat(d, na_rep="-") Out[80]: 0 aaa 1 bbb 2 c-c 3 ddd dtype: string

Concatenating a Series and an indexed object into a Series, with alignment#

For concatenation with a Series or DataFrame, it is possible to align the indexes before concatenation by setting the join-keyword.

In [81]: u = pd.Series(["b", "d", "a", "c"], index=[1, 3, 0, 2], dtype="string")

In [82]: s Out[82]: 0 a 1 b 2 c 3 d dtype: string

In [83]: u Out[83]: 1 b 3 d 0 a 2 c dtype: string

In [84]: s.str.cat(u) Out[84]: 0 aa 1 bb 2 cc 3 dd dtype: string

In [85]: s.str.cat(u, join="left") Out[85]: 0 aa 1 bb 2 cc 3 dd dtype: string

The usual options are available for join (one of 'left', 'outer', 'inner', 'right'). In particular, alignment also means that the different lengths do not need to coincide anymore.

In [86]: v = pd.Series(["z", "a", "b", "d", "e"], index=[-1, 0, 1, 3, 4], dtype="string")

In [87]: s Out[87]: 0 a 1 b 2 c 3 d dtype: string

In [88]: v Out[88]: -1 z 0 a 1 b 3 d 4 e dtype: string

In [89]: s.str.cat(v, join="left", na_rep="-") Out[89]: 0 aa 1 bb 2 c- 3 dd dtype: string

In [90]: s.str.cat(v, join="outer", na_rep="-") Out[90]: -1 -z 0 aa 1 bb 2 c- 3 dd 4 -e dtype: string

The same alignment can be used when others is a DataFrame:

In [91]: f = d.loc[[3, 2, 1, 0], :]

In [92]: s Out[92]: 0 a 1 b 2 c 3 d dtype: string

In [93]: f Out[93]: 0 1 3 d d 2 c 1 b b 0 a a

In [94]: s.str.cat(f, join="left", na_rep="-") Out[94]: 0 aaa 1 bbb 2 c-c 3 ddd dtype: string

Concatenating a Series and many objects into a Series#

Several array-like items (specifically: Series, Index, and 1-dimensional variants of np.ndarray) can be combined in a list-like container (including iterators, dict-views, etc.).

In [95]: s Out[95]: 0 a 1 b 2 c 3 d dtype: string

In [96]: u Out[96]: 1 b 3 d 0 a 2 c dtype: string

In [97]: s.str.cat([u, u.to_numpy()], join="left") Out[97]: 0 aab 1 bbd 2 cca 3 ddc dtype: string

All elements without an index (e.g. np.ndarray) within the passed list-like must match in length to the calling Series (or Index), but Series and Index may have arbitrary length (as long as alignment is not disabled with join=None):

In [98]: v Out[98]: -1 z 0 a 1 b 3 d 4 e dtype: string

In [99]: s.str.cat([v, u, u.to_numpy()], join="outer", na_rep="-") Out[99]: -1 -z-- 0 aaab 1 bbbd 2 c-ca 3 dddc 4 -e-- dtype: string

If using join='right' on a list-like of others that contains different indexes, the union of these indexes will be used as the basis for the final concatenation:

In [100]: u.loc[[3]] Out[100]: 3 d dtype: string

In [101]: v.loc[[-1, 0]] Out[101]: -1 z 0 a dtype: string

In [102]: s.str.cat([u.loc[[3]], v.loc[[-1, 0]]], join="right", na_rep="-") Out[102]: 3 dd- -1 --z 0 a-a dtype: string

Indexing with .str#

You can use [] notation to directly index by position locations. If you index past the end of the string, the result will be a NaN.

In [103]: s = pd.Series( .....: ["A", "B", "C", "Aaba", "Baca", np.nan, "CABA", "dog", "cat"], dtype="string" .....: ) .....:

In [104]: s.str[0] Out[104]: 0 A 1 B 2 C 3 A 4 B 5 6 C 7 d 8 c dtype: string

In [105]: s.str[1] Out[105]: 0 1 2 3 a 4 a 5 6 A 7 o 8 a dtype: string

Testing for strings that match or contain a pattern#

You can check whether elements contain a pattern:

In [131]: pattern = r"[0-9][a-z]"

In [132]: pd.Series( .....: ["1", "2", "3a", "3b", "03c", "4dx"], .....: dtype="string", .....: ).str.contains(pattern) .....: Out[132]: 0 False 1 False 2 True 3 True 4 True 5 True dtype: boolean

Or whether elements match a pattern:

In [133]: pd.Series( .....: ["1", "2", "3a", "3b", "03c", "4dx"], .....: dtype="string", .....: ).str.match(pattern) .....: Out[133]: 0 False 1 False 2 True 3 True 4 False 5 True dtype: boolean

In [134]: pd.Series( .....: ["1", "2", "3a", "3b", "03c", "4dx"], .....: dtype="string", .....: ).str.fullmatch(pattern) .....: Out[134]: 0 False 1 False 2 True 3 True 4 False 5 False dtype: boolean

Note

The distinction between match, fullmatch, and contains is strictness:fullmatch tests whether the entire string matches the regular expression;match tests whether there is a match of the regular expression that begins at the first character of the string; and contains tests whether there is a match of the regular expression at any position within the string.

The corresponding functions in the re package for these three match modes arere.fullmatch,re.match, andre.search, respectively.

Methods like match, fullmatch, contains, startswith, andendswith take an extra na argument so missing values can be considered True or False:

In [135]: s4 = pd.Series( .....: ["A", "B", "C", "Aaba", "Baca", np.nan, "CABA", "dog", "cat"], dtype="string" .....: ) .....:

In [136]: s4.str.contains("A", na=False) Out[136]: 0 True 1 False 2 False 3 True 4 False 5 False 6 True 7 False 8 False dtype: boolean

Creating indicator variables#

You can extract dummy variables from string columns. For example if they are separated by a '|':

In [137]: s = pd.Series(["a", "a|b", np.nan, "a|c"], dtype="string")

In [138]: s.str.get_dummies(sep="|") Out[138]: a b c 0 1 0 0 1 1 1 0 2 0 0 0 3 1 0 1

String Index also supports get_dummies which returns a MultiIndex.

In [139]: idx = pd.Index(["a", "a|b", np.nan, "a|c"])

In [140]: idx.str.get_dummies(sep="|") Out[140]: MultiIndex([(1, 0, 0), (1, 1, 0), (0, 0, 0), (1, 0, 1)], names=['a', 'b', 'c'])

See also get_dummies().

Method summary#