Guide des Expressions régulières

Auteur:A.M. Kuchling <amk@amk.ca>

Abstract

Ce document est un guide d’introduction à l’utilisation des expressions régulières en Python avec le module re. Il fournit une introduction plus abordable que la section correspondante dans le guide de référence des bibliothèques.

Introduction

Les expressions régulières (appelées RE, ou regexes, ou motifs regex) sont essentiellement un petit langage de programmation hautement spécialisé embarqué dans Python et dont la manipulation est rendue possible par l’utilisation du module re. En utilisant ce petit langage, vous spécifiez les règles pour l’ensemble des chaînes des caractères possibles pour lesquelles vous voulez une correspondance; cet ensemble contient des phrases en Anglais, ou des adresses de courriel, ou des commandes TeX, ou tout ce que vous voulez. Vous pouvez ensuite poser des questions tel que « Est-ce que cette chaîne de caractères correspond au motif ? », ou « Y-a-t’il une correspondance pour un motif dans le chaîne de caractères ? ». Vous pouvez aussi utiliser les RE pour modifier une chaîne de caractères ou la découper de différentes façons.

Les motifs d’expressions régulières sont compilés en bytecodes qui sont ensuite exécutés par un moteur de correspondance écrit en C. Pour une utilisation plus poussée, il peut s’avérer nécessaire de s’intéresser à la manière dont le moteur exécute la RE afin d’écrire une expression dont le bytecode est plus rapide. L’optimisation n’est pas traitée dans ce document, parce qu’elle nécessite d’avoir une bonne compréhension des mécanismes internes du moteur de correspondance.

Le langage d’expressions régulières est relativement petit et restreint donc toutes les tâches de manipulation de chaînes de caractères ne peuvent pas être réalisées à l’aide d’expressions régulières. Il y a aussi des tâches qui peuvent être réalisées à l’aide d’expressions régulières mais qui ont tendance à produire des expressions régulières très compliquées. Dans ces cas, il vous sera plus utile d’écrire du code Python pour réaliser le traitement; même si le code Python sera plus lent qu’une expression régulière élaborée, il sera probablement plus compréhensible.

Motifs simples

Nous commencerons par étudier les expressions régulières les plus simples. Dans la mesure où les expressions régulières sont utilisées pour manipuler des chaînes de caractères, nous commencerons par les tâches les plus communes: la correspondance de caractères .

Pour une explication détaillée sur la technologie sous-jacente des expressions régulières (automate d’état déterministe et non-déterministe), référez-vous à n’importe quel manuel sur l’écriture de compilateurs.

Caractères correspondants

La plupart des lettres ou caractères vont correspondre à eux-mêmes. Par exemple, l’expression régulière test correspond à la chaîne de caractère test précisément. (Vous pouvez activer le mode non-sensible à la casse qui permet à cette RE de correspondre à Test ou TEST également; plus à ce sujet dans la suite.)

Il y a des exceptions à cette règle; certains caractères sont des metacharacters spéciaux et ne correspondent pas à eux-mêmes. Au lieu de cela, ils signalent que certaines choses non ordinaires doivent correspondre, ou affectent d’autre portions de la RE en les répétant ou en changeant leur sens. Une grande partie de ce document est consacrée au fonctionnement de ces métacaractères.

Voici une liste complète des méta-caractères; leurs sens sera décrit dans la suite de ce guide.

. ^ $ * + ? { } [ ] \ | ( )

Les premiers méta-caractères que nous étudions sont [ et ]. Ils sont utilisés pour spécifier une classe de caractères, qui forme un ensemble de caractères dont vous souhaitez trouver la correspondance. Les caractères peuvent être listés individuellement, ou une plage de caractères peut être indiquée en fournissant deux caractères séparés par un “-“`. Par exemple, [abc] correspondra à n’importe quel caractère parmi a, b, ou c; ce qui est équivalent à [a-c], qui utilise une plage pour exprimer le même ensemble de caractères. Si vous voulez trouver une correspondance avec uniquement les lettres en minuscule, votre RE sera [a-z].

Les méta-caractères ne sont pas actifs dans les classes. Par exemple, [akm$] correspondra à n’importe quel caractère parmi 'a', 'k', 'm', ou '$'; '$' est habituellement un méta-caractère mais dans une classe de caractères, il est dépourvu de son caractère spécial.

Vous pouvez trouvez une correspondance avec les caractères non listés dans une classe en complimentant (complementing) l’ensemble. Ceci est indiqué en incluant un '^' en tant que premier caractère de la classe; '^' en dehors d’une classe de caractères correspondra au caractère '^'. Par exemple, [^5] correspondra à tout les caractères, sauf [^5].

Le méta-caractère qui est probablement le plus important est le backslash, \. Comme dans les chaînes de caractères en Python, le backslash peut être suivi par différents caractères pour signaler différentes séquences spéciales. Il est aussi utilisé pour échapper tout les méta-caractères afin d’en trouver les correspondances dans les motifs; par exemple, si vous devait trouver une correspondance pour [ ou \`, vous pouvais les précéder avec un **backslash** pour supprimer leur caractères spécial : ``\[ ou \\.

Certaines séquences spéciales commençant par '\' représentent des ensembles de caractères prédéfinis qui sont souvent utiles, tels que l’ensemble des chiffres, l’ensemble des lettres, ou l’ensemble des caractères qui ne sont pas des séparateurs.

Prenons un exemple: \w correspond à n’importe quel caractères alphanumérique. Si le motif regex est exprimé en octets, il est équivalent à la classe [a-zA-Z0-9_]. Si le motif regex est une chaîne de caractères, \w correspondra à tout les caractères identifiés comme lettre dans le base de données Unicode fourni par le module unicodedata. Vous pouvez utiliser la définition plus restrictive de \w dans un motif de chaîne de caractères en spécifiant le fanion re.ASCII lors de la compilation de l’expression régulière.

La liste de séquences spéciales suivante est incomplète. Pour une liste complète de séquences et définitions de classe étendue pour des motifs de chaînes de caractère s Unicode, reportez-vous à la dernière partie de la référence Syntaxe d’Expressions Régulières de la librairie standard. En général, les versions d’Unicode trouve une correspondance avec n’importe quel caractère présent dans le catégorie appropriée de la base de données Unicode.

\d
Correspond à n’importe quel caractère numérique ; ceci est équivalent à la classe [0-9].
\D
Correspond à n’importe caractère non numérique; ceci est équivalent à la classe [^0-9].
\s
Correspond à n’importe caractères d’espacement; ceci est équivalent à la classe [ \t\n\r\f\v].
\S
Correspond à n’importe caractère autre que d’espacement; ceci est équivalent à la classe [^ \t\n\r\f\v].
\w
Correspond à n’importe caractère alphanumérique; ceci est équivalent à la classe [a-zA-Z0-9_].
\W
Correspond à n’importe caractère non-alphanumérique; ceci est équivalent à la classe [^a-zA-Z0-9_].

Ces séquences peuvent être incluses dans une classe de caractères. Par exemple, [\s,.] est une classe de caractères qui correspond à tous les caractères d’espace, ou ',' ou '.'.

Le dernier méta-caractère dans cette section est .. Il correspond à tout les caractères, à l’exception du caractère de retour à la ligne; il existe un mode alternatif (re.DOTALL) dans lequel il correspond également eu caractère de retour à la ligne. . est souvent utilisé lorsque l’on veut trouver une correspondance avec « n’importe quel caractère ».

Répetitions

Being able to match varying sets of characters is the first thing regular expressions can do that isn’t already possible with the methods available on strings. However, if that was the only additional capability of regexes, they wouldn’t be much of an advance. Another capability is that you can specify that portions of the RE must be repeated a certain number of times.

The first metacharacter for repeating things that we’ll look at is *. * doesn’t match the literal character '*'; instead, it specifies that the previous character can be matched zero or more times, instead of exactly once.

For example, ca*t will match 'ct' (0 'a' characters), 'cat' (1 'a'), 'caaat' (3 'a' characters), and so forth.

Repetitions such as * are greedy; when repeating a RE, the matching engine will try to repeat it as many times as possible. If later portions of the pattern don’t match, the matching engine will then back up and try again with fewer repetitions.

Un exemple étape par étape mettra les choses au clair. Considérons l’expression a[bcd]*b. Elle correspond à la lettre 'a', suivi d’aucune ou plusieurs lettres de la classe [bcd]` et finira par un 'b'. Maintenant supposé que cette RE correspondra à la chaîne de caractères 'abcbd'.

Etape Correspond Explication
1 a Le a dans la RE correspondante.
2 abcbd The engine matches [bcd]*, going as far as it can, which is to the end of the string.
3 Failure Le moteur essaye de trouver une correspondance avec b mais la position courante est à la fin de la chaîne de caractères et échoue donc.
4 abcb Back up, so that [bcd]* matches one less character.
5 Failure Try b again, but the current position is at the last character, which is a 'd'.
6 abc Back up again, so that [bcd]* is only matching bc.
6 abcb Try b again. This time the character at the current position is 'b', so it succeeds.

The end of the RE has now been reached, and it has matched 'abcb'. This demonstrates how the matching engine goes as far as it can at first, and if no match is found it will then progressively back up and retry the rest of the RE again and again. It will back up until it has tried zero matches for [bcd]*, and if that subsequently fails, the engine will conclude that the string doesn’t match the RE at all.

Another repeating metacharacter is +, which matches one or more times. Pay careful attention to the difference between * and +; * matches zero or more times, so whatever’s being repeated may not be present at all, while + requires at least one occurrence. To use a similar example, ca+t will match 'cat' (1 'a'), 'caaat' (3 'a's), but won’t match 'ct'.

There are two more repeating qualifiers. The question mark character, ?, matches either once or zero times; you can think of it as marking something as being optional. For example, home-?brew matches either 'homebrew' or 'home-brew'.

The most complicated repeated qualifier is {m,n}, where m and n are decimal integers. This qualifier means there must be at least m repetitions, and at most n. For example, a/{1,3}b will match 'a/b', 'a//b', and 'a///b'. It won’t match 'ab', which has no slashes, or 'a////b', which has four.

You can omit either m or n; in that case, a reasonable value is assumed for the missing value. Omitting m is interpreted as a lower limit of 0, while omitting n results in an upper bound of infinity.

Readers of a reductionist bent may notice that the three other qualifiers can all be expressed using this notation. {0,} is the same as *, {1,} is equivalent to +, and {0,1} is the same as ?. It’s better to use *, +, or ? when you can, simply because they’re shorter and easier to read.

Utilisation des Expressions Régulières

Now that we’ve looked at some simple regular expressions, how do we actually use them in Python? The re module provides an interface to the regular expression engine, allowing you to compile REs into objects and then perform matches with them.

Compilations des Expressions Régulières

Regular expressions are compiled into pattern objects, which have methods for various operations such as searching for pattern matches or performing string substitutions.

>>> import re
>>> p = re.compile('ab*')
>>> p
re.compile('ab*')

re.compile() also accepts an optional flags argument, used to enable various special features and syntax variations. We’ll go over the available settings later, but for now a single example will do:

>>> p = re.compile('ab*', re.IGNORECASE)

The RE is passed to re.compile() as a string. REs are handled as strings because regular expressions aren’t part of the core Python language, and no special syntax was created for expressing them. (There are applications that don’t need REs at all, so there’s no need to bloat the language specification by including them.) Instead, the re module is simply a C extension module included with Python, just like the socket or zlib modules.

Putting REs in strings keeps the Python language simpler, but has one disadvantage which is the topic of the next section.

Le Maudit Backslash

As stated earlier, regular expressions use the backslash character ('\') to indicate special forms or to allow special characters to be used without invoking their special meaning. This conflicts with Python’s usage of the same character for the same purpose in string literals.

Let’s say you want to write a RE that matches the string \section, which might be found in a LaTeX file. To figure out what to write in the program code, start with the desired string to be matched. Next, you must escape any backslashes and other metacharacters by preceding them with a backslash, resulting in the string \\section. The resulting string that must be passed to re.compile() must be \\section. However, to express this as a Python string literal, both backslashes must be escaped again.

Caractères Stage
\section Text string to be matched
\\section backslash échappé pour re.compile()
"\\\\section" Escaped backslashes for a string literal

In short, to match a literal backslash, one has to write '\\\\' as the RE string, because the regular expression must be \\, and each backslash must be expressed as \\ inside a regular Python string literal. In REs that feature backslashes repeatedly, this leads to lots of repeated backslashes and makes the resulting strings difficult to understand.

The solution is to use Python’s raw string notation for regular expressions; backslashes are not handled in any special way in a string literal prefixed with 'r', so r"\n" is a two-character string containing '\' and 'n', while "\n" is a one-character string containing a newline. Regular expressions will often be written in Python code using this raw string notation.

In addition, special escape sequences that are valid in regular expressions, but not valid as Python string literals, now result in a DeprecationWarning and will eventually become a SyntaxError, which means the sequences will be invalid if raw string notation or escaping the backslashes isn’t used.

Regular String Chaîne de caractères brute
"ab*" r"ab*"
"\\\\section" r"\\section"
"\\w+\\s+\\1" r"\w+\s+\1"

Performing Matches

Une fois que nous avons un objet représentant une expression régulière compilée, qu’en faisons-nous ? Les objets motifs ont plusieurs méthodes et attributs. Seuls les plus significatifs seront couverts ici; consultez la documentation re pour la liste complète.

Méthode/Attribut Objectif
match() Determine if the RE matches at the beginning of the string.
search() Scan through a string, looking for any location where this RE matches.
findall() Find all substrings where the RE matches, and returns them as a list.
finditer() Find all substrings where the RE matches, and returns them as an iterator.

match() and search() return None if no match can be found. If they’re successful, a match object instance is returned, containing information about the match: where it starts and ends, the substring it matched, and more.

You can learn about this by interactively experimenting with the re module. If you have tkinter available, you may also want to look at Tools/demo/redemo.py, a demonstration program included with the Python distribution. It allows you to enter REs and strings, and displays whether the RE matches or fails. redemo.py can be quite useful when trying to debug a complicated RE.

This HOWTO uses the standard Python interpreter for its examples. First, run the Python interpreter, import the re module, and compile a RE:

>>> import re
>>> p = re.compile('[a-z]+')
>>> p
re.compile('[a-z]+')

Now, you can try matching various strings against the RE [a-z]+. An empty string shouldn’t match at all, since + means “one or more repetitions”. match() should return None in this case, which will cause the interpreter to print no output. You can explicitly print the result of match() to make this clear.

>>> p.match("")
>>> print(p.match(""))
None

Now, let’s try it on a string that it should match, such as tempo. In this case, match() will return a match object, so you should store the result in a variable for later use.

>>> m = p.match('tempo')
>>> m
<_sre.SRE_Match object; span=(0, 5), match='tempo'>

Now you can query the match object for information about the matching string. Match object instances also have several methods and attributes; the most important ones are:

Méthode/Attribut Objectif
group() Retourne la chaîne de caractères correspondant à la RE
start() Retourne la position de début de correspondance
end() Retourne la position de fin de correspondance
span() Return a tuple containing the (start, end) positions of the match

Trying these methods will soon clarify their meaning:

>>> m.group()
'tempo'
>>> m.start(), m.end()
(0, 5)
>>> m.span()
(0, 5)

group() returns the substring that was matched by the RE. start() and end() return the starting and ending index of the match. span() returns both start and end indexes in a single tuple. Since the match() method only checks if the RE matches at the start of a string, start() will always be zero. However, the search() method of patterns scans through the string, so the match may not start at zero in that case.

>>> print(p.match('::: message'))
None
>>> m = p.search('::: message'); print(m)
<_sre.SRE_Match object; span=(4, 11), match='message'>
>>> m.group()
'message'
>>> m.span()
(4, 11)

In actual programs, the most common style is to store the match object in a variable, and then check if it was None. This usually looks like:

p = re.compile( ... )
m = p.match( 'string goes here' )
if m:
    print('Match found: ', m.group())
else:
    print('No match')

Two pattern methods return all of the matches for a pattern. findall() returns a list of matching strings:

>>> p = re.compile(r'\d+')
>>> p.findall('12 drummers drumming, 11 pipers piping, 10 lords a-leaping')
['12', '11', '10']

The r prefix, making the literal a raw string literal, is needed in this example because escape sequences in a normal « cooked » string literal that are not recognized by Python, as opposed to regular expressions, now result in a DeprecationWarning and will eventually become a SyntaxError. See Le Maudit Backslash.

findall() has to create the entire list before it can be returned as the result. The finditer() method returns a sequence of match object instances as an iterator:

>>> iterator = p.finditer('12 drummers drumming, 11 ... 10 ...')
>>> iterator  
<callable_iterator object at 0x...>
>>> for match in iterator:
...     print(match.span())
...
(0, 2)
(22, 24)
(29, 31)

Module-Level Functions

You don’t have to create a pattern object and call its methods; the re module also provides top-level functions called match(), search(), findall(), sub(), and so forth. These functions take the same arguments as the corresponding pattern method with the RE string added as the first argument, and still return either None or a match object instance.

>>> print(re.match(r'From\s+', 'Fromage amk'))
None
>>> re.match(r'From\s+', 'From amk Thu May 14 19:12:10 1998')  
<_sre.SRE_Match object; span=(0, 5), match='From '>

Under the hood, these functions simply create a pattern object for you and call the appropriate method on it. They also store the compiled object in a cache, so future calls using the same RE won’t need to parse the pattern again and again.

Should you use these module-level functions, or should you get the pattern and call its methods yourself? If you’re accessing a regex within a loop, pre-compiling it will save a few function calls. Outside of loops, there’s not much difference thanks to the internal cache.

Fanion de Compilation

Compilation flags let you modify some aspects of how regular expressions work. Flags are available in the re module under two names, a long name such as IGNORECASE and a short, one-letter form such as I. (If you’re familiar with Perl’s pattern modifiers, the one-letter forms use the same letters; the short form of re.VERBOSE is re.X, for example.) Multiple flags can be specified by bitwise OR-ing them; re.I | re.M sets both the I and M flags, for example.

Here’s a table of the available flags, followed by a more detailed explanation of each one.

Option Signification
ASCII, A Makes several escapes like \w, \b, \s and \d match only on ASCII characters with the respective property.
DOTALL, S Fait en sorte que . corresponde à n’importe quel caractère, caractère de retour à la ligne inclus.
IGNORECASE, I Rechercher une correspondance non-sensible à la casse.
LOCALE, L Rechercher une correspondance sensible à la langue locale.
MULTILINE, M Correspondance multi-ligne, affectant ^ et $.
VERBOSE, X (for “extended”) Activer les RE verbeuses, qui peuvent être organisées de manière plus propres et compréhensibles.
I
IGNORECASE

Perform case-insensitive matching; character class and literal strings will match letters by ignoring case. For example, [A-Z] will match lowercase letters, too. Full Unicode matching also works unless the ASCII flag is used to disable non-ASCII matches. When the Unicode patterns [a-z] or [A-Z] are used in combination with the IGNORECASE flag, they will match the 52 ASCII letters and 4 additional non-ASCII letters: “İ” (U+0130, Latin capital letter I with dot above), “ı” (U+0131, Latin small letter dotless i), “ſ” (U+017F, Latin small letter long s) and “K” (U+212A, Kelvin sign). Spam will match 'Spam', 'spam', 'spAM', or 'ſpam' (the latter is matched only in Unicode mode). This lowercasing doesn’t take the current locale into account; it will if you also set the LOCALE flag.

L
LOCALE

Make \w, \W, \b, \B and case-insensitive matching dependent on the current locale instead of the Unicode database.

Locales are a feature of the C library intended to help in writing programs that take account of language differences. For example, if you’re processing encoded French text, you’d want to be able to write \w+ to match words, but \w only matches the character class [A-Za-z] in bytes patterns; it won’t match bytes corresponding to é or ç. If your system is configured properly and a French locale is selected, certain C functions will tell the program that the byte corresponding to é should also be considered a letter. Setting the LOCALE flag when compiling a regular expression will cause the resulting compiled object to use these C functions for \w; this is slower, but also enables \w+ to match French words as you’d expect. The use of this flag is discouraged in Python 3 as the locale mechanism is very unreliable, it only handles one « culture » at a time, and it only works with 8-bit locales. Unicode matching is already enabled by default in Python 3 for Unicode (str) patterns, and it is able to handle different locales/languages.

M
MULTILINE

(^ et $ n’ont pas encore été expliqués; ils seront introduit dans le section Plus de Méta-caratères.)

Usually ^ matches only at the beginning of the string, and $ matches only at the end of the string and immediately before the newline (if any) at the end of the string. When this flag is specified, ^ matches at the beginning of the string and at the beginning of each line within the string, immediately following each newline. Similarly, the $ metacharacter matches either at the end of the string and at the end of each line (immediately preceding each newline).

S
DOTALL

Makes the '.' special character match any character at all, including a newline; without this flag, '.' will match anything except a newline.

A
ASCII

Make \w, \W, \b, \B, \s and \S perform ASCII-only matching instead of full Unicode matching. This is only meaningful for Unicode patterns, and is ignored for byte patterns.

X
VERBOSE

This flag allows you to write regular expressions that are more readable by granting you more flexibility in how you can format them. When this flag has been specified, whitespace within the RE string is ignored, except when the whitespace is in a character class or preceded by an unescaped backslash; this lets you organize and indent the RE more clearly. This flag also lets you put comments within a RE that will be ignored by the engine; comments are marked by a '#' that’s neither in a character class or preceded by an unescaped backslash.

For example, here’s a RE that uses re.VERBOSE; see how much easier it is to read?

charref = re.compile(r"""
 &[#]                # Start of a numeric entity reference
 (
     0[0-7]+         # Octal form
   | [0-9]+          # Decimal form
   | x[0-9a-fA-F]+   # Hexadecimal form
 )
 ;                   # Trailing semicolon
""", re.VERBOSE)

Without the verbose setting, the RE would look like this:

charref = re.compile("&#(0[0-7]+"
                     "|[0-9]+"
                     "|x[0-9a-fA-F]+);")

In the above example, Python’s automatic concatenation of string literals has been used to break up the RE into smaller pieces, but it’s still more difficult to understand than the version using re.VERBOSE.

More Pattern Power

Jusqu’à présent nous avons seulement couvert une partie des fonctionnalités des expressions régulières. Dans cette section, nous couvrirons quelques nouveaux méta-caractères et comment utiliser les groupes pour récupérer des portions de textes correspondante.

Plus de Méta-caratères

There are some metacharacters that we haven’t covered yet. Most of them will be covered in this section.

Some of the remaining metacharacters to be discussed are zero-width assertions. They don’t cause the engine to advance through the string; instead, they consume no characters at all, and simply succeed or fail. For example, \b is an assertion that the current position is located at a word boundary; the position isn’t changed by the \b at all. This means that zero-width assertions should never be repeated, because if they match once at a given location, they can obviously be matched an infinite number of times.

|

Alternation, or the « or » operator. If A and B are regular expressions, A|B will match any string that matches either A or B. | has very low precedence in order to make it work reasonably when you’re alternating multi-character strings. Crow|Servo will match either 'Crow' or 'Servo', not 'Cro', a 'w' or an 'S', and 'ervo'.

To match a literal '|', use \|, or enclose it inside a character class, as in [|].

^

Matches at the beginning of lines. Unless the MULTILINE flag has been set, this will only match at the beginning of the string. In MULTILINE mode, this also matches immediately after each newline within the string.

For example, if you wish to match the word From only at the beginning of a line, the RE to use is ^From.

>>> print(re.search('^From', 'From Here to Eternity'))  
<_sre.SRE_Match object; span=(0, 4), match='From'>
>>> print(re.search('^From', 'Reciting From Memory'))
None

To match a literal '^', use \^.

$

Matches at the end of a line, which is defined as either the end of the string, or any location followed by a newline character.

>>> print(re.search('}$', '{block}'))  
<_sre.SRE_Match object; span=(6, 7), match='}'>
>>> print(re.search('}$', '{block} '))
None
>>> print(re.search('}$', '{block}\n'))  
<_sre.SRE_Match object; span=(6, 7), match='}'>

To match a literal '$', use \$ or enclose it inside a character class, as in [$].

\A
Matches only at the start of the string. When not in MULTILINE mode, \A and ^ are effectively the same. In MULTILINE mode, they’re different: \A still matches only at the beginning of the string, but ^ may match at any location inside the string that follows a newline character.
\Z
Correspond uniquement à la fin d’une chaîne de caractères.
\b

Word boundary. This is a zero-width assertion that matches only at the beginning or end of a word. A word is defined as a sequence of alphanumeric characters, so the end of a word is indicated by whitespace or a non-alphanumeric character.

The following example matches class only when it’s a complete word; it won’t match when it’s contained inside another word.

>>> p = re.compile(r'\bclass\b')
>>> print(p.search('no class at all'))
<_sre.SRE_Match object; span=(3, 8), match='class'>
>>> print(p.search('the declassified algorithm'))
None
>>> print(p.search('one subclass is'))
None

There are two subtleties you should remember when using this special sequence. First, this is the worst collision between Python’s string literals and regular expression sequences. In Python’s string literals, \b is the backspace character, ASCII value 8. If you’re not using raw strings, then Python will convert the \b to a backspace, and your RE won’t match as you expect it to. The following example looks the same as our previous RE, but omits the 'r' in front of the RE string.

>>> p = re.compile('\bclass\b')
>>> print(p.search('no class at all'))
None
>>> print(p.search('\b' + 'class' + '\b'))
<_sre.SRE_Match object; span=(0, 7), match='\x08class\x08'>

Second, inside a character class, where there’s no use for this assertion, \b represents the backspace character, for compatibility with Python’s string literals.

\B
Another zero-width assertion, this is the opposite of \b, only matching when the current position is not at a word boundary.

Regroupement

Frequently you need to obtain more information than just whether the RE matched or not. Regular expressions are often used to dissect strings by writing a RE divided into several subgroups which match different components of interest. For example, an RFC-822 header line is divided into a header name and a value, separated by a ':', like this:

From: author@example.com
User-Agent: Thunderbird 1.5.0.9 (X11/20061227)
MIME-Version: 1.0
To: editor@example.com

This can be handled by writing a regular expression which matches an entire header line, and has one group which matches the header name, and another group which matches the header’s value.

Groups are marked by the '(', ')' metacharacters. '(' and ')' have much the same meaning as they do in mathematical expressions; they group together the expressions contained inside them, and you can repeat the contents of a group with a repeating qualifier, such as *, +, ?, or {m,n}. For example, (ab)* will match zero or more repetitions of ab.

>>> p = re.compile('(ab)*')
>>> print(p.match('ababababab').span())
(0, 10)

Groups indicated with '(', ')' also capture the starting and ending index of the text that they match; this can be retrieved by passing an argument to group(), start(), end(), and span(). Groups are numbered starting with 0. Group 0 is always present; it’s the whole RE, so match object methods all have group 0 as their default argument. Later we’ll see how to express groups that don’t capture the span of text that they match.

>>> p = re.compile('(a)b')
>>> m = p.match('ab')
>>> m.group()
'ab'
>>> m.group(0)
'ab'

Subgroups are numbered from left to right, from 1 upward. Groups can be nested; to determine the number, just count the opening parenthesis characters, going from left to right.

>>> p = re.compile('(a(b)c)d')
>>> m = p.match('abcd')
>>> m.group(0)
'abcd'
>>> m.group(1)
'abc'
>>> m.group(2)
'b'

group() can be passed multiple group numbers at a time, in which case it will return a tuple containing the corresponding values for those groups.

>>> m.group(2,1,2)
('b', 'abc', 'b')

The groups() method returns a tuple containing the strings for all the subgroups, from 1 up to however many there are.

>>> m.groups()
('abc', 'b')

Backreferences in a pattern allow you to specify that the contents of an earlier capturing group must also be found at the current location in the string. For example, \1 will succeed if the exact contents of group 1 can be found at the current position, and fails otherwise. Remember that Python’s string literals also use a backslash followed by numbers to allow including arbitrary characters in a string, so be sure to use a raw string when incorporating backreferences in a RE.

For example, the following RE detects doubled words in a string.

>>> p = re.compile(r'\b(\w+)\s+\1\b')
>>> p.search('Paris in the the spring').group()
'the the'

Backreferences like this aren’t often useful for just searching through a string — there are few text formats which repeat data in this way — but you’ll soon find out that they’re very useful when performing string substitutions.

Non-capturing and Named Groups

Elaborate REs may use many groups, both to capture substrings of interest, and to group and structure the RE itself. In complex REs, it becomes difficult to keep track of the group numbers. There are two features which help with this problem. Both of them use a common syntax for regular expression extensions, so we’ll look at that first.

Perl 5 is well known for its powerful additions to standard regular expressions. For these new features the Perl developers couldn’t choose new single-keystroke metacharacters or new special sequences beginning with \ without making Perl’s regular expressions confusingly different from standard REs. If they chose & as a new metacharacter, for example, old expressions would be assuming that & was a regular character and wouldn’t have escaped it by writing \& or [&].

The solution chosen by the Perl developers was to use (?...) as the extension syntax. ? immediately after a parenthesis was a syntax error because the ? would have nothing to repeat, so this didn’t introduce any compatibility problems. The characters immediately after the ? indicate what extension is being used, so (?=foo) is one thing (a positive lookahead assertion) and (?:foo) is something else (a non-capturing group containing the subexpression foo).

Python supports several of Perl’s extensions and adds an extension syntax to Perl’s extension syntax. If the first character after the question mark is a P, you know that it’s an extension that’s specific to Python.

Now that we’ve looked at the general extension syntax, we can return to the features that simplify working with groups in complex REs.

Sometimes you’ll want to use a group to denote a part of a regular expression, but aren’t interested in retrieving the group’s contents. You can make this fact explicit by using a non-capturing group: (?:...), where you can replace the ... with any other regular expression.

>>> m = re.match("([abc])+", "abc")
>>> m.groups()
('c',)
>>> m = re.match("(?:[abc])+", "abc")
>>> m.groups()
()

Except for the fact that you can’t retrieve the contents of what the group matched, a non-capturing group behaves exactly the same as a capturing group; you can put anything inside it, repeat it with a repetition metacharacter such as *, and nest it within other groups (capturing or non-capturing). (?:...) is particularly useful when modifying an existing pattern, since you can add new groups without changing how all the other groups are numbered. It should be mentioned that there’s no performance difference in searching between capturing and non-capturing groups; neither form is any faster than the other.

A more significant feature is named groups: instead of referring to them by numbers, groups can be referenced by a name.

The syntax for a named group is one of the Python-specific extensions: (?P<name>...). name is, obviously, the name of the group. Named groups behave exactly like capturing groups, and additionally associate a name with a group. The match object methods that deal with capturing groups all accept either integers that refer to the group by number or strings that contain the desired group’s name. Named groups are still given numbers, so you can retrieve information about a group in two ways:

>>> p = re.compile(r'(?P<word>\b\w+\b)')
>>> m = p.search( '(((( Lots of punctuation )))' )
>>> m.group('word')
'Lots'
>>> m.group(1)
'Lots'

Named groups are handy because they let you use easily-remembered names, instead of having to remember numbers. Here’s an example RE from the imaplib module:

InternalDate = re.compile(r'INTERNALDATE "'
        r'(?P<day>[ 123][0-9])-(?P<mon>[A-Z][a-z][a-z])-'
        r'(?P<year>[0-9][0-9][0-9][0-9])'
        r' (?P<hour>[0-9][0-9]):(?P<min>[0-9][0-9]):(?P<sec>[0-9][0-9])'
        r' (?P<zonen>[-+])(?P<zoneh>[0-9][0-9])(?P<zonem>[0-9][0-9])'
        r'"')

It’s obviously much easier to retrieve m.group('zonem'), instead of having to remember to retrieve group 9.

The syntax for backreferences in an expression such as (...)\1 refers to the number of the group. There’s naturally a variant that uses the group name instead of the number. This is another Python extension: (?P=name) indicates that the contents of the group called name should again be matched at the current point. The regular expression for finding doubled words, \b(\w+)\s+\1\b can also be written as \b(?P<word>\w+)\s+(?P=word)\b:

>>> p = re.compile(r'\b(?P<word>\w+)\s+(?P=word)\b')
>>> p.search('Paris in the the spring').group()
'the the'

Lookahead Assertions

Another zero-width assertion is the lookahead assertion. Lookahead assertions are available in both positive and negative form, and look like this:

(?=...)
Positive lookahead assertion. This succeeds if the contained regular expression, represented here by ..., successfully matches at the current location, and fails otherwise. But, once the contained expression has been tried, the matching engine doesn’t advance at all; the rest of the pattern is tried right where the assertion started.
(?!...)
Negative lookahead assertion. This is the opposite of the positive assertion; it succeeds if the contained expression doesn’t match at the current position in the string.

To make this concrete, let’s look at a case where a lookahead is useful. Consider a simple pattern to match a filename and split it apart into a base name and an extension, separated by a .. For example, in news.rc, news is the base name, and rc is the filename’s extension.

Le motif de correspondance est plutôt simple:

.*[.].*$

Notice that the . needs to be treated specially because it’s a metacharacter, so it’s inside a character class to only match that specific character. Also notice the trailing $; this is added to ensure that all the rest of the string must be included in the extension. This regular expression matches foo.bar and autoexec.bat and sendmail.cf and printers.conf.

Now, consider complicating the problem a bit; what if you want to match filenames where the extension is not bat? Some incorrect attempts:

.*[.][^b].*$ The first attempt above tries to exclude bat by requiring that the first character of the extension is not a b. This is wrong, because the pattern also doesn’t match foo.bar.

.*[.]([^b]..|.[^a].|..[^t])$

The expression gets messier when you try to patch up the first solution by requiring one of the following cases to match: the first character of the extension isn’t b; the second character isn’t a; or the third character isn’t t. This accepts foo.bar and rejects autoexec.bat, but it requires a three-letter extension and won’t accept a filename with a two-letter extension such as sendmail.cf. We’ll complicate the pattern again in an effort to fix it.

.*[.]([^b].?.?|.[^a]?.?|..?[^t]?)$

In the third attempt, the second and third letters are all made optional in order to allow matching extensions shorter than three characters, such as sendmail.cf.

The pattern’s getting really complicated now, which makes it hard to read and understand. Worse, if the problem changes and you want to exclude both bat and exe as extensions, the pattern would get even more complicated and confusing.

A negative lookahead cuts through all this confusion:

.*[.](?!bat$)[^.]*$ The negative lookahead means: if the expression bat doesn’t match at this point, try the rest of the pattern; if bat$ does match, the whole pattern will fail. The trailing $ is required to ensure that something like sample.batch, where the extension only starts with bat, will be allowed. The [^.]* makes sure that the pattern works when there are multiple dots in the filename.

Excluding another filename extension is now easy; simply add it as an alternative inside the assertion. The following pattern excludes filenames that end in either bat or exe:

.*[.](?!bat$|exe$)[^.]*$

Modifying Strings

Up to this point, we’ve simply performed searches against a static string. Regular expressions are also commonly used to modify strings in various ways, using the following pattern methods:

Méthode/Attribut Objectif
split() Découpe la chaîne de caractère en liste, la découpant partout où la RE correspond
sub() Rechercher toutes les sous-chaînes de caractères où correspond la RE et les substituer par une chaîne de caractères différente
subn() Does the same thing as sub(), but returns the new string and the number of replacements

Splitting Strings

The split() method of a pattern splits a string apart wherever the RE matches, returning a list of the pieces. It’s similar to the split() method of strings but provides much more generality in the delimiters that you can split by; string split() only supports splitting by whitespace or by a fixed string. As you’d expect, there’s a module-level re.split() function, too.

.split(string[, maxsplit=0])

Split string by the matches of the regular expression. If capturing parentheses are used in the RE, then their contents will also be returned as part of the resulting list. If maxsplit is nonzero, at most maxsplit splits are performed.

You can limit the number of splits made, by passing a value for maxsplit. When maxsplit is nonzero, at most maxsplit splits will be made, and the remainder of the string is returned as the final element of the list. In the following example, the delimiter is any sequence of non-alphanumeric characters.

>>> p = re.compile(r'\W+')
>>> p.split('This is a test, short and sweet, of split().')
['This', 'is', 'a', 'test', 'short', 'and', 'sweet', 'of', 'split', '']
>>> p.split('This is a test, short and sweet, of split().', 3)
['This', 'is', 'a', 'test, short and sweet, of split().']

Sometimes you’re not only interested in what the text between delimiters is, but also need to know what the delimiter was. If capturing parentheses are used in the RE, then their values are also returned as part of the list. Compare the following calls:

>>> p = re.compile(r'\W+')
>>> p2 = re.compile(r'(\W+)')
>>> p.split('This... is a test.')
['This', 'is', 'a', 'test', '']
>>> p2.split('This... is a test.')
['This', '... ', 'is', ' ', 'a', ' ', 'test', '.', '']

The module-level function re.split() adds the RE to be used as the first argument, but is otherwise the same.

>>> re.split(r'[\W]+', 'Words, words, words.')
['Words', 'words', 'words', '']
>>> re.split(r'([\W]+)', 'Words, words, words.')
['Words', ', ', 'words', ', ', 'words', '.', '']
>>> re.split(r'[\W]+', 'Words, words, words.', 1)
['Words', 'words, words.']

Recherche et Substitution

Another common task is to find all the matches for a pattern, and replace them with a different string. The sub() method takes a replacement value, which can be either a string or a function, and the string to be processed.

.sub(replacement, string[, count=0])

Returns the string obtained by replacing the leftmost non-overlapping occurrences of the RE in string by the replacement replacement. If the pattern isn’t found, string is returned unchanged.

The optional argument count is the maximum number of pattern occurrences to be replaced; count must be a non-negative integer. The default value of 0 means to replace all occurrences.

Here’s a simple example of using the sub() method. It replaces colour names with the word colour:

>>> p = re.compile('(blue|white|red)')
>>> p.sub('colour', 'blue socks and red shoes')
'colour socks and colour shoes'
>>> p.sub('colour', 'blue socks and red shoes', count=1)
'colour socks and red shoes'

The subn() method does the same work, but returns a 2-tuple containing the new string value and the number of replacements that were performed:

>>> p = re.compile('(blue|white|red)')
>>> p.subn('colour', 'blue socks and red shoes')
('colour socks and colour shoes', 2)
>>> p.subn('colour', 'no colours at all')
('no colours at all', 0)

Empty matches are replaced only when they’re not adjacent to a previous match.

>>> p = re.compile('x*')
>>> p.sub('-', 'abxd')
'-a-b-d-'

If replacement is a string, any backslash escapes in it are processed. That is, \n is converted to a single newline character, \r is converted to a carriage return, and so forth. Unknown escapes such as \& are left alone. Backreferences, such as \6, are replaced with the substring matched by the corresponding group in the RE. This lets you incorporate portions of the original text in the resulting replacement string.

This example matches the word section followed by a string enclosed in {, }, and changes section to subsection:

>>> p = re.compile('section{ ( [^}]* ) }', re.VERBOSE)
>>> p.sub(r'subsection{\1}','section{First} section{second}')
'subsection{First} subsection{second}'

There’s also a syntax for referring to named groups as defined by the (?P<name>...) syntax. \g<name> will use the substring matched by the group named name, and \g<number> uses the corresponding group number. \g<2> is therefore equivalent to \2, but isn’t ambiguous in a replacement string such as \g<2>0. (\20 would be interpreted as a reference to group 20, not a reference to group 2 followed by the literal character '0'.) The following substitutions are all equivalent, but use all three variations of the replacement string.

>>> p = re.compile('section{ (?P<name> [^}]* ) }', re.VERBOSE)
>>> p.sub(r'subsection{\1}','section{First}')
'subsection{First}'
>>> p.sub(r'subsection{\g<1>}','section{First}')
'subsection{First}'
>>> p.sub(r'subsection{\g<name>}','section{First}')
'subsection{First}'

replacement can also be a function, which gives you even more control. If replacement is a function, the function is called for every non-overlapping occurrence of pattern. On each call, the function is passed a match object argument for the match and can use this information to compute the desired replacement string and return it.

In the following example, the replacement function translates decimals into hexadecimal:

>>> def hexrepl(match):
...     "Return the hex string for a decimal number"
...     value = int(match.group())
...     return hex(value)
...
>>> p = re.compile(r'\d+')
>>> p.sub(hexrepl, 'Call 65490 for printing, 49152 for user code.')
'Call 0xffd2 for printing, 0xc000 for user code.'

When using the module-level re.sub() function, the pattern is passed as the first argument. The pattern may be provided as an object or as a string; if you need to specify regular expression flags, you must either use a pattern object as the first parameter, or use embedded modifiers in the pattern string, e.g. sub("(?i)b+", "x", "bbbb BBBB") returns 'x x'.

Problèmes courants

Regular expressions are a powerful tool for some applications, but in some ways their behaviour isn’t intuitive and at times they don’t behave the way you may expect them to. This section will point out some of the most common pitfalls.

Use String Methods

Sometimes using the re module is a mistake. If you’re matching a fixed string, or a single character class, and you’re not using any re features such as the IGNORECASE flag, then the full power of regular expressions may not be required. Strings have several methods for performing operations with fixed strings and they’re usually much faster, because the implementation is a single small C loop that’s been optimized for the purpose, instead of the large, more generalized regular expression engine.

One example might be replacing a single fixed string with another one; for example, you might replace word with deed. re.sub() seems like the function to use for this, but consider the replace() method. Note that replace() will also replace word inside words, turning swordfish into sdeedfish, but the naive RE word would have done that, too. (To avoid performing the substitution on parts of words, the pattern would have to be \bword\b, in order to require that word have a word boundary on either side. This takes the job beyond replace()”s abilities.)

Another common task is deleting every occurrence of a single character from a string or replacing it with another single character. You might do this with something like re.sub('\n', ' ', S), but translate() is capable of doing both tasks and will be faster than any regular expression operation can be.

In short, before turning to the re module, consider whether your problem can be solved with a faster and simpler string method.

Greedy versus Non-Greedy

When repeating a regular expression, as in a*, the resulting action is to consume as much of the pattern as possible. This fact often bites you when you’re trying to match a pair of balanced delimiters, such as the angle brackets surrounding an HTML tag. The naive pattern for matching a single HTML tag doesn’t work because of the greedy nature of .*.

>>> s = '<html><head><title>Title</title>'
>>> len(s)
32
>>> print(re.match('<.*>', s).span())
(0, 32)
>>> print(re.match('<.*>', s).group())
<html><head><title>Title</title>

The RE matches the '<' in '<html>', and the .* consumes the rest of the string. There’s still more left in the RE, though, and the > can’t match at the end of the string, so the regular expression engine has to backtrack character by character until it finds a match for the >. The final match extends from the '<' in '<html>' to the '>' in '</title>', which isn’t what you want.

In this case, the solution is to use the non-greedy qualifiers *?, +?, ??, or {m,n}?, which match as little text as possible. In the above example, the '>' is tried immediately after the first '<' matches, and when it fails, the engine advances a character at a time, retrying the '>' at every step. This produces just the right result:

>>> print(re.match('<.*?>', s).group())
<html>

(Note that parsing HTML or XML with regular expressions is painful. Quick-and-dirty patterns will handle common cases, but HTML and XML have special cases that will break the obvious regular expression; by the time you’ve written a regular expression that handles all of the possible cases, the patterns will be very complicated. Use an HTML or XML parser module for such tasks.)

Utilisez re.VERBOSE

By now you’ve probably noticed that regular expressions are a very compact notation, but they’re not terribly readable. REs of moderate complexity can become lengthy collections of backslashes, parentheses, and metacharacters, making them difficult to read and understand.

For such REs, specifying the re.VERBOSE flag when compiling the regular expression can be helpful, because it allows you to format the regular expression more clearly.

The re.VERBOSE flag has several effects. Whitespace in the regular expression that isn’t inside a character class is ignored. This means that an expression such as dog | cat is equivalent to the less readable dog|cat, but [a b] will still match the characters 'a', 'b', or a space. In addition, you can also put comments inside a RE; comments extend from a # character to the next newline. When used with triple-quoted strings, this enables REs to be formatted more neatly:

pat = re.compile(r"""
 \s*                 # Skip leading whitespace
 (?P<header>[^:]+)   # Header name
 \s* :               # Whitespace, and a colon
 (?P<value>.*?)      # The header's value -- *? used to
                     # lose the following trailing whitespace
 \s*$                # Trailing whitespace to end-of-line
""", re.VERBOSE)

Ceci est beaucoup plus lisible que:

pat = re.compile(r"\s*(?P<header>[^:]+)\s*:(?P<value>.*?)\s*$")

Retour

Les expressions régulières sont un sujet compliqué. Est-ce que ce document vous a aidé à les comprendre ? Des parties ne sont pas claires, ou des problèmes que vous avez rencontrés ne sont pas traités ici ? Si tel est le cas, merci d’envoyer vos suggestions d’améliorations à l’auteur.

Le livre le plus complet sur les expressions régulières est certainement Mastering Regular Expressions de Jeffrey Friedl, publié par O’Reilly. Malheureusement, il se concentre sur les déclinaisons Perl et Java des expressions régulières et ne contient aucun contenu pour Python; il n’est donc pas utile d’en faire référence pour la programmation Python. (La première édition traitée du module Python:mod:!regex, maintenant supprimé, ce qui ne vous aidera pas beaucoup.) Pensez à le retirer de votre bibliothèque.