Skip to main content

rule-based algorithms converting Romanized text to original scripts

Project description

deromanize is a set of tools to aid in converting Romanized text back into original scripts.


deromanize requires Python 3.5 or better.

$ git clone
$ cd deromanize
$ pip3 install .

Or, to use the version in PyPI:

$ pip3 install deromanize

This assumes you’re working in a virtualenv, as you ought. Otherwise, use the --user flag with pip. There’s no reason ever to install this as root. Don’t do it.

Basic usage

The first step in working with deromanize is defining your decoding keys in data through a profile.

A profile has fairly simple format. It is a dictionary which contains dictionaries that have all the information needed to build up transliteration rules. It can easily be stored as JSON or any format can represent the same data structures as JSON. I like to use YAML because it’s easy to write.

The profile should contain at least one character group (the example below has two) and a keys section.

    - consonants
    - vowels

 ʾ: א
 b: ב
 v: ב
 g: ג
 d: ד
 h: ה
 ṿ: [וו, ו]
 z: ז
 : ח
 : ט
 y: [יי, י]
 k: כ
 kh: כ
 l: ל
 m: מ
 n: נ
 s: ס
 ʿ: ע
 p: פ
 f: פ
 ts: צ
 : ק
 r: ר
 ś: ש
 sh: ש
 t: ת

 i: י
 e: ['', י]
 a: ''
 o: [ו, '']
 u: ו

The letters in the arrays are reversed on this web page when viewed in most modern web browsers because of automatic bidi resolution. Most editors also pull these shenanigans, which is great for text, but not great for code. Emacs has options for this, and Vim doesn’t even try to fix bidi (though your terminal might). I don’t know what kind of options your favorite editor has for falling back to “stupid” LTR text flow when it screws up code readability.

Character groups:

Each character group is a dictionary containing the Romanized form character as a key, and the original form as the value. If a Romanized key can have multiple possible interpretations, they may be put in lists. The person defining the standard ought to put these replacements in the order they believe to most frequent in the actual language, as results will ultimately be sorted based on the index numbers of these lists.

Romanized forms can contain an arbitrary number of characters, so digraphs will be fine. You may even wish to define longer clusters to, for example, provide uniform handling of common morphological affixes. deromanize uses greedy matching, so the longest possible cluster will always be matched. There are also other uses for character groups involving pattern matching which will be covered later. (You can really stick any arbitrary data in this file that you think might be helpful later; aside from two keys, keys and char_sets, nothing will be processed automatically)


keys is a dictionary of objects that allow you to compose the different character groups in different ways. For one-to-one transliteration standards, you’d theoretically only need one key (and probably not need to mess around with this framework, though it would get the job done just fine).

In this case, we create one key called base and a list of the groups it will contain, consonants and vowels.

Given the above configuration, we can do something like this:

>>> # KeyGenerators only deal with python objects, so we have to
>>> # deserialize it from our chosen format.
>>> import deromanize as dr
>>> import yaml
>>> PROFILE = yaml.safe_load(open('above_profile.yml'))
>>> keys = dr.KeyGenerator(PROFILE)

From here, we can start sending words to the base key and see what comes out.

>>> parts = keys['base'].getallparts('shalom')
>>> parts
[ReplacementList('sh', [(0, 'ש')]), ReplacementList('a', [(0, '')]), ReplacementList('l', [(0, 'ל')]), ReplacementList('o', [(0, 'ו'), (1, '')]), ReplacementList('m', [(0, 'מ')])]
>>> # looks a little silly.
>>> shalom = dr.add_rlists(parts)
>>> shalom
ReplacementList('shalom', [(0, 'שלומ'), (1, 'שלמ')])
>>> # conversion to a string provides a more readable version
>>> print(shalom)
0 שלומ
1 שלמ

So, basically, the .getallparts() method takes a string as input and decodes it bit by bit, grabbing all possible original versions for each Romanization symbol. You can get all the possible version of the word together. Ignore the numbers for now. They have to deal with sorting. This is just to demonstrate the most basic use-case. The Hebrew-speakers may observe that neither of these options is correct (because it doesn’t account for final letters), so we’ll dive a bit deeper into the system to see how more complex situations can be dealt with.

Building Complex Profiles

Let’s take a look at a more complex profile, bit by bit. (See the profile in its entirety here.)

Defining Keys

      - consonants
      - vowels
      - other
      - clusters
      - infrequent: 10

    parent: base
      - beginning
      - beginning patterns

    parent: base
    groups: final
    suffix: true

The first thing to know is that there are a few configuration shortcuts if a key only contains a list, that list is automatically assigned to groups. Therefore:

  - consonants
  - vowels
  - other
  - clusters
  - infrequent: 10

is the same as…

    - consonants
    - vowels
    - other
    - clusters
    - infrequent: 10

The other shortcut is that base is actually a special key name. If it is defined, all other character groups will inherit the default character set from it as a prototype which you can selectively override and extend with other character groups to build all the groups you need.


  - beginning
  - beginning patterns

... is the same as…

  base: base
    - beginning
    - beginning patterns

If you don’t want this behavior for any of your keys, you can simply choose not to define base. If you find it useful, but you want to get out of it for a particular key, you can set it to None (which happens to be spelled null in JSON and YAML).

  base: null
  groups: (some groups here)

You can, of course, use any other key as your base and get into some rather sophisticated composition if you wish. Just don’t create a dependency cycle or you’ll end up in a never-ending loop. (Well, I guess it will end when Python hits its recursion limit.)

One last thing you may notice that’s odd in this section is that one of the groups in base is infrequent: 10. This is a way to manipulate the sort order of results. It might be a good time to explain that in a little more detail.

Sorting and “Weight”

Each possible replacement for any Romanization symbol or cluster may have one or more possible replacements, and therefore can be given as lists. As shorthand, if there is only one possible replacement, it may be a string, but it will be converted to a list containing that one item at runtime.

As the items are added, they are assigned a weight. In the common case, that weight is simply the index number in the list.

Let’s go back and pretend that are working with the simple profile at the top of this README. We have a line like this in the file:

y: [יי, י]

When we run this through the KeyGenerator instance we can see what happens to it:

>>> key['base']['y']
ReplacementList('y', [(0, 'יי'), (1, 'י')])
>>> key['base']['y'][0], 'יי')

Basically, each item is explicitly assigned its weight. When you add two Replacement instances together, their weights are added, and their strings are concatenated.

>>> key['base']['y'][0] + key['base']['o'][0], 'ייו')

Likewise, when two ReplacemntList items are added together, the Romanized strings are concatenated, and all the permutations of their original forms are combined as well:

>>> print(key['base']['y'] + key['base']['o'])
 0 ייו
 1 יי
 1 יו
 2 י

As you may observe, the ReplacementList comes with pretty formatting when used with print() for easier debugging.

After all the variations have been generated, the resulting ReplacementList can be sorted with its .sort() method according to these weights, from least to greatest.

However (coming back to the real config file), certain normalizations may appear infrequently, so that one wants to try everything else before resorting for that. These may be rare cases as is the case with my infrequent character group, or it may be a way to hedge bets against human error in input data.

What infrequent: 10 does is tell the KeyGenerator instance to add 10 to the index number of each Replacement to generate its weight. Groups used in this way will not overwrite groups that already values that already exist in the key. Instead, the replacement list will be extended to include these values. This will drag less likely options to the bottom of the list.

>>> shalom = add_rlists( key['base'].getallparts('shalom'))
>>> print(shalom)
 0 שלומ
 5 שלמ
10 שלאמ
10 שאלומ
15 שאלמ
20 שאלאמ

A couple of colleagues pointed out to me that this weighting system seems very arbitrary in and it should be based on values between 0 and 1 for a more scientific and statistical approach. However, the purpose of the weighting system is simply to allow the person defining to have a greater control over how results are sorted and have nothing to do with science or statistics. If you want to sink items in a particular group lower in the final sort order, stick a big fat number besides the replacement value. This is the only meaning the numbers have.

However, if you need to have these numbers look more scientific to use with a statistical framework, they can be converted at any point:

>>> shalom.makestat()
>>> print(shalom)
0.6855870895937674 שלומ
0.11426451493229456 שלמ
0.06232609905397886 שלאמ
0.06232609905397886 שאלומ
0.04284919309961046 שאלמ
0.03264700426636988 שאלאמ

Also note that weights can arbitrary be added to any replacement directly when it is defined. We could get a similar result for the word above if, instead of using the infrequent group, we had defined the letters like this:

a: ['' [10, 'א']]
o: [ו, '', [10, א]]

Here are those bidi shenanigans I mention earlier. Paste into Vim or something to see the correct character order.

Any replacement that is a list or tuple of two beginning with an integer will use that integer as its weight assignment. In this way, one can have very direct control over how results are sorted.

This is also what is done for the case when o should be replaced with the empty string. It is manually weighted at 5.

Using Suffix Keys

Those of you who know Hebrew have noticed, dobutless, that we are still unable to generate the word שלום as it is supposed to look, with a proper final mem. Suffix keys are used to deal with word endings, such as final letters (common in Semitic writing systems but also found in Greek) or perhaps common morphological suffixes.

A suffix group is defined like this:

  groups: [ some list of groups ]
  suffix: true

This will create a reversed tokenizer that begins looking for tokens at the end of the word and moves forward. It can be used to deal with endings separately.

>>> suffix, remainder = keys['end'].getpart('shalom')
>>> suffix
ReplacementList('m', [(0, 'ם')])
>>> remainder
>>> front = add_rlists(keys['base'].getallparts(remainder))
>>> shalom = front + suffix
>>> print(shalom)
 0 שלום
 5 שלם
10 שלאם
10 שאלום
15 שאלם
20 שאלאם

We’ve also seen the .getpart() method of a key for the first time. This method takes a string as input returns a replist for the first matching token (or the last matching token, if suffix was specified) as well as the remaining string. This is useful if you want to have different rules about the beginning, middle and end of a word, as I typically do.

Pattern-Based Replacement Generation

deromanize profiles also allow the user to generate large numbers of replacements from pattern-based definitions. Patterns rely on the use of special characters that will generate sets of characters defined elsewhere in the profile.

This somewhat analogous to ranges of characters like \w or \s in regex. However, unlike regex, which characters will be treated as special are not yet defined (nor are there values). To create these character sets and their aliases, the char_sets group must be defined in the profile.

    key: base
    chars: consonants
    key: front
    chars: consonants

What this says is that C will be an alias for all the characters defined in the group consonants and replacements will be drawn from the base key. Likewise F will stand for the same character set, consonants, but replacements will be drawn from the key called front. The value of chars may also be a list of literal characters instead of the name of a character group. key, however must be a key defined in the keys group. If no base is defined for the character set alias, it defaults to the base key. Likewise, if the value of any character alias is not a dictionary (containing at least a chars value), its value will be assigned to for chars, so a shorthand for the above is:

  C: consonants
    key: front
    chars: consonants

Also note that the character aliases themselves (C and F above) can be arbitrary length. You should try to chose sequences that cannot possibly appear in your transliteration. Capitals have no meaning in the standard I’ve defined, so I use them, but you could also use something like \c and \v if you needed. Just note that there is no mechanism for escaping special characters once defined.

When it comes to actually using these in replacement definitions, it goes something like this…

beginning patterns:
  FiCC: [\1\2\3, \1י\2\3]
  FoCC: [\1ו\2\3, \1\2\3]
  FeCC: [\1\2\3]

Each alias character becomes something like a ‘capture group’ in regex, and can be recalled int the replacement string with a backslashed number (like regex). The appropriate replacements will be generated for all characters in the group.

Please be aware that you can generate a LOT of replacements this way (the above groups, with the rest of this config file, generate over 50,000 new replacements). This can take a few seconds to chug through.

A Little Hidden Metadata

Each Replacement in a ReplacementList has an attribute called keyvalue. This is a tuple where each item a two-tuple of the token found and how it was interpreted in the case of the specific Replacment. Continuing with our shalom variable from previous examples:

>>> shalom[0], 'שלום')
>>> shalom[0].keyvalue
(('sh', 'ש'), ('a', ''), ('l', 'ל'), ('o', 'ו'), ('m', 'ם'))

This can be useful for various things. Say we wanted to generate another transliteration standard from this. Some outside source has verified that the generated option שלום is the correct Hebrew form of shalom, but now we want to create a more detail transliteration that will show that the /o/ vowel was marked with the letter vav. Because we can go back and specifically see that /o/ was realized as vav in this case, it is easy to generate something like šalôm if we want to.

Additionally, this can be a way to detect errors in the transliteration.

In the system we use, the letter ק is supposed to be written as , using the diacritic to distinguish it from hard kaf (כ). However, sometimes people make mistakes. Assuming we have defined a fault-tolerant standard which understands that sometimes people will write k instead of ḳ, we can generate something like this:

>>> shuk = 'shuk' # oops! should be "shuḳ"
>>> shuk = add_rlists(keys['base'].getallparts(shuk))
 0 שוכ
20 שוק

When it has been verified that שוק is the correct Hebrew form, we can look at how it was built up:

>>> shuk[1].keyvalue
(('sh', 'ש'), ('u', 'ו'), ('k', 'ק')

At this point it is trivial for the computer to see that ק was incorrectly transcribed as k, and it can easily go back and correct the source if necessary. There is a function to aid in using this key-value data in generating new forms in the cacheutils module. See the following section for links to documentation about that.

Note that some of this data may be lost for tokens generated with patterns if the keys have been cached with cached_keys and recalled. cached_keys should only be used to speed-up small utilities where this information is not needed.

Extras: Caching Helpers, Miscellaneous Utilities, and Microservice

At the end of the day, deromanize is just a helpful tool taking data in one script and generating all possible equivalents in another script. For conversion between any systems that don’t have one-to-one correspondence. It’s up to the user figure out how the correct alternative will be selected. However, deromanize.cacheutils has some simple utilities that can help with recall once the correct form has been selected. has some other helper functions that have been very useful to me while working with deromanize on real data in different languages and scripts – helpers to strip punctuation, remove diacritics, correct mistakes in the source text, as well as a decoder function that will work well with complex profiles which have different rules for the beginning, middle and end of a word.

If you’re using deromanize, there is a good chance you’ll want this kind of stuff. Check out the docs on those modules!

Additionally, there is another package you can use to spin up deromanize as a microservice, microdero. This primarily for people who are interested using deromanize, but cannot or do not wish to have most of their project in Python, such web app that uses the generated data on the client or a mature project in another language that would like to integrate deromanize.

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Source Distributions

No source distribution files available for this release.See tutorial on generating distribution archives.

Built Distribution

deromanize-0.9-py3-none-any.whl (25.3 kB view hashes)

Uploaded py3

Supported by

AWS AWS Cloud computing and Security Sponsor Datadog Datadog Monitoring Fastly Fastly CDN Google Google Download Analytics Microsoft Microsoft PSF Sponsor Pingdom Pingdom Monitoring Sentry Sentry Error logging StatusPage StatusPage Status page