Translating HTML content using a plain text supporting machine translation engine

At Wikimedia, I am currently working on ContentTranslation tool, a machine aided translation system to help translating articles from one language to another. The tool is deployed in several wikipedias now and people are creating new articles sucessfully.

The ContentTranslation tool provides machine translation as one of the translation tool, so that editors can use it as an initial version to improve up on. We used Apertium as machine translation backend and planning to support more machine translation services soon.

A big difference in editing using ContentTranslation, is it does not involve Wiki Markup. Instead, editors can edit rich text. Basically it is contenteditable HTML elements. This also means, what you translate is HTML sections of articles.

The HTML contains all possible markups that a typical Wikipedia article has. This means, the machine translation is on HTML content. But, not all MT engines support HTML content.

Some MT engines, such as Moses, output subsentence alignment information directly, showing which source words correspond to which target words.

$ echo 'das ist ein kleines haus' | moses -f phrase-model/moses.ini -t
this is |0-1| a |2-2| small |3-3| house |4-4|

The Apertium MT engine does not translate formatted text faithfully. Markup such as HTML tags is treated as a form of blank space. This can lead to semantic changes (if words are reordered), or syntactic errors (if mappings are not one-to-one).

$ echo 'legal <b>persons</b>' | apertium en-es -f html
Personas <b>legales</b>
$ echo 'I <b>am</b> David' | apertium en-es -f html
Soy</b> David 

Other MT engines exhibit similar problems. This makes it challenging to provide machine translations of formatted text. This blog post explains how this challenge is tackled in ContentTranslation.

As we saw in the examples above, a machine translation engine can cause the following errors in the translated HTML. The errors are listed in descending order of severity.

  1. Corrupt markup – If the machine translation engine is unaware of HTML structure, they can potentially move the HTML tags randomly, causing corrupted markup in the MT result
  2. Wrongly placed annotations – The two examples given above illustrate this. It is more severe if content includes links and link targets were swapped or randomly given in the MT output.
  3. Missing annotations – Sometimes the MT engine may eat up some tags in the translation process.
  4. Split annotations -During translation a single word can be translated to more than one word. If the source word has a mark up, say <a> tag. Will the MT engine apply the <a> tag wrapping both words or apply to each word?

All of the above issues can cause bad experience to translators.

Apart from potential issues with markup transfer, there is another aspect about sending HTML content to MT engines. Compared to plain text version of a paragraph, HTML version is bigger in terms of size(bytes). Most of these extra addition is tags and attributes which should be unaffected by the translation. This is unnecessary bandwidth usage. If the MT engine is a metered engine(non-free, API access is measured and limited), we are not being economic.

An outline of the algorithm we used to transfer markups from source content to translated content is given below.

  1. The input HTML content is translated into a LinearDoc, with inline markup (such as bold and links) stored as attributes on a linear array of text chunks. This linearized format is convenient for important text manipulation operations, such as reordering and slicing, which are challenging to perform on an HTML string or a DOM tree.
  2. Plain text sentences (with all inline markup stripped away) are sent to the MT engine for translation.
  3. The MT engine returns a plain text translation, together with subsentence alignment information (saying which parts of the source text correspond to which parts of the translated text).
  4. The alignment information is used to reapply markup to the translated text.

This make sure that MT engines are translating only plain text and mark up is applied as a post-MT processing.

Essentially the algorithm does a fuzzy match to find the target locations in translated text to apply annotations. Here also content given to MT engines is plain text only.

The steps are given below.

  1. For the text to translate, find the text of inline annotations like bold, italics, links etc. We call it subsequences.
  2. Pass the full text and subsequences to the plain text machine translation engine. Use some delimiter so that we can do the array mapping between source items(full text and subsequences) and translated items.
  3. The translated full text will have the subsequences somewhere in the text. To locate the subsequence translation in full text translation, use an approximate search algorithm
  4. The approximate search algorithm will return the start position of match and length of match. To that range we map the annotation from the source html.
  5. The approximate match involves calculating the edit distance between words in translated full text and translated subsequence. It is not strings being searched, but ngrams with n=number of words in subsequence. Each word in ngram will be matched independently.

To understand this, let us try the algorithm in some example sentences.

  1. Translating the Spanish sentence <p>Es <s>además</s> de Valencia.</p> to Catalan: The plain text version is Es además de Valencia.. And the subsequence with annotation is  además . We give both the full text and subsequence to MT. The full text translation is A més de València.. and the word  además  is translated as a més. We do a search for a més in the full text translation. The search will be successfull and the <s> tag will be applied, resulting <p>És <s>a més</s> de València.</p>.The seach performed in this example is plain text exact search. But the following example illustrate why it cannot be an exact search.
  2. Translating an English sentence <p>A <b>Japanese</b> <i>BBC</i> article</p> to Spanish. The full text translation of this is Un artículo de BBC japonés  One of the subsequenceJapanese will get translated as Japonés. The case of J differs and search should be smart enough to identify japonés as a match for Japonés. The word order in source text and translation is already handled by the algorithm. The following example will illustrate that is not just case change that happens.
  3. Translating <p>A <b>modern</b> Britain.</p> to Spanish. The plain text version get translated as Una Gran Bretaña moderna.  and the word with annotation modern get translated as  Moderno. We need a match for moderna and Moderno. We get <p>Una Gran Bretaña <b>moderna</b>.</p>. This is a case of word inflection. A single letter at the end of the word changes.
  4. Now let us see an example where the subsequence is more than one word and the case of nested subsequences. Translating English sentence <p>The <b>big <i>red</i></b> dog</p> to Spanish. Here, the subsequnce Big red is in bold, and inside that, the red is in italics. In this case we need to translate the full text, sub sequence big red and red. So we have,   El perro rojo grande as full translation, Rojo grande and Rojo as translations of sub sequences. Rojo grande need to be first located and bold tag should be applied. Then search for Rojo and apply Italic. Then we get <p>El perro <b><i>rojo</i> grande</b></p>.
  5. How does it work with heavily inflected languages like Malayalam? Suppose we translate <p>I am from <a href=”x”>Kerala<a></p> to Malayalam. The plain text translation is ഞാന്‍ കേരളത്തില്‍ നിന്നാണു്. And the sub sequence Kerala get translated to കേരളം. So we need to match കേരളം and കേരളത്തില്‍. They differ by an edit distance of 7 and changes are at the end of the word. This shows that we will require language specific tailoring to satisfy a reasonable output.

The algorithm to do an approximate string match can be a simple levenshtein distance , but what would be the acceptable edit distance? That must be configurable per language modules. And the following example illustrate that just doing an edit distance based matching wont work.

Translating <p>Los Budistas no <b>comer</b> carne</p> to English. Plain text translation is The Buddhists not eating meat. Comer translates as eat. With an edit distance approach, eat will match more with meat than eating. To address this kind of cases, we mix a second criteria that the words should start with same letter. So this also illustrate that the algorithm should have language specific modules.

Still there are cases that cannot be solved by the algorithm we mentioned above. Consider the following example

Translating <p>Bees <b>cannot</b> swim</p>. Plain text translation to Spanish is Las Abejas no pueden nadar and the phrase cannot translates as Puede no. Here we need to match Puede no andno pueden which of course wont match with the approach we explained so far.

To address this case, we do not consider sub sequence as a string, but an n-gram where n= number of words in the sequence. The fuzzy matching should be per word in the n-gram and should not be for the entire string. ie. Puede to be fuzzy matched with no and pueden, and no to be fuzzy matched wth no and pueden– left to right, till a match is found. This will take care of word order changes as welll as inflections

Revisiting the 4 type of errors that happen in annotation transfer, with the algorithm explained so far, we see that in worst case, we will miss annotations. There is no case of corrupted markup.

As and when ContentTranslation add more language support, language specific customization of above approach will be required.

You can see the algorithm in action by watching the video linked above. And here is a ascreenshot:

Translation of a paragraph from Palak Paneer article of Spanish Wikipedia to Catalan. Note the links, bold etc applied in correct position in translation at right side

If anybody interested in the code, See https://github.com/wikimedia/mediawiki-services-cxserver/tree/master/mt – It is a javascript module in a nodejs server which powers ContentTranslation.

Credits: David Chan, my colleague at Wikimedia,  for extensive help on providing lot of example sentences with varying complexity to fine tune the algorithm. The LinearDoc model that make the whole algorithm work is written by him. David also wrote an algorithm to handle the HTML translation using an upper casing algorithm, you can read it from here. The approximation based algorithm explained above replaced it.

Parsing CLDR plural rules in javascript

English and many other languages have only 2 plural forms. Singular if the count is one and anything else is plural including zero.

But for some other languages, the plural forms are more than 2. Arabic, for example has 6 plural forms, sometimes referred as ‘zero’, ‘one’, ‘two’, ‘few’, ‘many’, ‘other’ forms. Integers 11-26, 111, 1011 are of ‘many’ form, while 3,4,..10 are ‘few’ form.

While preparing the interface messages for application user interfaces, grammatically correct sentences are must. “Found 1 results” or “Found 1 result(s)” are bad interface messages. For a developer, if the language in the context is English or languages having similar plural forms, it may be a matter of an if condition to conditionally choose one of the messages.

But that approach is not scalable if we want to deal with lot of languages. Some applications come with their own plural handling mechanism, probably by a module that tells you the plural form, given a number, and language. The plural forms per language and the rules to determine it is defined in CLDR. CLDR defines the plural rules in a markup language named LDML and releases the collections frequently.

If you look at the CLDR plural rules table you can easily understand this. The rules are defined in a particular syntax. For example, the Russian plural rules are given below.

One need to pass the value of the number to the variable in the above expressions and evaluate. If the expression evaluates to a boolean true, then the corresponding plural form should be used.

So, an expression like  n = 0 or n != 1 and n mod 100 = 1..19 mapped to ‘many’ holds true if the value of n=0,119, 219, 319. So we say that they are of ‘few’ plural form.

But in the Russian example given above, we don’t see n, but we see variables v, i etc. The meaning of these variables are defined in the standard as:

Symbol Value
n absolute value of the source number (integer and decimals).
i integer digits of n.
v number of visible fraction digits in n, with trailing zeros.
w number of visible fraction digits in n, without trailing zeros.
f visible fractional digits in n, with trailing zeros.
t visible fractional digits in n, without trailing zeros.

Keeping these definitions in mind, the expression v = 0 and i % 10 = 1 and i % 100 != 11 evaluates true for 1,21,31, 41 etc and false for 11. In other words, number 1,21,31 are of plural form “one” in Russian.

A module to support the plural forms for any language can manually(or semi automatically) convert this expressions to programming language one time and use it. Twitter-cldr a CLDR abstraction library by twitter follows this method. It converted the above given plural rules to the following javascript expression using a compiler.

This works. But CLDR updates the plural rules in every releases. Most of the time, it contains additional language support. Sometimes the rules are changed or improved too. The maintainer of the module need to recompile them to javascript expression in such cases.

If we can write a compiler to generate javascript from this expressions, can’t we write a parser-evaluator for the expressions? So that we just need to pass the rule and the number to that evaluator  and it returns the plural form?

CLDRPluralRuleParser

 CLDR Plural Rule Evaluator
CLDR Plural Rule Evaluator

CLDRPluralRuleParser is that evaluator. I wrote this parser when we at Wikimedia foundation wanted a data driven plural rule evaluation for the 300+ languages we support. It started as a free time project in June 2012. Later it became part of MediaWiki core to support front-end internationalization. We wanted a PHP version also to support interface messages constructed at server side. Tim Starling wrote a PHP CLDR plural rule evaluator.

It is javascript library that takes the standardized plural rule and an integer and returning true or false depending on whether the rule pass for the given integer. It is written with UMD/common.js pattern and available as a node module too.

The node module comes with command line interface, just to experiment with rules.

$ cldrpluralruleparser 'n is 1' 0

false

The module does not self contain the plural rules collection or data. Developers need to have that collection as an xml or json inside the application and need to pass to the module. In that sense, one cannot offload the whole i18n message processing task to this module. For a more handy internationalization with javascript, that takes care of plural, gender, grammar etc, you may consider jquery.i18n which contains CLDRPluralRuleParser.

An example showing how to use the CLDR supplied plural rule data and this library is included in the repository. You can play with that application here.

License: Initially the license of the module was GPL, but as per some of the collaboration discussion between Wikimedia, cldrjs, jQuery.globalize, moment.js, it was decided to change the license to MIT.

Mediawiki moves to json based localisation file format

Mediawiki is moving from PHP array based localisation file format to json format.

This is based on the RFC: https://www.mediawiki.org/wiki/Requests_for_comment/Localisation_format

A lot of extensions were also migrated to the new localisation format, thanks to Siebrand Mazeland

The json based localization file format was first introduced in our frontend javascript i18n library https://github.com/wikimedia/jquery.i18n

If you are interested in seeing some of the sample json files see https://gerrit.wikimedia.org/r/#/c/122787/ , claimed as “largest patch set in the history of MediaWiki” 

Brackets, my favorite javascript IDE

I use Brackets for web development. I had tried several other IDEs but Brackets is my current favorite IDE. A few things I liked is listed below

Some extensions I use with Brackets are:

  1. Markdown Preview for easy editing of markdown
  2. Brackets Git for git integration
  3. Themes for Brackets For Monokai Darksoda theme I use
  4. Brackets Linux UI
  5. Interactive Linter realtime JSHint/JSLint/CoffeeLint reports into brackets as you work on your code
  6. WD Minimap for SublimeText like code overview
  7. Beautify for automatic code formatting as you save using jsbeautify

Beautify extension helps me a lot because most of the MediaWiki related code I write needs be as MediaWiki javascript coding convention. I never get it right if I format manually. The convention is a bit different from usual js code formatting. In general you need to use a lot of whitespaces. This extension was using a default jsbeautify formatting configuration and I wanted it to be customizable per project so that I can write my own .jsbeautifyrc file to get my code formatted as per conventions.

There was an enhancement bug for this. I wrote a patch for handling project specific jsbeautifyrc and Martin Zagora merged it to the repo. Here is my .jsbeautifyrc for MediaWiki https://gist.github.com/santhoshtr/9867861

Brackets is in active development and I look forward for more features. The most important bug I would like to get fixed, that all code editors I tried suffer including brackets is support of pain free complex script editing and rendering. Brackers uses CodeMirror for the code editor and I had reported this issue . It is not trivial to fix and root cause is related to the core design. Along with js,css,html, php etc I have to work with files containing all kind of natural language text and this feature is important to me.