لوحة مفاتيح عربية
Ergonomic Arabic Keyboard Layout
This is work in progress and contributions are welcome. Head over to GitHub to see where you can help.
Other languages using the Arabic alphabet (regional dialects, Urdu, Persian) are explicitly not supported.
xmodmap ar-lulua.xmodmap%localappdata%\klavaro (Windows) or
~/.config/klavaro/ (Linux), select Keyboard →
(Custom), then lulua_ar.28 letters make up the Arabic alphabet and quite a few extra symbols are required for proper text input, like the hamzah (ء) in its different shapes أ إ آ ء ئ ؤ, tāᵓ marbūṭah (ة), ᵓalif maqṣūrah (ى) and various diacritics for vowelized texts. Since the performance of a keyboard layout depends on the text entered it is necessary to study its mono-, di- and trigraph frequencies first. The novel corpus built for the following analysis consists of
| مصدر | كلام | حروف | |
|---|---|---|---|
| الجزيرة | ٥٤٧٬١١٠ مقالة | ١٨٩٫٤ مليون | ١٫٢ مليار |
| ويكيبيديا العربية | ٨٥٧٬٣٨٦ مقالة | ١١٠٫٨ مليون | ٧٠٢٫٢ مليون |
| بي بي سي العربية | ١٤٩٬٩٠١ مقالة | ٣٦٫٩ مليون | ٢٣٣٫٣ مليون |
| هنداوي | ٢٬٢٢٧ كتاب | ٩٠٫٧ مليون | ٦٦٥٫٠ مليون |
| خريطة الشارع المفتوحة Arabic Labels | ٣٧٦٬١٤٨ labels | ٩١٠٫١ ألف | ٥٫٥ مليون |
| ORPUS OpenSubtitles 2018 | ٩٤٬٠٩٣ فيلم | ٣٦٣٫٩ مليون | ٢٫١ مليار |
| القرآن (من tanzil.net) | ٧٨٫٢ ألف | ٧٠٣٫٨ ألف | |
| الأمم المتحدة Parallel Corpus v1.0 | ١١٦٬٧٥٤ ملف | ٤٢٤٫٣ مليون | ٢٫٩ مليار |
| مجموع | ١٫٢ مليار | ٧٫٨ مليار |
| Source | Words | Characters | |
|---|---|---|---|
| Al-Jazeera | 547,110 articles | 189.4 million | 1.2 billion |
| Arabic Wikipedia | 857,386 articles | 110.8 million | 702.2 million |
| BBC Arabic | 149,901 articles | 36.9 million | 233.3 million |
| hindawi.org | 2,227 books | 90.7 million | 665.0 million |
| OpenStreetMap Arabic Labels | 376,148 labels | 910.1 thousand | 5.5 million |
| ORPUS OpenSubtitles 2018 | 94,093 movies | 363.9 million | 2.1 billion |
| tanzil.net Quran | 78.2 thousand | 703.8 thousand | |
| United Nations Parallel Corpus v1.0 | 116,754 documents | 424.3 million | 2.9 billion |
| Total | 1.2 billion | 7.8 billion |
The chosen Quran representation does not include all quranic diacritization symbols, like other datasets. This makes comparison fairer, since most keyboards presented below do not include any of them.
The plot below shows ᵓalif (ا), lām (ل), yāᵓ (ي), mīm (م), wāw (و) and nūn (ن) can be considered the most frequently used letters in the Arabic language. Together they account for more than 55% of all letters in the corpus.
Arabic letter frequency distribution
Throughout this evaluation color coding is used to identify fingers:
Asymmetry is defined as the difference between left and right hand button usage \(b_{left/right}\) and includes the thumb:
$$a = \frac{b_{left}}{b_{total}} - \frac{b_{right}}{b_{total}}$$The layout proposed uses four shift layers in a way inspired by Neo2. Thus it assumes a 102/105 key ISO keyboard common in Europe – but also available in Arab countries – to accommodate for the necessary shift keys. These are in order: Shift on the left and right, caps lock on the left and the rightmost key in the middle row, the key right to the left shift key and the key labeled Alt Gr to the right of the spacebar. Symbols are assigned to the four layers by their function: characters, punctuation, diacritics, other.
Apple, for instance, provides an Arabic hardware keyboard with this physical layout. But both variants, 101/104 key and 102/105 key devices, seem to exist in the Arab world.
The first layer was optimized using an extended reimplementation of carpalx. From several runs with 100.000 iterations each the layout which had good scores and looked reasonable to the human eye was picked. Afterwards the third layer was optimized using the same process, but only using data from the Hindawi corpus, because it is the only one with at least some fully diacriticised texts. Finally the different brackets were arranged by hand and the remaining symbols algorithmically distributed on the second layer using the raw Wikitext from the Arabic Wikipedia dataset.
This is a common way of arranging brackets, because most algorithms ignore human desire for symmetry.
The most frequent letters have all been assigned to the home row, which makes them easily accessible. ᵓAlif (ا) and lām (ل) are typed with different hands, balancing the load on hands almost evenly. The index and middle finger of both hands share the majority of the typing load, but naturally the left middle finger is used more frequently due to its assignment to the letter ᵓalif (ا).
The layout targets Quaranic and Modern Standard Arabic (MSA), also called Fuṣḥa (الفصحى), only. Dialectical Arabic (العامية) is mainly a spoken language, although with the rise of social media sites like Twitter and Facebook this is changing. For now however it’s not an optimization target due to the lack of a good, representative corpus.
Designing the layout to be compose-based has both benefits and disadvantages. Compose-based mainly means the hamzah (ء) is treated like an optional diacritic for ᵓalif (ا), wāw (و) and yāᵓ (ي) instead of viewing ᵓalif-hamzah (أ), wāw-hamzah (ؤ) and yāᵓ-hamzah (ئ) as precombined, atomic units. Although ᵓalif-hamzah and ᵓalif are not the same, the hamzah can be dropped if the writer’s intention is unambigiously inferable from context. Thus it makes sense to provide hamzah as a combining character on the keyboard. Additionally it uses two keys less than precombining it with its stems, allowing the entire alphabet plus hamzah diacritic to fit on a single keyboard layer. However, there is a cost to this approach: All hamzah variants account for 3.0% of button combinations. Splitting hamzah and from its stem means doubling the total number of button combinations and thus button presses, decreasing scores like words per minute (WPM) slightly. Splitting ᵓalif (ا) and ᵓalif-hamzah (أ) could also reduce pressure on left middle finger and allow for more even distribution, since 1⁄5th of all ᵓalif (ا) uses are with hamzah (ء).
See for example section 3.3 of Buckwalter’s Issues in Arabic Morphological Analysis.
This section explores existing keyboard layouts made for the Arabic language and analyzes their usability. Comparing them with the proposed layout above is difficult at best, because the layouts presented below cover different character sets. Some lack numbers, some do not include short vowels and others provide no way to type symbols. Therefore no individual score is assigned to each layout, but an analysis of each layout’s features is given.
Carpalx tries to minimize the effort of typing in blocks of three consecutive keystrokes, triads, and thus a good layout in that sense should make typing frequent triads easy. The figure above plots cumulative triad frequency on the x-axis and weighted cumulative effort on the y-axis. Thus, at an x value of 0.5 the y-axis is the sum of triad frequencies multiplied by their effort for all triads responsible for 50% of the typing process. Standard layouts are the layouts from ASMO, Linux and OSX whereas usable lists only those which are actually relevant for typing. As we can see the layout presented above meets the optimization goal. Only the top 5% of all triads are “easier” to type with Malas’ layout, because lulua splits hamzah (ء) from its ᵓalif (ا) stem. As expected the phonetic layout is one of the worst ones, because QWERTY is not optimized for Arabic letter frequencies.
The following sections provide details about these layouts.
Trying to unify existing layouts, the Arab Standardization and Meterology Organization (ASMO), now part of AIDMO, published an Arabic keyboard layout in 1987 as standard 663 titled Arabic terminal keyboard layout. The layout has several shortcomings: Most notably it clusters letters by their visual similarity (ض ص، س ش، ح ج خ) and not frequency. Also it overuses the right index finger by assigning the four high-frequency letters ᵓalif (ا), taᵓ (ت), wāw (و) and tāᵓ marbūṭah (ة) to it.
Mac OS X’s Arabic keyboard layout makes a few small changes to ASMO 663 by moving the tāᵓ marbūṭah (ة) to a hard to reach spot on the right of the top row. It also moves the short vowels from the first to the top row of the second layer and replaces them with symbols. The bottom row keys are aditionally shifted to the right, beginning with raᵓ (ر).
A more common layout is the one used on Linux, which also exists on Windows with minor changes to the first layer. While its top and center row barely differ from ASMO 663 the bottom row now contains a separate key for the ligature lām-ᵓalif (لا) , likely inherited from early typewriter layouts. But at the cost of pushing punctuation characters to the second layer, dāl (د) into the top and ḏāl (ذ) even further into the number row.
In contrast to the layouts presented so far the following layouts claim to be optimized for the Arabic language and were, at least partially, created algorithmically.
One of the eariest accounts can be found in the article Statistical Distribution of Arabic Letters Aids to the Design of a New Keyboard by Al-Ramly et al published in 1980. The Arabic half of the proposed bi-lingual layout seems to be hand-optimized based on several metrics including character frequencies – without mentioning a source for them though – visual similarity (“letter groups”) and their position on previous layouts. It tries to balance load between hands, assign more work to index and middle fingers and place common letters in the home row. However the asymmetry given in the article, 0.032, cannot be reproduced here. For the most part the layout lacks combining and pre-combined characters, a task that is left to “machine intelligence” making it hard to use nowadays.
While the layout distributes load between fingers quite well it favors the left hand by assigning ᵓalif (ا) and lām (ل) to it. The decision to place ṯaᵓ (ث) in a very prominent spot seems weird, given it only accounts for 0.5% of all symbols, even in their own analysis.
Another article from the early days of computers published in 1990 is Design of Arabic Keyboard Layout Based on Statistical Properties of Arabic Characters by Idlebi et al. They present two examples of programmatically optimized layouts and account for character and bigram frequencies based on a corpus of 100.000 characters, finger movement time of unknown origin and finger load.
Unfortunately the results use 12 keys per row and are not suitable for use with current European keyboards, which usually feature only 11 keys in the bottom row. Thus the layout displayed below lacks the Arabic question mark and comma in the bottom right. Probably due to their unusual assumption that middle- and ring-finger rest in the top row their results are suboptimal, placing both ᵓalif (ا) and yāᵓ (ي) in the top row. Their analysis notices this and suggests improved positions for both characters, but these are not actually implemented. The big asymmetry is caused by placing ᵓalif (ا), lām (ل), yāᵓ (ي) and wāw (و), four of the five most frequent letters, on the right hand side.
About 20 years later (2008) Malas et al. presented an alternative layout generated by a genetic algorithm in their article Toward Optimal Arabic Keyboard Layout Using Genetic Algorithm. They used a snapshot of the Arabic Wikipedia probably from around 2008 and optimized for typing speed only, claiming 35% faster typing compared to the currently used layouts. However the decision to put yāᵓ (ي) in the top row seems odd. Assigning the same left index finger to ᵓalif (ا), yāᵓ (ي) and wāw (و), which are three of the most frequent letters, heavily strains this particular finger.
In figure 8 of the paper A new optimal Arabic keyboard layout using genetic algorithm published in 2010 Khorshid et al. present yet another layout. They reject the idea of using contemporary text sources like newspapers or books and consequently their only input text is a copy of the Quran, whose distribution of mono- and digraphs is quite different from modern, non-religious texts. A 36% improvement over the standard keyboard based on their criteria for ergonomic layouts is claimed in their conclusion as well. However their algorithm seems to favor the bottom row instead of the easier to use top row since it places the letters baᵓ (ب), taᵓ (ت) and raᵓ (ر) there.
In 2015 patent 9,041,657 B2 was filed in the US, presenting yet another computer-generated layout. Its genetic algorithm was seeded with just 54 Arabic e-books consisting of 7 million characters in total. Overall it claims to be 9% faster than default layouts. This layout rips off most of the standard layout’s second layer, but amusingly fails to include a question mark, while it does provide three single-quote marks ’ and two Arabic semicolon ؛. Additionally it places yāᵓ (ي) in an even worse position than Malas’ layout.
Qtaish et al presented this layout in An Improved Arabic Keyboard Layout in 2021. They use a novel corpus of 5 million words or roughly 66 million letters consisting of newspapers, (now defunct) social networks and blogs, as well as six dictionaries (see remarks). Then letters were classified into three categories based on their frequency, which apparently were used to populate home, top and bottom row (in this order). Additionally bigrams were somehow used to arrange letters and avoid placing them on the same or adjacent fingers, making rolling finger movements incentiviced by carpalx impossible. Ultimately it looks like the layout was designed by hand and not through an automated process.
Although the authors do not provide a number row, it has been added for fair comparison. However to be actually usable the layout would need punctuation symbols and diacritics.
Being first released in 2021 Ergoarabic is the most recent optimized layout for Arabic. On the first layer it combines the positions of 17 keys from the standard PC layout with brackets and punctuation symbols from QWERTY, aiming to retain compatibility with both. The remaining keys have been re-arranged or moved to the shift layer by hand for improved ergonomics.
The Arabic Phonetic Keyboard simply maps the QWERTY layout to Arabic letters, based on their sound. Thus Q becomes qāf (ق), Y becomes yāᵓ (ي) and so on. It claims to be optimized for writing vowelized texts, especially Quranic Arabic, and thus includes quite a few combining characters and special symbols. Although it claims to make frequently used letters easily available – based on the work of Intellaren – it makes no effort to arrange letters according to their usage frequency.
While technically speaking not a layout but alternative input method, Intellark by Intellaren is worth mentioning. It is based on repeatedly pressing the same button to modifiy the current character. For example pressing A on the QWERTY keyboard cycles through the alternatives ا أ إ آ and ء. Obviously this is slow, error-prone and violates Dvorak’s guidelines for keyboard layout designs.