Bengali is the language people were shot for, and the only one whose dead are commemorated by a United Nations observance. That is the first half of this page. The second half follows the language east, along a dialect continuum that runs from Kolkata through Dhaka and Chittagong and does not stop at the Myanmar border, and asks a question the first half makes unavoidable: what happens to the far end of a continuum when the machinery that reads it was built for the near end. The layout sits in the keyboard gallery; the argument that grows out of the second module belongs to Codex Lexpressia.
The script that hangs from a line
Most writing systems sit on a baseline. Bengali hangs from one. The horizontal stroke running along the top of a word, the মাত্রা, is the first thing a reader notices and the last thing a typeface can get away with mishandling, because it is not decoration; it is the rail the letters are suspended from. Everything else about the script follows from a set of decisions taken between the ninth century and the middle of the nineteenth, and the last and most consequential of them were taken by two men cutting metal punches in a river town and a schoolmaster writing a primer.
It is an abugida rather than an alphabet. A consonant letter carries a vowel already, so ক is not k but ko; other vowels attach as diacritics that can sit before, after, above or below the consonant they follow in speech. The inherent vowel is অ, and it has two realisations selected by harmony and context rather than written down. Consonants meeting without a vowel between them fuse into conjuncts, of which there are several hundred, some transparently made of their parts and some not.
The descent runs from Brahmi through Siddham and then Gaudi. Bengali did not come from Devanagari and Devanagari did not come from Bengali; they are siblings that separated out of the same stock and were clearly distinct by the twelfth century, which is where Britannica puts the alphabet as nearly complete.
Unicode encodes the script at U+0980 to U+09FF and has done since the first version of the standard, in October 1991. The block is called Bengali, and the chart carries its own note about that: "The Bengali script is also known as Bangla. In Assam, the preferred name of the script is Asamiya or Assamese." Bengali and Assamese are the same writing system separated by two letters, ৰ and ৱ, and the block is named after one of them. It is the first instance on this page of a pattern the second module is entirely about: a technical decision, taken for sound administrative reasons, which quietly settles whose language a thing is.
Bengali had been printed before 1778, in Lisbon, in roman letters. What happened at Hooghly was different. Nathaniel Brassey Halhed needed a grammar of Bengali for the East India Company's servants, and a grammar of Bengali needs Bengali type. Charles Wilkins cut the face; Panchanan Karmakar, a blacksmith and engraver, cut the punches with him. Karmakar went on to Serampore, cut the type for William Carey's Bengali New Testament in 1801, and in 1803 produced the first Devanagari type cut in India. The printed shape of two of the subcontinent's major scripts was settled, in the first instance, by a village metalworker whose dates of birth and death are both approximate.
Type fixes shapes. It does not fix what counts as a letter, or what order the letters come in, and for that the date is 1855 and the book is Barnaparichay, published in two parts by Ishwar Chandra Vidyasagar. It is a children's primer. It settled the order of the alphabet and pared an inventory that had been carrying letters inherited from Sanskrit and used for nothing else, and every Bengali child since has learned the letters in the order a nineteenth-century reformer decided a child should learn them.
Through the same period Bengali was writing itself in two registers at once. সাধু ভাষা, the elegant style, is heavily Sanskritised and takes its verbs and pronouns in full; চলিত ভাষা, the current style, contracts them and reaches more readily for the inherited and the Perso-Arabic. The distance between them is a diglossia rather than a matter of taste, and for most of the nineteenth century only one of them was thought fit to write in.
The turn is usually dated to April 1914 and the first issue of সবুজ পত্র, edited by Pramatha Chaudhuri, which argued the colloquial case and printed in it; Tagore contributed steadily and moved to the colloquial himself within about a decade. But 1914 is the pivot and not the finish. Banglapedia supplies the cleaner end marker: since March 1965, many Bengali newspapers have adopted the current form and discarded the elegant one. Half a century, to stop writing in a register nobody spoke.
In March 1948, at the Race Course Maidan in Dhaka and again at Curzon Hall three days later, Muhammad Ali Jinnah told East Bengal that the state language of Pakistan would be Urdu and only Urdu. The province that heard him was the majority of the country's population and spoke Bengali. What followed took four years to come to a head.
On 21 February 1952, with section 144 in force and public assembly banned, students gathered at Dhaka University at eleven in the morning and marched toward the Assembly. Police fired. The government reported three dead; Bengali accounts name Rafiq Uddin Ahmed, Abdul Jabbar, Abul Barkat, Abdus Salam, and a nine-year-old boy, Ohiullah, on the day, and Shafiur Rahman and others the following morning. Students raised a monument on the spot overnight on the twenty-third and twenty-fourth. Police pulled it down on the twenty-sixth. The permanent Shaheed Minar was inaugurated on 21 February 1963, destroyed by the Pakistan Army in 1971, and rebuilt in 1973.
Bengali was recognised alongside Urdu in Pakistan's first constitution in 1956. Nineteen years later, on 17 November 1999, the UNESCO General Conference proclaimed 21 February International Mother Language Day, first observed in 2000 and welcomed by the General Assembly in 2002. Language protests are not rare and language killings are not unique to Bengali; there were eleven more of them at Silchar in 1961, over the same language on the other side of the border. What is unique is the observance. A crowd of students in Dhaka is the reason there is a day in the international calendar for everybody else's mother tongue.
Article 3 of the Constitution of Bangladesh reads, in its entirety: "The state language of the Republic is Bangla." Nine words, and no second sentence. Across the border, Bengali is a scheduled language of India under the Eighth Schedule and an official language of West Bengal, Tripura and the Barak Valley of Assam, but it is not an official language of the Indian Union, which has only Hindi and English. The Union added Bengali to its list of classical languages in October 2024.
The continuum, and where it stops being counted
Bengali is not one thing spoken in two countries. It is a continuum, and the useful way to picture it is as a line drawn from Kolkata through Dhaka, down through Sylhet and Chittagong, and on across the Naf River into Rakhine State, where the people speaking the last stretch of it are called Rohingya and are not, according to the state they live in, from there. Nothing on that line is a natural break. The breaks that exist on it were put there by administrators, and one of them was put there in 2007.
Chatterji, in 1926, grouped the dialects on the old regional divisions of Radha, Varendra, Kamarupa and Vanga, and observed that the differences between them are slight except for the south-eastern group, where they are not. Banglapedia's later classification gives four Bangladesh groups, with Chittagong and Noakhali and the Hill Tracts making up South Bengal. Britannica's summary is blunter: Sylhet and Chittagong are mostly unintelligible to other speakers.
The map that matters most here is the oldest. Grierson, in the Linguistic Survey of India of 1903, grouped the dialects of Chittagong under Southeastern Bengali alongside those of Noakhali and Akyab. Akyab is the colonial name for Sittwe, the capital of Rakhine State. The standard reference survey of the subcontinent put the Arakan varieties inside Bengali a century before anyone was killed over the question.
Against all that variety there is a single written standard, and both countries use it. It derives from the west-central speech of the Shantipur and Shilaidaha region, which is why it is called the Nadia dialect on one side of the border and the Kushtia standard on the other. Banglapedia puts the position precisely: although the written language of West Bengal and Bangladesh is more or less similar, spoken Bengali differs widely.
Where the two have drifted, the drift is lexical, and its main axis is communal rather than national. Water is পানি or জল; a bath is গোসল or স্নান; the first of each pair carries Perso-Arabic weight and the second Sanskrit, and the choice tracks the speaker's community at least as closely as their passport.
In November 2022 Localization Lab and Internews published a review of Facebook's and YouTube's content moderation policies as translated into Amharic, Arabic, Bengali and Hindi. Its Bengali findings are the most useful thing anyone has published on this subject, because they are about the document that tells a user what may be reported and how, which sits upstream of every classifier and every moderator.
The translations were machine-produced. "Immediate family members" came out as অবিলম্ব পরিবারের সদস্য, which the reviewers gloss as closer to "prompt" or "quick member of family", and which does not name any social category a Bengali speaker uses. "Marijuana" was transliterated as মারিজুয়ানা when the everyday words are গাঁজা and ভাং. And "places of worship" was rendered উপাসনাস্থল, which is a Hindu place of worship; the mosque, মসজিদ, does not appear.
The reviewers name the cause. The register throughout is Standard Central Bengali with terms specific to West Bengal, which suggests, in their words, a bias toward localising content for Bengali speakers in India. The document is read overwhelmingly in Bangladesh, where roughly nine people in ten are Muslim. Before any model has been trained or any moderator hired, the rules have been localised for the wrong Bengal, and everything downstream inherits it.
Meta added a Bengali hate speech classifier in 2020, two years after Hindi. An internal presentation disclosed in 2021 put the preceding position in one line: the lack of Hindi and Bengali classifiers means much of this content is never flagged. One classifier now covers the whole continuum above, and Meta has never published its accuracy in any variety of it, nor how many people it employs who read Bengali at all. Amnesty International asked in March 2026 and was told the request could not be answered inside the two weeks allowed; Bangladesh's own cyber security agency asked in December 2025 and got no figure either.
The assumption underneath a single classifier is that the continuum is close enough to itself to be treated as one language. That assumption has been tested, though not by Meta. In 2018 Translators without Borders surveyed more than four hundred Rohingya adults in the Cox's Bazar camps, where the humanitarian response had been built on the premise that Chittagonian was near enough to Rohingya to be understood. More than a third could not understand a simple sentence in it. The aid sector checked, because lives depended on the answer, and the premise broke at better than one in three.
Rohingya and Chittagonian shared one ISO code until 2007, when cit was retired and split into ctg and rhg. The separation is nineteen years old. It is an administrative act, not an ancient finding, and before it the international standards register said what Grierson had said: that the language spoken in northern Rakhine was a Bengali one.
That is also, word for word, what the Myanmar state says, and it says it in order to conclude that the speakers came from Bengal and are therefore not indigenous. Rohingya are not among the hundred and thirty-five recognised national races; the 1982 citizenship law makes indigeneity turn on presence before 1823, so the entire question of who these people are is conducted as an argument about a date. Calling the language Bengali is not a classification error. It is the argument, and the reference works have been supplying it with cover for a century.
What the far end of the continuum actually has is this. A script, Hanifi, devised around 1980, which waited until June 2018 for a Unicode block. No machine translation anywhere, and no place in the hundred and ten languages Google Translate added in 2024. No annotated hate speech dataset, in a catalogue that lists Bengali and Hindi and Urdu and twenty more. Burmese is not in that catalogue either, which is the detail that finishes the thought: the most notorious case of platform-enabled ethnic violence of the last decade produced no dataset in the victims' language and none in the perpetrators'.
Ramu, 2012: mobs destroyed twenty-four Buddhist temples and monasteries after an image of Quran desecration appeared on the timeline of an account bearing a Buddhist man's name. He had not posted it. Nasirnagar, 2016: a Hindu fisherman, Rasraj Das, was accused of posting a composite of Shiva on the Kaaba; a crowd of some three thousand vandalised nineteen temples. Forensic examination found no such image had been uploaded from his device and his account had been compromised. Gangachara, 2025: thousands looted homes after a post from a fake account carrying a Hindu student's name.
In each case the inciting post was fabricated or the account was not the accused person's. The mechanism is not that a member of the minority said something. It is that a post attributable to a member of the minority is sufficient, and the attribution does not have to be true. Nothing operating on the text of a post can detect that, in any language, at any accuracy, because the falsehood is not in the language.