English to Bengali, and Bengali to English
Both directions work the same way. A Word file is edited where it stands: every paragraph style, table, list, header, footer and page setting stays the file’s own and only the words change, so a Bengali circular comes back as an English circular, and an English notice as a Bengali one, with its centred heading, numbered list and bordered table where they were. A PDF with selectable text is first turned into a Word file and then translated, so the result is a Word file you can edit, and how closely it follows the PDF is the PDF-to-Word conversion’s doing.
Translated Bengali is set in Noto Sans Bengali, as real Unicode text you can search and edit; if that font is not on your computer, choose any Bengali font in Word. Numbers, dates, section references, amounts, times of day, e-mail addresses and web addresses are taken out before a sentence is sent and put back afterwards, and an answer in which one does not come back exactly once is not accepted. A sentence that still cannot be translated stays in the source language and is highlighted yellow; it is never dropped, and never filled in with a guess.
What to read carefully in the Bengali
The address was kept exactly, but the word in front of it was not right: the translation says to write to help@example.gov.in নম্বরে, “at the number”, where Bengali would say ঠিকানায়, “at the address”. The machine sees a protected item, not an e-mail address, and guessed what kind of thing it was. Read the words around each address, phone number and reference.
The fee became 250 টাকা. টাকা is how Bengali in India writes rupees, so the amount is unchanged; read back into English it was “250 rupees”.
What was measured for Bengali
Before Bengali was switched on, the production pipeline translated the test notice from English into Bengali and back. In both directions 6 of 6 protected items came back exactly: the section reference, the year, the date, the amount, the e-mail address and the web address. The Bengali side was written in the Bengali script, and the English read back close to the original. A Word notice with a heading, a centred line with a bold date, a numbered list, a table, a tab, a header and a footer was also translated into Bengali and back, and its paragraph settings, table grid, page setup, header and footer were identical to the source, on the same number of pages.
That is a check of the machinery, not of the language. It says the tool kept your numbers and your layout; it does not say every sentence reads the way a Bengali writer would put it.
Bengali in a scan or an old font
A scanned PDF has no text to translate, so it is refused rather than guessed at. Read it first with Bengali OCR, correct what needs correcting, and translate the Word file that comes out. Bengali set in a Bijoy font such as SutonnyMJ pastes as Latin letters; convert it to Unicode with the legacy converter first, then translate.
Footnotes, endnotes and comments are not translated yet, and the result says so. Text inside a picture is not translated. Bengali OCR reads scans of printed Bengali, and its Word file can be translated straight away.