🧐 Ah, the timeless debate of Byte Pair Encoding vs. #Unigram 🤔—because nothing screams "cutting-edge linguistics" like retrofitting the English language into a robotic token soup. 🤖 Nick hoped to revolutionize quarantine productivity, but instead, we get a blog post beating a dead token horse. 🐴📉
https://ndingwall.github.io/blog/tokenization #BytePairEncoding #Linguistics #TechDebate #QuarantineProductivity #Tokenization #HackerNews #ngated
Tokenization for language modeling: Byte Pair Encoding vs Unigram Language Modeling

Tokenizers used by the best-performing language models (Bert, GPT-2, etc.) poorly reflect the morphology of English text. I had hoped to use some quarantine time to design one that more closely aligns to relationships between wordforms. But Kaj Bostrom and Greg Durrett beat me to it and so this blog post materialized instead. I add some additional motivation, evaluate both methods against ‘gold standard’ tokenizations, and speculate about what might come next.

Nick Dingwall

Tokenization for language modeling: BPE vs. Unigram Language Modeling (2020)

https://ndingwall.github.io/blog/tokenization

#HackerNews #Tokenization #LanguageModeling #BPE #Unigram #NLP

Tokenization for language modeling: Byte Pair Encoding vs Unigram Language Modeling

Tokenizers used by the best-performing language models (Bert, GPT-2, etc.) poorly reflect the morphology of English text. I had hoped to use some quarantine time to design one that more closely aligns to relationships between wordforms. But Kaj Bostrom and Greg Durrett beat me to it and so this blog post materialized instead. I add some additional motivation, evaluate both methods against ‘gold standard’ tokenizations, and speculate about what might come next.

Nick Dingwall
@freedomscientific @mastoblind Finally, even if you're not impacted by this bug, please feel free to boost this thread to ensure greater visibility throughout the community. I--and many other #JAWS users--would greatly appreciate it. Thank you. #Unigram #Telegram #WhatsApp #Blind #VisuallyImpaired #LowVision #BlindMasto #BlindMastodon #BlindFedi
@freedomscientific @mastoblind To be clear, #JAWS is configured to report spelling errors by default. Additionally, #NVDASR and Narrator correctly report spelling errors in these applications. Therefore, I am certain that this is a bug in JAWS itself. #Unigram #Telegram #WhatsApp #Blind #VisuallyImpaired #LowVision #BlindMasto #BlindMastodon #BlindFedi
Hello, #JAWS users. If you use #Unigram for #Telegram or #WhatsApp and have noticed that spelling errors are not reported by JAWS in these applications, please feel free to contact @freedomscientific and request a fix. I reported this bug over a year and 9 months ago after discovering it in Unigram, and it has since appeared in the WhatsApp UWP application for which Freedom Scientific has written JAWS scripts. #Blind #VisuallyImpaired #LowVision #BlindMasto #BlindMastodon #BlindFedi @mastoblind
Stories for Channels, Your Music in Stories and Much More

Telegram for Android and iOS have been updated to version 10.1.

⚠️ Other Telegram clients will receive the new version soon. The update will likely be rolled out to users gradually.

Stories from channels
• Now Telegram users can give (https://t.me/betainfoen/1419) their favorite channels the opportunity to publish stories.
• Telegram Premium subscription provides one boost, which you can give to any channel.
• The more boosts a channel has, the higher its level, and each new level allows you to publish 1 additional story per day for all subscribers.
• Users can transfer their boost to another channel once a day.
• Channels can collect boosts using special links like t.me/tginfoen?boost.
• If a channel receives the required number of boosts, administrators with the appropriate rights (https://t.me/betainfoen/1423) and an active Premium subscription will be able to publish a story on its behalf by clicking on the create story icon in the channel profile.
• You can find a link to boost for your channel and see how many boosts are left until the next level on the Channel page > More > Statistics > Boosts.

Reaction stickers in stories
• Users and channels can add reaction (https://t.me/betainfoen/1421) stickers to their stories, which allow them to send a variety of emoji with one tap.
• In stories from channels, such stickers display the number of users who have chosen a particular emotion.
• You can add 1 reaction sticker to a story. Premium users have access to more reaction stickers - up to 5 at a time.
• To add a reaction sticker, just click on the 💭 icon in the sticker panel and select the appropriate emoji. Premium subscribers also have access to any emoji from the author's sets.

Stories with your music
• Stories now support loading audio (https://t.me/betainfoen/1422) files from the device's memory - so you can add any music or voice accompaniment to a photo or video.
• To do this, just click on “Sound” in the sticker panel and select a file, and then select the desired fragment.
• In video stories, you can leave the original audio track and choose any moment for a new one to appear - for example, insert a sound effect in the middle or add music at the end.

View media files once
• The Telegram team has updated the interface and added a one-time media viewing feature.
• In any personal chat, just click on the one-time viewing icon in the photo editor to select the period during which the photo or video file will be available - from one-time viewing to 30 seconds.
• If you select the “One view” option, the media file will be deleted from the chat immediately after the end of viewing.
• If View Once mode is enabled, the recipient will not be able to save the file or take a screenshot.

New login notifications
• Now such notifications will appear not in the Telegram service dialog, but above the list of chats - so that you definitely notice them.
• If you were not the one who logged into your account, select “No, not me!” to secure your account (the new session will be terminated immediately).
• You can check the list of all devices that currently have access to your messages in Settings > Devices.

Other innovations
• Sending video recordings has been accelerated (on iOS).
• The loading bar more accurately displays the submission status (on iOS).
• Devices (Android) will now do a better job of displaying animations on hidden text and hidden media.
• The button for adding additional photos in the story editor is placed in a separate “Photo” sticker.

Article: telegram.org/blog/channel-stories/

Android: Google Play (http://play.google.com/store/apps/details?id=org.telegram.messenger), APK from the official website (http://telegram.org/dl/android/apk) or verified channel (http://t.me/TAndroidAPK).
iOS: App Store (http://itunes.apple.com/ru/app/telegram-messenger/id686449807?mt=8).
Desktop: from the official website (https://desktop.telegram.org/) or from GitHub (https://github.com/telegramdesktop/tdesktop/releases).
macOS: App Store, from the official website (https://macos.telegram.org/) or from a verified channel (https://t.me/macos_stable_updates_files).
Unigram for Windows 10+: Microsoft Store (https://www.microsoft.com/store/productId/9N97ZCKPD60Q) | Direct download (https://t.me/unigramappx) | Installation instructions (https://t.me/betainfo/651)

#update #Android #iOS #Desktop #macOS #Unigram

#telegram #update
Beta Info English

⭐ Boost System Now in Telegram In the Beta version of Telegram for Android, Premium subscribers can now allocate their boost to one of the channels they are subscribed to. For a channel to have the ability to post stories, it must reach level 1. To achieve this, the channel needs to accumulate a certain number of boosts from its readers, approximately 1/250 (0.4%) of all channel subscribers. This coefficient may change in the future. It's likely that boosts given by subscribers to a channel will not be permanent but rather valid for a limited period. Information about users and the duration of their boosts will be displayed in a special tab within the channel's statistics section. There, you can also find out what percentage of your total subscribers are Premium users. For more details on how the boost system in Telegram channels might work, you can read this article. ⚠️ Currently, the boost mechanism is not fully operational: boosts that users give to their favorite Telegram channels are not yet counted by the messenger. #Android

Telegram
Okay, I keep getting added to spam groups without permission on #Telegram and don't know how to actually leave said groups. Blind folks in particular, can I use #tweesecake to do this, or do I have to use #Unigram?
#Unigram users. If you are having the bug where the emoji and stickers pannel is stuck open, press control windows down arrow in the unigram chat window. As just accidentally discovered by me, this somehow removes the pannel from the window and stops it from automatically showing
@nortix Nein! Ich lasse das auch sofort wieder. Weder weiß ich was ein UWP-Device ist, noch finde ich unter uwpx.org/ einen Link, um das Teil zu downladen. Ein Grund mehr, mal wieder von Telegram und der WindowsApp #Unigram zu schwärmen. Open Source und #Barrierefrei. Ja, das gibts tatsächlich, und man muss kein Nerd sein, um das zu nutzen. ;-)
Manche hier behaupten ja, es läge an mangelnder Zeit und vor Allem an mangelndem Geld, dass so viele Apps etwa für #Mastodon #Matrix oder #XMPP so gar nicht #Barrierefrei sind. Das Beispiel der App #Unigram für #Telegram beweist, dass das nicht stimmt. Eine App aus der Community und so weit ich weiß sogar open Source! Es geht also, wenn man nur den Willen dazu hat.