Show HN: Chrome extension that summarizes text content using DistilBART model | Hacker News Reader

Show HN: Chrome extension that summarizes text content using DistilBART model

chrome.google.com

64 pointsignatif24 comments

Highlight any text, double tap 'Enter' and get a summary of any text on web

Use it to summarize wikipedia articles, scientific papers, books or anything

(special credits to Sam Shleifer and Huggingface team for the model and infrastructure)

Are you experiencing any cost on your part for utilizing the Huggingface platform ? I see they have something like $25/1M char lookups so just wondering and am excited to see what you have done (possibly at your own expense!). I have been following Semantic Scholar [1] who are part of the Paul Allen AI for Good program and they are also working on NLP summarization -- definitely interesting things happening in this space! Thanks for your contribution.

[1] https://www.semanticscholar.org/

Yeah I'm currently on a free trial of startup plan :)

Dunno what I'm gonna do next. Interested in becoming an angel? :D

best tested at articles like this https://www.sciencedirect.com/topics/psychology/serotonin

Chrome 92.

Double tap "enter" does nothing. Feedback link on extension leads to nowhere. No description page on plugin. No instructions.

you might have it already mapped by another extension

> Beware: summarisations might not be accurate due to the nature of the model

I'd like more information about this before installing it.

Why are there no examples? The one example on the page is too small to read.

What personal information does the plugin collect?

Not trying to be negative. The plugin might be useful. It's difficult to make an assessment without this information.

Plugin collects users Google OpenID and number of characters used for summarization. No emails or other personal information is collected

Summarization request gets proxied to the Huggingface Inference API. According to https://huggingface.co/inference-api they claim to protect the data and not share it with third parties.

On examples - thank you for the comment, I'm thinking of recording a video and improving screenshots etc. Right now it's the earliest launch to hear people opinions

I'm not sure if the model can dismiss facts. From BART model paper:

"Because BART has an autoregressive decoder, it can be directly fine tuned for sequence generation tasks such as abstractive question answering and summarization. In both of these tasks, information is copied from the input but manipulated, which is closely related to the denoising pre-training objective."

"Information is copied but manipulated" - idk if it can lead to the factual mistakes and disinformation. Knowledgable opinions are best welcomed!

here's an example:

http://www.paulgraham.com/hwh.html

There are three ingredients in great work: natural ability, practice, and effort . To do the best work you need all three: you need great natural ability and to have practiced a lot and to be trying very hard . The most basic level of which is simply to feel you should be working without anyone telling you to .

Thanks!

Pretty good. I remember reading that article ang finding it difficult to get through. He needs an editor.

I love the idea! How do you get it to work though? I'm using Chrome 92.0.4515.159, I've disabled all other extensions, and highlighting text and double tap enter does nothing. When I right click on the text there's no context menu option, which is what I would expect.

Can you do the same for Firefox?

I tried it on general news and the outputs were quite good actually.

I encourage you to edit this title (or ask the mods to do so, if it won't let you, just write them at the footer Contact link). It doesn't say what it does, and has a lot of capital letters, and is missing the specific keyword for what you're doing: Summarizing. Consider instead "Show HN: Chrome extension that summarizes page content using the DistilBART model" instead. You could try to wedge TL;DR in there as "Tl;dr" but the product name is honestly not a high priority when I'm considering if something is curious/interesting.

thank you

How large of a model download would be required to remove the dependency on a cloud service?

No more than 1GB, and inference would be dependent on the speed but no more than 10 seconds. Completely reasonable as a standalone program.

Would that require running a separate server locally? That's all kinds of non-techie hostile :(

Is there a guide on running it locally? I am currently using some classic algorithms to summarize my RSS feeds, but they aren’t very good.

Sure! Here's the weights https://huggingface.co/sshleifer/distilbart-cnn-12-6 , here's the model https://huggingface.co/transformers/model_doc/bart.html?#tra... . Uses standard pytorch loading, so write your python program around that and it should be fine.

Pretty sure it’s DistilBERT, not DistilBART: https://huggingface.co/transformers/model_doc/distilbert.htm...

Not sure why it doesn't summarize the entire article

This is awesome!

Reply on news.ycombinator.com