3 ms·

Makes sense! I've actually had a similar feature request before: https://github.com/lindylearn/unclutter/issues/297 https://github.com/lindylearn/unclutter/iss

by phgn 4y ago

Makes sense! I've actually had a similar feature request before: https://github.com/lindylearn/unclutter/issues/297 https://github.com/lindylearn/unclutter/issues/297 Have you tried some of the existing solutions mentioned in the ticket? I'm curious if they solve the problem for you, and if not, what Unclutter could do better.

clnq 4y ago
I took time to try all four examples in the feature request. BlinkNotes didn't work for a technical reason, TLDR; it did not have AI capabilities - only crowdsourced summaries. TLDR This was good but seemed to use an NLP algorithm they call "AI." So instead of summarizing articles, it seems to produce a bullet list of the essential text fragments. This is in contrast to the much easier-to-read conversational style summaries GPT-3 can provide. TLDR This also did poorly with some of the technical articles I have authored, sometimes extracting pieces of C++ code as meaningful text. Summari worked very well for me. It summarized even quite technical articles well because it uses a conversational AI. Thanks for the recommendation. If I had to name one key criterion that differentiates good summaries from bad, in my opinion, it would be the effective use of language. Conversational/language models like GPT-3 are proficient at absorbing much contextual information and synthesizing a short and effective summary. NLP algorithms are good at throwing away superfluous context, which is common in casual writing, but they do not seem to work well for technical writing or texts whose purpose is to explain concepts and where there is little superfluous context to throw away. Perhaps for something like Unclutter, if the users mainly read news sites, then an NLP approach could be appropriate (it would be cheaper and works well for such content). But the ideal implementation for an article summarizer for me needs that summary-from-a-lot-of-context synthesizing capability.
- phgn 4y ago
  Thank you for doing the research! The big problem with current transformer-based models is the input limit. We'd probably need to split large articles and then summarize the summarizations of these chunks. And yeah I agree, Unclutter should work on any kind of article, not just on factual news. Please let me know if you're still using Summari in a bit! They seem to focus on their hosted version, I'm curious why.

by phgn 4y ago

clnq 4y ago

I took time to try all four examples in the feature request. BlinkNotes didn't work for a technical reason, TLDR; it did not have AI capabilities - only crowdsourced summaries. TLDR This was good but seemed to use an NLP algorithm they call "AI." So instead of summarizing articles, it seems to produce a bullet list of the essential text fragments. This is in contrast to the much easier-to-read conversational style summaries GPT-3 can provide. TLDR This also did poorly with some of the technical articles I have authored, sometimes extracting pieces of C++ code as meaningful text. Summari worked very well for me. It summarized even quite technical articles well because it uses a conversational AI. Thanks for the recommendation. If I had to name one key criterion that differentiates good summaries from bad, in my opinion, it would be the effective use of language. Conversational/language models like GPT-3 are proficient at absorbing much contextual information and synthesizing a short and effective summary. NLP algorithms are good at throwing away superfluous context, which is common in casual writing, but they do not seem to work well for technical writing or texts whose purpose is to explain concepts and where there is little superfluous context to throw away. Perhaps for something like Unclutter, if the users mainly read news sites, then an NLP approach could be appropriate (it would be cheaper and works well for such content). But the ideal implementation for an article summarizer for me needs that summary-from-a-lot-of-context synthesizing capability.

phgn 4y ago
Thank you for doing the research! The big problem with current transformer-based models is the input limit. We'd probably need to split large articles and then summarize the summarizations of these chunks. And yeah I agree, Unclutter should work on any kind of article, not just on factual news. Please let me know if you're still using Summari in a bit! They seem to focus on their hosted version, I'm curious why.