3 ms·

It's an interesting idea. I'm not sure if that information is usually structured enough to distill objective facts and criticisms. Are we talking about using NL

by jameslk 10y ago

It's an interesting idea. I'm not sure if that information is usually structured enough to distill objective facts and criticisms. Are we talking about using NLP to pull out those details? I think the tech might not be quite there yet, but maybe there's a better way that I'm not familiar with.

Mathnerd314 10y ago
Yeah, NLP. I was thinking of a version of http://newsblaster.cs.columbia.edu/ http://newsblaster.cs.columbia.edu/ (now open-source: https://github.com/kedz/newsblaster https://github.com/kedz/newsblaster). That code doesn't use sentiment analysis or word/phrase vectors so there's significant room for improvement by borrowing ideas from other research. Of course, the amount of improvement is uncertain, so I'm not sure if it translates into a startup per se. It's probably lower-risk to target a specific segment and do in-depth reviews, e.g. https://examine.com/about/ https://examine.com/about/ is apparently quite profitable (http://www.solorwell.com/examine-com-is-grown-up-700000-in-the-past-year/ http://www.solorwell.com/examine-com-is-grown-up-700000-in-t...), and Grokstyle is doing interior design (http://www.ldv.co/blog/2016/5/30/grokstyle-wins-ldv-vision-summit-2016 http://www.ldv.co/blog/2016/5/30/grokstyle-wins-ldv-vision-s... https://www.cs.cornell.edu/~sbell/pdf/siggraph2015-bell-bala.pdf https://www.cs.cornell.edu/~sbell/pdf/siggraph2015-bell-bala...). But that sort of site will never cover everything, or even a significant fraction of products, so there's lots of room for value-added reviews. I envision a site like https://reviewmeta.com/ https://reviewmeta.com/, but with more emphasis on information extraction (& actual reviews) instead of machine learning (ML metrics don't help much when comparing products; to really corner the market the flow needs to be review site -> Amazon rather than the reverse).
- Mathnerd314 10y ago
  For an concrete example let's take garlic peelers (a typical "long-tail" product). The top search results are e-commerce sites, affiliate reviews, Cook's Illustrated, WikiHow, a 1996 NYT article describing its invention, and a Kickstarter campaign to reinvent the garlic peeler. Finally somewhere on page 2-3 is this review site: https://moo.review/garlic-peelers/ https://moo.review/garlic-peelers/. They clearly spent some effort on this; it's all original content, nice pictures, good layout, etc. But as a person buying a garlic peeler, it's pretty much unusable due to the lack of editing; spelling errors such as "realy", the majority of the page taken up by a how-to, and tons of fluff such as "Some of the peelers presented more problems than they solved." There's about one paragraph per item. Meanwhile, Amazon reviews are composed mostly of worse fluff ("amazing", "excellent", "I use it every day", etc.), and there's more of it. But there is also the occasional interesting complaint such as "The interior has to be cleaned after 2-3 uses due to garlic juice." and "The rolling motion gave my friend arthritis". Definitely more than a paragraph of info, maybe even two or three. So the info is there, it's just a problem of extracting it.

by jameslk 10y ago

Mathnerd314 10y ago

Yeah, NLP. I was thinking of a version of http://newsblaster.cs.columbia.edu/ http://newsblaster.cs.columbia.edu/ (now open-source: https://github.com/kedz/newsblaster https://github.com/kedz/newsblaster). That code doesn't use sentiment analysis or word/phrase vectors so there's significant room for improvement by borrowing ideas from other research. Of course, the amount of improvement is uncertain, so I'm not sure if it translates into a startup per se. It's probably lower-risk to target a specific segment and do in-depth reviews, e.g. https://examine.com/about/ https://examine.com/about/ is apparently quite profitable (http://www.solorwell.com/examine-com-is-grown-up-700000-in-the-past-year/ http://www.solorwell.com/examine-com-is-grown-up-700000-in-t...), and Grokstyle is doing interior design (http://www.ldv.co/blog/2016/5/30/grokstyle-wins-ldv-vision-summit-2016 http://www.ldv.co/blog/2016/5/30/grokstyle-wins-ldv-vision-s... https://www.cs.cornell.edu/~sbell/pdf/siggraph2015-bell-bala.pdf https://www.cs.cornell.edu/~sbell/pdf/siggraph2015-bell-bala...). But that sort of site will never cover everything, or even a significant fraction of products, so there's lots of room for value-added reviews. I envision a site like https://reviewmeta.com/ https://reviewmeta.com/, but with more emphasis on information extraction (& actual reviews) instead of machine learning (ML metrics don't help much when comparing products; to really corner the market the flow needs to be review site -> Amazon rather than the reverse).

Mathnerd314 10y ago
For an concrete example let's take garlic peelers (a typical "long-tail" product). The top search results are e-commerce sites, affiliate reviews, Cook's Illustrated, WikiHow, a 1996 NYT article describing its invention, and a Kickstarter campaign to reinvent the garlic peeler. Finally somewhere on page 2-3 is this review site: https://moo.review/garlic-peelers/ https://moo.review/garlic-peelers/. They clearly spent some effort on this; it's all original content, nice pictures, good layout, etc. But as a person buying a garlic peeler, it's pretty much unusable due to the lack of editing; spelling errors such as "realy", the majority of the page taken up by a how-to, and tons of fluff such as "Some of the peelers presented more problems than they solved." There's about one paragraph per item. Meanwhile, Amazon reviews are composed mostly of worse fluff ("amazing", "excellent", "I use it every day", etc.), and there's more of it. But there is also the occasional interesting complaint such as "The interior has to be cleaned after 2-3 uses due to garlic juice." and "The rolling motion gave my friend arthritis". Definitely more than a paragraph of info, maybe even two or three. So the info is there, it's just a problem of extracting it.