Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Show HN: Testing HN titles against a neural network

github.com

Post reply on HN
Print view View original

201–210 of 225 posts

Re: Show HN: Testing HN titles against a neural network

#201

Post by thaumasiotes »

lm28469 wrote: post #161
thaumasiotes wrote: post #123

Earlier quoted context omitted.

> I'd argue that titles, as long as they're neutral and accurate, have nothing to do with the upvotes you'll get. This is nonsense. If the title is accurate, then it is closely related to the content, and the content has a lot to do with the upvotes you get.

Well yeah, that's exactly what I said. Gaming the system by tweaking your title according to a random github "neural network" won't help you if your content is shit, and if it's not shit the title will be good enough.

You're assuming that the use case for this tool is "I'm submitting an article; what should I title it?"

Don't overlook "Should I submit this article?"

Re: Show HN: Testing HN titles against a neural network

#202

Post by DoctorOetker »

mehdix wrote: post #164
trulyrandom wrote: post #160

Earlier quoted context omitted.

"I plan to rewrite Linux in C# - Linus Torvalds" Bad: 0.1897 - Good: 0.7988 Looks like we have a winner.

Clojure scores exactly the same.

as does

in assembly, in VHDL, in brainfuck, in America, in McDonalds, in clickbait

but "in legalese" scores surprisingly high

Re: Show HN: Testing HN titles against a neural network

#203

Post by wodenokoto »

GistNoesis wrote: post #154

Congratulations on getting first rank on front page. Congratulations on getting your hands dirty and doing everything yourself like computing gradients manually, badly shuffling (non Fisher-Yates), badly js transpose (double swapping), it is a great way to learn. Congratulations on completing a full pipeline, that's the hard part then it just swapping pieces for better pieces. I advise non-technical readers not to at…

This really puts the work in perspective. Thank you for the summary!

Re: Show HN: Testing HN titles against a neural network

#204

Post by DoctorOetker »

brudgers wrote: post #173
michjedi wrote: post #131

Current top 10: 1. "Apple introduces 16-inch MacBook Pro, the world’s best pro notebook" Bad: 0.9964 - Good: 0.0038 2. "Developing open-source FPGA tools" Bad: 0.3381 - Good: 0.6652 3. "Show HN: Can a neural network predict if your HN post title will get up votes?" Bad: 0.0598 - Good: 0.9307 4. "How internet ads work" Bad: 1.0000 - Good: 0.0000 5. "More Intel speculative execution vulnerabilities" Bad: 0.7413 - Good:…

new mac seven characters, 0.0060 bad, 0.9940 good

"I know"

6 characters Bad: 0.0085 - Good: 0.9915

EDIT: find a higher score than yours at one character less

"I went"

6 characters Bad: 0.0025 - Good: 0.9975

for 5 characters:

"I won" Bad: 0.0031 - Good: 0.9976

===

at 4 characters:

"J ML" Bad: 0.0002 - Good: 0.9998

Re: Show HN: Testing HN titles against a neural network

#205

Post by segfaultbuserr »

michjedi wrote: post #131

Current top 10: 1. "Apple introduces 16-inch MacBook Pro, the world’s best pro notebook" Bad: 0.9964 - Good: 0.0038 2. "Developing open-source FPGA tools" Bad: 0.3381 - Good: 0.6652 3. "Show HN: Can a neural network predict if your HN post title will get up votes?" Bad: 0.0598 - Good: 0.9307 4. "How internet ads work" Bad: 1.0000 - Good: 0.0000 5. "More Intel speculative execution vulnerabilities" Bad: 0.7413 - Good:…

I think the problem is that the title is not a good indicator for current-event related submissions. "More Intel speculative execution vulnerabilities" may be a bad blogpost, but it's an important current event, so it still gets to the top regardless of the title selection.

Categorizing submissions to different types, and repeat the experiment, you'll find the program may predict blog/article and "Show HN" submissions with higher accuracy.

Re: Show HN: Testing HN titles against a neural network

#208

Post by frozenport »

skykooler wrote: post #171
jansan wrote: post #134

Earlier quoted context omitted.

So the answer "Can a neural network predict if your HN post title will get up votes?" is a clear "no", at least for this tool.

The only one it predicts well...is itself.

Awfully mean

Re: Show HN: Testing HN titles against a neural network

#209

Post by objektif »

codesushi42 wrote: post #195
GistNoesis wrote: post #154

Congratulations on getting first rank on front page. Congratulations on getting your hands dirty and doing everything yourself like computing gradients manually, badly shuffling (non Fisher-Yates), badly js transpose (double swapping), it is a great way to learn. Congratulations on completing a full pipeline, that's the hard part then it just swapping pieces for better pieces. I advise non-technical readers not to at…

This. Exactly this. No sophisticated tokenization. No interesting architecture using attention. And the author is completely clueless about overfitting... and even cross entropy loss. He could have gotten better results just using a bag of words approach. But this ends up on frontpage anyway. Welcome to HN.

What tools would you use to detect overfitting in this case and in general?
Post reply on HN