Hacker News Reader

Comment by timtadh | Hacker News Reader

What I noticed using this to read a couple Ars Technica articles was it worked really well for short words. But, if a longer complicated word appeared, at say 600 wpm, I would miss it. It seems like an adaptive algorithm based on word length would improve the speed even more allowing it to go faster on short common words and slow down on longer unusual words.

Also, there is bug that sometimes causes two words to appear at once. Sometimes they are overlapping vertically which is basically impossible to read. Try reading http://pragdave.me/blog/2014/03/04/time-to-kill-agile/ to see what I mean.

View on HN

Hey timtadh! Spritz Inc's blog (http://www.spritzinc.com/blog/) has some interesting things to say about word recognition, primarily that it's related to a word's unique shape more than length. Words that are less than 4 characters long are actually harder to recognize due to their having more visual analogues than longer words, and words over 7 characters are start looking "long" more than any other shape characteristic. I've got some ideas about how to incorporate this into my app without doing a bunch of image processing, so, until then, keep your eyes peeled ;) Thanks for the feedback!

That may be true for familiar words, but a long, uncommon word is unavoidably going to take more processing. Flashing "unforgettable" is probably fine, but I'd want a split second pause on "periodontia" or "imbroglie".

I agree, If the app could maintain a frequency table containing how common words are in normal usage and slow down more for the most uncommon words that would be awesome.

Until then slowing down for longer words at least would definitely help. This could at least be an option if the authors aren't convinced its a good idea yet, then A/B test the amount of usage for people who use the option compared with those that don't.

This is a great observation. After reading Spritz' blog post, I've been thinking about using word shape uniqueness as the main signal for how long to pause--i.e., "soliloquy" would be quite unique thanks to its ascenders and descenders, whereas "excessive" is less unique, owing to its relatively featureless outline. Combining shape with word frequency (across language) seems even more promising.

I'm not sure ascenders and descenders are the only things factoring in here. I guess doing a visual blur on an image of a word could help to spot characteristics of words. Admittedly, the ascenders and descenders would have a probably primary role in aiding classifying. So maybe it's an 80%/20% thing.

At a guess words with unique clusters of vowels would also stand out say 'Hawaiian'. Or other words like 'cameraman', 'minimum', 'consciousness', 'neuroscience', that all probably have a certain shape or density, even if they are flat.

This regex dictionary may be useful: http://www.visca.com/regexdict/ /^[aeiuocmnrsvwxz]*$/

Slow down on long words would be very helpful, I just tried it with some german news text[1] at 600wpm: Worked great except for words that occured like

    Kinderbetreuungsmöglichkeiten
    US-Geheimdienstausschusses
    Bundestags-Innenausschuss
    Bundeswirtschaftsminister
    Bundesarbeitsministeriums

[1] http://www.deutschlandfunk.de/nachrichten.353.de.html

Do you think german words like this would benefit from automatic syllable colorization -- slightly changing the shade of grey to make the syllable boundaries more obvious?

Not sure. E.g. 'Kinderbetreuungsmöglichkeiten' means 'child care possibilities', that's three words in one. I suppose the brain just needs more time to parse that. But just counting characters to determine delay won't cut it, since 'Mississippi' is quite quick to parse.

Then maybe colorize nouns in a compound noun word instead of syllables ?

Couldn't you split compound words so the components appear in sequence? e.g.

Kinder- -betreuungs- -möglichkeiten

One could, but most likely one would need a proper morphological analysis[2] to make it work correctly for the endless possibilities to combine words in German. Pretty hard to implement in the browser, but maybe an 80:20 like approach using a good dictionary could help for common terms in combination with slowdown on terms it doesn't know how to split.

[1] for example, one could use wiktionary data and try to extract how to properly separate common terms: https://de.wiktionary.org/wiki/Schifffahrt

Edit: [2] German demo of how that usually looks http://www.tagh.de/demo.php (not affiliated)

What 'bout making it simpler and instead of colorizing nouns you colorize a horizontal gradient like http://www.beelinereader.com/

Syllable recognition is still an unsolved problem, no? Is anybody familiar with the state of the art?

For German (and maybe English too?), I think splitting up compound words into their constituent noun stems could work.

LOL. This is on the todo list :) Thanks for the feedback!

Also doesn't work for Japanese text. It shows pretty much a whole paragraph at a time.

It's also doesn't work well for hyphenated words: 'step-by-step' and 'hastily-implemented' are flashed up all at once.

(Regardless of whether or not those are properly hyphenated.)

Edit: But this is still very cool.

That's what I did with Spree [1] (also posted below). Just a simple word.length*8 seemed to have quite a nice effect on readability.

1. https://chrome.google.com/webstore/detail/spree/aehoaolhojlm...

Nice, I've been trying a bunch of these things recently (my bookmarks bar now has OpenSpritz, RSVP and squirt along with 2 chrome extensions) and your extensions the nicest implementatiion I've found so far. (partly due to https errors with the other bookmarklet approaches and partly due to nicer design). Thanks.

Agreed. They need to slow down the speed for longer words. Otherwise, it would be perfect.

It seems that Spritz splits long words into multiple parts, using - on the ends of the components. That seems easy enough to implement.

Reply on news.ycombinator.com