Comment by kerkeslager | Hacker News Reader

Comment by kerkeslager | Hacker News Reader

Parent Full thread

I can understand why someone would want to run WhatFreeWords if What3Words is already a standard: it's dumb to pay for 238-lines-of-Python-as-a-service. The added latency alone is ridiculous.

My question is, why would anyone commit to What3Words as a standard? There are much simpler ways to map geographic data to three-word combinations.

One obvious way, dividing things up into rough "squares" of approximately even size, has the added benefit that shared or alphabetically proximal words would indicate geographical proximity: with What3Words, wounds.client.face is feet away from to jumped.storm.nest, but hundreds of miles from wounds.client.fact.

If you don't think that's useful, then perhaps it's useful to use one of the other word lists which are on many computers, to share space, such as the PGP Word List[1] or the word list from RFC 1760[2]. One of the touted benefits of What3Words is that it is algorithmic and therefore doesn't require a large mapping file, so it's appropriate for systems where storage comes at a premium. Yet they used a more complicated algorithm and didn't reuse an existing word list?

It seems like their algorithm design was basically driven to try to make it hard to reverse engineer, and it didn't even achieve that. The only reason I can see why this became anything resembling a standard is that some people really want to pay for services.

[1] https://en.wikipedia.org/wiki/PGP_word_list

[2] https://tools.ietf.org/html/rfc1760

W3W is a nasty rent-seeking operation. They have invented a flawed solution looking for a problem and are using lots and lots of dubious PR to try and get it mainstream so they can then start charging for the use of their stupid algorithm.

Their PR attempts are absolutely pathetic but sadly seem to work on the masses. Their most frequent angle is someone contacting emergency services but somehow can't figure out his location, and W3W comes along and magically saves the day. The problem is that if I need a phone app & GPS signal to figure out my W3W, why can't I just use the GPS coordinates directly?

While I agree that W3W is a nasty rent-seeking operation, I do think that in broad strokes the location-as-words concept could be useful.

> The problem is that if I need a phone app & GPS signal to figure out my W3W, why can't I just use the GPS coordinates directly?

Think about verbal/handwritten transmission and mental/handwritten storage. It's not easy to transcribe "41.022755, -73.913102" or remember it. "medium.axed.hobby" is much easier. Copy/paste doesn't always work well on phones, and doesn't work at all across devices.

There are other open standards such as Plus Codes, but I actually disagree that (what it looks like to me) a hash-based system (which is more or less what W3W is) is better than a numeric system like GPS coordinates.

41, -73 (removed all decimals from your example location) still gives me the general location that it's somewhere in New York. My understanding is that removing a word from W3W completely invalidates it as it's like a hash.

Incomplete GPS coordinates can be augmented by other information. W3W doesn't look like it can, if your phone cuts out before you said the last word of your W3W to the emergency dispatcher you're screwed. If your phone cuts out before you said the last few decimals of the GPS coordinates the dispatcher still has a chance of finding you.

> 41, -73 (removed all decimals from your example location) still gives me the general location that it's somewhere in New York.

I think it would be hard to overstate how useless this data is rounded that far. 41.022755, -73.913102 is almost a two hour drive away from 41, -73 according to Google Maps, and that "drive" starts with you swimming a tenth of a mile to shore. Even if you round better (round -73.913102 to -74) we're talking about a 6.9 mile drive that will take 15 minutes. With different examples (anywhere which is close to _.5 for both coordinates, i.e. -73.5, 41.5) the distances are going to be even larger.

Rounded to two decimal points, you're at a point where the accuracy is probably good enough for most purposes, but now you're looking at 41.02, -73.91, which is clearly harder to transmit over the phone than "medium.axed.hobby".

The purpose of W3W not having nearby locations have similar words is because of transmission errors. They're pushing it in the UK at the moment having got several emergency services on board, and in that context if you say to, say, Norfolk Police call handlers that you're at bamboo cheese elephant and that shows up as somewhere in the middle of the Pacific they know they got a transmission error pretty much instantly and can ask for clarification. Although it'd be nice maybe to have soundalike tools in the search so it can say "hang on maybe they meant this place".

Two conflicting needs really - as it stands, W3W is incredibly good for simple transmission by voice, but it's awful for giving you clues to geographic proximity.

I wrote a short Python script which cross-references the W3W word list with a homophones list[1], and found this:

    census,senses
    choral,coral
    clairvoyance,clairvoyants
    collard,collared
    confectionary,confectionery
    disburse,disperse
    epic,epoch
    equivalence,equivalents
    hurdle,hurtle
    incidence,incidents
    incite,insight
    incompetence,incompetents
    independence,independents
    innocence,innocents
    instance,instants
    intense,intents
    lightening,lightning
    ordinance,ordnance
    overdo,overdue
    parse,pass
    pokey,poky
    precedence,precedents,presidents
    purest,purist
    recede,reseed
    sari,sorry
    senses,census
    variance,variants
    verses,versus

Those aren't sound-alikes, they're straight-up homophones. If we cross-referenced against a list which contained sound-alikes, I'm sure we'd end up with a much larger list of collisions.

EDIT: This is the script:

    with open('word_data.txt') as f:
        words = set(l.strip() for l in f.readlines())

    with open('homophones.txt') as f:
        for l in f.readlines():
            l = l.strip()
            homophone_group = (h.strip('*') for h in l.split())
            filtered_homophone_group = tuple(h for h in homophone_group if h in words)
            if len(filtered_homophone_group) > 1:
                print(','.join(filtered_homophone_group))

The `strip('*')` bit is needed because my copy/paste into homophones.txt pulled in some asterices(sp?) and I didn't want to go through and fix them manually.

[1] http://homophonelist.com/homophones-list/

I have to say, I don't think W3W is actually that great for simple transmission by voice. It doesn't seem like they've put in any effort at all to avoid similar-sounding words. I did a few searches on the word list and found a bunch of groups with very little effort:

* pace, placed, plates, pays

* blush, plush, brushed, brushes

* crave, craved, chafe, safe

* inks, mink, pink, pinks, kinks, link, wink, winks, think, thinks, blink, brink, blinks, clinks, slinks, rinks, drinks

* yappy, happy, nappy, snappy, sappy, strappy, scrappy

I kind of just stopped searching for each of these groups, I'm sure there's more for each one.

If I can find sound-alikes this easily, just by probability there are going to be similar-sounding adjacent locations.

I'm also not convinced that geographic proximity versus transmission accuracy is a tradeoff we need to make.

One very simple solution would be to add an error-checking word, which is essentially a fast-hash of the first three words. I'm not convinced given what we know about memory with regard to numbers, that three words is somehow the ideal number--storing four words in short- or long-term is not noticeably harder to me.

> Their PR attempts are absolutely pathetic but sadly seem to work on the masses.

I guess you just described marketing. It's frustrating but if you want to sell to non-niche markets, you need to lower the standards to make it work.

It always seemed to me that this should’ve been restricted to, say, 100 common words. That gives 1 million unique points, covering about 20 sq miles, each. If you want to be more precise, you could allow size, color, and texture adjectives on each word. 5 of each (including ‘no adjective’) would increase precision to about a 16x16 foot area.

EDIT: to be clear: I think there should be 3 distinct lists of 100 words, so that the set is order free.

A lot of good ideas here, but I'd shoot for a slightly larger number here than 100. You get a polynomial effect here, where a word added to list A gets you (the length of list B x the length of list C) more combinations--there's a lot of payoff to even adding just a few more words to each list.

The surface area of the earth is about 200 million sq miles, so 1 million unique points gets you 200 sq miles each, not 20. That's 5,575,680,000,000,000 square feet.

40,000 words per list (the length of the W3W list) gets you down to about a 87 square foot (around 9ft x 9ft) area with three words.

I tend to agree with you that 9ft x 9ft is smaller than necessary for most purposes. If we target a 16ft x 16ft area (5m x 5m for those using metric), we get that with three word lists of 28,000 words. Using six words (three nouns for general area, plus three adjectives for precision), word lists of 168 words are adequate.

I'm not completely convinced that the larger areas are ever useful, however. By going to just 4 words, we can achieve the 16ft x 16ft area with four lists of 2200 words. Those lists are short enough to plausibly check for sound-alikes, enough words exist to have four separate word lists, and have a shorter resulting location phrase.

I'm also not convinced that the set being order-free is really necessary. Transposition errors seem more likely in longer data sets than three words (hence even/odd word lists in the PGP word list).

Reply on news.ycombinator.com