Hacker News Reader

Order the modifiers to maximize the product of their successive restrictive effects.

Can you explain this? A lot of the discussion here seems to be based on the fact that the problem is with interpreting the importance of an adjective in relation to others written out in succession (attached to a noun). Then comes the way it flows in speech.

The general 'rule' (for a lack of a better term), as far as I can tell, is interpreting the importance/value of each adjective, and, thus, recommends a placement order. This importance is based on customary usage.

If you mean to say that we should place superlatives ahead of all adjectives, what if the modifiers include more than one superlative? E.g., the greatest, darkest horse. To give another example: the most beautiful, darkest horse. More than one permutation seem to work. Doesn't this undermine any inherent value one order may have over another?

Also, what does 'impotent' mean? Effective? Isn't this begging the question? If it does mean 'effective', then aren't you essentially saying the order recommended is the most effective because it is the most effective? So, that begs the question: why is it more effective than any other order? And what is effective in terms of adjective order? Fluidity (in terms of pronunciation)? Or physicality/nature of the attributes (black vs. white, etc.)? Or something different altogether?

The way I see it, there is very little rhyme and reason in the order of adjectives. There is no intrinsic value in brick modern house or modern brick house. The rules were created based on arbitrary custom/tradition. But now we are getting into semantics, which really complicates everything.

My apologies if I'm way off here.

Sure, let me try to explain. My model is built as follows:

1. People speak and write not by following rules but their ears.

2. Since the ear hears words in sequence, the interpretation of those words also occurs in sequence.

3. Thus the phrase “great green dragon” is interpreted first as “great _”, then as “great green _”, and then finally as “great green dragon”, with the hole “_” representing the expectation of something concrete to come, something that the mind’s eye can eventually see.

4. This expectation is satisfied only when each new word reduces the possibilities for what will fill the hole. With each new word, the hole must shrink toward a single point, and rapidly.

5. Therefore, the most-pleasing ordering of N modifiers will tend to be the one that most rapidly and evenly converges to a point.

6. Therefore, we can find the most-pleasing ordering by maximizing the geometric mean of the shrinkages for all steps in the sequence.

6. For fixed N, the geometric mean and the product are interchangeable for optimization purposes, so we can just maximize the product of the shinkages.

And that’s pretty much it.

Now to your questions:

By “impotent” I mean without effect. Any word that you put after a word that has already reduced the hole to a near-point is going to seem worthless and spoil the whole sequence.

Re. begging the question, a sequence isn’t effective because it’s effective; it’s effective because, given our mental probability models for word sequences, it rapidly and evenly converges upon a concrete interpretation.

To take your “darkest/greatest” as an example, here’s the rapid-and-even-convergence scores for the two possible orderings:

    $ { echo darkest greatest;
        echo greatest darkest; } | ./modorder.py
            170047264509 darkest greatest
           3409987166088 greatest darkest

So, without any knowledge that “horse” will follow, we can predict that “greatest darkest” will be the best ordering of the modifiers.

(This is from a crude approximation of the model that I just whipped up. It uses a small database of 2-gram and 3-gram frequencies.)

No offense, but this sounds pretty tautological to me.

Can you show me the tautology?

I don't know how to come up with a reasonable model for how people are likely to interpret words without accounting for the fact that people interpret words in large part by how they have been interpreted in the past. But I don't see how accounting for prior knowledge of word frequencies makes the model tautological.

In points 2-3. Obviously we're used to hearing things in a certain order, and tend to copy that, but the question is why we put things in a certain order to start with. How is a 'great X Y' any more concrete than a 'green X Y'? Obviously concrete things have size, but then abstractions aren't exactly known for their coloring either :-)

But "the question", as you offer it, is based on the fallacy that there is some underlying natural law to the way we use words. There isn't. There is no blank slate "to start with," no axioms from which everything else follows. Words have always been used in the context of how they have been used.

Therefore, any reasonable model of how people use words will have some representation of how people have used words. My model uses simple word frequencies, yes, but it doesn't predict that one modifier sequence will be preferred to another because "that's the way it is." Rather, it predicts a sequence will be more preferred because it more rapidly and more evenly converges on a meaning.

How is that prediction tautological?

The original question was 'what is the rule for ordering adjectives [in English.]' The first thing you said was: How about this rule? Order the modifiers to maximize the product of their successive restrictive effects.

I'm saying that your observation people copy what they know is tautological, and tells us nothing whatever about the restrictive effects of a given adjective.

How is Order the modifiers to maximize the product of their successive restrictive effects equivalent to "people copy what they know"? The restrictive effects of a given adjective are an input to the model because that issue was never the question. The ordering was the question, and that's what my proposed model attempts to predict, given prior knowledge of restrictive effects. The interesting prediction it makes is that people prefer orderings that converge rapidly and evenly.

In information theory terms, if you ignore the current expectations of ordering and try to construct an ordering rule based on the information content of each ordering, all orderings have equal information content. Spreading out the information content so that it arrives a little bit with each word instead of in bursts and trickles doesn't get you anything, except that maybe the human brain can process it more easily in smaller chunks. Is this a theory?

This actually jives with the fact that languages with less information per phoneme (Japanese, Spanish) are spoken at a faster rate in phonemes per second than languages with more information per phoneme (Mandarin), so that most languages contain a similar information rate.

"Across-Language Perspective on Speech Information Rate" http://muse.jhu.edu/journals/language/summary/v087/87.3.pell...

> Order the modifiers to maximize the product of their successive restrictive effects.

Compare: "The great grey-green greasy Limpopo river, all set about with fever trees."

So we'd have to decide on the fly which is more restrictive of great rivers, grey-green or greasy. Perhaps sometimes it's just about how the sounds sound.

How about in the case of:

1. "A Chinese vegetarian lawyer"

2. "A Vegetarian Chinese lawyer"

According to the referenced paper[1], #1 is considered awkward, but #1 follows your ordering logic, assuming that there are more Chinese than Vegetarians.

[1] http://www.lingref.com/cpp/wccfl/25/paper1473.pdf

Those examples are hard for me to reason about. In my limited experience with lawyers, my prior for vegetarian is probably higher than for Chinese, but my sample size is too low to be statistically significant.

I also find either ordering awkward for categorizing a professional by their county of origin and their dietary restrictions :)

#1 Would suggest to me a lawyer who is Chinese and specialises in vegetarians someone while #2 would suggest a layer who is Chinese (or who deals in Chinese laws) and is a vegetarian.

Maybe just me?

Yes, that is also how I interpreted the two phrases.

There needs to be a comma in between 'Vegetarian' and 'Chinese' in both phrases to clarify that whichever is the former modifies 'lawyer' as opposed to the latter.

No, because the two adjectives aren't from the same category, are they? That would be the only reason why you would use a comma.

A hyphen also helps...

A Chinese vegetarian lawyer= a lawyer, who is both ethnic Chinese and vegetarian, where the set of vegetarian lawyers is more generally recognized than the set of Chinese lawyers

A Chinese vegetarian-lawyer= ethnic Chinese who practises law of vegetarian issues ("vegetarian" is a compound word, i.e. spoken quicker, at higher pitch, and with less gap before the following word)

A Chinese, vegetarian, lawyer= an ethnic Chinese who just happens to also be a vegetarian, ("vegetarian" is a non-restrictive phrase, i.e. spoken at lower pitch, with a longer gap both before and after)

Even better, #2 is unclear if the person practices Chinese law or is Ethnically Chinese.

Am I supposed to know how to parse those? I'm not sure how, but I really want to. Is that a standard representation I'm just not familiar with (or even one I am familiar with but having trouble recognizing)?

I'm just making this stuff up based on what seems sensible. For all I know, my hypothesis has been made countless times before and refuted by practicing linguists. But, from what I can see so far in this conversation, the hypothesis seems plausible.

So what is all the extra markup supposed to indicate?

  ([_] - [great _]) * ([great _] - [great green _])

What is the first [_] ? Why do you alternate between *s and -s? What is the point of "[x _] denotes the expected number of substitutions admitted by the hole (_) in the context x." if everything seems to have the same number of substitutions?

When I first saw the markup without examining it closely I was excited that I might learn a new markup/sentence diagramming format. Upon closer inspection it seems its just a bunch of characters that when interrogated closely add very little value.

I think smackfu was correct: substituting different adjectives of the same type should not change the order of the adjectives. How are people supposed to discuss things they are not familiar with? What should I do if I do not know what portion of houses are brick? What proportion are colonial versus modern? Or what proportion of craftsman style houses are brick?

It is beyond me why this is the lead thread under a pretty interesting and out of the ordinary link on HN.

I defined [x _] to denote the size of the hole _ in the context x. By "size" I mean some measure of the width of the distribution of things that can sensibly fill the hole. One reasonable measure would be entropy. So, for example, we could say that [_] means the entropy of the distribution of things that could sensibly fill a hole in the empty context, which is to say start a new phrase. Similarly, [great _] denotes the entropy of the phrases that could sensibly follow "great". Does the notation make sense now?

Update: The * and - symbols are just the usual arithmetic operators applied to the sizes.

Look into LKB, if you haven't already. I used it extensively in my master's (computational linguistics) and though I've forgotten most of what I knew, you might find it an interesting rabbit hole in terms of sentence representation formats. Especially if you dig Lisp.

Wouldn't that imply that substituting in a different adjective of the same type (color / size) could change the order?

But in practice, that doesn't seem to be the case.

Personally, I've always mentally preferred an ordering with the noun first followed by increasing levels of restrictiveness. In my head, that's how I think. I then have to translate that into proper English grammar.

It makes sense to me because I'm providing as much detail as necessary, but giving as much detail up front as possible. For example, when saying "the [adjective_detailed], [adjective_general] [noun]", what if the last word were lost? Then you'd have no idea what the phrase was about. And for the listener, you start with the most useless of detail and have to remember it, applying it to more useful of information as it comes in, and you have no idea what's being spoken about until you've heard the end. So then you have to backup and apply the adjectives once you've actually heard the noun. It makes (IMO) bad structure for parsing. Conversely, with the most general detail first, you drill down to the necessary level of detail as you speak, and continuously describe something the listener already heard.

The order you describe reminds me of inventory or shipping labels, particularly military-ese: "QTY 5, pistol, M1911, A2". However, in my experience, Western spoken languages are not at all optimized for dropping the last bits (so often that's where the predicate / object lives).

You made me think of two similar (pet peeves) of mine:

Address and US Dates

In the latter, obviously Mar 15, 2013 is expressed as 3/15/2013. I prefer to encode as 2013 03 15 (insert whatever syntax you like). There are enough techie people here I doubt I need to say much more on that.

But the former, addresses, really bother me. Why do we start with the smallest modifier first? It probably relates to the way postal routes were first created, maybe for ease of use for the delivery person, would would already be in a restricted area.

Seems that addresses should be ordered: Country, City, ZIP/subregion, Street, Number, Person

If you assume that most post travels between points that are close to one another, putting the most specific modifier first makes sense. Most of the time you only have to think about a small part of the address to route correctly.

That breaks down if the assumption is incorrect, and also to some extent if you have many different places with the same name that all receive lots of mail from you.

So you're essentially right: It's an artifact of how the postal system used to work. Though both assumptions still apply more often than you might think.

Good point - likewise email addresses and domain names.

In Chinese, everything is "big to small". It is very logical. dates are 2013-03-15, addresses are USA, New York, Manhattan, names are Last Name, First Name, etc.

I really like this. I can't imagine that the cognitive process humans use to order adjectives in everyday speech involves categorizing each adjective and memorizing an arbitrary ordering on those categories.

On the other hand, the category-ordering approach does have the advantage that it's an O(log n) operation, whereas solving the restrictive effect maximization problem is harder: a heuristic greedy search in the solution space might be O(n), and a true solution might be O(n!).

Some humans categorize every noun by gender and every verb by conjugation behavior... and even memorize multiple definitions and connotations for thousands of words.

For example, we would prefer "modern brick house" to "brick modern house" because, in the second ordering, [brick modern _] is about the same as [brick _], making the second term in

This bit makes no sense to me. Are you assuming that most brick houses are modern? Or that most modern houses are made of brick? How do you describe a brick house in the colonial style? Craftsman? Greek Revival? I dont even want to think about Brick Gothic...

I'm saying that "brick _" is more restrictive on the phrases that can plausibly follow it than is "modern _". By choosing to place brick first, you narrow the possibilities so much that the following word has little work to do and seems out of place.