If you want to see how far linguistics lags behind physics in scientific understanding, turn to the Proceedings of the National Academy of Sciences for a paper by Richard Futrell, Kyle Mahowald, and Edward Gibson, titled “Large scale evidence of dependency length minimalization in 37 languages.”
Imagine that Newton had not announced his law of gravity, but trumpeted instead the finding that apples fall to earth every time. Chance alone suggests that half the apples tossed from my bed should fall up, but way fewer than that do so.
That level of understanding is where the Futrell et al paper advances us, and the sad thing is it marks real progress.
It also threatens to send researchers running down an imaginary alley.
Suppose our alternative Newton had said that the tendency to fall down to earth was a universal property of matter. Wait, we don’t have to imagine an alternative Newton because Aristotle did say there was a universal property of matter that drove it to fall toward the center of the earth.
Some Aristotelian thinking was still present in Newton who imagined that all matter had a property that drove it to fall toward other matter. It took a couple of more centuries before Einstein cleared that part up.
So, the “universal property” of syntax that I am about to report should be labeled, “Do not inhale.” If ingested, it could delay linguistic progress for many years.
The proposed universal says, “dependency lengths are shorter than chance.”
Dependency lengths measure the distance between a word (what the authors call the head (H)) and a second word (D) that depends on the head. If sentences were organized randomly, the distance between H and D would average half the length of the sentence. But the average length between H and D is well shy of that. Indeed, the longer the sentence, the greater the difference between average dependency lengths and a chance length.
Of course, nobody believes sentences are organized randomly, so the really astonishing finding would be that dependency lengths match chance lengths.
Furthermore, anybody who thinks for a moment about language is likely to realize that words are organized in phrases that lump related words together. So naturally related words will be closer than chance would expect.
The authors know all this and gladly grant the point. They also grant that for some time researchers have assumed that language users tend to minimize dependency length (the DLM hypothesis). They have studied sentences in 37 languages (23 of them Indo-European) and find that the DLM hypothesis holds across the board, although there are some interesting differences.
Amusingly, one of the languages with the least variance from random lengths is Latin while its direct descendent, modern Italian, is one of the most minimized. Other Romance languages examined—Catlan, French, Portuguese, Romanian, and Spanish—are also much more minimized than Latin. I also note that ancient Greek is much less minimized than modern Greek. So these comparisons may be contrasting the language of an educated, literary elite with that of a mass-literacy population.
Latin scholars will recall that word order is freer in Latin, thanks to the language’s extensive use of morphology (special markings to indicate relationships). The authors report that there is a difference between languages that tend to put heads at the ends of phrases (head-final languages) and those that do not:
“In particular, the head-final languages such as Japanese, Korean, and Turkish show much less minimization than more head-initial languages…. Head-final languages typically have highly informative word morphology … [that] might give more freedom in their dependency lengths because it makes long dependencies easier to identify.”
That sounds like a real finding, so why did I begin this post in such a carping tone?
My chief complaints are both technical and personal. The technical concern is that this universal is called a “syntactic property of languages.” This claim repeats Aristotle’s error of assigning the property to the thing observed, whereas the effect is much more likely to come from something outside.
If we take the DLM hypothesis literally, we might suppose that minimization is a goal in itself. That mistake was exactly the one that Aristotelians made when they thought falling to earth was a goal-directed activity. Reduced to the ridiculous, a person might even argue that when generating sentences there is a step in the process that says count the average dependency lengths in possible sentence x and possible sentence y, and then go with the shorter length.
My personal complaint is that I have already proposed a way of thinking about language production in which the paper’s findings are implicit. In my paper on Attention-Based Syntax I argued that language works by creating focal points of attention that are bound into whole perceptions. In that process, dependency lengths will naturally be constrained to less than chance, and they will be less constrained when morphology provides greater rhetorical freedom.*
The observations reported by Futrell et al are important and I am happy to pass along the news, but they are side-effects of unexplored processes, not properties of what is observed.
* In that paper I say: “A fundamental principle of attention-based syntax says the speaker must keep attention focused on one [noun phrase] until it has been specified.” [p. 8] Classical grammarians have long recognized rules against breaking up noun phrases—you cannot say, for example, Jack went up the hill and Jill. But I have a reason for speaking in phrases: it would confuses matters if attention was broken. So, naturally the sentence's dependency will be closer than chance predicts.
I also say, “Part of [some demonstrated] rhetorical freedom probably reflects the fact that [adverbs] are indicated by the morpheme –ly, so there is usually no ambiguity about which syntactic element the word [modifies].”  Thus, languages heavy in morphology are likely to be less minimized than morphology-poor languages like English.
This observation, by the way, makes me think about an English grammar rule that has irritated many a student. Fast, as in That was a fast visit, is an adjective, whereas quickly—e.g., That visit passed quickly—is an adverb. Therefore, you should say he ran quickly rather than he ran fast. The supposedly wrong form is perfectly intelligible and has inspired generations of grumbling, leading bright scholars to imagine grammar is a conspiracy to make them feel stupid. But compare: He ran along the twisting, cobblestone lane quickly with He ran along the twisting, cobblestone lane fast. Thanks to morphology, interpreting the first sentence seems to take a bit less deliberate work. So the rule is not entirely crazy, just overstressed.