A few months ago, I posted about how I was relatively unimpressed
by a paper arguing that the observed Zipfian distribution of words in a corpus is due to "preferential attachment"
aka the Matthew Effect aka the rich get richer. The author of that paper is apparently also a co-author of a paper called "Languages cool as they expand: Allometric scaling and the decreasing need for new words
." The writeup in Inside Science
summarizes it like this:
[A] recent analysis has found that as a language grows over time, it becomes more set in its ways. New words are always being added, according to this study, but few become widely used and part of the standard vocabulary.
My linguist hackles immediately raised at this statement, and that's because there is a large and fundamental difference between what a linguist understands the term "language" to refer to, and what the authors of the column and paper understand it to refer to. What the physicists and the reporter mean by "language" is roughly "a set of words," and in the context of the paper, they almost seem to mean "the set of words which have been published."
This "language is words" axiom is part of most people's folk
linguistics that we have to train people out of when they take Intro to Linguistics. That's why it's a little hard to take the work of these physicists seriously at first glance. It is as if they were trying to write a serious paper on biological evolution with the assumption that traits acquired by an organism during its life were inheritable.
But there is an aspect of linguistic knowledge relating to the set of words and morphemes a speaker knows, which linguists call the "lexicon". So, I'll just go ahead and reread the paper mentally replacing each instance of "language" with "lexicon" in order to get through it.
This paper seems to be a relatively competent (modulo Mark Liberman's concerns about OCR errors
) description of the statistical properties of large corpora. But that's really as far as I think any of the claims can go. I am totally unconvinced that their results shed any light on language change, development, evolution, etc. I'm not even sure that the simplest statement that "the lexicon of languages has grown over the past 200 years" can be supported by the results reported.
The key problem that I see with the paper is the conflation of "new to the corpus" and "new to the lexicon." Here's how the problem of sampling language was describe to me, and I believe it goes back to Good (1953) and is key to Good-Turing Smoothing
. Say you are a entomologist working in a rain forest, trying to make a survey of insect life. You put out your net for a night to collect a sample, then count up all the species in your net. Some bug species are going to be a lot more frequent than others. You'll have some species that show up many times in the net, but even more species will show up in the net with only one member. Now, let's say that you come back to the same rain forest two years later, and repeat the sample. You are nearly guaranteed to observe new species in your net this time around, but the key question is whether they are just new to the net, or are they new to the rain forest. If they're new to the rain forest, did they migrate in, or are they hybrids of two other species, or has a species you saw previously evolved really rapidly so that you're seeing it as different now?
These are really interesting and important questions for our entomologist to answer, but you cannot arrive at a definitive answer based simply on the fact that this new species has now showed up in your net. In fact, depending on a few factors, the answer with the highest probability is that the new species is simply new to your net. The Good-Turing estimate of the probability that the very next bug you catch will be a new species is that it's roughly equivalent to the proportion of bugs you've already caught that belong to a species you've only seen once.
The situation gets even more confusing if you come back to the same rain forest two years later with a net twice the size.
The paper has a figure plotting the increase in lexicon size over time. My first thought when I saw it was that it must be the case that the overall size of the corpus at each time point must also be going up. Coming back to the entomologist in the rain forest, the number of species in his net is merely a sample
of how many species there are in forest. In the same exact way, the number of words in a lexicon can only be estimated by the words which people happened to write down. As you increase the size of the net, you're going to find more species which were already in the forest, but not in your net. As you increase the size of your corpus, you're going to find more words which were already in the lexicon, but not in the corpus.
Now, you need to add to this that at any given point in time, the true maximum number of possible
words you could potentially observe in any given language is ∞. Yes, in fact, the whole reason language is interesting to study is because given a finite set of mental objects, and a finite set of operations to combine them, you can come up with an infinite set of stings, and that goes for words too, not just sentences. In 1951, "iPod" was a possible word of English, it just wasn't used, or at least not for the same purpose it is now.
Regarding the question of whether the "active" (as I'll call it) lexicons of languages have grown over the past 200 years, well, indeed, the overall number of printed words has also increased. Almost all of their results seem to have more to do with the technological development of publishing
than it does with any other linguistic or cultural development. It is as if the entomologist said that over the past decade, the biodiversity in his rainforest has exploded, when really what's going on is his nets have been getting progressively larger.
Now, it might be the case that the active lexicon has grown more than would be expected given the increase in the size of the corpus year over year, but as far as I can tell, the authors did not try to estimate whether this was the case.
What about this cooling down?
The "cooling" effect referred to by the paper is the suggestion that as a language "grows" (which as I just said is dubious), the frequency with which particular words are used becomes more stable. Some words are more frequent than others, but words are less likely to move up and down in frequency over time/as the lexicon grows. Back to entomology, the suggestion is that as more species cram into a rainforest, each species is less likely to become more or less populous.
Again, though, the frequency, even relatively frequency, of a word in a corpus is merely an estimate of its true frequency. As the size of the corpus increases, so should the reliability of its frequency estimates, and we would predict decreasing volatility of those frequency estimates. The authors check for this, and find exactly this relationship between corpus size and frequency volatility, but I can't tell whether there was excess "cooling" left over. I wish they had said, "there was x proportion of cooling left unaccounted for by simply accounting for the size of the corpus," but I think this is perhaps another symptom of the assumption that the corpus=the lexicon=the language that I complained about before.
The Allure of Big Data
The reporter who wrote the Inside Science article did what it appears that the editors of Scientific Reports did not, asked a linguist to comment on the paper. Bill Kretzschmar was "underwhelmed," saying that most of these results are not new to linguists. I would take this as a word of warning about the allure of big data. The results discussed in this paper are not, by and large, new, but rather have never been done with data of this scale. But unfortunately, a fact which is already known does not get more interesting when it is reestablished with data 100 or 1000 times larger than before.