A Highly Improbable Data Point

Several recent papers (Silverman 2010; Kaplan 2011; Wedel et al. 2013a,b) have presented evidence in favor of the longstanding functional load hypothesis (Martinet 1952; Hockett 1967), arguing that phoneme pairs that distinguish many words are unlikely to undergo merger, and that this phenomenon can be detected in sufficiently large datasets. In contrast with some other approaches (e.g., King 1967), these recent studies assume that functional load operates as a statistical tendency: homophony avoidance is one factor among many that influences the course of sound change, and does not by itself predict whether a given pair of sounds will merge or not.

Geoffrey Sampson observes that many of the sound changes that have occurred in the history of Chinese involve merger, and that the result of these mergers is homophony on a massive scale. He notes that homophone creation of this magnitude is unexpected if there is indeed a tendency for pairs of sounds with high functional load to avoid merger.

To get a sense of just how unusual the Chinese case is, consider the following examples. Table 1 lists several phoneme pairs from the dataset used in Wedel et al. (2013b). For each pair, the table gives the raw number of minimal pairs distinguished by those phonemes (in a particular position, if relevant) and the probability of merger between those phonemes as predicted by the logistic regression model of Wedel et al. (2013b).¹ Note that we do not claim that these estimates represent, in any concrete sense, the actual probability that a particular pair of sounds will merge; I include them here only to clarify the relative predicted probabilities of various types of mergers.

Table 1.

Probability of merger for various pairs, as predicted by the model in Wedel et al. (2013b)

The first group of entries in the table gives the three mergers that were predicted by the model to have the highest probability of actually occurring, and that also had at least one minimal pair. These mergers (as it happens, all in Korean) have all in fact occurred; the model assigns a probability of .511 to the most likely one.

The second group of entries gives the actually occurring mergers that were predicted by the model to be least likely. The most unlikely merger that has actually taken place, Slovak l ˜ λ, is assigned a probability of .0136; of the actual mergers, the one associated with the largest number of minimal pairs is French e ˜ ε, with 419. (The model includes a random effect of System – essentially, language plus phonological context – with the result that predicted probability is not monotonically decreasing as the number of minimal pairs increases when we make comparisons across systems.)

The third group of entries gives some of the phoneme pairs in the database with the largest number of minimal pairs. Unsurprisingly, none of these pairs correspond to actual mergers. These non-mergers have minimal pairs numbering in the thousands, and the model assigns them vanishingly small probabilities of actually occurring. (Note that the figure of 2.2 × 10^-16 for Slovak is inaccurate, the result of underflow in R.)

In this context, it is quite clear that the Chinese case is indeed an outlier. Sampson estimates that the merger of /k k^h x/ with /ts ts^h s/, for example, may have resulted in as many as ten thousand homophones. Even if we suppose that Sampson’s estimate is two or three times too large, we are still left with a merger that produced an order of magnitude more homophones than any of the actual mergers in Wedel et al. (2013b). We would predict that mergers of this type ought to be vanishingly rare.

Now, ‘vanishingly rare’ is of course not the same as ‘impossible’. In fact, if the effect of functional load really is a statistical tendency (rather than an absolute restriction on sound change), then we should fully expect to find a few instances of homophone creation on an unusually large scale – as long as there aren’t too many such cases. Moreover, it is widely accepted that functional...