Stronger Affinity…than Could Have been Produced by Accident: A Probabilistic Comparison of Old Chinese and Tibeto-Burman
更强的同源性 … 比偶然可能产生的：古汉语及藏缅语族概率性比较
William H. Baxter 白一平

Abstract 摘要

1. INTRODUCTION
In recent years there has been increasing interest in the possibility of tracing distant language relationships. The discussion has even reached the popular press, which usually pays little attention to linguistics. One attraction of hypotheses about language relationships is that they might tell us something about human population movements in prehistoric times – something we otherwise have relatively little evidence about. Some of the bolder recent proposals for distant linguistic relationships include the putative families Nostratic (Illič-Svityč 1971-1984, Dolgopolsky 1964) and Amerind (Greenberg 1987). As for proposals involving Chinese, the association of Chinese and Tibeto-Burman in a Sino-Tibetan family, though widely accepted, is not uncontested; the position of Thai and related languages is still debated (traditionally part of Sino-Tibetan, but assigned by Paul Benedict to Austro-Thai, along with Austronesian). Sergei Starostin (1982, 1984, 1991b) has argued for a Sino-Caucasian family including Sino-Tibetan, Yeniseian, and North Caucasian, and his colleague Nikolaev (1989) has further proposed a Dene-Caucasian family including Starostin’s Sino-Caucasian and the Na-Dene family of North America. (this proposal ties in with Greenberg’s Amerind hypothesis, according to which Na-Dene and Eskimo-Aleut represent more recent incursions from Asia into the Americas, compared with the older Amerind family incorporating All other native American languages.) More recently, Laurent Sagart (1993) has given arguments for a genetic relationship between Chinese and the Austronesian family.

The discussion of these hypotheses among linguists has been understandably lively, but at times also surprisingly acrimonious: proposals which some linguists regard as established scientific fact are dismissed by others as irresponsible nonsense. The source of the problem, I believe, is the lack of consensus among linguists about how hypotheses like these can be evaluated objectively. In the absence of such consensus, the attitudes of individual linguists towards a controversial hypothesis often reflect more about their individual temperaments, and the habits of their respective academic microcultures, than about the evidence for or against any particular hypothesis.

This paper attempts to remedy this situation by proposing objective methods for evaluating hypotheses about remote linguistic relationships, and to illustrate these methods by applying them to the hypothesis that Chinese and the Tibeto-Burman languages are genetically related. The fundamental assumption of this approach is that when two languages show phonological correspondences in their lexicons which are too great to attribute to chance, this fact calls for some explanation. The only plausible explanations other than a genetic relationship are (1) lexical borrowing and (2) a non-arbitrary relationships between sound and meaning. If these two explanations cannot account for the correspondences, then a genetic relationship is the only plausible explanation left. This may be the case even if no ancestral proto-language has actually been reconstructed, and if the phonological histories of the languages are still poorly understood.

The use of greater-than-chance correspondences as a criterion for genetic relatedness is at least as old as Sir William Jones’s famous observation in 1786 that Sanskrit. Greek, and Latin show a stronger affinity, both in the roots of verbs and in the forms of grammar, than could have been produced by accident; so strong that no philologer could examine Sanskrit, Greek, and Latin, without believing them to have sprung from some common source, which perhaps, no longer exists. (Quoted in Ruhlen 1994:12; emphasis added)

If this approach is to move beyond relying on the intuition of the ‘philologer’, it is crucial that one be able to decide on principled grounds how likely it really is that the correspondences observes are the result of chance. The theory of probability theory to test possible genetic relationship is by no means new, but this approach is still not widely understood or used. The approach described here is essentially the same as that of Justeson and Stephens (1980). Possibly novel aspects of this study include (1) the emphasis on using a fully explicit algorithm, which can be implemented on a computer, to identify phonological matches; (2) using a computer to actually simulate repeated random trials; and (3) applying the method to Chinese and Tibeto-Burman.
The approach used here can be summarized as follows:

1. Parallel word lists are independently assembled for the languages being compared: in this study I use a 35-item list (of which I exclude two) compiled by S. E. Jaxontov; from this list are constructed parallel word lists for Old Chinese and Tibeto-Burman. These lists are specified and discussed more fully below.

2. An algorithm is constructed which will decide whether any particular pair of words will be counted as a phonological match. (An algorithm is an explicit procedure which can be applied mechanically by a computer, and which always gives an answer.)

3. This algorithm is used to count how many of the items on the list match when they are paired according to their meaning. We may call this the observed score; it will be compared with the scores obtained when items are paired at random.

4. With the aid of a computer, a large number of random trials are performed in which one of the lists is mechanically scrambled and then matched against the other list, using the same algorithm. The computer counts and remembers how many phonological matches are obtained on each trial. The proportion of random trials whose scores are as high as the observed score is an estimate of the probability that the observed score could have occurred by chance. If this probability is below a certain level, then the phonological matches observed are judged too numerous to ‘have been produced by accident’. In some cases it is also relatively easy to estimate this probability by a mathematical formula.

It should be kept in mind that the observed score has no significance in itself; it will tend to be higher or lower depending on the strictness of the criteria for a phonological match. Seven matches in a test with very strict criteria may be strong evidence of a relationship, while fifteen matches on another test may not be significant at all. The observed score is like the raw score on a standardized test: observed scores from different tests are standardized by reference to the probability in each particular test of getting so high a score by chance.

Section 2 describes in some detail two experiments I conducted in this fashion on word lists from Old Chinese and reconstructed Tibeto-Burman (as presented in Benedict 1972). Section 3 gives some further comments and cautions about how to design such experiments; section 4 discusses the interpretation of results.

2. THE EXPERIMENT
2.1 Word lists
2.2 An algorithm for Old Chinese-Tibeto-Burman correspondences
2.3 Results

3. HOW TO CHEAT
3.1 Stacking the deck
3.2 Inflating the score
3.3 But I didn’t cheat much…

4. INTERPRETING THE RESULTS
4.1 Interpreting a positive result: getting lucky
4.2 Interpreting a negative result