Last November, Hannah Haynie and I published a paper in the Proceedings of the National Academy of Sciences on color term systems in Pama-Nyungan. In it, we used phylogenetic methods to show that color term systems can both gain and lose terms, and that while they do so mostly in accordance with prior work on color term systems (Berlin and Kay, Kay and Maffi, and colleagues), we also found evidence for ‘exceptional’ systems that appeared not to conform to the B&K system. We used data from the Chirila database and fairly standard phylogenetic methods of ancestral state reconstruction.
For an analysis of this type to be correct, several assumptions must be satisfied:
- sample data need to be representative of the languages as a whole;
- sample data need to be correct;
- the analytical tools need to be applicable to what’s being studied;
- the analyses need to be interpreted correctly.
Over the last six months, Hannah and I have been in correspondence with David Nash about many of these points, particularly those involving sampling, the correctness of the underlying data, and judgments about what is a color term. In particular, in the original version of Table S1, a data conversion error resulted in words from several languages being associated with the wrong row in the table (particularly Wargamay and Warlmanpa). This did not affect the analyses reported in the paper, as the error was introduced when spreadsheets were converted to Microsoft Word documents for uploading to the journal’s online submission site. [The corrected table is available here.]
The discussions with Nash revolved around several issues already identified both in our paper and the supplementary materials:
- the difficulty of determining whether a color term is genuinely absent from the language, or simply not recorded;
- the difficulty of establishing the ranges of color terms glossed in English by non-native speakers of the language;
- the issue of polysemy, for example, whether a term glossed as “unripe, green” is truly a color term, or whether “green” here is meant solely in the sense of “unripe, not ready for eating” (and therefore not glossing a true color term).
Coding decisions of this type are based on a careful philological analysis of each individual source, and while phylogenetic analyses are usually robust to individual errors, systematic errors may bias the results. In general, where Hannah and I were unsure, we tended to include rather than exclude; this applies especially to terms for ‘green’ and terms for ‘red’ based on words meaning ‘blood’ (which could be interpreted as the descriptive adjective ‘bloody’ rather than a true color term). For ‘green’ terms, many languages have a word that is glossed as ‘green’ or ‘unripe’; while some of these terms do appear to be real color terms (in that they can refer to items that aren’t unripe, like shirts), others aren’t — they refer to the ripeness of fruit, not directly to its color. (We had a similar problem with ‘grey’, which was often ambiguously glossed as a color term or a word referring only to grey hair.)
Another issue is the extent to which we make use of data from closely related languages in determining the color inventory of a particular language variety. For example, if a particular variety appears to lack a term for ‘blue’, but a term is present in other languages in the subgroup, are we justified in treating the lack of a term as a true omission? In our analyses, we treated such cases as absent rather than indeterminate, because we did not want to omit true variation in the color inventories of languages. But it would also be a possible argument to claim that color inventories are unlikely to vary so much between dialects of the same language (or closely related languages in a subgroup), so unrecorded colors are probably omissions from data collection rather than genuine absences from the language.
We suspect that some terms were not recorded because of the linguists’ expectations about what items are present (or not) in a language. For example, Australian languages are stereotypically claimed to lack color terms beyond black, white, red, and yellow; this can lead researchers not to ask for terms like blue or purple.
Finally, data for this paper came from the Chirila database (Bowern 2016), which while extensive (800,000+ items), is by no means exhaustive. Nash brought to our attention several cases where color terms had been recorded in sources which are not in Chirila. These are also noted in the revised supplementary table and reflected in the newly uploaded analysis files.
In order to assess the impact of our coding decisions, as well as the impact of terms which were missing from Chirila and hence recorded as absent from the languages, we re-ran all analyses. We ran two sets of updated analyses. One simply corrected errors resulting from data missing from Chirila. The other also used Nash’s alternative judgments about presence/absence of color terms like ‘green’. In neither case were our main conclusions affected. That is, we still find support for both color gain and color loss. While, as is expected, the numerical values of individual results changed somewhat, our inferences and conclusions stand. Color loss is possible (under this model), though it’s substantially less common than color gain.
I am currently working on a new update to Chirila and many of these revised sources will be available there.