NOTE-406: Alfano, Cheong & Curry 2024, moral universals in 256 societies

Contribution

The paper (LIT-512) builds a LIWC-format dictionary for the seven domains of morality-as-cooperation (MAC-D). It runs the dictionary over the eHRAF Ethics paragraphs of every society that has any (256 of 331), and checks the machine codes against the hand codes of the sixty-society study (LIT-047). It is the first use of the programme’s domains on the whole HRAF corpus, and the first published MAC dictionary.

Key insight

A dictionary can measure how much a paragraph talks about a domain, but not whether it praises or condemns. So a machine-read replication of a universality claim can scale up presence. It cannot scale up valence, which is the part of the claim that made the sixty-society result look strong.

Assumptions

• Construction (§3).
  • The authors and five colleagues “already familiar with MAC” brainstormed words, stems and n-grams for seven virtue and seven vice categories: 371 virtue and 148 vice items.
  • WordNet expanded these. Two co-authors culled the false positives (κ = .70 virtue, .65 vice).
  • Antonyms were swapped between the virtue and vice lists by consensus.
  • Final sizes run from 35 items (fairness vice) to 477 (group virtue) (Table 1).
• Corpus. Only paragraphs labelled Ethics, unlike LIT-047, which also used Norms and a keyword search (footnote 10). In all: 1,620,644 words, 1,389 documents, 8 regions, 9 subsistence types.
• Measure. The percentage of a paragraph’s words found in each sub-dictionary. Sub-dictionaries differ in size and base rate, so domains cannot be compared with one another, only across paragraphs (footnote 13).
• Heroism is coded as bravery or generosity, where LIT-047 coded bravery only, so the two are “not directly comparable” (footnote 4).

Key results

• Distribution. Median and mode are 0% for every domain. Means run from 0.04% (fairness) to 0.92% (family) of words (Table 2).

• PSF vs non-PSF. Some differences are significant, but all are small (r ≤ .084).

• Regions (Table 3). All seven differ significantly (p < .001), with ε² from .004 (property) to .055 (heroism). Europe is high on group and heroism, and Asia on family.

• Subsistence (Table 4). Six of seven differ significantly, all with small effects. Reciprocity does not differ (p = .47).

• Validation against the hand codes (§4.2).

  • Each sub-dictionary is the best predictor of its own hand code (Table 5; odds ratios 1.82–4.43).
  • Dichotomised, the models are right about 80% of the time. But positives are rare (about 3%), so false positives swamp true ones: for family, 469 false against 110 true.
  • κ: family .25, group .10, reciprocity .21, heroism .13, deference .16, fairness .08, property .23.

• Predicted prevalence (Table 6, PSF hand codes → PSF predicted). The dictionary overstates every domain except property:

  • family 78 → 88%;
  • group 73 → 95%;
  • reciprocity 72 → 88%;
  • heroism 53 → 92%;
  • deference 58 → 92%;
  • fairness 15 → 58%;
  • property 90 → 85%.

  The non-PSF predictions (78, 88, 69, 78, 81, 43, 76%) are read as overestimates “roughly the same” as in the PSF. The correction is asserted, not computed.

• Spot-checks (§4.3).

  • Confirming examples for all seven domains.
  • Counterexamples: Navajo “no conception of a moral obligation to reciprocate”; !Kung San who “attach no value to fighting”; admired theft among Bedouin, Greek (Sarakatsani), Pashtun and Wogeo.
  • Two moral molecules, filial piety and honour.
  • Two dilemmas, gratitude against prestige and feud against harmony.

Claims

id	claim	strength	support
C1	Most of the seven morals are present in most of 256 societies, in all regions	weak–moderate	dictionary presence at κ .08–.25 against hand codes, with systematic over-detection
C2	The 196 non-PSF societies have about the same prevalence as the 60 PSF societies	weak	inferred from similar paragraph-level means and an assumed common error; not hand-checked
C3	Variation by region and subsistence is minor	moderate	significant but small effect sizes (Tables 3–4)
C4	MAC-D is “validated”	weak	best-predictor-of-own-code holds; agreement is slight to fair, and the authors call the overall fit “fairly low” (§5)
C5	The results are “evidence for moral universalism” (significance statement)	weak	the method cannot detect valence, and universalism of valence is the claim at issue
C6	The counterexamples reflect “competition between the morals … rather than outright devaluation”	assertion	the same reading as LIT-047’s footnote 3; no criterion separates the two

Concepts

• MAC-D. Seven virtue and seven vice sub-dictionaries in LIWC’s .dic format, released for LIWC, R (quanteda) and Python.
• PSF60. The HRAF Probability Sample Files societies hand-coded in LIT-047.

Connections

• The sixty-society test (LIT-047, NOTE-034). The paper restates LIT-047’s results, including “equal frequency across six cultural regions”, as support for MAC. Its own regional tests then find significant differences. It bears on NOTE-034’s verdict as the LIT entry sets out:
  • valence is untested;
  • the spot-checks found the negative cases the original design rarely registered, and the paper absorbs them as conflicts;
  • regional equality is not borne out;
  • non-cooperative content is still uncoded.
• Purity (LIT-101). There is no purity, sexual-conduct or harm dictionary, so the paper is silent on the commentaries’ objection. The authors defer an MFT comparison until HRAF has hand codes for it (§3.1).
• Moral molecules (LIT-509). Filial piety and honour are reported as examples, not counted.
• The MAC-Q (LIT-496). Proposed as the instrument for the ecological-variation test the paper could not run (§5).

Bearing on the record

There is no instruction for ML practice. The method is lexicon counting from 2020 with WordNet expansion, and the authors point to BERT and LLMs as future work, so it is not an anthology candidate. For the morality-as-cooperation thread it should not be cited as a replication of the 99.9% valence result. It replicates presence, roughly. It is better evidence of variation by region and subsistence than of its absence.

Limitations

• Valence-blind by design, as the authors state (§5).
• Poor agreement with the hand codes (κ ≤ .25), and systematic over-detection.
• Ethics paragraphs only. This narrows the corpus relative to LIT-047.
• Spot-checks are anecdotes. Counterexamples were not counted. With the same reading, LIT-047’s one negative might have been several.
• Ethnographers’ words, mostly in English, as the authors note (§5).

Open questions

• With a valence-aware classifier validated against blind hand codes, what share of HRAF mentions of each domain are negative?
• How many societies admire some form of theft, cowardice or ingratitude? Does MAC predict which ones, beyond calling them conflicts?
• Run the same way, how much purity, sexual and dietary moral content does HRAF contain?

Corrections

• none to a seeded skim (there was no seed)
• The citation. The owner’s citation is correct: Alfano, Cheong and Curry (2024), Heliyon 10(6):e25940.
• Internal inconsistencies.
  • §4.1 compares the PSF60 with “the additional 256 non-PSF60 societies”; there are 196 (abstract, Table 6).
  • §4.3 cites the sixty-society paper as “Curry et al. (2018)”; the reference list gives 2019.
  • The text gives ε² up to .056 for regions; Table 3’s largest is .055.