FOIL in Anthropology: Automated Discovery of Kinship Structures

TL;DR

Social anthropology is often data-heavy but "computation-light." This paper bridges that gap by applying Inductive Logic Programming (ILP)—specifically the FOIL algorithm—to automate the extraction of complex social rules from genealogical data. Unlike previous attempts, this method handles noisy, recursive, and culturally diverse relationship structures, turning raw field notes into rigorous logical definitions.

Background: The Anthropological Bottleneck

For decades, anthropologists have recorded massive amounts of quantitative data. However, the rules governing social interactions—marriage taboos, inheritance, and descent—are rarely recorded as "laws." Instead, they are implicit in a sea of examples.

The author argues that social anthropology is one of the least "mathematized" social sciences. Traditional machine learning (like decision trees or early neural nets) fails here because it requires "flat" attribute-value tables. But how do you put an "uncle" into a spreadsheet when the number of children and siblings varies for every person?

The Core Insight: Relational Learning

The paper proposes that Kinship is inherently relational, not attribute-based.

Using FOIL (First-Order Inductive Learner), the system treats data as a collection of Prolog-like facts:

• parent(jack, ann)
• male(jack)
• husband(jack, jill)

By providing a target (e.g., "What is a father?"), FOIL induces a Horn Clause: father(A, B) :- parent(A, B), not(female(A)).

Key Methodology: The Kinship Predicate Set

To capture the richness of non-Western societies, the author defines a minimum predicate set:

• Biological ties: Father, mother, son, daughter.
• Attributes: Sex, age, status (living/deceased).
• Social Class: Tribe, moiety, or clan membership.

Note: The system represents the kinship structure as a directed graph where nodes are individuals and arcs are specific relations.

Case Studies & Experiments

1. Age-Determinate Relationships

In many cultures (e.g., Confucian), a "Brother" is not enough; one must distinguish "Older Brother." FOIL successfully induced: older_brother(A, B) :- age(A, C), age(B, D), C > D, brother(A, B). The experiment showed that neglecting "male" predicates could lead the AI to assume a sister could be an "older brother" if the negative examples weren't sufficiently diverse—a crucial reminder of the importance of Negative Constraints in social science AI.

2. Disambiguating Multiple Definitions

In certain Surinamese dialects, the same term is used for "wife" and "mother’s brother’s daughter." FOIL proved powerful enough to create multiple rules for a single relation, effectively mapping these distinct biological paths to the same social label.

3. The Limits of "Absence"

A fascinating finding was FOIL's struggle with Exogamy (permissible marriage). While it can learn what is incest (a list of forbidden relations), it struggles to define "unrelated" or "less related than a second cousin" because symbolic logic is better at finding connections than proving their total absence in a sparse graph.

Sample of how kinship data was structured before moving into predicate logic.

Critical Analysis: Why This Matters

The value of this work isn't just in "solving" kinship. It's in Rule Verification. Anthropologists often report "idealized" rules (e.g., "you must marry a paternal cousin"). FOIL allows researchers to run these algorithms on actual marriage data to see if the "ideal" matches the "real." If the AI induces a different rule than what the tribe says, we've found a point of cultural tension.

Limitations

• Domain Theory: FOIL cannot easily "re-use" concepts it just learned unless we move to its successor, FOCL.
• Counting: FOIL struggles with rules like "A chief is a man with more than three wives" unless explicit "count" predicates are added manually.

Conclusion: Toward a Logical Anthropology

Sally Jo Cunningham demonstrates that ILP is far more than a "toy" for Western genealogy. It is a robust tool for sifting through the noise of fieldwork to find the underlying logic of human civilization. For the future of Digital Humanities, this symbolic approach offers a level of interpretability that black-box deep learning models currently lack.