Compressible kinship terminologies are more learnable than less compressible alternatives

Sara Rolando, Simon Kirby, Kenny SmithProceedings of the 12th International Conference on the Evolution of Language (Evolang12) 2018 · DOI

Abstract

Different languages partition meanings into different semantic categories, labelled with words or morphemes. The scope of variation in these partitions is wide, as systems of semantic categories can differ in both the number of labels used and in the strategies used to group meanings into one category. However, this variation is constrained—not all theoretically-possible partitions are found in natural languages, and similar meanings are encountered in unrelated languages.

Introduction

Why do human languages categorize the world in such similar ways? Across the thousands of languages spoken on Earth, the way we name family members (kinship terminology) isn't random. While there is variation, it is highly constrained. For example, many unrelated languages use the same word for "brother" and "sister" (underspecifying gender), but almost none use the same word for "paternal grandfather" and "maternal uncle."

Researchers like Kemp and Regier (2012) have hypothesized that these systems are shaped by a trade-off between informativeness (being precise enough to distinguish people) and simplicity (minimizing cognitive load). While the "informativeness" side is well-documented, the "simplicity" side—specifically the claim that simpler systems are actually easier for humans to learn—needed a direct mechanistic test. This paper provides that test using an artificial language experiment to prove that kinship systems are shaped by a "simplicity bias."

Experimental Design

The authors conducted an artificial language learning experiment to see if "compressibility" correlates with "learnability." Participants were tasked with learning one of two artificial kinship systems, both using 12 labels to describe 16 possible family relationships.

Comparison of simple (left) and complex (right) kinship labeling systems used in the experiment.

The study utilized two primary conditions:

  • Simple Condition: A system with high compressibility. It used fewer rules to define categories (e.g., grouping relatives by generation but ignoring gender).
  • Complex Condition: A system with low compressibility. It required more rules because it grouped seemingly unrelated meanings together (e.g., using the same label for a paternal grandfather and a maternal uncle).

Complexity was quantified using the length of the grammar required to describe the system—an information-theoretic approach to measuring cognitive "rules." Participants first learned a family tree using basic "parent/child" primitives. Once they mastered the tree, they were trained and tested on the specific kinship labels for their assigned condition.

Key Results and Insights

The core finding of the study is that compressibility equals learnability. The mathematical definition of simplicity (grammar length) accurately predicts human performance. If a system can be "zipped" into a smaller set of rules, the human brain acquires it significantly faster.

The Regularization Bias

Interestingly, participants in both groups didn't just learn the input; they tended to simplify it further. They often collapsed distinctions present in the training data, producing final languages with fewer than the original 12 labels. In the Complex condition, participants struggled to maintain the "illogical" groupings, often reverting to simpler, more rule-based categorizations.

Simplicity vs. Precision

When humans learn a new system, they are willing to sacrifice communicative functionality (precision) if it makes the system easier to represent mentally. This suggests that the simplicity pressure in language evolution is a primary force. Without a counter-pressure for accuracy (like the need to communicate to achieve a specific goal), languages naturally drift toward extreme simplicity.

Discussion and Future Directions

While this experiment demonstrates that learning favors simplicity, it focused on individual learning in a vacuum. In the real world, languages are used for communication, which favors informativeness. The researchers are currently investigating whether the "near-optimal" trade-off seen in natural languages emerges from learning alone, or if it requires a "closed-loop" of both learning and communicative use.

Furthermore, while the study used a controlled set of 16 relationships, future work will need to scale these results to more massive semantic domains and explore how they interact with other cognitive biases, such as the biological importance of certain relatives.

Summary updated

Abstract

Different languages partition meanings into different semantic categories, labelled with words or morphemes. The scope of variation in these partitions is wide, as systems of semantic categories can differ in both the number of labels used and in the strategies used to group meanings into one category. However, this variation is constrained -not all theoretically-possible partitions are found in natural languages, and similar meanings are encountered in unrelated languages.
BibTeX
@inproceedings{rolando2018compressible,
  author = {Sara Rolando and Simon Kirby and Kenny Smith},
  title = {{Compressible kinship terminologies are more learnable than less compressible alternatives}},
  booktitle = {Proceedings of the 12th International Conference on the Evolution of Language (Evolang12)},
  year = {2018},
  doi = {10.12775/3991-1.102},
}
Figure