When learners surpass their models: mathematical modeling of learning from an inconsistent source

Area of Science:

• Cognitive Science
• Computational Linguistics
• Developmental Psychology

Background:

• Learners often encounter inconsistent linguistic input from native speakers.
• Both children and adults can regularize and modify this input to varying degrees.

Purpose of the Study:

• To present a novel algorithm modeling the learning process from inconsistent linguistic sources.
• To investigate how learners master grammatical or lexical forms despite input imperfections.
• To demonstrate a learner's capacity to equal or exceed the quality of its source.

Main Methods:

• Developed a new algorithm inspired by reinforcement learning and drift-diffusion models.
• Applied the algorithm to model a child's acquisition of American Sign Language (ASL) from imperfect input.
• Utilized existing psychological and computational modeling principles.

Main Results:

• The algorithm exhibits psychologically relevant properties: fidelity, robustness, discounting, and computational simplicity.
• Demonstrated the algorithm's 'frequency boosting' property, where common forms are amplified by the learner.
• Successfully modeled key features of Simon's ASL acquisition from imperfect parental input.

Conclusions:

• The developed algorithm effectively models language acquisition from inconsistent sources.
• Learners can not only master but also refine and improve upon imperfect linguistic input.
• The model provides insights into the mechanisms underlying language regularization and acquisition in development.