Information theory and artificial grammar learning: inferring grammaticality from redundancy

Area of Science:

• Cognitive Psychology
• Artificial Grammar Learning
• Linguistics

Background:

• Artificial grammar learning (AGL) experiments typically assess rule induction by testing participants' ability to differentiate grammatical from ungrammatical strings.
• This ability is commonly interpreted as evidence of successful implicit rule learning.

Purpose of the Study:

• To investigate the actual basis of grammaticality judgments in AGL experiments.
• To determine whether participants are truly inducing grammatical rules or relying on alternative heuristics.

Main Methods:

• Participants studied letter strings generated by a specific grammar.
• They then judged novel strings as grammatical or ungrammatical.
• Analysis focused on the relationship between local redundancy within strings and grammaticality judgments, including a transfer test with novel symbols.

Main Results:

• Judgments of grammaticality were strongly predicted by the local redundancy of the test strings.
• This effect persisted even in a transfer test using different letters than the training set.
• Grammaticality itself was not the primary predictor of participant judgments.

Conclusions:

• The common interpretation of AGL findings as evidence of rule induction is flawed due to confounding variables.
• Participants likely rely on a heuristic of 'pattern goodness' or local redundancy rather than abstract grammatical rules.
• The study highlights the importance of carefully controlling stimuli in AGL research and understanding attribute substitution in judgment tasks.