Grammar Induction (en.wikipedia.org)

Grammar induction (grammatical inference) is the process of learning a formal grammar or automaton from observed discrete structures (strings, trees, graphs), producing a compact model of the generative rules behind data. Historically focused on learning finite-state machines (regular languages), the field has expanded to context-free and mildly context-sensitive formalisms (e.g., MCFGs), stochastic grammars and categorial grammars. It matters to AI/ML because induced grammars provide interpretable structure for natural language, program synthesis, compression, anomaly detection and semantic parsing—enabling unsupervised or weakly supervised discovery of hierarchical patterns that neural models may struggle to expose explicitly. Technical approaches and trade-offs are diverse: passive learning from positive examples; active/query models (Angluin’s membership/counterexample framework); hypothesis-testing trial-and-error; evolutionary methods that evolve tree-structured production rules (genetic programming); and greedy algorithms that build compact CFGs online or offline (LZW, Sequitur, byte-pair encoding). More recent distributional and pattern-learning algorithms give provable, efficient recovery for large subclasses of grammars. Practical limits remain—finding the smallest grammar is NP-hard, so real systems rely on heuristics and probabilistic models (PCFGs, Bayesian induction). For researchers, grammar induction offers a bridge between symbolic structure and statistical learning, providing interpretable inductive biases and compression-aware representations useful across NLP, program induction and anomaly detection.