Boosting Unsupervised Grammar Induction by Splitting Complex Sentences on Function Words

The statistical-structural algorithm for unsupervised language acquisition, ADIOS (for Automatic DIstillation Of Structure), developed by Solan et al. (2005), has been shown capable of learning precise and productive grammars from realistic, raw and unannotated corpus data, including transcribed children-directed speech from the CHILDES corpora, in languages as diverse as English and Mandarin. This algorithm, however, does not deal well with grammatically complex texts: the patterns it detects in complex sentences often combine parts of different clauses, negatively affecting performance. We address this problem by employing a twostage learning technique. First, complex sentences are split into simple ones around function words, and the resulting corpus is used to train an ADIOS learner. Second, the original complex corpus is used to complete the training. We also show how the function words themselves can be learned from the corpus using unsupervised distributional clustering.