zoobereq / richness-of-the-stimulus Goto Github PK

This project attempts to reproduce an experiment conducted by Reali and Christiansen (2005). The original paper is also included in the repository for reference.

Chomsky's Poverty of the Stimulus (henceforth POS) thesis famously postulates that children are not exposed to sufficient linguistic data in their environments to be able to acquire every feature of the target langauge. Consequently, langauge acquisition requires an alternative cognitive mechanisim (i.e. the universal grammar) in order to be successful. Reali and Christiansen (2005) undermine this POS assumption by demonstrating that a simple statistical language model is sufficient to induce well-formed utterances, and that it can do so in spite of a lack of adequate amounts of linguistic data. The authors illustrate their claim by investigating auxiliary inversion in English polar interrogatives containing subject-modifying relative clauses.

The experiment relies on the corpus of spontaneous adult-child conversations developed by Bernstein-Ratner (available here). get_raw_data iterates over the corpus isolating the utterances of adult speech directed at 13 through 21-month-old children and outputs them into a separate file (raw_data.txt). Thus obtained sub-corpus is then cleaned, normalized with clean_data.py, and output into a separate file (normalized_data.txt). The latter is used by evaluate.py to develop a bigram language model with Knesser Ney interpolation. The model is used to calculate the entropy values of the 100 grammatical (Bigrams_Richness_Experiment_1_grammatical.txt) and ungrammatical (Bigrams_Richness_Experiment_1_ungrammatical.txt) utterance pairs. The utterance with a lower entropy score is judged to be grammatical based on the bigram language model.

The accuracy of the langauge model reported by Reali and Christiansen was 96%. The accuracy reached by the current model is 88%, meaning that the model correctly evaluates 88 out of 100 utterance pairs.