Which distributional cues help the most? Unsupervised contexts selection for lexical category acquisition

Starting from the distributional bootstrapping hypothesis, we propose an unsupervised model that selects the most useful distributional information according to its salience in the input, incorporating psycholinguistic evidence. With a supervised Parts-of-Speech tagging experiment, we provide preliminary results suggesting that the distributional contexts extracted by our model yield similar performances as compared to current approaches from the literature, with a gain in psychological plausibility. We also introduce a more principled way to evaluate the effectiveness of distributional contexts in helping learners to group words in syntactic categories.