Enhancing the input to models of lexical acquisition using techniques from speech recognition
Speaker: Engin Ural
Abstract:
This talk describes a research project I'm beginning that applies techniques from speech recognition to computational models of human language acquisition. I'll explain how current computational models on unsupervised word segmentation assume that their inputs are relatively artificial, and I'll describe how I hope to use forced alignment techniques to produce more realistic input for these models. We expect that this richer, more realistic input will change the nature of the learning problem (making it both harder and easier), and I will sketch ways in which we might change our word segmentation/lexical acquisition models in the face of this more realistic input.