Machine Learning Of Natural Language And Speech

this paper we take the broadest possible view of language. What have we learnt about language in the last few millenia. Much in every way. But I will argue that, most significant of all, we have started to learnt how to analyze it. Just as, in Physics, the theories of one century have bowed before the breakthroughs of the next, so in Linguistics, we are in a period of rapid change, where traditional approaches give way to new traditions and ad hoc methodologies give way to scientific discipline. We have analytic methods which allows us to objectively determine the phonology of a language, we have theoretical results which tell us how much work is involved in parsing or learning a language, and we share with the rest of science a heritage of scientific methodology based on conjecture and refutation, generalization and parsimony. And what of learning? Given sufficient information, in the right form, there are no bounds on what we can learn. Given no a priori assumptions, however, there is no way to judge what really is a pattern (leading to what has recently been termed the "conservation law", namely that there will be as many patterns unsuccessfully identified as are successfully identified). As with language, we take the broadest possible view of learning, and in particular evolutionary theory and scientific methodology both fit into the arsenal available for us to bring to bear. The way we think and the way we use language has fitted us to comprehend and analyze the world the way we have. Conversely, the capability to understand and deal with the world clearly involves a learning bias which we use not just in learning language, but in creating it and dynamically moulding it to fit a changing world. The spectrum of unsupervised to supervised learning paradigms, thus tak..