Blackwell Reference Online

Extract

Artificial neural networks (ANNs) have been used in a variety of ways in the study of language. In this chapter we will focus on work in NLP within the framework of statistical modeling. For statistical modeling, the most useful ANN architecture is the multi-layered perceptron (MLP). MLPs can be used for probability estimation and feature induction, and have been extended for modeling both sequential and structured data. Their most successful applications in NLP have been language modeling and parsing. MLPs have also been inspirational for much current research in machine learning methods, and can be reinterpreted in terms of approximations to latent variable models. In this chapter we will first cover background material, and then discuss contemporary research. The emphasis will be on the usefulness of ANNs as an engineering tool, but theoretical motivations and context will be given wherever possible. ANNs have the advantages of being robust in training and testing, of being fast in testing, and of requiring little prior knowledge of the domain. ANNs are also interesting because they discover compact feature-based representations specific to the task they are trained on. The term ‘artificial neural network,’ or often just ‘neural network,’ refers to a variety of computational models which share certain properties inspired by the networks of neurons found in the brain. They consist ... log in or subscribe to read full text