Adaptive acquisition of language

Language understanding and subsequential transducer learning

“…Pieraccini, Levin and Vidal (1993) addressed the problem of performing sequential transductions through stochastic finite state networks. The information-theoretic connectionist approaches for language learning and understanding presented by Gorin, Levinson, Gertner and Goldman (1991) and Gorin (1995) are based on the basic idea of mapping input sentences into semantic actions, although they are not properly based on the concept of transduction. More recently, a statistical approach to language understanding based on the source-channel paradigm has been presented which explicitly models the understanding process as a transduction from natural language into a formal language, through a statistical translator (Epstein, Papineni, Roukos, Ward & Della Pietra, 1996; Della Pietra, Epstein, Roukos & Ward, 1997).…”

Modeling and Interpreting Multimodal Inputs: A Semantic Integration Approach

“…In principle the mutual information network is capable of learning incrementally during actual use, as demonstrated by Gorin et al [3]. The MS-MIN inherits this capability; however, we need to conduct more experiments to determine to what degree this is true in practice, when the task domains are complex and a lot of training data must be used to achieve adequate coverage.…”

“…The output with the highest activation is selected as the most probable hypothesis given the input sequence. This network architecture was first proposed by Gorin et al [3].…”

“…Although the simplifying independence assumption does not usually hold in practice, this mutual information network has been shown to learn input-output associations quite successfully [3][9] [14]. The input independence assumption implies that classification does not depend on the order of the input tokens.…”

“…Multi-State Mutual Information Network Architecture rence probabilities observed in the training data. These probabilities can be estimated by a simple counting procedure [3]. In Section 5.5 we will describe how the weights can be automatically generated from a grammar-based input model.…”