A goal of computational linguistics is to automate the kind of reasoning that linguists do. Given text in a new language, can we determine the underlying morphemes and the grammar rules that arrange and modify them? The Bayesian strategy is to devise a joint probabilistic model that is capable of generating the descriptions of new languages. Given data from a particular new language, we can then seek explanatory descriptions that have high prior probability. This strategy leads to fascinating and successful algorithms in the case of morphology. Yet the Bayesian approach has been less successful for syntax. It is limited in practice by our ability to (1) design accurate models and (2) solve the computational problem of posterior inference. I will demonstrate some remedies: build only a partial (conditional) model, and use synthetic data to train a neural network that simulates correct posterior inference.
展开▼
