Exploring an auxiliary distribution based approach to domain adaptation of a syntactic disambiguation model
Barbara Plank, Gertjan van Noord · 2008
We investigate auxiliary distributions (Johnson and Riezler, 2000) for domain adaptation of a supervised parsing system of Dutch. To overcome the limited target domain training data, we exploit an original and larger out-of-domain model as auxiliary distribution. However, our empirical results exhibit that the auxiliary distribution does not help: even when very little target training data is available the incorporation of the out-of-domain model does not contribute to parsing accuracy on the target domain; instead, better results are achieved either without adaptation or by simple model combination. 1