Is bad structure better than no structure? unsupervised parsing for realisation ranking

In natural language generation using symbolic grammars, state-of-the-art realisation rankers use statistical models incorporating both language model and structural features. The rankers depend on multiple structures produced by the particular large-scale symbolic grammars to rank the output; for languages with smaller resources and in-development grammars, we look at the feasibility of an alternative source of structural features, unsupervised parsers. We show that, in spite of their lower quality of structure, raw sets of unsupervised parse features can be helpful with smaller language models; and that the parses do contain particular elements that can be highly useful, improving performance on our classification task by up to 10% on 60% of the test set leading to an overall improvement under a back-off model.