Tree Linearization in English: Improving Language Model Based Approaches

We compare two approaches to dependency tree linearization, a task which arises in many NLP applications. The first one is the widely used ’overgenerate and rank’ approach which relies exclusively on a trigram language model (LM); the second one combines language modeling with a maximum entropy classifier trained on a range of linguistic features. The results provide strong support for the combined method and show that trigram LMs are appropriate for phrase linearization while on the clause level a richer representation is necessary to achieve comparable performance.