Heterogeneous Parsing via Collaborative Decoding

There often exist multiple corpora for the same natural language processing (NLP) tasks. However, such corpora are generally used independently due to distinctions in annotation standards. For the purpose of full use of readily available human annotations, it is significant to simultaneously utilize multiple corpora of different annotation standards. In this paper, we focus on the challenge of constituent syntactic parsing with treebanks of different annotations and propose a collaborative decoding (or co-decoding) approach to improve parsing accuracy by leveraging bracket structure consensus between multiple parsing decoders trained on individual treebanks. Experimental results show the effectiveness of the proposed approach, which outperforms state-of-the-art baselines, especially on long sentences.