Constituency Parse Reranking for Morphologically Rich Languages

Abstract

In this article we introduce a constituent parsing system which can achieve stateof-the-art results on morphologically rich languages. Our system consists of a Probabilistic Context Free Grammars (PCFG) and n best reranking steps. We compare two methods to handle lexical sparsity in a PCFG parser. The n best reranking step, the discriminative reranker extracts large amount of features from n best parses of the PCFG parser and selects the best tree from these parses. We introduce three feature templates which extend the standard feature set of rerankers. We propose to extract features from Brown clustering-which is a context-based clustering over the words-and analyze the effect of dependency-based and morphology-based feature templates. The effects of these techniques are evaluated on datasets of eight morphologically rich languages.