Parser lexicalisation through self-learning

Proceedings of NAACL-HLT 2013, pages 391–400, Atlanta, Georgia, 9–14 June 2013. c⃝2013 Association for Computational Linguistics Parser lexicalisation through self-learning Marek Rei Computer Labratory University of Cambridge United Kingdom Marek.Rei@cl.cam.ac.uk Ted Briscoe Computer Laboratory University of Cambridge United Kingdom Ted.Briscoe@cl.cam.ac.uk Abstract We describe a new self-learning framework for parser lexicalisation that requires only a plain-text corpus of in-domain text. The method first creates augmented versions of de- pendency graphs by applying a series of mod- ifications designed to directly capture higher- order lexical path dependencies. Scores are assigned to each edge in the graph using statis- tics from an automatically parsed background corpus. As bilexical dependencies are sparse, a novel directed distributional word similar- ity measure is used to smooth edge score es- timates. Edge scores are then combined into graph scores and used for reranking the top- n analyses found by the unlexicalised parser. The approach achieves significant improve- ments on WSJ and biomedical text over the unlexicalised baseline parser, which is origi- nally trained on a subset of the Brown corpus. 1 Introduction Most parsers exploit supervised machine learning methods and a syntactically annotated dataset (i.e. treebank), incorporating a wide range of features in the training process to deliver competitive perfor- mance. The use of lexically-conditioned features, such as relations between lemmas or word forms, is often critical when choosing the correct syntac- tic analysis in ambiguous contexts. However, util- ising such features leads the parser to learn infor- mation that is often specific to the domain and/or genre of the training data. Several experiments have demonstrated that many lexical features learnt in one domain provide little if any benefit when pars- ing text from different domains and genres (Sekine, 1997; Gildea, 2001). Furthermore, manual creation of in-domain treebanks is an expensive and time- consuming process, which can only be performed by experts with sufficient linguistic and domain knowl- edge. In contrast, unlexicalised parsers avoid using lex- ical information and select a syntactic analysis us- ing only more general features, such as POS tags. While they cannot be expected to achieve optimal performance when trained and tested in a single do- main, unlexicalised parsers can be surprisingly com- petitive with their lexicalised counterparts (Klein and Manning, 2003; Petrov et al., 2006). In this work, instead of trying to adapt a lexicalised parser to new domains, we explore how bilexical features can be integrated effectively with any unlexicalised parser. As our novel self-learning framework re- quires only a large unannotated corpus, lexical fea- tures can be easily tuned to a specific domain or genre by selecting a suitable dataset. In addition, we describe a graph expansion process that captures selected bilexical relations which improve perfor- mance but would otherwise require sparse higher- order dependency path feature types in most ap- proaches to dependency parsing. As many bilex- ical features will still be sparse, we also develop an approach to estimating confidence scores for de- pendency relations using a directional distributional word similarity measure. The final framework in- tegrates easily …

Content truncated to 500 words