The Effects of Semantic Annotations on Precision Parse Ranking

First Joint Conference on Lexical and Computational Semantics (*SEM), pages 228–236, Montr´eal, Canada, June 7-8, 2012. c⃝2012 Association for Computational Linguistics The Effects of Semantic Annotations on Precision Parse Ranking Andrew MacKinlay♦♥, Rebecca Dridan♠∗, Diana McCarthy♣† and Timothy Baldwin♦♥ ♦ Dept. of Computing and Information Systems, University of Melbourne, Australia ♥ NICTA Victoria Research Laboratories, University of Melbourne, Australia ♠ Department of Informatics, University of Oslo, Norway ♣ Computational Linguistics and Phonetics, Saarland University, Germany amack@csse.unimelb.edu.au, rdridan@ifi.uio.no, diana@dianamccarthy.co.uk, tb@ldwin.net Abstract We investigate the effects of adding semantic annotations including word sense hypernyms to the source text for use as an extra source of information in HPSG parse ranking for the English Resource Grammar. The semantic an- notations are coarse semantic categories or en- tries from a distributional thesaurus, assigned either heuristically or by a pre-trained tagger. We test this using two test corpora in different domains with various sources of training data. The best reduces error rate in dependency F- score by 1% on average, while some methods produce substantial decreases in performance. 1 Introduction Most start-of-the-art natural language parsers (Char- niak, 2000; Clark and Curran, 2004; Collins, 1997) use lexicalised features for parse ranking. These are important to achieve optimal parsing accuracy, and yet these are also the features which by their nature suffer from data-sparseness problems in the training data. In the absence of reliable fine-grained statis- tics for a given token, various strategies are possible. There will often be statistics available for coarser categories, such as the POS of the particular token. However, it is possible that these coarser represen- tations discard too much, missing out information which could be valuable to the parse ranking. An intermediate level of representation could provide valuable additional information here. For example, ∗This research was conducted while the second author was a postdoctoral researcher within NICTA VRL. †The third author is a visiting scholar on the Erasmus Mundus Masters Program in ‘Language and Communication Technologies’ (LCT, 2007–0060). assume we wish to correctly attach the prepositional phrases in the following examples: (1) I saw a tree with my telescope (2) I saw a tree with no leaves The most obvious interpretation in each case has the prepositional phrase headed by with attaching in dif- ferent places: to the verb phrase in the first example, and to the noun tree in the second. Such distinctions are difficult for a parser to make when the training data is sparse, but imagine we had seen examples such as the following in the training corpus: (3) Kim saw a eucalypt with his binoculars (4) Sandy observed a willow with plentiful foliage There are few lexical items in common, but in each case the prepositional phrase attachment follows the same pattern: in (3) it attaches to the verb, and in (4) to the noun. A conventional lexicalised parser would have no knowledge of the semantic similarity between eucalypt and tree, willow and tree, binoc- ulars and telescope, or foliage and leaves, so would not be able to …

Content truncated to 500 words