Robust PCFG-Based Generation Using Automatically Acquired LFG Approximations

Proceedings of the 21st International Conference on Computational Linguistics and 44th Annual Meeting of the ACL, pages 1033–1040, Sydney, July 2006. c⃝2006 Association for Computational Linguistics Robust PCFG-Based Generation using Automatically Acquired LFG Approximations Aoife Cahill1 and Josef van Genabith1,2 1 National Centre for Language Technology (NCLT) School of Computing, Dublin City University, Dublin 9, Ireland 2 Center for Advanced Studies, IBM Dublin, Ireland {acahill,josef}@computing.dcu.ie Abstract We present a novel PCFG-based archi- tecture for robust probabilistic generation based on wide-coverage LFG approxima- tions (Cahill et al., 2004) automatically extracted from treebanks, maximising the probability of a tree given an f-structure. We evaluate our approach using string- based evaluation. We currently achieve coverage of 95.26%, a BLEU score of 0.7227 and string accuracy of 0.7476 on the Penn-II WSJ Section 23 sentences of length ≤20. 1 Introduction Wide coverage grammars automatically extracted from treebanks are a corner-stone technology in state-of-the-art probabilistic parsing. They achieve robustness and coverage at a fraction of the development cost of hand-crafted grammars. It is surprising to note that to date, such grammars do not usually figure in the complementary operation to parsing – natural language surface realisation. Research on statistical natural language surface realisation has taken three broad forms, differ- ing in where statistical information is applied in the generation process. Langkilde (2000), for ex- ample, uses n-gram word statistics to rank alter- native output strings from symbolic hand-crafted generators to select paths in parse forest repre- sentations. Bangalore and Rambow (2000) use n-gram word sequence statistics in a TAG-based generation model to rank output strings and ad- ditional statistical and symbolic resources at in- termediate generation stages. Ratnaparkhi (2000) uses maximum entropy models to drive generation with word bigram or dependency representations taking into account (unrealised) semantic features. Valldal and Oepen (2005) present a discriminative disambiguation model using a hand-crafted HPSG grammar for generation. Belz (2005) describes a method for building statistical generation mod- els using an automatically created generation tree- bank for weather forecasts. None of these prob- abilistic approaches to NLG uses a full treebank grammar to drive generation. Bangalore et al. (2001) investigate the ef- fect of training size on performance while using grammars automatically extracted from the Penn- II Treebank (Marcus et al., 1994) for generation. Using an automatically extracted XTAG grammar, they achieve a string accuracy of 0.749 on their test set. Nakanishi et al. (2005) present proba- bilistic models for a chart generator using a HPSG grammar acquired from the Penn-II Treebank (the Enju HPSG). They investigate discriminative dis- ambiguation models following Valldal and Oepen (2005) and their best model achieves coverage of 90.56% and a BLEU score of 0.7723 on Penn-II WSJ Section 23 sentences of length ≤20. In this paper we present a novel PCFG-based architecture for probabilistic generation based on wide-coverage, robust Lexical Functional Gram- mar (LFG) approximations automatically ex- tracted from treebanks (Cahill et al., 2004). In Section 2 we briefly describe LFG (Kaplan and Bresnan, 1982). Section 3 presents our genera- tion architecture. Section …

Content truncated to 500 words