Automatically Extracting Templates from Examples for NLP Tasks

Automatically Extracting Templates from Examples for NLP Tasks * Ethel Ong, Bryan Anthony Hong, Vince Andrew Nuñez College of Computer Studies, De La Salle University, Manila, Philippines onge@dlsu.edu.ph, bashx5@yahoo.com, link7488933@yahoo.com Abstract. In this paper, we present the approaches used by our NLP systems to automatically extract templates for example-based machine translation and pun generation. Our translation system is able to extract an average of 73.25% correct translation templates, resulting in a translation quality that has a low word error rate of 18% when the test document contains sentence patterns matching the training set, to a high 85% when the test document is different from the training corpus. Our pun generator is able to extract 69.2% usable templates, resulting in computer-generated puns that received an average score of 2.13 as compared to 2.7 for human-generated puns from user feedback. Keywords: Template Extraction, Machine Translation, Joke Generation. 1. Introduction Templates have been used in IE as extraction patterns to retrieve relevant information from documents (Muslea, 1999), and in NLG as forms that can be filled in to generate syntactically correct and coherent text for human readers. They have also been used in machine translation (Cicekli and Guvenir, 2003) and (McTait, 2001), and in pun generation (Ritchie et al, 2006). In this paper, we present two NLP systems. TExt (Go et al, 2006 and Nunez, 2008), a bi- directional English-Filipino machine translator, extracts translation templates from a bilingual corpus, and together with a bilingual lexicon, uses these templates to translate an input text to another language. T-Peg (Hong, 2008) utilizes semantic and phonetic knowledge to capture the wordplay used in a training set of human jokes, resulting in templates that contain variables, tags, and word relationships that are used to generate punning riddles. Because manually creating templates can be tedious and time consuming, several researches have worked on automatically extracting templates from training examples that have been preprocessed. In our previous example-based MT, SalinWika (Bautista et al, 2005), templates are extracted from a bilingual corpus that has been pre-tagged and manually annotated with word features, resulting in a long training process. Its successor, TExt (Go et al, 2006), did away with a tagger and instead requires a parallel bilingual corpus and an English-Filipino lexicon to align pairs of untagged sentences to extract translation templates. Our pun generator, T-Peg (Hong, 2008), on the other hand, subjects the training examples through a pre-processing stage to identify nouns, verbs and adjectives. Instead of manually annotating the example set, the training algorithm relies on existing linguistic resources and tools to perform its task. * Acknowledgments: TExt Translation is part of the “Hybrid English-Filipino Machine Translation System” project that is funded by the Department of Science and Technology – Philippine Center for Advanced Science and Technology Research and Development (DOST-PCASTRD). It was developed by Kathleen Go, Manimin Morga, Vince Nunez, and Francis Veto as their undergraduate thesis. Copyright 2007 by Ethel Ong, Bryan Anthony Hong, and Vince Andrew Nuñez 452 22nd Pacific Asia Conference on Language, Information and …

Content truncated to 500 words