Almost Flat Functional Semantics for Speech Translation

Proceedings of the 22nd International Conference on Computational Linguistics (Coling 2008), pages 713–720 Manchester, August 2008 Almost Flat Functional Semantics for Speech Translation Manny Rayner1, Pierrette Bouillon1, Beth Ann Hockey2, Yukie Nakao3 1 University of Geneva, TIM/ISSCO 40 bvd du Pont-d’Arve CH-1211 Geneva 4, Switzerland Emmanuel.Rayner@issco.unige.ch Pierrette.Bouillon@issco.unige.ch 2 UCSC UARC, Mail Stop 19-26 NASA Ames Research Center Moffet Field, CA 94035 bahockey@ucsc.edu 3 University of Nantes, LINA 2, rue de la Houssini`ere BP 92208 44322 Nantes Cedex 03 yukie.nakao@univ-nantes.fr Abstract We introduce a novel semantic represen- tation formalism, Almost Flat Functional semantics (AFF), which is designed as an intelligent compromise between linguis- tically motivated predicate/argument se- mantics and ad hoc engineering solutions based on flat feature/value lists; the cen- tral idea is to tag each semantic element with the functional marking which most closely surrounds it. We argue that AFF is well-suited for medium-vocabulary speech translation applications, and describe sim- ple and general algorithms for parsing, generating and performing transfer using AFF representations. The formalism has been fully implemented within a medium- vocabulary interlingua-based Open Source speech translation system which translates between English, French, Japanese and Arabic. 1 Introduction Many speech translation architectures require some way to represent the meaning of spoken ut- terances, but even a brief review of the literature reveals a serious divergence of opinion as to how this may best be done. At risk of oversimplifying a little, there are two competing heritages. On the one hand, there is the mainstream computational semantics approach, which ultimately goes back to philosophers like Montague, Russell and Frege and views predicate calculus as the paradigm rep- resentation language. On this view of things, a c⃝ 2008. Licensed under the Creative Commons Attribution-Noncommercial-Share Alike 3.0 Unported li- cense (http://creativecommons.org/licenses/by-nc-sa/3.0/). Some rights reserved. suitable way to represent meaning is to use com- plex structures, in which components and relation- ships are based on deep grammatical functions. Typical ways to realise this strategy are unscoped logical forms, neo-Davidsonian semantics, mini- mal recursion semantics, and similar formalisms. Thus a sentence like “I want a pepperoni pizza” might be represented as something like want1(E, X, Y), ref(X, pronoun(i)), quant(Y, indef), pizza1(Y), pepperoni1(Z), nn(Z, Y) Approaches based in the linguistic tradition were dominant about 10 to 15 years ago, when they were used in major systems like Germany’s Verbmo- bil (Wahlster, 2000) and SRI’s Spoken Language Translator (Rayner et al., 2000). They are still reasonably popular today, as exemplified by major systems like PARC’s XLE (Riezler et al., 2002). The competing heritage has its roots in engi- neering approaches to spoken language systems, which historically have been intimately connected with Machine Learning. On this view of things, a typical semantic representation is a flat list of feature-value pairs, with the features represent- ing semantic concepts: here, “I want a pepperoni pizza” would be represented as something like [utterance_type=request, food=pizza, type=pepperoni] It is interesting to see how little contact there has been between these two traditions. Writers on formal semantics usually treat ad hoc feature- value …

Content truncated to 500 words