The TALP Ngram-based SMT System for IWSLT’05

The TALP Ngram-based SMT System for IWSLT’05 Josep M. Crego, Adri`a de Gispert and Jos´e B. Mari˜no TALP Research Center Universitat Polit`ecnica de Catalunya, Barcelona {jmcrego|agispert|canton}@gps.tsc.upc.es Abstract This paper provides a description of TALP-Ngram, the tuple-based statistical machine translation system devel- oped at the TALP Research Center of the UPC (Univer- sitat Polit`ecnica de Catalunya). Briefly, the system per- forms a log-linear combination of a translation model and additional feature functions. The translation model is es- timated as an N-gram of bilingual units called tuples, and the feature functions include a target language model, a word penalty, and lexical features, depending on the lan- guage pair and task. The paper describes the participation of the system in the second international workshop on spoken lan- guage translation (IWSLT) held in Pittsburgh, October 2005. Results on Chinese-to-English and Arabic-to- English tracks using supplied data are reported. 1. Introduction and overview of the system During the last several years, statistical machine trans- lation (SMT) has gained much attention within the re- search community. This is mainly due to its relatively easy development in terms of human effort, its robustness in face of non-grammatical input data (such as recog- nised speech), and its good results against rule-based and transfer-based approaches. The statistical approach to machine translation is based on the assumption that every sentence t in the target language is a possible translation of a given sentence s in the source language, and the main difference between two translation hypotheses is a probability assigned to each, which is to be learned from a bilingual corpus. The first SMT systems were based on the noisy channel approach on a word-based basis, modeling the translation of a tar- get language sentence t given a source language sentence t as a translation model probability p(s|t) times a target language model probability p(t) [1]. Recently, word-based translation models have been replaced by phrase-based translation models [2, 3], which are estimated from aligned bilingual corpora by using rel- ative frequencies. On the other hand, according to the maximum en- tropy framework [4], we can define the translation hy- pothesis t given a source sentence s, as the target sen- tence maximizing a log-linear combination of feature functions, as described in the following equation: ˆtI 1 = arg max tI 1 � M � m=1 λmhm(sJ 1 , tI 1) � (1) where λm correspond to the weighting coefficients of the log-linear combination, and the feature functions hm(s, t) to a logarithmic scaling of the probabilities of each model. Following this approach, the translation system de- scribed in this paper implements a log-linear combina- tion of one translation model and four additional fea- ture models. In contrast with standard phrase-based ap- proaches, our translation model is expressed in tuples as bilingual units. Given a word alignment, tuples define a unique and monotonic segmentation of each bilingual sentence, building up a much smaller set of units than with phrases and allowing N-gram estimation to account for the history of the translation …

Content truncated to 500 words