The TALP ngram-based SMT system for IWSLT 2007

The TALP Ngram-based SMT System for IWSLT 2007 Patrik Lambert, Marta R. Costa-juss`a, Josep M. Crego, Maxim Khalilov, Jos´e B. Mari˜no, Rafael E. Banchs, Jos´e A.R. Fonollosa and Holger Schwenk1 TALP Research Center Universitat Polit`ecnica de Catalunya, Barcelona {lambert|mruiz|jmcrego|khalilov|canton|rbanchs|adrian}@talp.upc.edu 1 LIMSI-CNRS, BP 133 91403 Orsay Cedex schwenk@limsi.fr Abstract This paper describes TALPtuples, the 2007 N-gram-based statistical machine translation system developed at the TALP Research Center of the UPC (Universitat Polit`ecnica de Catalunya) in Barcelona. Emphasis is put on improvements and extensions of the system of previous years. Mainly, these include optimizing alignment parameters in function of translation metric scores and rescoring with a neural network language model. Results on two translation directions are reported, namely from Arabic and Chinese into English, thoroughly explaining all language-related preprocessing and translation schemes. 1. Introduction Ngram-based Machine Translation (MT), originally based in the Finite-State Transducers approach to Statistical MT (SMT) [1, 2], has proved to be a competitive alternative to phrase-based and other state-of-the-art systems in previous evaluation campaigns, as shown in [3, 4]. Efforts have been focused on improving translation ac- cording to human evaluation by further developing different stages of the SMT system: alignment and rescoring. As in previous years, we aligned the training corpus us- ing Giza++ software. However, instead of keeping the de- fault parameters, we performed a minimum translation error training procedure to adjust Giza++ smoothing parameters to the task. This procedure had been successful with an align- ment system based on discriminative training [5]. For the rescoring we incorporate a neural network lan- guage model as previously experienced in [6]. The neural network language model mainly is able to produce a better generalization in the translation system. This paper is organized as follows. Section 2 briefly re- views last year’s system, including tuple definition and ex- traction, translation model and feature functions, decoding tool and reordering and optimization criterion. Section 3 describes the alignment translation-minimum-error training procedure. Section 4 focuses on rescoring using a neural lan- guage model (NNLM). Next, Section 5 reports on all exper- iments carried out from Arabic and Chinese into English for IWSLT 2007. Finally, Section 6 sums up the main conclu- sions from the paper. 2. Baseline description 2.1. N-gram-based Machine Translation The TALP Ngram-based SMT system performs a log-linear combination of a translation model and additional feature functions (see further details in [7, 8]). In contrast to phrase- based models, our translation model is estimated as a stan- dard n-gram model of a bilingual language expressed in tu- ples. In this way, it approximates the joint probability be- tween source and target languages capturing bilingual con- text, as described by the following equation: p(S, T) = K � k=1 p((˜s, ˜t)k|(˜s, ˜t)k−N+1, ..., (˜s, ˜t)k−1) (1) where s refers to source, t to target, and (˜s, ˜t)k to the kth tuple of a given bilingual sentence pair segmented in K tuples. 2.2. Tuple extraction Given a certain word-aligned parallel corpus, tuples are ex- tracted according to the following constraints [9]: • a …

Content truncated to 500 words