Sehda S2MT: Incorporation of Syntax into Statistical Translation System

Sehda S2MT: Incorporation of Syntax into Statistical Translation System Yookyung Kim, Jun Huang, Youssef Billawala, Demitrios Master, Farzad Ehsani Sehda Inc. 465 N. Fairchild Drive, Suite 123 Mountain View, CA 94043, USA {kim,jun}@Sehda.com Abstract This paper describes Sehda’s S2MT (Syntactic Statistical Machine Translation) system submitted to the Korean-English track in the evaluation campaign of the IWSLT-05 workshop. The S2MT is a phrase-based statistical system trained on linguistically processed parallel data. 1. Introduction Sehda’s S2MT (Syntactic Statistical Machine Translation) system is a hybrid system which incorporates linguistic knowledge into statistical learning. The system learns phrase- to-phrase mapping and syntactic ordering separately. A feasibility test of the system is performed on the translation task presented by International Workshop on Spoken Language Translation (IWSLT) for Korean-to-English “Supplied+Tools” data track. We show that syntactic phrases are useful units to handle the structural problems of statistical Machine Translation and reduce the need for huge parallel text corpora. 2. Overview of S2MT Our system capitalizes on the intuition that language is broadly divided into two levels: structure and vocabulary. Structure is the syntax or relation among phrases that govern the formation of complex structures in a language. Vocabulary is the word-level representation of individual concepts in a language. In traditional approaches to Statistical Machine Translation (SMT), the system learns both types of information simultaneously. By separating the acquisition of structural information from the acquisition of vocabulary, however, an SMT system can learn both levels more easily and more efficiently. By modifying the existing corpus to isolate structure and vocabulary, we are able to take full advantage of all of the information content of the bilingual corpus, ultimately producing higher quality machine translation with less training data. We separate the two levels of translation information by “chunking” [1] [2] the sentences in the bilingual corpus. Chunking is the process of separating the sentences into contiguous, structurally significant groups, such as noun phrases, verbal clusters, and prepositional phrases.1 In contrast to full syntactic parsing employed in [5] [6], chunking is flexible enough to handle the ungrammaticalities of 1 We use “chunks” and “phrases” interchangeably, unless otherwise noted. conversational data and provide us with syntactic information useful in handling structural issues [3]. Two learning passes are then performed2: one at the sentence level composed of phrase sequences to handle phrase reordering, and the other at the phrase level composed of word sequences to learn phrase translation properties. The results of the two learning passes are merged in the decoding step to produce translations, as shown in Figure 1. Translated Text Chunk Text Input Text Translate Phrases Reorder Phrases Decode Language Model Sentence-level Phrase-level Figure 1: Overview of S2MT The system is composed of four modules: (1) linguistic analyzer (2) phrase translation module (3) phrase reordering module (4) decoding module which integrates phrases into a sentence The linguistic analyzer consists of a parts-of-speech (POS) tagger, a morphological analyzer3, a chunk parser, and a head word extractor. Except for a Korean morphological analyzer and Brill’s English POS tagger, we use in-house …

Content truncated to 500 words