A Smorgasbord of Features for Statistical Machine Translation

A Smorgasbord of Features for Statistical Machine Translation Franz Josef Och USC/ISI Daniel Gildea U. of Rochester Sanjeev Khudanpur Johns Hopkins U. Anoop Sarkar Simon Fraser U. Kenji Yamada Xerox/XRCE Alex Fraser USC/ISI Shankar Kumar Johns Hopkins U. Libin Shen U. of Pennsylvania David Smith Johns Hopkins U. Katherine Eng Stanford U. Viren Jain U. of Pennsylvania Zhen Jin Mt. Holyoke Dragomir Radev U. of Michigan Abstract We describe a methodology for rapid exper- imentation in statistical machine translation which we use to add a large number of features to a baseline system exploiting features from a wide range of levels of syntactic representation. Feature values were combined in a log-linear model to select the highest scoring candidate translation from an n-best list. Feature weights were optimized directly against the BLEU eval- uation metric on held-out data. We present re- sults for a small selection of features at each level of syntactic representation. 1 Introduction Despite the enormous progress in machine translation (MT) due to the use of statistical techniques in recent years, state-of-the-art statistical systems often produce translations with obvious errors. Grammatical errors in- clude lack of a main verb, wrong word order, and wrong choice of function words. Frequent problems of a less grammatical nature include missing content words and incorrect punctuation. In this paper, we attempt to address these problems by exploring a variety of new features for scoring candidate translations. A high-quality statistical translation system is our baseline, and we add new features to the exist- ing set, which are then combined in a log-linear model. To allow an easy integration of new features, the base- line system provides an n-best list of candidate transla- tions which is then reranked using the new features. This framework allows us to incorporate different types of fea- tures, including features based on syntactic analyses of the source and target sentences, which we hope will ad- dress the grammaticality of the translations, as well as lower-level features. As we work on n-best lists, we can easily use global sentence-level features. We begin by describing our baseline system and the n-best rescoring framework within which we conducted our experiments. We then present a selection of new fea- tures, progressing from word-level features to those based to part-of-speech tags and syntactic chunks, and then to features based on Treebank-based syntactic parses of the source and target sentences. 2 Log-linear Models for Statistical MT The goal is the translation of a text given in some source language into a target language. We are given a source (‘Chinese’) sentence f = f J 1 = f1, . . . , fj, . . . , fJ, which is to be translated into a target (‘English’) sentence e = eI 1 = e1, . . . , ei, . . . , eI Among all possible target sentences, we will choose the sentence with the highest probability: ˆeI 1 = argmax eI 1 {Pr(eI 1|f J 1 )} (1) As an alternative to the often …

Content truncated to 500 words