Automatic Determination of Number of clusters for creating Templates in Example-Based Machine Translation

Automatic Determination of Number of clusters for creating templates in Example-Based Machine Translation Rashmi Gangadharaiah rgangadh@cs.cmu.edu Ralf D. Brown ralf@cs.cmu.edu Language Technologies Institute, Carnegie Mellon University, Pittsburgh, PA 15213. Jaime Carbonell jgc@cs.cmu.edu Abstract Example-Based Machine Translation (EBMT), like other corpus based methods, requires substantial parallel training data. One way to reduce data requirements and improve translation quality is to generalize parts of the parallel corpus into translation templates. This automated generalization process requires clustering. In most clustering approaches the optimal number of clusters (N) is found empirically on a tune set which often takes several days. This paper introduces a spectral clustering framework that automatically estimates the optimal N and removes unstable oscillating points. The new framework produces significant improve- ments in low-resource EBMT settings for English-to-French (≈1.4 BLEU points), English-to-Chinese (≈1 BLEU point), and English-to-Haitian (≈2 BLEU points). The translation quality with templates cre- ated using automatically and empirically found best N were almost the same. By discarding “incoherent” points, a further boost in translation scores is observed, even above the empirically found N. 1 Introduction An EBMT system uses a parallel corpus to trans- late new input source sentences. In the Translation Model (TM), the input sentence to be translated is matched against the source sentences. When a match is found, the corresponding translation in the target language is obtained through sub- sentential alignment. In our EBMT system, the final target translation is obtained from these par- tial target translations with a beam-search decoder using a target Language Model (LM). EBMT sys- tems require large amounts of data to function well (Brown, 2000). c⃝ 2010 European Association for Machine Translation. Generalization using equivalence classes (Veale and Way, 1997; Brown, 2000) reduces the amount of pre-translated text required and improves trans- lation quality. Translation templates (or short reusable sequences) are generalizations of source and target sentences where sequences of one or more words are replaced by variables. Various methods have been proposed to create such tem- plates in EBMT and differ in the way the tem- plates are created. Some systems required a full template to match the input source sentence for target sentence generation (Cicekli and G¨uvenir, 2001), others adopted a fragmentary translation algorithm (Kaji et al., 1992) where the target sentence generation is similar to that adopted in Transfer-based MT systems. Somers (1994) sug- gests using ‘hooks’ using alignment information and Block (2000) uses a simple translation algo- rithm that can join only single variable target frag- ments. Gaijin (Veale and Way, 1997) performs phrase chunking based on the marker hypothesis. Templates resemble transfer rules used in Rule- Based MT (Lavie, 2008) but have fewer con- straints. Syntax-based SMT (Yamada and Knight, 2001) also uses transfer rules on parsed input that contain only non-terminal symbols and use lexical transfer rules for translating the source words. In EBMT, lexical transfer rules consist only of termi- nal symbols and a generalization template contains non-terminal and terminal symbols. Templates provide a way to reorder target phrasal matches that are not …

Content truncated to 500 words