Robust Sub-Sentential Alignment of Phrase-Structure Trees

Robust Sub-Sentential Alignment of Phrase-Structure Trees Declan Groves School of Computing Dublin City University Dublin 9, Ireland dgroves@computing.dcu.ie Mary Hearne School of Computing Dublin City University Dublin 9, Ireland mhearne@computing.dcu.ie Andy Way School of Computing Dublin City University Dublin 9, Ireland away@computing.dcu.ie Abstract Data-Oriented Translation (DOT), based on Data- Oriented Parsing (DOP), is a language-independent MT engine which exploits parsed, aligned bitexts to produce very high quality translations. How- ever, data acquisition constitutes a serious bottleneck as DOT requires parsed sentences aligned at both sentential and sub-structural levels. Manual sub- structural alignment is time-consuming, error-prone and requires considerable knowledge of both source and target languages and how they are related. Au- tomating this process is essential in order to carry out the large-scale translation experiments necessary to assess the full potential of DOT. We present a novel algorithm which automatically in- duces sub-structural alignments between context-free phrase structure trees in a fast and consistent fash- ion requiring little or no knowledge of the language pair. We present results from a number of experi- ments which indicate that our method provides a se- rious alternative to manual alignment. 1 Introduction Approaches to Machine Translation (MT) using Data-Oriented Parsing (DOP: (Bod, 1998; Bod et al., 2003)) require ⟨source,target⟩ tree fragments aligned at sentential and sub-sentential levels. In previous approaches to Data-Oriented Translation (DOT: (Poutsma, 2000; Hearne and Way, 2003)), such fragments were produced manually. This is time-consuming, error-prone, and requires consid- erable expertise of both source and target languages as well as how they are related. The obvious solu- tion, therefore, is to automate the process of sub- sentential alignment. However, while there are many approaches to sentential alignment e.g. (Kay and R¨oscheisen, 1993; Gale & Church, 1993), no methods exist for aligning non-isomorphic phrase- structure (PS) tree fragments at sub-sentential level for use in MT. (Matsumoto et al., 1993) align ⟨source,target⟩ dependency trees, with a view to re- solve parsing ambiguities, but their approach can- not deal with complex or compound sentences. Other researchers (Imamura, 2001) also use phrase- alignment in parsing but in DOT the translation fragments are already in the form of parse-trees. (Eisner, 2003) outlines a computationally expensive structural manipulation tool which he has used for intra-lingual translation but has yet to apply to inter- lingual translation. (Gildea, 2003) performs tree-to- tree alignment, but treats it as part of a generative statistical translation model, rather than a seperate task. The method of (Ding et al., 2003) can cope with a limited amount of non-isomorphism, but the algorithm is only suitable for use with dependency trees. We develop a novel algorithm which automati- cally aligns translationally equivalent tree fragments in a fast and consistent fashion, and which requires little or no knowledge of the language pair. Our ap- proach is similar to that of (Menezes and Richard- son, 2003), who use a best-first approach to align dependency-type tree structures. We conduct a number of experiments on the English-French section of the Xerox HomeCentre corpus. Using the manual alignment of …

Content truncated to 500 words