Grammar Formalisms beyond Context-Free Grammars and their use for Machine Learning Tasks

点击翻译按钮获取中文摘要

The aim of this project is to develop new models and algorithms forprobabilistic parsing and for statistical machine translation, usingformalisms such as Linear Context-Free Rewriting Systems (LCFRSs),i.e., extensions of Context-Free Grammars that combine aspects ofsynchronous grammars with the capacity to describe discontinuities.Both, parsing and machine translation, face the challenge that theyhave to deal with non-local dependencies, to put it differently, withdiscontinuous units in a sentence that belong together. One way toaccount for these non-local phenomena is to use formalisms that candescribe more than just context-free structures. In the first part ofthe project, this has already been successfully done in parsing, wherewe use LCFRS. We plan to continue this line of research in the secondpart of the project. Concerning syntax-based statistical machinetranslation, we have shown in the first part of the project thatsynchronous LCFRS can describe the sometimes complex alignments oneencounters in parallel corpora. In the second part of the project wetherefore aim at completing the (already started) implementation of astatistical machine translation system based on synchronous LCFRS.