tailieunhanh - Báo cáo khoa học: "An Exploration of Forest-to-String Translation: Does Translation Help or Hurt Parsing?"

Syntax-based translation models that operate on the output of a source-language parser have been shown to perform better if allowed to choose from a set of possible parses. In this paper, we investigate whether this is because it allows the translation stage to overcome parser errors or to override the syntactic structure itself. We find that it is primarily the latter, but that under the right conditions, the translation stage does correct parser errors, improving parsing accuracy on the Chinese Treebank. . | An Exploration of Forest-to-String Translation Does Translation Help or Hurt Parsing Hui Zhang University of Southern California Department of Computer Science hzhang@ David Chiang University of Southern California Information Sciences Institute chiang@ Abstract Syntax-based translation models that operate on the output of a source-language parser have been shown to perform better if allowed to choose from a set of possible parses. In this paper we investigate whether this is because it allows the translation stage to overcome parser errors or to override the syntactic structure itself. We find that it is primarily the latter but that under the right conditions the translation stage does correct parser errors improving parsing accuracy on the Chinese Treebank. 1 Introduction Tree-to-string translation systems Liu et al. 2006 Huang et al. 2006 typically employ a pipeline of two stages a syntactic parser for the source language and a decoder that translates source-language trees into target-language strings. Originally the output of the parser stage was a single parse tree and this type of system has been shown to outperform phrase-based translation on for instance Chinese-to-English translation Liu et al. 2006 . More recent work has shown that translation quality is improved further if the parser outputs a weighted parse forest that is a representation of a whole distribution over possible parse trees Mi et al. 2008 . In this paper we investigate two hypotheses to explain why. One hypothesis is that forest-to-string translation selects worse parses. Although syntax often helps translation there may be situations where syntax or at least syntax in the way that our models use it can impose constraints that are too rigid for good-quality translation Liu et al. 2007 Zhang et al. 2008 . For example suppose that a tree-to-string system 317 encounters the following correct tree only partial bracketing shown 1 NP jingji zengzhang de sudu economy growth DE rate