The Trouble with Multiple Annotations Sample Clauses

The Trouble with Multiple Annotations. ‌ When the set of translation rules that can be extracted from a sentence pair depends on both a word alignment and a syntactic parse tree, it is important that the two annotations agree on what consitutes a valid phrase. For example, in Figure 1.3, the English phrase “The first step” both has a coherent phrasal Chinese translation (“ ,— t”) and forms a constituent in the English parse tree. Thus, it is possible to extract a corresponding syntactic translation rule. However, lack of agreement between the parse tree and the word alignment is a frequently recurring problem. For example, in the same sentence pair, the VP in red cannot appear as a translation rule because the corresponding English phrase (“select team members”) does not map nicely onto any Chinese phrase according to the word alignments. Conversely, there is a correspondence between the English phrase “to select” and the Chinese word “ ”, and a phrase-based system would extract that phrase pair, but since “to select” does not make up a constituent in the English parse tree, no simple syntactic rule can be extracted here either.5 There has been a large amount of previous work that attempts to address this problem of incompatibility between word alignments and syntactic annotations. Some systems at- tempt to directly model the divergence between source and target language syntax (Yamada and Knight, 2001; Xxxxxx, 2003), but most recent work focuses on reducing the sensitiv- ity of the rule-extraction procedure to the constituency decisions made by 1-best syntactic 5It is possible to extract a more complicated rule for a VP that needs the object NP to be filled, but the simpler rule is more widely applicable and thus would also be independently useful. parsers, either by using forest-based methods for learning translation rules (Mi and Xxxxx, 0000; Zhang et al., 2009), or by learning rules that encode syntactic information but do not strictly adhere to constituency boundaries (Zollmann et al., 2006; Marton and Xxxxxx, 2008; Chiang, 2010). Instead of making improvements to specific syntactic MT systems, the approach we take in the present work is to directly modify the input annotations in order to improve downstream MT performance in a way that is independent of the specific syntactic MT system being used. There has also been previous work of this sort, particularly related to word alignments. XxXxxx and Xxxxx (2007) built a word alignment model that conditioned on target-side syntax, while May an...