Phrase-Projection for Cross-Lingual JAMR training

Adi Renduchinala
View Count
Creative Commons CC BY 4.0
We are given spans of the target text which align to concepts in the AMR graph.These alignment do not cover every token in the target sentnce. Typically function words are not aligned to any graph fragment. Next, we obtain word alignments between the target sentence and source sentence. Since we have word alignments between target and source, and phrase alignments between target and AMR graph, we must convert the word alingments into phrase alignments. The phrases on the source side will then be projected to the AMR concepts via the target sentence
Phrase-Projection for Cross-Lingual JAMR training