Aligning English Strings with Abstract Meaning Representation Graphs

Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 425–429, October 25-29, 2014, Doha, Qatar. c⃝2014 Association for Computational Linguistics Aligning English Strings with Abstract Meaning Representation Graphs Nima Pourdamghani, Yang Gao, Ulf Hermjakob, Kevin Knight Information Sciences Institute Department of Computer Science University of Southern California {damghani,yanggao,ulf,knight}@isi.edu Abstract We align pairs of English sentences and corresponding Abstract Meaning Repre- sentations (AMR), at the token level. Such alignments will be useful for downstream extraction of semantic interpretation and generation rules. Our method involves linearizing AMR structures and perform- ing symmetrized EM training. We obtain 86.5% and 83.1% alignment F score on de- velopment and test sets. 1 Introduction Banarescu et al. (2013) describe a semantics bank of English sentences paired with their logical meanings, written in Abstract Meaning Represen- tation (AMR). The designers of AMR leave open the question of how meanings are derived from English sentences (and vice-versa), so there are no manually-annotated alignment links between English words and AMR concepts. This paper studies how to build such links automatically, us- ing co-occurrence and other information. Auto- matic alignments may be useful for downstream extraction of semantic interpretation and genera- tion rules. AMRs are directed, acyclic graphs with labeled edges, e.g., the sentence The boy wants to go is represented as: (w / want-01 :arg0 (b / boy) :arg1 (g / go-01 :arg0 b)) We have hand-aligned a subset of the 13,050 available AMR/English pairs. We evaluate our automatic alignments against this gold standard. A sample hand-aligned AMR is here (“˜n” speci- fies a link to the nth English word): the boy wants to go (w / want-01˜3 :arg0 (b / boy˜2) :arg1 (g / go-01˜5 :arg0 b)) This alignment problem resembles that of statisti- cal machine translation (SMT). It is easier in some ways, because AMR and English are highly cog- nate. It is harder in other ways, as AMR is graph- structured, and children of an AMR node are un- ordered. There are also fewer available training pairs than in SMT. One approach is to define a generative model from AMR graphs to strings. We can then use EM to uncover hidden derivations, which align- ments weakly reflect. This approach is used in string/string SMT (Brown et al., 1993). How- ever, we do not yet have such a generative graph- to-string model, and even if we did, there might not be an efficient EM solution. For exam- ple, in syntax-based SMT systems (Galley et al., 2004), the generative tree/string transduction story is clear, but in the absence of alignment con- straints, there are too many derivations and rules for EM to efficiently consider. We therefore follow syntax-based SMT custom and use string/string alignment models in align- ing our graph/string pairs. However, while it is straightforward to convert syntax trees into strings data (by taking yields), it is not obvious how to do this for unordered AMR graph elements. The ex- ample above also shows that gold alignment links reach into the internal …

Content truncated to 500 words