AMR Quality Rating with a Lightweight CNN

Proceedings of the 1st Conference of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 10th International Joint Conference on Natural Language Processing, pages 235–247 December 4 - 7, 2020. c⃝2020 Association for Computational Linguistics 235 AMR Quality Rating with a Lightweight CNN Juri Opitz Dept. of Computational Linguistics Heidelberg University 69120 Heidelberg opitz@cl.uni-heidelberg.de Abstract Structured semantic sentence representations such as Abstract Meaning Representations (AMRs) are potentially useful in various NLP tasks. However, the quality of automatic parses can vary greatly and jeopardizes their usefulness. This can be mitigated by models that can accurately rate AMR quality in the ab- sence of costly gold data, allowing us to in- form downstream systems about an incorpo- rated parse’s trustworthiness or select among different candidate parses. In this work, we propose to transfer the AMR graph to the domain of images. This allows us to create a simple convolutional neural network (CNN) that imitates a human judge tasked with rating graph quality. Our experi- ments show that the method can rate quality more accurately than strong baselines, in sev- eral quality dimensions. Moreover, the method proves to be efficient and reduces the incurred energy consumption. 1 Introduction The goal of sentence meaning representations is to capture the meaning of sentences in a well-defined format. One of the most prominent frameworks for achieving this is Abstract Meaning Represen- tation (AMR) (Banarescu et al., 2013). In AMR, sentences are represented as directed acyclic and rooted graphs. An example is displayed in Figure 1, where we see three equivalent displays of an AMR that represents the meaning of the sentence “The baby is sleeping well”. In AMR, nodes are vari- ables or concepts, while (labeled) edges express their relations. Among other phenomena, this al- lows AMR to capture coreference (via re-entrant structures) and semantic roles (via :argn relation). Furthermore, AMR links sentences to KBs: e.g., predicates are mapped to PropBank (Palmer et al., 2005; Kingsbury and Palmer, 2002), while named w s b sleep-1 baby arg0 mod ( s / sleep-01 :arg0 (b / baby ) :mod (w / well )) Penman notation well { <s, instance, sleep>, <s, arg0, b>, <b, instance, baby>, <w, instance, well>, <s, mod, w> } Triples Figure 1: Equivalent representations of the AMR for “The baby is sleeping well”. (p4 / possible-01 :arg1 (d5 / destabilize-01 :arg0 [:arg1] (c3 / country :quant (w2 / whole))) :condition (e1 / economy [:poss c3] :arg0-of (f0 / function-01 [:pol -] ))) Figure 2: Parse of Without a functioning economy, the whole country may destabilize with errors outlined. entities are linked to Wikipedia. From a logical perspective, AMR is closely related to first-order logic (FOL, see Bos (2016, 2019) for translation mechanisms). Currently, AMRs are leveraged to enhance a va- riety of natural language understanding tasks. E.g., they have enhanced commonsense reasoning and question answering (Mitra and Baral, 2016), ma- chine translation (Song et al., 2019), text summa- rization (Liao et al., 2018; Dohare et al., 2017) and paraphrasing (Issa et …

Content truncated to 500 words