Abstract Meaning Representation for Multi-Document Summarization

Proceedings of the 27th International Conference on Computational Linguistics, pages 1178–1190 Santa Fe, New Mexico, USA, August 20-26, 2018. 1178 Abstract Meaning Representation for Multi-Document Summarization Kexin Liao Logan Lebanoff Computer Science DepartmentFei Liu University of Central Florida, Orlando, FL 32816, USA {ericaryo,loganlebanoff}@knights.ucf.edu feiliu@cs.ucf.edu Abstract Generating an abstract from a collection of documents is a desirable capability for many real- world applications. However, abstractive approaches to multi-document summarization have not been thoroughly investigated. This paper studies the feasibility of using Abstract Meaning Rep- resentation (AMR), a semantic representation of natural language grounded in linguistic theory, as a form of content representation. Our approach condenses source documents to a set of sum- mary graphs following the AMR formalism. The summary graphs are then transformed to a set of summary sentences in a surface realization step. The framework is fully data-driven and flexi- ble. Each component can be optimized independently using small-scale, in-domain training data. We perform experiments on benchmark summarization datasets and report promising results. We also describe opportunities and challenges for advancing this line of research. 1 Introduction Abstractive summarization seeks to generate concise and grammatical summaries that preserve the mean- ing of the original; further, they shall abstract away from the source syntactic forms. The task often in- volves high-level text transformations such as sentence fusion, generalization, and paraphrasing (Jing and McKeown, 1999). Recent neural abstractive summarization studies focus primarily on single-document summarization (Paulus et al., 2017; See et al., 2017). These approaches are limited by the availability of training data, and large datasets for multi-document summarization can be costly to obtain. Generating abstractive summaries for sets of source documents thus remains a challenging task. Traditional approaches to abstractive summarization often condense the source documents to a set of “semantic units,” then reconstruct abstractive summaries from these semantic units. Previous work has investigated various forms of content representation. Examples include noun/verb phrases (Genest and Lapalme, 2011; Bing et al., 2015), word-occurrence graphs (Ganesan et al., 2010), syntactic parse trees (Cheung and Penn, 2014; Gerani et al., 2014), and domain-specific templates (Pighin et al., 2014). Nonetheless, generating summary text from these heuristic forms of representation can be difficult. There is an increasing need to exploit a semantic formalism so that condensing source documents to this form and generating summary sentences from it can both be carried out in a principled way. This paper explores Abstract Meaning Representation (AMR, Banarescu et al., 2013) as a form of con- tent representation. AMR is a semantic formalism based on propositional logic and the neo-Davidsonian event representation (Parsons, 1990; Schein, 1993). It represents the meaning of a sentence using a rooted, directed, and acyclic graph, where nodes are concepts and edges are semantic relations. Figure 1 shows an example AMR graph. A concept node can be a PropBank frameset (“state-01”), an English word (“warhead”), a special keyword (“date-entity”), or a string literal (“Japan”). A relation can be either a core argument (“ARG0,” “ARG1”) or a modification relationship (“mod,” “time” ). The AMR representation abstracts …

Content truncated to 500 words