English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Pre-trained sequence-to-sequence language models have led to widespreadsuccess in many natural language generation tasks. However, there has beenrelatively less work on analyzing their ability to generate structured outputssuch as graphs. Unlike natural language, graphs have distinct structural andsemantic properties in the context of a downstream NLP task, e.g., generating agraph that is connected and acyclic can be attributed to its structuralconstraints, while the semantics of a graph can refer to how meaningfully anedge represents the relation between two node concepts. In this work, we studypre-trained language models that generate explanation graphs in an end-to-endmanner and analyze their ability to learn the structural constraints andsemantics of such graphs. We first show that with limited supervision,pre-trained language models often generate graphs that either violate theseconstraints or are semantically incoherent. Since curating large amount ofhuman-annotated graphs is expensive and tedious, we propose simple yeteffective ways of graph perturbations via node and edge edit operations thatlead to structurally and semantically positive and negative graphs. Next, weleverage these graphs in different contrastive learning models with Max-Marginand InfoNCE losses. Our methods lead to significant improvements in bothstructural and semantic accuracy of explanation graphs and also generalize toother similar graph generation tasks. Lastly, we show that human errors are thebest negatives for contrastive learning and also that automatically generatingmore such human-like negative graphs can lead to further improvements. Our codeand models are publicly available at https://github.com/swarnaHub/ExplagraphGen

Explanation Graph Generation via Pre-trained Language Models: An Empirical Study with Contrastive Learning