Latent Constraints on Unsupervised Text-Graph Alignment with Information Asymmetry

Unsupervised text-graph alignment (UTGA) is a fundamental task that bidirectionally generates texts and graphs without parallel data. Most available models of UTGA suffer from information asymmetry, common phenomenon include additional invisible to each other. On the one hand, these fail supplement asymmetric effectively due lack ground truths. other it challenging indicate with explicit indicators because cannot be decoupled data directly. To address challenge posed by we propose assumption encoded in unobservable latent variables only affects one-way generation processes. These corresponding should obey prior distributions recovered approximately original Therefore, first taxonomy variable classifies into transferrable (TV) non-transferable (NTV) further distinguish NTV as dependent (DV) independent (IV). Next, three VAE-based regularizations on TV, DV, IV constrain their well-designed introduce enhance preservation shared contents. Finally, impose proposed constraints cycle-consistent learning framework, back-translation (BT), named ConstrainedBT. Experimental results tasks demonstrate effectiveness ConstrainedBT information-asymmetric challenge.