Multimodal topic modeling aims to learn interpretable topics from collections that combine text with other modalities. A persistent difficulty is lexical incompleteness: the text often records only part of the evidence needed to describe a sample, and concepts that are visible in images may have no direct path into the learned word distributions. In many multimodal topic models, visual features mainly help infer document-level topic mixtures, whereas the decoder is still trained on the incomplete observed text. We introduce TIGNTM, a graph completion framework that turns visual evidence into lexical supervision. TIGNTM builds a corpus graph over textual and visual representations, treats related samples as weak lexical witnesses, and produces sparse completion targets for words missing from the observed text. These targets supervise the same decoder that generates topic words, so visual cues can shape the word distributions that define each topic. To match the geometry of normalized multimodal embeddings, TIGNTM adopts hyperspherical neural topic learning. Experiments on multiple multimodal datasets show consistent improvements in coherence, diversity, and overall topic quality compared with strong neural and multimodal baselines.

“Relation Extraction (RE) task aims to discover the semantic relation that holds between two entitiesand contributes to many applications such as knowledge graph construction and completion. Reinforcement Learning (RL) has been widely used for RE task and achieved SOTA results, whichare mainly designed with rewards to choose the optimal actions during the training procedure,to improve RE’s performance, especially for low-resource conditions. Recent work has shownthat offline or online RL can be flexibly formulated as a sequence understanding problem andsolved via approaches similar to large-scale pre-training language modeling. To strengthen theability for understanding the semantic signals interactions among the given text sequence, thispaper leverages Transformer architecture for RL-based RE methods, and proposes a genericframework called Transformer Enhanced RL (TERL) towards RE task. Unlike prior RL-basedRE approaches that usually fit value functions or compute policy gradients, TERL only outputsthe best actions by utilizing a masked Transformer. Experimental results show that the proposedTERL framework can improve many state-of-the-art RL-based RE methods.”