Retrieve, Caption, Generate: Visual Grounding for Enhancing Commonsense in Text Generation Models

Steven Y. Feng, Kevin Lu, Zhuofu Tao, Malihe Alikhani, Teruko Mitamura, Eduard H Hovy, Varun Gangal

Keywords: NLG, Language generation, Concept-to-text, CommonGen, Commonsense, Grounding, Multimodal, Retrieve and generate, Transformers

TL;DR: We introduce a visually-grounded approach to enhance the commonsense of text generation models called VisCTG: Visually-Grounded Concept-to-Text Generation to improve upon BART and T5 baselines on the CommonGen task.

Abstract: We investigate the use of multimodal information contained in images as an effective method for enhancing the commonsense of Transformer models for text generation. We perform experiments using BART and T5 on concept-to-text generation, specifically the task of generative commonsense reasoning, or CommonGen. We call our approach VisCTG: Visually Grounded Concept-to-Text Generation. VisCTG involves captioning images representing appropriate everyday scenarios, and using these captions to enrich and steer the generation process. Comprehensive evaluation and analysis demonstrate that VisCTG noticeably improves model performance while successfully addressing several issues of the baseline generations, including poor commonsense, fluency, and specificity.

Community Implementations: [ 2 code implementations](https://www.catalyzex.com/paper/retrieve-caption-generate-visual-grounding/code)