Recent works on form understanding mostly employ multimodal transformers or large-scale pre-trained language models. These models need ample data for pre-training. In contrast, humans can usually identify key-value pairings from a form only by looking at layouts, even if they don't comprehend the language used. No prior research has been conducted to investigate how helpful layout information alone is for form understanding. Hence, we propose a unique entity-relation graph parsing method for scanned forms called LAGNN, a language-independent Graph Neural Network model. Our model parses a form into a word-relation graph in order to identify entities and relations jointly and reduce the time complexity of inference. This graph is then transformed by deterministic rules into a fully connected entity-relation graph. Our model simply takes into account relative spacing between bounding boxes from layout information to facilitate easy transfer across languages. To further improve the performance of LAGNN, and achieve isomorphism between entity-relation graphs and word-relation graphs, we use integer linear programming (ILP) based inference. Code is publicly available at https://github.com/Bhanu068/LAGNN

The ability to recognize analogies is an important factor that is closely related to human intelligence. Verbal analogies have been used for evaluating both examinees at university entrance exams as well as algorithms for measuring relational similarity. However, relational similarity measures proposed so far are confined to measuring the similarity be- tween pairs of words. Unfortunately, such pairwise approaches ignore the rich relational structure that exists in real-world knowledge bases containing millions of entities and semantic relations. We pro- pose a method to efficiently identify analogous en- tity tuples from a given entity-relation graph. First, we present an efficient approach for extract- ing potential analogous tuples from a given entity- relation graph. Second, to measure the structural similarity between two tuples, we propose two types of kernel functions: vertex-feature kernels, and edge-feature kernels. Moreover, we combine those kernels to construct composite kernels that simultaneously consider both vertex and edge fea- tures. Experimental results show that our proposed method accurately identifies analogous tuples and significantly outperforms a state-of-the-art pairwise relational similarity measure, extended to tuples.