Converting Structured Forms to Storable Databases using Form2DB
A. Anbarasi, S. Jagadeesan, K. C. Nithyasree · 2023
Document understanding is a set of related tasks involving understanding of semantic syntactic as well as visual structure of document images. Form understanding is a subset of document understanding, which specifically deals with semantic entity relationship between elements in a form, with no prior information on form template. There are multiple models that have been developed to solve this task, however they tend to be languagespecific and rely heavily on OCR/HWR for functioning. Thus, we create a pipeline for converting form images into relational database (like SQL) data entries. We further try to implement the training process in a language-agnostic fashion. The questionanswer pairs are extracted from forms and passed to LLM, which can handle complex text-to-SQL conversion queries. We compare the performance of various modules in a qualitative manner, analysing the different errors and issues arising. We also use existing benchmarks to compare the performances of various models.