Named Entity Recognition in Semi Structured Documents Using Neural Tensor Networks

Information Extraction and Named Entity Recognition algorithms derive major applications related to many practical document analysis system. Semi structured documents pose several challenges when it comes to extract relevant information from these documents. The state-of-the-art methods heavily rely on feature engineering to perform layout-specific extraction of information and therefore do not generalize well. Extracting information without taking the document layout into consideration is required as a first step to develop a general solution to this problem. To address this challenge, we propose a deep learning based pipeline to extract information from documents. For this purpose, we define 'information' to be a set of entities that have a label and a corresponding value, e.g., application number: ADNF8932NF and submission date: 15FEB19. We form relational triplets by connecting one entity to another via a relationship, such as (max temperature, is, 100 degrees) and train a neural tensor network that is well-suited for this kind of data to predict high confidence scores for true triplets. Up to 96% test accuracy on real world documents from publicly available GHEGA dataset demonstrate the effectiveness of our approach.