Evaluating the Impact of OCR Quality on Short Texts Classification Task

Abstract

AbstractThe majority of text classification algorithms have been developed and evaluated for texts written by humans and originated in text mode. However, in the contemporary world with an abundance of smartphones and readily available cameras, the ever-increasing amount of textual information comes from the text captured on photographed objects such as road and business signs, product labels and price tags, random phrases on t-shirts, the list can be infinite. One way to process such information is to pass an image with a text in it through an Optical Character Recognition (OCR) processor and then apply a natural language processing (NLP) system to that text. However, OCR text is not quite equivalent to the ‘natural’ language or human-written text because spelling errors are not the same as those usually committed by humans. Implying that the distribution of human errors is different from the distribution of OCR errors, we compare how much and how it affects the classifiers. We focus on deterministic classifiers such as fuzzy search as well as on the popular Neural Network based classifiers including CNN, BERT, and RoBERTa. We discovered that applying spell corrector on OCRed text increases F1 score by 4% for CNN and by 2% for BERT.KeywordsNLPOCRText classificationMulti-class classificationCNNBERTRoBERTaFuzzy searchShort texts