Hybrid Context Aware Gujarati Spell Correction Using Norvig Algorithm, GRU, and IndicBERT

Abstract

Numerous applications in the domain of Natural Language Processing (NLP) rely on spelling and grammatical checks, including email, opinion mining, text summarization, chatbots, and countless more. An individual’s credibility, cybersecurity efforts, legal ambiguities, and NLP application performance can all take a hit if they make a mistake when dealing with regional languages such as Assamese, Gujarati, Hindi, etc. In order to lessen the frequency of spelling errors, this article examines and concentrates on Gujarati. In addition to a thorough examination of issues related to the Gujarati language, this article provides up-to-date strategies for fixing spelling mistakes based on context of the word. A novel hybrid approach ensures top-notch Gujarati context aware spelling verification. After thoroughly considering all the suggestions, we used a two-layer GRU network and the IndicBERTv2-SS model, which was fine-tuned only on our curated Gujarati dataset of about 20,000 sentences (70/15/15 split into training, validation, and test), to choose the best correction while keeping the context in mind. Normalization for Gujarati (diacritics, compound characters, and numbers), regex-based tokenization, and edit-distance candidate creation were all part of preprocessing. We used accuracy, precision, and recall to assess the test split. Our proposed IndicBERT-GUJBRIJAPU tool got 93.49% accuracy, 94.46% precision, and 91.59% recall, which is much better than other approaches for context-aware correction.