Joint Intent Classification and Slot Tagging on Agricultural Dataset for Indic Languages

Agriculture makes up for about 50% of the total workforce in India while contributing to approximately 20% of the Gross Domestic Product (GDP) as of 2021. Due to the lack of resources in regional languages for farmers, there is a dire need for a corpus that spans multiple topics relating to agriculture. The corpus can be developed by collecting data via web scraping. The existence of such a corpus must be accompanied by a classification system that could direct the farmers along the database. For this objective, the proposed work concentrates on designing an intent classification dataset, Query-based Agricultural Data System (QuADS) along with a high accuracy model that could be implemented for identifying what the input intent and slots are, and further be expanded to provide output using voice recognition. This paper mainly focuses on comparing the Long Short-Term Memory (LSTM) and Bidirectional Encoder Representations from Transformers (BERT) models for intent recognition, which returned an accuracy of 93.89% and 98.32% on QuADS respectively. It further demonstrates the application of intent classification in the domain of a query-based voice assistant using google interfaced with python APIs for ease of farmer interaction.