Robustness to Spelling Errors for Intent Detection
Ahmet Birim, Mustafa Erden · 2022 30th Signal Processing and Communications Applications Conference (SIU) · 2022
Intent detection is determining the requested action of the user from the query text in chatbot applications. Generally, the query for this task is comprised of only a few words. In real life use cases there can be spelling errors when the communication is through a written channel. The employed text classification method is based on fine tuning of the pre-trained BERT model on the specific problem. It is proposed to augment the training data with automatically generated erroneous samples to compensate the performance degradation caused by the queries containing spelling errors. Test sets are generated by replacing one or two words in query with its corresponding realistic human errors. Intent detection accuracies are reduced by %6.9 and %15.3 relative to the clean set accuracy respectively for single and double error test sets. The proposed error augmentation approach compensated the performance decrease caused by spelling errors %70,3 and %66,0 relatively.