Sentiment Classification of Indian Banks' Customer Complaints

Abstract

Sentiment Analysis of the customer complaints of Indian banks is an interesting task and critical for appropriate business functionality and improvement. Applying machine learning (ML) techniques on these raw textual data is increasingly gaining traction. Towards pre-processing the raw textual data, we employed techniques like document term matrix (DTM) driven by Term Frequency - Inverse Document Frequency (TF-IDF), embedding model likeWord2Vec and psycho-linguistic method like Linguistic Inquiry and Word Count (LIWC). For the purpose of classification, the raw textual data of complaints is labeled as “moderate” or “extreme” by the three human annotators. Results indicate that the LIWC in combination with Random Forest and Naïve Bayes techniques performed the best in three banks datasets. The results were statistically corroborated with a t-test.