Having around 16+ years of professional experience and was working in the Technology Services division of my company which provides software and hardware technical support for our customers. Read further to learn about my journey with Great Learning’s PGP Artificial Intelligence and Machine Learning Course.
It was discovered that the customers were generous in providing higher satisfaction ratings (ranging from 1-5). However, there were scenarios where feedback in the free text field was not as positive. This was a potential area that could be addressed by the organization to improve customer satisfaction and thereby prevent unexpected revenue erosion.
We used ‘Sentiment Analysis’ to get more insights into customer feedback apart from the rating provided. For training data, we had to create a training set manually based on the feedback that was historically provided by the customer and categorized as “Positive or Neutral” or “Negative.” We used tokenizer & glove embeddings for data pre-processing and trained a Bidirectional LSTM model using Adam optimizer. We selected this model as we found that this performed comparatively better with lesser complexity. The model was used to predict if the feedback was “Positive or Neutral” or “Negative.”
The challenge of building this model was mainly to manually create large training data set based on the historical data available. Also, we had to do data clean-up before pre-processing to remove special characters and typos. We also had to fine-tune the training set, and its volume as initial results on the validation data set was not very satisfactory.
Based on the feedback category prediction by the model, additional engagements were initiated with the top customers. We are expecting this initiative to improve our share of wallet with those customers.