Review Language Processing System - Towards a zero touch customer care
CRM – reviews – sentiment analysis – similar documents – document classification – Deep Neural Networks – conversational interface – Text analytics – Prediction accuracy
Reviews are an important source of customer relationship and experience mgmt. for any digital company. Rightly captured and acted upon review helps in differentiating a company’s position in market place. Besides automated text analytics helps improve the cost effectiveness of operations. Text analytics, natural language processing provides techniques to support above strategy.
Identified areas are defined in 5 use cases –
- product category prediction,
- sentiment prediction
- similar reviews feedback
- Topic modelling, and
- Gender/age prediction
We collected historical customer reviews labelled data over several preceding years. We defined and engineered features from the raw datasets that heuristically synthesize and summarize information. We set the transaction data sets by extracting word and document embedding. We researched a variety of statistical and Deep Neural Networks based models. The pipelines were duly built for data pre-processing as well as online processing for new review mgmt. in a conversational dialogue.
Architectures were defined.We trained multiple deep learning models and received varying prediction accuracy.
Overall, we built the complete pipelines for data ingestion, data cleaning, feature engineering, ongoing training, so that a commercial review and CRM system can be served online reviews.
Model has successfully learnt predicting bi-directional Long short term memory (Bi-directional LSTM). While we will back-tested and took care to avoid overfitting, the model performance underperforms our expectations for being used as a decision support system for a trading desk. Our Prediction model lacks incorporation of higher and varied topic reviews to be a general purpose review processing system. Surveying various literatures as well as applying our own experience, we see that incorporation of same should improve the prediction accuracy substantially.