Accelerate product reviews with Deep Learning and Natural Language Processing

Estimated reading time: 2 mins

A plethora of products are available in a highly consumerized world. As it is a ‘buyer’s market’, people prefer to read the product online reviews before purchasing them. Product review analysis per se also helps companies get consumer feedback and add value to their product. Though in both the cases, reading massive amounts of inputs provided by consumers in an unstructured format is an extremely lengthy and cumbersome process. Artificial Intelligence (AI) technologies such as Natural Language Processing (NLP) and Deep Learning help to analyze vast amounts of product review data and decipher consumer sentiment as positive, negative, or neutral.

A modular approach:

The solution to this business challenge comprises of four modules, which greatly simplify the review task:

• Review classification: By using Word Embedding and Deep Learning, the module classifies the product review in to four categories – Query, Complaint, Praise, and Suggestion. Word Embedding uses Word2Vec and GloVe mechanism.
  
  
• Feature capture: By using Advanced NLP and Topic Modeling, the module extracts the feature from the review.
  
  Topic capture: By using Latent Dirichlet Allocation (LDA), the module extracts the topic from the review.
  
  
• Sentiment analysis: By using Deep Learning methods such as Recurrent Neural Network (RNN) and Long Short Term Memory (LSTM), the module captures the sentiment including sarcasm.

Together, the four modules help classify, capture the feature and topic, as well as assign sentiment to each product review and make a lengthy and cumbersome task very easy.

How does Word Embedding, Topic Modelling, and Long Short Term Memory work?

• Word Embedding:
  It represents similar words having similar representations. It is done by using Word2Vec and GloVe mechanisms.
  
  Word2Vec is a Neural Network based training method. It uses Continuous Bag of Word (CBOW) method to learn the embedding from the context word and predicts the current word. It also uses Skip-Gram model, where a target word is used to predict the embedding of the context word.
  
  GloVe is a modification of Word2Vec model. It takes the advantage of local context based learning in Word2Vec and adds the matrix factorization technique such as Latent Semantic Analysis (LSA).

• Topic Modelling:
  A topic model captures the topic, say whether it is 10% about sports and 90% about cinema, by counting the number of words from the corresponding jargon. The aim of this solution is to apply topic modeling for a variable number of topics. The model combines the advantages of LDA with genetic algorithms, which works for variable number of topics.

• Long Short Term Memory:
  Long Short Term Memory (LSTM) is the extension of Recurrent Neural Network (RNN). The limitation of RNN is that it does not capture long term dependency. To capture long term dependency, LSTM uses a cell unit, input gate, forget gate, and output gate.

In summary:

Product review analysis includes multiple tasks such as classification, feature extraction, topic extraction and sentiment analysis. The solution powered by Deep Learning, Topic Modeling, and NLP techniques overcomes the limitations of the traditional method and quickly classifies, captures the feature and topic, as well as assigns a sentiment to each and every review in a mega product review project. Where earlier a product review analysis took days, the task is now completed within minutes.