Computers understanding human language is a fascinating area and has piqued interest since the 1950s. 70 years later we still are continuing to solve the problem. This could be because human expressions are very varied and possibly the advancement in algorithms has yet to catch up with the human evolution of language and its context. One key area where NLP is extensively used is eCommerce. Reviews are a significant part of online buying cycles today. Though the “vocal minority” is few, the number of users who are impacted by reviews is significantly large. One study found that 63% of users prefer online sites that have reviews. Customers who visit review pages have an astounding 105% more chance of buying from the website. Mining these reviews gives insights to both the online service provider as well as the seller who has listed the product. Other than knowing whether the customer is happy or not, we can also know how users feel about each feature in the product. Sometimes reviewers write a lot about their lifestyle and the use case they have found for the product as well. This can provide insights into things like the product-market fit or the value proposition for the product. This can later be used in brand communications for the product. We can also find opportunities or gaps in a category and hence get the “voice of the customer” to create a new product or even start a new business (Sri 2021).
Role of sentiment Analysis in reviews: (Kang and Park 2014) proved the voice of customers in reviews contains higher information about user sentiment than the actual rating provided by the user. There are different ways of analyzing user sentiment in reviews – primarily supervised learning and unsupervised learning. Unsupervised learning involves using a list of lexicons and mining the sentiment from there on using heuristic approaches. Supervised on the other hand requires a corpus of labeled data to train the machine learning models. There are also various degrees of sentiment analysis. For instance, detecting subjectivity or objectivity in a review could provide us possibly some insight into how useful a review could be. Highly opinionated reviews may not be perceived to be very useful. Sentiment in reviews could also be used to separate out the fake from genuine reviews in eCommerce. Fake reviews could be more polarized towards one sentiment and “genuine” reviews possibly could be a mix of positive and negative sentiments.
Mother of all sentiment frameworks: If a researcher is interested in even more granular sentiment, then they should consider using “Plutchik’s wheel of emotions” (Kim and Klinger 2018). As you can see there are granular emotions across a wide variety of sentiments. The wheel of emotions at the core consists of eight basic emotions: joy, sorrow, anger, fear, trust, disgust, surprise, and anticipation. The intensity reduces as one moves towards the outer wheel and intensifies as one moves towards the core of the wheel.
In our tutorial in ODSC APAC 2021, “NLP in eCommerce,” we will discuss 3 methods with codes on how to crack the sentiment analysis problem for reviews in eCommerce. The examples in the tutorial are taken from my recently published book – “Practical Natural Language Processing with Python.”
Mathangi Sri has 17+ years of proven track record in building world-class data sciences solutions and products. She has 11 patent grants and 20+ patents published in the area of intuitive customer experience, indoor positioning, and user profiles. She has recently published a book with Apress, Springer – “Practical Natural Language Processing with Python” She is currently heading the data organization of GoFood, Gojek. In the past, she has built data science teams across large organizations like Citibank, HSBC, GE, and tech startups like 247.ai, PhonePe. She is an active contributor in the Data Science community – through lectures, talks, blogs, and advisory roles. She is recognized as one of “The Phenomenal SHE” by Indian National Bar Association in 2019.