Introduction: Is Customer Feedback Enough?
The importance of feedback analysis in the success of a product or a service cannot be ignored. For most new-age businesses, feedback is a continuous process critical to keeping the offerings relevant to the customers. Until recently, the typical customer feedback mechanism on products and services was a single-click rating question (‘How many stars out of 5?’ Or, something like ‘How was your experience? – excellent, very good, good, average, poor’).
Businesses now realize that for the feedback to actually shape their products, services, marketing, distribution, positioning, packaging, delivery, and so on, it is pertinent to reach beyond mere ‘rating’ and ‘ranking’ to comprehensive analysis of the components of the customer’s experience.
Let us take an example. Suppose a customer takes a ride from his office to home. At the end of the ride, the customer will first rank his ride on a scale from excellent to poor. Next, he would go a step further and explain what worked / didn’t work for him: Was the driver good? Was the cab good? And so on.
The typical feedback analysis today reaches only till this stage. But businesses today are asking deeper and more relevant questions – Why was the driver good / not good? Was he polite? Did he reach on time? Was the driving smooth? Why was the cab good / not good? Was it clean? Was it comfortable? Is a specific model of cab being preferred by customers? Are there other reasons?
What is Opinion Mining?
While ratings and scores tend to polarize the customer experience into ‘good’ and ‘bad’, opinions depolarize and add richness through perceptive dimensions of the customer experience. ‘Opinion Mining’ stems from the need for businesses to identify the exact problems that a customer faces. Also known as ‘Sentiment Analysis’, it uses natural language processing to track the mood, attitude, perceptions, and sentiments of the target market about a particular product / service, using automatic detection of relevant opinion variables.
This is critical in identifying the source of customer satisfaction / dissatisfaction, and is especially important when multiple parties are involved in creating a customer experience (in the above case of a customer taking a can ride, the different parties who create the customer experience are 1. The driver 2. The cab manufacturer 3. The cab owner 4. The cab aggregator and app interface provider) Why is a customer happy / unhappy? More importantly, does a ‘happy’ customer like everything about the product? Or, does an ‘unhappy’ customer dislike everything about the product? Opinion mining gives a constructive feedback that enables managers to pinpoint the source of the problem, or identify the factors that contributed to the customer’s satisfaction, which can then be replicated across customers.
Even if multiple parties are not involved in a transaction, the breakdown of components still signifies the importance of mapping the actual experience of the customers. Why? Because if businesses want to improve, they need to know exactly where they need to improve.
The Analytics Challenge – Making Sense of Unstructured Opinion Data
In real world, opinion mining is easier said than done. Most businesses with some form of digital platform or the other have access to opinions of its customers and even potential customers, but the challenge is – how to analyze these opinions and derive meaning from it? To begin with, opinions are unstructured textual or voice data which are no fit for traditional analytics tools. Secondly, even if stand-alone technologies are able to mine and process opinions, the insights derived are not married to the traditional databases of businesses. This leads to a sheer disconnect. Thirdly, opinions are extremely subjective, time-dependent, and location-dependent, – this leads to a problem of normalization.
Let us look at the intricacy of the problem of normalization using an example.
Let’s say a multinational fast-food chain operates through franchisees across multiple cities in India. By analyzing the opinions it was found that –
- Some customers preferred the dim lighting currently used, but others did not.
- Majority of customers in Gujarat disliked the fact that non-veg menu was being served, whereas majority of customers in West Bengal cited this as a key reason for satisfaction.
- The same customer who visited the chain on a weekday had a different opinion when he visited during the weekend, although services and staff were the same (was the customer in a better mood on weekend?)
Only powerful analytics tools that can normalize opinions and contextualize them to derive meaning, can resolve the conflicting sentiments that arise from opinion mining. Why? Because just ‘knowing’ opinions is not enough – to let opinions shape the strategy of a business, action points must emanate from it. Further, these action points will have to be reconciled with traditional databases that businesses use.
Lastly, opinions tend to come loaded with a lot of noise. For example, let us look at the following sample review by a customer about his cab ride: “I took a cab from Airport to hotel. There was a lot of traffic, and the driver was constantly speaking on the phone, not focusing on the drive – we missed two signals just like that, and I got late for office. What a bad experience!”
The noise element here is the ‘traffic’ on which the cab service provider has no control. The relevant opinion here is ‘driver was constantly speaking on phone’, and the overall feedback is ‘bad experience’ – this is relevant and actionable. But, the challenge is, that analytics tools have to be equipped enough to eliminate noise and process the important stuff.
MECBOT for Intelligent Opinion Mining
At FORMCEPT, we believe that opinion analysis can be made easy for businesses, if they are equipped with the right analytics tools. Our flagship product MECBOT can perform state-of-the-art opinion mining with superior tools such as:
- Powerful natural language processing
- Automatic opinion variable detection
- Noise elimination
- Automatic Normalization techniques
- Seamless connectivity of opinion data flow to organization work flow and traditional databases used (CRM, product databases, etc.)
The following flowchart shows how MECBOT performs Opinion Mining on reviews compiled from users.
Let us understand this process with the help of an example. For example, let us take a sample customer review statement:
“The screen of the phone is too small, but it has a good display.”
MECBOT performs co-referencing on this statement to find all the references in a text that refer to the same entity. In this case, it reads the above sentence as ‘the screen of the phone is too small, but the phone has a good display.” It then performs sentence segmentation and boundary detection to break the sentence into relevant components and establish relationships, if any. By determining word dependencies such as subject-verb, verb-object, etc. it comes up with a dependency graph as a sentence analyzer. Lastly feature extraction is done by forming opinion pairs, usually ‘adjective+noun’, e.g. ‘small+screen’, ‘good+display’.
Feature extraction across thousands of such reviews is carried out by capturing frequent patterns that occur in the reviews.The features are then filtered to eliminate any noise from the data (false positives) (e.g. if the user was looking at the phone in a noisy retail store, then the customer’s statements / opinions about the store will be filtered out as it is irrelevant, only opinion on the mobile phone will be captured). The features are then grouped by possible opinions across the reviews in order to arrive at the popular opinion garnered from all the reviews about a particular feature.