Article

AI-Powered Opinion Mining - How Sentiment Analysis Revolutionizes Customer Experience Management
June 2021 Nikos Tsantanis,
Senior Product Marketing Manager, Intracom Telecom

Ioanna Sanida,
Senior Data Scientist, Intracom Telecom

Nikos Anastopoulos,
Senior Product Manager & Solutions Architect, Intracom Telecom
Introduction

As CSPs constantly strive to stay ahead of the competition, they draw upon their customers’ perception – positive or negative – for their products which, perhaps, is the ultimate satisfaction indicator. Customers, in turn, often do realize the power that their feedback has and opt to express their feelings in writing, either to draw the operator’s attention on their opinion or even to have their acquaintances involved.

In some occasions, customer care online forms, emails, chatbots or IVR are the digital touchpoints that convey textual feedback spontaneously offered by clients, directly to the CSP. In other instances, it is the CSP that encourages customers to provide feedback via questionnaires or surveys. There is, also, a third option, where a happy or dissatisfied customer may resort to uploading posts on all types of social media, acting like a social influencer.

Evidently, if an operator was able to sift through the abundance of internal and external customer produced texts referring to its products, then immense business value would come out of it!

CSP diagram
Sentiment Analysis: Value and Limitations
image of Sentiment Analysis Processes

Making sense of textual customer feedback has been a long-standing pursuit for CSPs, usually addressing it on a best effort basis and through manual labor. Not long ago, however, the advances in both Natural Language Processing (NLP) and Machine Learning (ML) have boosted Sentiment Analysis: the automatic process that helps identify the topics for which opinions are expressed in texts, and classify them into the proper sentiment tag (positive, neutral or negative).

In simple terms, Sentiment Analysis follows two processes:

  • In the Training Process, the NLP model learns to associate each textual input to the corresponding sentiment tag based on already characterized samples. The feature extractor module transforms input into an intermediate representation, a feature vector. Each feature vector, along with its associated annotation, is then fed into the ML algorithm to generate a classifier model.
  • In the Prediction Process, feature vectors of unseen text are fed into the model to infer their sentiment tag score.

Once properly deployed, a Sentiment Analysis solution can automatically process large quantities of text in near real-time, unearthing otherwise hidden, hard to detect insights. Having said that, two key challenges cannot stay unnoticed:

  1. Lack of native language corpus. A prerequisite for achieving highly accurate Sentiment Analysis results in a given language is the availability of the respective corpus: a large collection of natural language text organized into datasets that is used by the NLP system to find out how human language is represented. Corpus enables words with similar meaning to have a similar representation by the creation of ‘word embeddings’. Due to the very limited research on the Greek language, and some other leastspoken, nonEnglish languages, there is lack of respective corpora.
  2. Insufficient annotated data. For the Sentiment Analysis system to correctly classify sentiment and learn the textual differences between sentences in various categories, the use of supervised ML is typically required. Creating structured training data that enables machines to understand human speech is heavily dependent on natural language annotation. Through this process, the model is trained with data already annotated in terms of both the product that they refer to and the sentiment tag. It is implied that Sentiment Analysis can only function properly when the available dataset is both large and annotated with respect to its sentiment tag score.
Cognitiva™ Sentiment Analysis For Telcos

Intracom Telecom’s Cognitiva™ Sentiment Analysis is an NLP-based solution that analyzes textual information in Greek and extracts meaningful information regarding customer feedback. It empowers CSPs to deep dive into customer-generated textual feedback, ‘listen’ to their conversations and take into account their messages.

Cognitiva™ Sentiment Analysis addresses the aforementioned key challenges, by making use of a native language corpus into the training process, either annotated with accuracy using highly automated tools or reusing the already annotated textual CSP data verified by experts. This process can be applied not only to Greek language, but to other non-English ones, such as Bulgarian and Serbian.

Cognitiva™ Sentiment Analysis is an approach that can be applied on a wide product range of vertical industries, such as Telecommunications, Banking and Retail. In abstract terms, the solution is structured along three pillars:

  • Enumeration of Customer Satisfaction KPIs by quantifying them in terms of polarity (positive/negative) and prediction confidence.
  • Extraction of topics/issues mentioned in the respective texts, helping thus narrow down the subject to which a conversation refers to (such as specific products or services).
  • Flexible integration and visualization through powerful features such as welldefined APIs, automation capabilities, modular micro services architecture and customizable dashboards.
image of solution structured

The deployment of Cognitiva™ Sentiment Analysis is based on five conceptual steps, each executed in different fre-quency.

  1. Raw Data Gathering. Collect a reasonable amount of CSP textual data containing customer feedback, taken from both internal (customer care chat interactions, speech-to-text transcripts of voice conversations) and external (social media posts) sources.
  2. Data Preprocessing. Clean and label a sample dataset of this input where each sentence belongs to a different class (1 for positive, 0 for neutral, -1 for negative) for training purposes.
  3. Model Training. Create a classification algorithm that, once trained, classifies correctly the sentiment of each sentence. Evaluate the results and repeat if needed.
  4. Dashboard Visualization. Evaluate the quantified outcome of the sensitivity analysis process via statistical analytics and information-rich dashboards full of easy to grasp, color coded information.
  5. Maintenance. Retrain the model and revisit key system parameters, typically on a monthly basis.
image Sentiment Analysis in five conceptual Steps
Use Case: Sentiment Analysis For Churn Prevention

Network degradations, billing dissatisfactions or bad customer support experiences are often among the top reasons for churn. In those cases customers may choose one of the following means to interact with the CSP in free-text form: send a message directly to the CSP describing their dissatisfaction, respond to a CSP-initiated survey providing their feedback, or even post their experience in public through social media.

Cognitiva™ Sentiment Analysis receives customer input, regardless of its origin, and performs the following tasks:

  • Derives the sentiment tag score of the message
  • Detects the primary reason of dissatisfaction
  • Identifies the service or product the message is referring to
  • Identifies the geographic location the message is possibly referring to
  • Routes the message to the most appropriate function(s) within the organization.

For example, Cognitiva™ Sentiment Analysis will automatically identify frustration in a user message and detect that the reason is a network issue affecting IPTV. Based on the user’s location, Cognitiva™ Sentiment Analysis will initiate a trouble ticket and route it to the corresponding network Operations/Monitoring Department for immediate corrective actions. At the same time, the ticket will also be routed to the Campaigns Department to propose a recovery offer to the customer.

Churn Prevention Diagram