A matter of opinion – sentiment analysis for businesses
By : Dr. Utpal Sikdar
Social media has emerged as a popular sounding board for customers to express their experiences with a brand. Keeping a track of all customer’s emotions will help brands to better understand their feedbacks and opinions. Businesses can leverage NLP-powered sentiment analysis to gain actionable insights from the unstructured data available from social media channels and third party websites. And these insights in turn can drive effective business decisions and strategies.
With the rise of the era of social media, customers are always connected and are empowered like never before. They commend. They condemn. They share feedbacks and opinions through social media channels in form of tweets, reviews, chats and comments. Enterprises realize value of capturing and acting on this ‘voice of the customer’.
Sentiment analysis is a text analysis process that uses Natural language processing (NLP) to identify and analyze a given text in the document, sentence, or entity/phrase level collected from various channels, according to the underlying tone of the expression. In simpler terms, using sentiment analysis, we can decide whether a document, a sentence or an entity/phrase is having a tone that is positive, negative or neutral. The entities/phrases are called aspect terms and identifying the sentiment for each aspect term is called aspect based sentiment analysis (ABSA). ABSA gives us customer opinions/sentiments on different aspect of a product or a service. Nowadays, researchers are giving more attention for each aspect terms’ sentiment rather than focusing on overall polarity (e.g. positive or negative). This type of analysis gives a relatively more nuanced overview of sentiments.
Importance and applications of Aspect Based Sentiment Analysis
E-commerce is a good platform where people can express their views/opinions about services and products they access. In this context, aspect-based sentiment analysis specifies aspect terms (e.g. price, service, network, etc.) and their sentiment (positive/negative/neutral) which can help customers to decide what to consume and what to avoid. From the business point of view, it gives a better understanding of products or services and business management can improve the quality of products or services based on deeper insights obtained from reviewers’ comments.
The goal here is to identify aspect terms from given reviews and classify them into sentiment categories (positive/negative/neutral). For example, in case of e-commerce domain, the product laptop could have various aspects associated with it, such as price, design, battery life, processor, etc. For instance, if we consider the review “Easy to start up and does not overheat as much as other laptops.” Here ‘start up’ and ‘overheat’ are the two aspect terms and the sentiments of both aspect terms are ‘positive’. In case of telecom domain, ‘app’ and ‘bill amount’ are two aspect terms of the review “Hopeless app. The current bill amount is never updated.” Here the sentiments of both aspect terms are negative.
The following visualizations depict an application of this ASBA model on a set of comments provided by random users who attended the Mobile World Congress (MWC) event at Barcelona in 2018. Each bubble represents an aspect and its size represents how many comments are present for this aspect. MWC is green as people have talked about ‘MWC’ mostly in a positive way. Traffic bubble is red as people have already talked about ‘traffic’ mostly in a negative way.
‘Booths’ bubble is shown as blue as it is a new aspect which people have not talked about earlier.
Aspect wise sentiment detection after user has entered the comment.
Flytxt’s research work on Aspect Based Sentiment Analysis (ABSA)
Flytxt has developed a supervised Machine Learning approach for Aspect Based Sentiment Analysis (ABSA). Conditional Random Fields (CRFs) are being used to deploy the supervised model for extracting aspect terms and identifying sentiments from customer reviews and comments. A large number of features are extracted from the reviews and most of the features are domain independent. CRFs assign a well-defined probability distribution over possible labeling, trained by maximum likelihood estimation. For ABSA, in first step, aspect terms are extracted using CRFs with the large set of features (e.g. word itself, context words, part of- speech tags, word frequency, etc.). In second step, each aspect terms’ sentiment/opinion is identified using CRFs.
The supervised model is applied on three domains – i) Laptop, ii) Restaurant and iii) Amazon product reviews (e.g. coffee machine, cutlery, microwave, toaster, etc.) The goal of this project is to build a generic model which can be applied to any domain to discover relevant aspect terms and sentiments. We are also building a hybrid model using unsupervised and supervised approaches towards each of the discovered aspect. You can access the white paper here.
With the ever-expanding data sets in today’s world, tools like sentiment analysis open many gateways for analyzing this data to derive meaningful insights and gain a greater business value. However, there are many challenges in the path of implementing effective sentiment analysis.
The emotions expressed by the customers may not be always having a direct tone and can be very complex in nature, like irony or sarcasm. This complicates the process of identifying a clear sentiment. Though, advancement in technology will overcome these challenges.
The bottom line is that sentiment analysis is all about converting data into meaningful and actionable information in hands of companies. No matter how complex it is, its benefits are massive.