Sentiment Analysis – Making the most of your research
Sentiment analysis refers to the use of natural language processing, text analysis and computational linguistics to systematically identify, extract, quantify, and study affective states and subjective information.
We love open ended question for adding colour and depth to quantitative research but getting the most out of these can be difficult.
How can it add value to your research?
At Data Pad we use Machine Learning to extract subjective impressions from the text of open-ended questions. Sentiment Analysis has often been used to infer public opinions based on tweets or product reviews. Now we can use it as an invaluable tool for modern market research.
We have combined both Deep Learning and Sentiment Analysis and leveraged recent advances in the field of Natural Language Processing to build a state-of-the-art model which can predict the sentiment of free text responses to our surveys.
Deep Learning is an area of Machine Learning which uses Neural Networks, inspired by the human brain, to build especially accurate and powerful models. Deep Learning has had successful applications from personal assistants such as Siri and Alexa, recommendation systems used by Netflix and Amazon, and even the colourisation of old black and white images.
How do we use Artificial Intelligence to analyse free text?
First, we need to teach an AI to understand English. If our AI can predict the next word in a sentence then it clearly has a solid grasp of language. AI spent days repeatedly reading some 30,000 Wikipedia articles (over 75 billion words!), learning not only English language but also some facts about the world – for example “Prime Minister Theresa” is very likely to be followed by “May”; this is very similar to the predictive text in a smart phone.
This is great, but we knew we could do better. Free text responses are often short and use informal language or textspeak. Using Transfer Learning we fine-tuned the 33 million parameters in our AI’s Neural Network to be especially well suited to the kind of responses it will see in open ended survey questions. Finally, the model is retrained to predict the sentiment of text, rather than predicting the next word.
An Example: What words or phrases would you use to describe people with tattoos, specifically thinking about values and outlook.
We asked our panel, OpinionHive, to describe people with Tattoos in a free text question. Sentiment analysis allowed us to understand not only the type of sentiment – positive, negative or neutral, but also how this varied by demographics.
The graph below shows the mean sentiment of the responses by age. The strength of the positive response correlated strongly with age, the most positive being the youngest age group. The only group with a mean negative response was the oldest age group.
Looking in more detail at the text, we are able to create word clouds to show both sentiment and frequency of the words used. The size of the word is an indicator of frequency of the word in the answers and the colour shows the sentiment in the context it was used; red words being negative and green being positive.