Before the advent of social media, voicing documented opinion was the preserve of a few people, generally called the "opinion makers".
But in today's day and age, anyone anywhere with an internet connection can register her likes, dislikes, endorsements at minimal cost and effort.
This information, when collected on a large enough scale, can help in building models that help gauge what these "opinion makers" are thinking.
And these people matter because they are the consumers - of everything from Hotstar subscriptions to airline tickets to financial products.
The product that we have conceived is web tool that lets users know the sentiment of viewers for a movie trailer on Youtube from the date of its launch to the date of release of the movie.
It will help people as well as production houses monitor what people are thinking about upcoming movies.
Though it sounds like a simple task, the complication with Youtube is that the comments are not necessarily structured using proper grammar and vocabulary.
This makes it tricky to deduce the sentiment from the comment. Also, in India most of the comments are in Hinglish, which coupled with the bad grammar and varied vocabulary make it even harder to wring the sentiment out of the comments.
The above graph depicts the behaviour of top five positive and top five negative words based on their respective L2 penalties over their weights.
L2 penalty is a proxy for estimating the affect a word has on the sentiment of sentence.
As you can see when L2 is close to zero the weights were spread over a large range for the respective negative and positive words.
But this is not scalable for finding complicated decision boundaries (a problem very intrinsic for sentiment analysis on Youtube comments) for large datasets.
The decision boundaries establish the difference between the positive and negative character of the words. This characteristic in turn lends the sentiment to a sentence.
As we move to the center region of the graph, the words have a smaller spread based on their weights. This is good for scaling up a model for large and complicated dataset such as the one we are dealing with and building a deeper neural net.
Also, as the words have lesser spread it becomes easier to extend our vocabulary and mark a large dataset for supervised sentiment analysis..
Meanwhile, the right hand side part of the graph depicts L2 weights converging to zero, which is as expected.
Although, this is the most popular version of regularisation used by AI practitioners we are working on other novel methods such as dynamically rotating penalty parameters using Langrange multipliers.