Vyrill is a recent B2B startup based in San Francisco that's working on product aimed at digital marketing teams. Their platform develops tools for intelligently analyzing video content and data to find marketing and branding insights, aggregating video data from multiple sources, in particular social media, and performs analytics on the data. The bulk of the product makes heavy use of computer vision and NLP. Towards that end, the Vyrill project is to take video data of product reviews, unboxings, etc., video metadata, and captions / comments and build a sentiment analysis model to determine what product the video is in reference to and whether the overall reception is positive and negative.
The project goes beyond basic sentiment analysis and requires a quantitative metric of how positive or negative the review is, and should be able to point to specific comments or sections of the video where the sentiment was expressed. Moreover, the analysis should be nuanced enough to identify not only specific segments corresponding to a sentiment score but should also be able to produce a vector of sentiment scores corresponding to a set of distinct product features (eg cost, durability, etc.). Ideally these features would be learned by the model itself, and so the project would involve both unsupervised and supervised learning.