Something related to textual analysis:
The New York Times is researching how to contextualize which ads they show with the feelings an article is likely to inspire. I’m not a fan. They claim having learned that ads perform better on emotional articles won’t influence the newsroom, but we’ll see. At least they’re being transparent about doing this work. They’ve published an article with information on how they developed their sentiment analysis algorithm (link below).
There’s an explanation of the types of models they used and why. The initial steps were linear and tree-based textual analysis models, followed by a deep learning phase intended to “focus on language patterns that signaled emotions, not topics.” This outperformed the linear models some of the time, but not all of the time.
From what I can tell, the training set used a survey showing articles with images to establish a baseline, but the linear predictive models focus purely on text. I may be misunderstanding this or information may be missing. I expect that image selection can enhance or diminish the emotionality of an article. Perhaps sensational or graphic images would prove to drive more (or fewer) ad clicks. Despite the buffer the NYT cites between their newsroom and marketing arms, this feels like morally hazardous territory. So to answer the question in the title of the NYT piece, this article makes me feel disturbed. But I still didn’t click an ad.
It’s a quick read. Check it out.