I recently sent a quick one-paragraph e-mail about some security concerns. I intended it to be sensible and precautionary, and thought it made clear that I was still investigating and had found only theoretical problems. I specifically tried to avoid showing fear or crying wolf. But my dad was convinced I’d gone insane. Which almost became a self-fulfilling prophecy.
Is this the sort of misunderstanding that text sentiment analysis could improve the chance of avoiding? I’m choosing this as the place to ask because people could – and I would – also use it for mood tracking.
If it’s doable, what are the hardware requirements? Ideally, what I’d like would be real-time interactive feedback while I was composing in Gmail, even when I was on my Nexus 7 in offline mode (as I was this time). But I’m guessing that even when I’m on my Core i7, it will probably have to happen with GPU acceleration, or in the cloud, or with GPU acceleration in the cloud.
Most sentiment analysis is actually super simple and based on word lists (happy = +5, enraged = -3, etc). More advanced stuff needs training data. You can generate some yourself, or use a pretrained system, which usually costs a reasonable amount of money. GPU acceleration is really overkill for sentiment analysis on short pieces of text.
I had a few minutes, so I decided to make you a Chrome extension. It’s super easy. Just highlight some text, and then click on the extension button to the right of your url bar. It will tell you the sentiment of the highlighted text.
Thanks! I tried it with the messages I sent; the sentiment score was -1.075 for the first message in the thread, and +5.310 for the message that my dad has since said was the tipping point of concern. Three “control” messages I wrote that same month scored +1.067, -1.961 and +2.206. Three more recent ones score -0.704, -0.883 and -1.869. A recent explanation I gave my dad of why his reaction upset me (something he didn’t seem to understand) scored -3.003.
Three of my dad’s replies about the incident score +7.407 (his first reply, offering help “in person and in situ” but without describing his idea of the problem), -5.085 (criticizing a particular misinformed remark as racist) and +4.054 (where he says he’s concerned about my mental health and that my expectations of help from him aren’t realistic). Some messages from my mom about the incident score +2.100 and +3.448, and two of her messages to me on other subjects score -1.667 and +7.742.
It looks like sentiment analysis probably won’t be a very useful tool on individual messages; and if it will, the magnitude will be more reliable than the +/- sign. But I’ll look into using it in weekly or monthly aggregates as a mood-tracking tool.