What, besides figuring out that trolls (some of whom work here at bloomfield knoble) rule the Internet, can Tay teach us about social media?
By now you’ve probably heard about the rise (and very dramatic fall) of Microsoft’s Tay, an artificially intelligent bot on Twitter. Microsoft was hoping to show off that it had made significant strides in the world of artificial intelligence while attempting to build a real understanding of how a specific subset of society interacts. Unfortunately for Microsoft, they didn’t actually create an AI, they simply created a chat bot – a program that repurposed the content it received in a way that would seem to emulate the subset of society it was trying to emulate (in this case, the personality of a teenage girl). The impact is that a chat bot doesn’t know “right” or “wrong” just nouns, verbs, adjectives, etc. For some unknown reason, Microsoft decided to let Tay learn courtesy of the interwebs – and the interwebs obliged in only the way the interwebs know how.
I’m not going to jump on the “how dumb are engineers at Microsoft for not knowing how the Internet works?” bandwagon (too easy), because we may have actually learned something very interesting from this experiment.
If you can analyze the sentiment of large-scale populations, then you can ensure that things like public policies are effective. In fact, a recent paper by Annabelle Wenas from the University of Indonesia titled Measuring happiness in large population addresses just this. She writes,
“Governing complex modern societies requires some basic measurements in the societal level. These measurements will ensure that public policies are effective and meet the ever changing demand. However, currently, the most common aggregate measures of societies are economic measures such as economic growth. Yet, as modern societies grow more complex, there is a need to develop other measurements beyond economic measures especially for psychological measures that can capture subjective well-being. It is reasonable to think that a combination of economic and psychological measures can provide more comprehensive view of a society which, in turn, will be useful for formulating better public policies and their evaluations.”
Wenas proposes an approach to measure psychological characteristics for large populations based on text data (like Twitter). The authors also note that this concept isn’t exactly new, in fact,
“Our focus is on the measurement of emotional states and we follow [J.A. Russell from Psychological Review] who asserted that emotion, mood and other emotionally charged events are states that are simply combinations between feeling good or bad and energized or worn out. Russell addressed these emotional states as core affect, and mapped its structure into circumplex model. Horizontal axis of circumplex model is valence, which is a measure of emotion ranging from negative to positive emotions. Whereas its vertical axis is arousal, a measure of emotional intensity. Thus, for example, anger is a negative emotion with high intensity and lethargic is a negative emotion with low intensity. On the other hand of the spectrum, excited and calm are positive emotions with high and low intensity respectively. Note that happiness is a positive emotion with moderate intensity.”
In a nutshell, the author scoured Twitter for keywords that generally reflect happiness and measured not only word valence, but also the measurement of arousal dimension. The reason to include arousal is because positive valence is necessary but not a sufficient component of happiness, because there are either states that have positive valence like excitement and calm. Thus, the level of arousal is the key to differentiate excitement, happiness and calm. Three of them indeed have positive valence, yet their arousal level are variable from high, moderate to low (respectively). The author provides the formula and proof of their test and admits that there approach has the potential to be used as a measure of emotions for large population in multi domains. Further development of their approach will include tests for sensitivity, robustness and also the inclusion of other psychological measures, such as moral judgments, values and personality.
So, what does a paper about studying happiness have to do with Tay? Nothing and everything. The concept behind Tay was, I suspect, not to generate an AI that can pass the Turing test, but rather a chat bot that would better understand a specific subset of culture. Understanding that subset of culture would help identify trends, patters, concepts and could then, theoretically, be used to identify future activity. IBM did something similar when Watson announced that SteamPunk was the next big thing. This concept – understanding and predicting – is what every agency (like bloomfield knoble) – is trying to achieve for clients. It’s why we spend so much time gathering and analyzing big data (yes, I said it) – we want to spend money where it will be most effective. Bottom line. And if a chat bot can learn enough about a specific subset to help us identify best use of ad dollars, then so be it.
A STEM (Science / Technology / Engineering / Math) graduate and COO of bloomfield knoble, Thomas exemplifies the view that advertising is becoming an engineering discipline. He leads the integrated insights and strategic planning group in a way consistent with bloomfield knoble’s goal of bringing a strong analytical foundation to uncover fresh and innovative insights and business opportunities.