AI or what people really mean these days, machine learning, is often considered a distant, complex technology that many campaigns lack the skills to execute or perhaps do not know how to apply to their everyday work. The incredible power of machine learning seems to exist behind an inaccessible veil requiring profound expertise and technical know-how. But this year, that has changed with the advent of tools such as those available through huggingface.com. Combined with the release of several open source tutorials and politics-specific natural language processing libraries, we are truly in a golden era of machine learning. The tools are now available to build custom solutions that can interpret some of the hardest speech and text out there, including from social media, email, and text streams and conversations generated by political and advocacy campaigns.
Python enables us to customize existing language models using large data sets with the help of GPU-accelerated systems and incredible solutions such as PyTorch that make the process of tokenizing, categorizing, and tagging natural language rapid. Rapid enough, in fact, that they can be used by advocacy organizations in a timely fashion to intelligently respond to conversations happening online in real-time.
We are using a lot of jargon here, so let’s step back and talk about why we think this new technology is important and how it solves a crucial problem. That problem is the age-old uselessness of sentimentality and other machine-generated metrics in our current social media listening tools.
Current social media listening tools are really good at explaining in basic terms what happened, but they are terrible at explaining why it happened. Tools like Meltwater or Crimson Hexagon rely solely on broad sentimentality, word clouds, top tweets etc. to give you an idea of the context of the conversation.
While helpful, those tools are limited in gleaning actionable lessons from the texts analyzed. This comes back to why we analyze conversations on social media in the first place. We would like to be the drivers of those conversations, compelling activists to take action and make their voices on the most salient issues of the day. To do so, we should be looking specifically for how different segments of the audience are talking about the issues in their posts, separating those with large followings from those without. We should be training an “AI” model to go through thousands of messages to correlate certain phrases with contextually positive and negative sentiments and policy stances so that you can start to separate messages out into two camps.
Once you accomplish this deep, context-driven sorting, you can glean great insights into how people are talking about any subject. A great example of this is the child tax credit, where language that would normally come up as negative to a normal sentimentality model is actually the way average Americans talk about their NEED for the child tax credit. They are terrified of LOSING their enhanced benefit. They are afraid they will go WITHOUT food, electricity, transportation, etc. These are negative frames to describe why they desperately need the enhanced benefits; to express their support FOR the CTC..
The way that many organizations were talking about the CTC was the opposite; highlighting the positives. Groups talked about how many children the CTC would bring out of poverty; how the program could help so many families; and how necessary it is for them to succeed in the future. They messaged that the CTC will LIFT families OUT OF poverty. These are positive words and would have a positive sentimentality score, but since it is not how people are talking about the CTC, they are not effective words.
This is the point, organizations talking about the CTC are doing it wrong. They are talking about it amorphously, without a serious connection to the sense of loss and fear of having something taken AWAY instead of GIVEN to them. This disconnect means that those organizations motivated people less, reached fewer people, and, most importantly organized less people because the language they used was foriegn to the very people who they were trying to support and uplift.
In short, we can apply machine learning to accurately take the temperature of the policy conversation on social media, gleaning insights into how people are feeling about policy and the specific phrases they are using to express themselves. We can then plug in to that conversation by delivering content that will resonate with, as opposed to fighting against, the conversation on social media. This is not just a good communications strategy, this is good organizing strategy, and it is the type of organizing we must embrace to succeed in achieving our public policy objectives.