Some political speeches are ingrained in our minds, either by the significance of the occasion on which they were delivered, the eloquence of the speaker, or both. There’s something about a rousing speech that can stir emotions in a seemingly indefinable way.

But Nick Beauchamp, an assistant professor at Northeastern University’s Department of Political Science, in Boston, USA, believes he may have found a way to define the indefinable. He has developed an algorithm that, he claims, can build more compelling, convincing political speeches.

Beauchamp initially set out to get a better understanding of what makes people support or oppose issues by breaking down the political discourse around those issues, and working out which elements are likely to be welcomed or rejected by the public.  

He chose to focus on the US’s Affordable Care Act – also known as ‘ObamaCare’ – when developing his algorithm, as he says it’s one on which many Americans still have fluid opinions. He started by feeding around 2,000 sentences from a pro-ObamaCare website into a model.

“The algorithm is given a large collection of raw sentences, and its job is to choose more and more persuasive three-sentence combinations,” explains Beauchamp. “Just choosing sentences at random will neither make much progress nor give us insight into why those sentences happen to be more persuasive, so we need some way of systematising those raw sentences.”

First, the algorithm identifies the underlying topics in the sentences – clusters of words that tend to co-occur. For example, for ObamaCare, one topic has words such as ‘pre-existing’, ‘condition’, ‘coverage’, ‘plans’, while another has ‘employee’, ‘employer’, ‘business’ and ‘mandate’. The former, says Beauchamp, is a topic that tends to occur in sentences about how – if you have a pre-existing condition – you can no longer be denied coverage, while the latter tends to appear in sentences about being able to keep healthcare if you already receive it through your employer.

“Each sentence is then encoded – for example, as 20% Topic 1, 15% Topic 2, etc – and the machine then tries to find the most persuasive distribution of topics to guide which sentences to choose.”

Beauchamp then tests his sentence combinations; in the ObamaCare experiment, respondents were asked to read various combinations and indicate their approval of the issue on a scale of 1 to 9. Using this feedback iteratively, the system returned to its pools of topics to find more and more effective sentence combinations – which were, in turn, sent on to new groups for testing. 

Beauchamp found that the two topics mentioned earlier (pre-existing conditions and keeping an employer plan) tended to be the most persuasive in favour of ObamaCare – those who read them tended to have higher approval levels when surveyed afterwards. Two other topics – about state/ federal relations and laws and rights – tended, instead, to push people away from approving of the issue. Within 90 minutes, Beauchamp had a collection of sentences with a 30% higher approval rating than the original text. 

Taking a general approach

In theory, the algorithm can be fed any sort of sentences and have any outcome question you can think of. “It’s a very general approach, not at all restricted to politics,” says Beauchamp.

“One could just as easily feed the machine a bunch of sentences about a product and ask the subjects to rate their approval of the product as the outcome variable. The algorithm knows nothing about content – it would just go ahead and find the topics, then systematically vary them until it found the subset that best increased the outcome the researcher was interested in.”

What the method doesn’t cover – deliberately, Beauchamp insists – is rhetorical manipulations such as tone, emotion and syntax. In his experience, the effect of the tone is very much dependent on the subject matter and who is delivering the speech. 

However, there is still room for the algorithm to incorporate the ‘wow’ factor that is so often present in the most powerful speeches.  

“At the moment, this method doesn’t write the original sentences, it just collects them from elsewhere,” says Beauchamp. “So there’s still plenty of room for human-inspired ‘wow’. The ideal is that, while the process may be relatively boring for the participants – although, as it stands, each person only sees one chunk of text – the outcome could end up being something as surprising as anything crafted by a person. For instance, you could ask people how much ‘wow’ and surprise they found in the text, or how unformulaic it seemed, then have the machine optimise that.  

“There’s no reason you couldn’t end up with something even more impressive than the standard product of a copywriter. Although, as with all computer-generated content, it’s unlikely that the automated approach will beat the best human practitioners any time soon.”