Plenty of studies have examined the negative side of social media – toxic language, harassment, division – but equally as important is considering how these platforms can help people build connections and better understand each other.
A new study by Josephine Lukito, Bin Chen, Gina M. Masullo, and Natalie Jomini Stroud presents a novel approach to identifying “connective language” that brings people together on social media.
What is Connective Language?
The study defines connective language as that which “facilitates engagement, understanding, and conversation.” People using this type of language often show that they’re open to hearing other people’s perspectives, even if those perspectives are not aligned with their own beliefs.
Examples of connective phrases include: “In my opinion,” “I respectfully disagree,” “You’re right about…,” “I see where you’re coming from,” “That’s an interesting perspective,” “I never thought about it like that,” and “Can you clarify?”
As the study shows, connective language is different from concepts like politeness and empathy. One example provided is that saying “please” is polite, but not connective. Additionally, posts that include profanity might be connective – but those that include hate speech or that demonize or disrespect a person, or their point of view, are not. For language to be connective, it should present opinions in ways that invite people to engage with each other in a productive way.
Study Details
The study considered (1) the best way to identify connective language and (2) how connective language is different from other related concepts. Content was collected from Reddit, Twitter, and Facebook users who were posting about topics where people were likely to disagree. This included political topics, like who people should vote for, and apolitical topics, like whether pineapple should be a pizza topping.
To determine the best way to identify connective language on social media, the study evaluated two classifiers built on popular language models:
- BERT (Bidirectional Encoder Representations from Transformers): A method believed to excel due to its deep understanding of context and language nuances,1 making it useful for complex tasks such as detecting connective language in texts.
- Generative AI (GPT- 3.5 Turbo): The most recently available version of OpenAI’s language model at the time, known for its enhanced speed and accuracy, making it ideal for real-time text classification tasks.
To demonstrate how connectivity is a unique concept, the study compared connective language detection with several other concepts, including politeness, civility, and attributes related to political discussion quality (such as constructiveness, justification, relevance, and reciprocity).
Results and Takeaways
When comparing the two types of classifiers, the BERT model outperformed the GPT-3.5 Turbo model in accurately identifying connective language in social media posts. The study authors suggest several takeaways from these results: 1) the BERT classifier could be used to test whether people using connective language have more deliberative conversations; 2) it could be used to evaluate the effects of exposure to social media posts that contain connectivity; and 3) it could be used to examine practical ways of increasing connectivity, such as helping people understand alternative views.
In looking at the uniqueness of connective language, the study found that it was distinct from related concepts such as politeness, toxicity, constructiveness, and reciprocity, among others. The study authors suggest that connectivity “captures elements of communication not fully addressed by traditional metrics.” They go on to stress that recognizing this distinct concept is vital to better understanding the structural and relational dynamics often neglected in the study of online discussions.
- (Shen and Liu, 2021; Shushkevich et al., 2022; Moreira et al., 2023)[↩]