MIT and Harvard have presented language models that can predict public opinion.
Traditional survey-based approaches to measuring public opinion have limitations, but public opinion reflects and influences the behavior of society. Questions need to be explored about the extent to which AI can understand and accept relationships based on human language. Solving these problems is becoming increasingly important as huge language models evolve and become more widely used through recent work such as GPT3, PaLM, ChatGPT, Claude and Bard.
Recent work by MIT and Harvard University follows in the footsteps of other recent advances in natural language processing software that generalize large datasets to help human decision making. They present a novel method for investigating media diet patterns, which are modified language models that mimic the attitudes of subpopulations based on their consumption of certain media (such as internet news, TV shows, or radio shows).
To predict what the audience’s opinion is on a particular issue, AI uses the dependence of the target audience’s opinion on the media diet of the audience.
There are three steps in developing a media diet model:
- 1️⃣ A language model is developed or used to predict missing words in a document. In this work, they mainly use a pre-trained BERT model.
- 2️⃣ Modifying the language model by training it on the media diet dataset includes content from various media spanning a specific time period. Researchers use television and radio to show transcripts and internet news. This modification allows the model to get fresh data while updating its internal knowledge representations.
- 3️⃣ By asking these models questions to see if their distribution of responses reflects the distribution of populations with different eating patterns based on the media they consume. They analyze the responses to the survey questions by querying the media diet model.