Prescience

This respository provides an introductory tutorial and code to quantify prescient ideas in natural language as published in:

Paul Vicinanza, Amir Goldberg, Sameer B Srivastava, A deep-learning model of prescient ideas demonstrates that they emerge from the periphery, PNAS Nexus, Volume 2, Issue 1, January 2023, pgac275, https://doi-org.stanford.idm.oclc.org/10.1093/pnasnexus/pgac275

Overview

We propose a simple measure to identify prescient ideas in natural language using finetuned transformer models (e.g., BERT). By finetuning separate model on a temporally split corpus, we learn time-specific discourse. We identify prescient politicians, judges, and firms as individuals and groups who use language that is poorly predicted by a model trained on contemporaenous text but well-predicted on a model trained on future text. The following figure is an illustration of our model:

The principle advantage of this model is that it enables researchers to study variation in a given word within a specific time period and between actors. For example, here are the most and least prescient uses of the word race by politicans in the 1961 to 1963 United States Congress.

Implementation

We provide a easy to follow tutorial in jupyter notebook demonstrating how to compute prescience on your own corpora.

1. Finetune custom BERT models over a temporally split corpus
2. Compute prescience using the finetuned BERT models
3. Analyze prescience to understand what the model deems prescient

GPU Acceleration

This code has PyTorch dependencies and is best run on GPU. You should install the correct version of PyTorch from their website. It has been tested on torch=1.6.0 and cuda=10.2.