3 ms·

Using 200M pages of 13,000 US local newspapers and ML to predict GDP [pdf]

• helsinkiandrew 3y ago
  Interesting: > As our corpus is 95 times larger than Wikipedia, scaling up the estimate linearly would give us a range of around $237,500 to $4,750,000 to use the BERT on our corpus. ChatGPT is the latest and very powerful generative AI tool in text analysis. Again, it is not yet possible to use ChatGPT for a corpus as big as ours. > Furthermore, compared to BERT and ChatGPT, which are essentially black-box methodologies, our approach allows us to validate and examine the output of intermediate steps (as shown in the word cloud and word polarity figures). Thus, our methodology, which uses Word2vec and Sentprop, is well suited for a large-scale empirical application as ours, requiring custom-build word embeddings from scratch for a specific task in a parsimonious and transparent manner.
• JumpCrisscross 3y ago
  > One of the important features of the text-based sentiment plotted in Figure 4a is the downward trend that begins in the 1970s, roughly coinciding with the secular stagnation period studied in the macroeconomic literature Once the authors corrected for the increasing "general negativity bias in reporting macroeconomic news," they found the data "no longer exhibits a significant downward trend over the last several decades."