Another forensic exercise! This time, we’re focusing on finding traces of the
decisions made in training the model (i.e. the parameter settings), using the
validation techniques we talked about and also by querying the model and looking
at clusters. Please make notes in the section for your group in the shared notes
document.
For this exercise, first choose a few of the experimental
training models to compare (you can do this in the Sandbox if you
prefer):
- Run clusters and see how they differ from model to model
- Choose a few query terms and run the same query in each of the
models
What do you find in the experimental training models that
reveals the impacts of the following?
- Model trained with a very small window (think about the effect this might
have on the model’s behavior when you query it)
- Model trained with a very large number of dimensions
- Model trained on a very small corpus
- Bonus points if you can find evidence of a model trained with a high/low
number of vectors
Then try the same process in RStudio on the model you trained overnight (or use
one of the sample models). Do you see evidence of any of the training-related
phenomena you noticed above?
Based on these observations, consider the following:
- Which parameters and limitations would have the greatest impact on your
research and how?
- What adjustments to the parameters are you interested in and why? What
effects would you expect?
- Where might the standard validation (with common word pairs) need to be
revised for your data?
- Are there any additional insights about the impact of data preparation,
or changes to the data preparation for your corpus that you might
try?
Word Vectors: Hands-on Practice and Group Work, Intensive Research-focused slide 08
of 10
© 2019 Syd Bauman, Julia Flanders, Sarah Connell, and the Women Writers Project This
TEI-encoded XML file is available under the terms of the Creative Commons Attribution-ShareAlike
3.0 (Unported) license.
