Repo containing the code for Assessing Incrementality in sequence-to-sequence models [TODO: Add link].
To use the code in this repo, install all dependencies first:
pip install -r requirements.txt
Models are then trained using the machine library (refer to the library's documentation for more information).
After finishing model training, you can use the following scripts inside the repo:
| Script | Purpose |
|---|---|
evaluate.py |
Evaluate a single model given some metrics |
test_incrementality.py |
Evaluate multiple models and compute metric scores per model type |
plot_correlation.py |
Create correlation scatter plots and heat maps based on a list of models and metrics |
qualitative_analysis.py |
Perform a qualitative model analysis using the Integration Ratio |
The command line arguments of every individual script can be inspected using the -h or --help argument.
Metrics to measure the incremental processing capabilities of a model are defined in incremental_metrics.py and comprise
the following:
- (Average) Integration Ratio: Indicates whether the model prefers to integrate new information about the current input at every time step or is inclined to maintain a representation about the previous tokens.
- Diagnostic Classifier Accuracy: Quantifies to what extend information about previous tokens is contained within the current hidden representation.
- Weighed Dianostic Classifier Accuracy: The same as above, but models that are able to maintain information about inputs that occurred much earlier in the sequence are scored higher.
- Representational Similarity: Measure how much hidden representations after encoding the same subsequence of tokens resemble each other. Resemblance is quantified using a distance measure like cosine similarity or euclidean distance.