Open julianbernauer opened 6 years ago
We should consider two types of evaluation: a) when using general pre-trained embeddings to show that they capture relevant properties in the pol-sci domain b) when using our trained embeddings that they capture both pol-sci relevant properties and general linguistic properties
Good point. Having both type of trained data would surely make use more immune to critique.
Something for construct validation: qualitative assessment of high-scoring sentences
For instance by predicting new keywords?