compare verb-prediction to simple shuffle embeddings

[source-data]

No description provided.

[source-data]

verbs not better than shuffle3 on entity recognition.

[source-data]

verbs a tiny bit better than shuffle on roles:

[source-data]

With determinant/article predicting embeddings:

[source-data]

• Differences are small.
• Shuffle3 has the advantage that is is smaller and faster as it does not have attention.
• Suprisingly, the best could be the determinant/article predicting embeddings. It is the only that managed to push both intervention and assayed above 80% and 85% resp.