Skip to main content

POS vs NER for Coref

Created on August 25|Last edited on August 30

Multidomain - Multitask

Target Coref: DWIE;

Aux Source: Ontonotes
So in effect, we're trying to find that when training together, which works better:
  1. Ontonotes POS
  2. Ontonotes NER
  3. Ontonotes Coref
When the target task is Coref on DWIE dataset


10203040Step00.20.40.60.81
task_2.names: ["pos"]
task_2.names: ["coref","pruner"]
task_2.names: ["ner"]
10203040Step00.20.40.60.81
task_2.names: ["pos"]
task_2.names: ["coref","pruner"]
task_2.names: ["ner"]
10203040Step00.20.40.60.81
task_2.names: ["pos"]
task_2.names: ["coref","pruner"]
task_2.names: ["ner"]
Run set
456



Observations

  1. NER and POS seem almost the same
  2. Coref in some settings is better, in some settings is worse
These are just the first experiments. We have not done hyperparameter optimisation, or error analysis just yet. They're up next.


Single Domain - Multitask - TRIM

NOTE: This experiment is done with only 50 instances in each dataset for quick turnaround time.

Dataset: DWIE



Run set
456



Observations

There seems to be little evidence here that NER is more beneficial than POS, infact.

Dataset: Ontonotes

NOTE: This experiment is done with only 50 instances in each dataset for quick turnaround time.


Run set
4