neurosys-pl / coreference-resolution Goto Github PK

Thank you for your solutions with regards to resolving cataphora.
I have a discussion point:
"Bayern Munich stalwart had an excellent outing against Wolfsburg. Robert Lewandowski scored five goals in nine minutes."

resolves as:
"Bayern Munich stalwart had an excellent outing against Wolfsburg. Bayern Munich stalwart scored five goals in nine minutes."

Expected:
"Robert Lewandowski had an excellent outing against Wolfsburg. Robert Lewandowski scored five goals in nine minutes."

How can we try and resolve this? I am trying to use the longest proper noun mention but above case fails this approach.

Hi NeuroSYS,
Congratulations on the great work with Coreference Resolution model.

Unfortunately, I do not have Ontonotes dataset and am using my .txt file. I am unable to find any useful link to convert .txt file into conll 2012 format. I tried using conll u format for training but did not succeed. It would be great if you can answer the following questions:

1.) Which tool can be used to annotate the text to match the coreferences
2.) Can your packages handle custom training

Looking forward to your response.

Thanks and Regards
Marur Srikanta

Many thanks for sharing your coreference resolution code and blog series, it's fantastic! I was hoping to follow up to ask your advice on other possible coreference resolution packages that I could use to adapt the code. We're hoping to adapt it in two ways:

1. I’ve been trying to only replace mentions that include a key term (“roomba"). I'd been thinking perhaps I could modify your solution to redundant clusters so that it only returns key term clusters as meaningful clusters, skipping over the rest. Just in case it'd be a useful illustration (putting code in words isn't always the easiest...), I attached a screenshot of the adapted code. Unfortunately, when I adapted it I haven't been able to get it to return just those clusters. Do you happen to know of any packages that only replace clusters if they contain a key term?

2. Very similarly, I'm trying to make sure that if the key term is contained in a cluster, it's returned as the head of the cluster (e.g. in a cluster of "this item," "it", and "roomba," I'm trying to replace everything with "roomba" instead of "this item"). I was thinking I could adapt the get_cluster_head function, but unfortunately I again can't get it to work. I was wondering if you happen to know of any packages that prioritize a key term as the head of a cluster?

A downside of the approach I was trying out is that it ties the two together (it would use the clusters from 1 in 2), but I'm sure there's also ways to approach it where you could do 2 but not 1 (or vice-versa).

Thanks a ton for your advice on this and again for sharing these resources, it's very appreciated!

Hey,
Nice work with the repo and blog post. This is really helpful.

I guess I do have a question regarding this particular line in the code base.

It seems unlikely that this condition is ever true. So is it meant to be

if key != vals[0]:
#because if it's the same value then we don't need to add it to the swap dict.

Just wanted to make sure I understood the work properly. Thanks again for the work!

I am doing bootstrap sampling on my text data and finding relevant coref solution using your updated "improvements to allennlp_cr", sometimes it is abruptly getting terminated, don't know the issue. Followed all the steps given by you it will work for some of the texts.

Any solution regarding this ASAP.

When I use pip install spacy==2.1.0, predictor crashes. Allennlp installs different spacy, therefore en_core_web_sm becomes incompatible. Even you download compatible en_core_web_sm, code crashes anyway. Predictor doesn't work. As a solution, you should update the installation part in Readme file as below.

Use pip install spacy~=3.3.0 instead of pip install spacy==2.1.0

Also add pip install allennlp-models to installation part.