To improve a chatbot's performance I need to modify the user's question to include contextual information before passing it to the chatbot. For this task I planned on using a neural coreference resolution model such as spaCy's "en-coreference-web-trf" model. I can just use GPT for this but it is heavy weight and not suitable for my application, but spaCy's model is just fine for this issue.
Now I came across another problem, that is, when there are no explicit references available to resolve in the user's question. For example, consider the following conversation:
Q: How to treat for blast pest?
A: To treat for blast pest in rice, you can treat the seeds with Pseudomonas fluorescens
Q: What about brinjal?
I need the question "What about brinjal" to include contextual information and be modified to "What about blast pest in brinjal". The coreference in the user's question here is implicit.
If there is an explicit coreference term such as "it", it can be resolved/replaced by "blast pest" by the coreference resolution model.
Q: What about it in brinjal?
Any suggestions on how these cases can be handled? I don't think a neural coreference resolution model can deal with this. I know we can solve this through the use of large LLMs but I need something lightweight with fast inference time.
How to resolve implicit coreferences such as in the above examples?