You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Please provide a use case to help us understand your request in context
phrases provides functionality to tune the SR model by providing phrases that the model should recognize / give context to the model about what is being talked about. We should try to improve this wherever possible. I.e. choosing phrases that most improve model performance and context.
Solution
Please describe your ideal solution
We provide the minutes items to the model as just a list of string and then the "clean_phrases" function simply returns the first 500 characters for each of the first 100 phrases of the full phrases list.
However, we could run TFIDF or some n-gram indexing to find the most valuable (specific to that meeting) n-grams in the minutes items and just provide those. This has a cold-start problem but we can probably get around that.
There may also be better methods for phrase selection to improve the model.
Along this entire issue should be tracking and reporting model performance. We could probably add ASV benchmarking as a way of monitoring performance.
The text was updated successfully, but these errors were encountered:
evamaxfield
changed the title
Configure GoogleCloudSRModel phrases to use tfidf phrases
Configure GoogleCloudSRModel phrases to use tfidf found n-grams
May 24, 2021
Use Case
Please provide a use case to help us understand your request in context
phrases
provides functionality to tune the SR model by providing phrases that the model should recognize / give context to the model about what is being talked about. We should try to improve this wherever possible. I.e. choosing phrases that most improve model performance and context.Solution
Please describe your ideal solution
We provide the minutes items to the model as just a list of string and then the "clean_phrases" function simply returns the first 500 characters for each of the first 100 phrases of the full phrases list.
However, we could run TFIDF or some n-gram indexing to find the most valuable (specific to that meeting) n-grams in the minutes items and just provide those. This has a cold-start problem but we can probably get around that.
There may also be better methods for phrase selection to improve the model.
Along this entire issue should be tracking and reporting model performance. We could probably add ASV benchmarking as a way of monitoring performance.
The text was updated successfully, but these errors were encountered: