2

I'm making a quote-bot which requires NER/extraction of suburbs/postcodes/states in Australia, which, perhaps unsurprisingly, seems harder to find than equiv for US etc.

Example utterances would be:

"Hi i want to ship 3 boxes from [suburb1] [postcode1] to [suburb2] [postcode2] they're 1 by 1 by 1m and 2kgs"

or

"How much to send from [suburb1] [state1] to [suburb2] [postcode2]"

I have a dataset of all the locations/postcodes etc so really all i need is a good sized set of diverse sentence templates for these sorts of requests.

I have purposely avoided discussing my reasons for approaching it this way vs other ways (i.e. pattern/fuzzy matching) because it doesn't seem the right SE website for it, but suffice to say I have explored a variety of options.

The main hurdle in this approach is that the number of suburb/postcode pairs, 16k, is far larger than the number of example sentences i can think of. In a lot of cases i really do need to show the model the suburb at least once as some are very weird/unique. This means i am re-using sentences a lot and i fear this will lead to overfitting, hence the desire for a larger diverse set of template sentences.

Part of the issue is that, for example in the first sentence, there are ordinals which are postcodes and which are not (quantity). I need datasets which include both these kinds of things so i can have it ignore the non-postcodes

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.