Possibly Improved Sequence Tagger

[ghost]

Hi,
This is a pretty awesome project, thanks for posting it! I've been experimenting with Structured Prediction methods, and decided to use this project to compare CRFs and Learning to Search methods. Pending a more rigorous evaluation (fingers crossed) I'm seeing roughly 96% per-token accuracy and 95.55% sentence-level accuracy with L2S and vw, taking 22 minutes total for read+train+test. This is an 80/20 split on the full dataset, using the output of bin/generate. Once I've cleaned up the source, I'll be happy to send over a pull request.

- Arthur

[tettoffensive]

@Zintinio did you ever clean up the source? I'm curious to see your improvements. I'm seeing a few things where I'll get "name": "Salt and pepper" instead of two separate ingredients. Wondering if your improvements would help with this sort of problem?

Also, ingredients like "Basil" and "Basil leaves". I think it would be better if they were both recognized as the same ingredient. But that might be much more challenging ;)

[ghost]

Yeah I'll dig it up. It's actually just using Vowpal Wabbit + Dagger, with the same features.

…

On Fri, Dec 1, 2017 at 2:02 PM Stuart Tett ***@***.***> wrote: @Zintinio <https://github.com/zintinio> did you ever clean up the source? I'm curious to see your improvements. I'm seeing a few things where I'll get "name": "Salt and pepper" instead of two separate ingredients. Wondering if your improvements would help with this sort of problem? Also, ingredients like "Basil" and "Basil leaves". I think it would be better if they were both recognized as the same ingredient. But that might be much more challenging ;) — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#9 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AAnkEEvrXJUjhUYyzTxHL6VTRWIiaUgzks5s8E1IgaJpZM4KeCLe> .