Question on your paper.

[kwonmha]

Hello, I'm here to ask a question. It's not about code but about your paper.

To my understanding, the last term of the training objective should be logP(yt_dep | Vg, X) instead of logP(yt_dep | Pg, X). Or Pg should be removed.

Because predicting predicates with POS tagging precedes parsing dependency. And Pg is the answer for yt_dep.

I think it doesn't make sense to use Gold Parse(Pg) to predict yt_dep.

I'll look forward to your gold explanation. :) TY

[strubell]

We do condition on the gold parse edges to predict dependency labels (following Dozat and Manning, 2017). This isn't necessary, you could use predicted edges, but we started with their implementation and didn't experiment with using the predicted edges.

[kwonmha]

I see. Parsing edges and predicting dependency labels are separate process. But If you used parse edges from Dozat and Manning's work, I thought it is corresponding to LISA + D&M. Then, I thought vanilla LISA model seems to predict parse edges and do labeling. I can't catch the difference of LISA and LISA+D&M.

And one more thing irrelevant with my original question. It's impossible for me to get PTB dataset. So I decided to run with CoNLL 2012 dataset but I couldn't find how to preprocess CoNLL 2012 dataset. When will the instructions be uploaded?

[strubell]

Ah, I thought the instructions were already available! I'll try to upload them this weekend. Thanks for pointing this out!

On Tue, Jan 15, 2019 at 6:53 AM M. H. Kwon [email protected] wrote:

I see. Parsing edges and predicting dependency labels are separate process.

And one more thing irrelevant with my original question. It's impossible for me to get PTB dataset. So I decided to run with CoNLL 2012 dataset but I couldn't find how to preprocess CoNLL 2012 dataset. When will the instructions be uploaded?

— You are receiving this because you commented.

Reply to this email directly, view it on GitHub https://github.com/strubell/LISA/issues/3#issuecomment-454366908, or mute the thread https://github.com/notifications/unsubscribe-auth/ADHZtw-J3ld9Z-BM8VQRNb7thzddLVjVks5vDcFOgaJpZM4Z-GTS .

[strubell]

I should also say that, like PTB, CoNLL-2012 is unfortunately not freely available (if that's the reason you're unable to get the PTB).

On Thu, Jan 17, 2019 at 7:45 PM Emma Strubell [email protected] wrote:

Ah, I thought the instructions were already available! I'll try to upload them this weekend. Thanks for pointing this out!

On Tue, Jan 15, 2019 at 6:53 AM M. H. Kwon [email protected] wrote:

I see. Parsing edges and predicting dependency labels are separate process.

And one more thing irrelevant with my original question. It's impossible for me to get PTB dataset. So I decided to run with CoNLL 2012 dataset but I couldn't find how to preprocess CoNLL 2012 dataset. When will the instructions be uploaded?

— You are receiving this because you commented.

Reply to this email directly, view it on GitHub https://github.com/strubell/LISA/issues/3#issuecomment-454366908, or mute the thread https://github.com/notifications/unsubscribe-auth/ADHZtw-J3ld9Z-BM8VQRNb7thzddLVjVks5vDcFOgaJpZM4Z-GTS .

[kwonmha]

That's too bad for me. Thanks for the informing.

[kwonmha]

And I think you got my comment that is not up to date via e-mail. Maybe I should have added comment instead of modifying comment.

Could you explain me the difference of LISA and LISA+D&M? I thought vanilla LISA model seems to predict parse edges and do labeling. While LISA+D&M use D&M's parser to predict parse edges and do labeling. And you said you didn't experiment with using the predicted edges.