Best way to correct dependency parsing mistakes

[jama1017]

Hi,

I am using SpaCy's built-in en_core_web_trf to parse out texts that describe a motion instruction, for example "Move the black rock upward by 50 miles" and "Move the black rock north by 50 miles"

In the first example, spaCy would correctly label the preposition "by" as "prep" depending on the verb "Move". However, in the second example, since "upward" has been changed to the noun "north", spaCy would label "by" as "prep" depending on the noun "north", instead of "move" like it previously did.

There are many other examples like this in my data, and I wonder how I can correct these parsing mistakes. Is it through training from scratch, or fine-tuning, or pseudo rehearsal? I tried to search for tutorials on how to fine-tune the DependencyParser but didn't find any up-to-date ones. Any ideas and pointers would be helpful. Thanks so much!

[jama1017]

Just to add onto this - I wonder if it's possible to train a dependency parser from scratch with OntoNotes 5.0 (I believe this is what the out of the box en_core_web_trf was being trained on) but also with some custom data? Would love some guidance on this if people have done this before!!

[ivan-kleshnin]

I second this question. The largest amount of mistakes Spacy does, in my experience, are related to sentence dependencies.
For example, in my case, "co-founder" is constantly mislabeled in a way that founder is subservient to co, which is clearly should be vice-versa, or sometimes it mislables them such as both words are unrelated. sm, md, lg models – all share this mistake.

I would like more guidance on how:

• add manual exceptions or corrections for quick fixes
• more guidance on model retraining for sentence parts. The only doable approach I see at the moment is to work from existing examples, e.g I find a correct sentence closest to some of my incorrect cases and get clues from there. Because manually setting ALL the data seems an extremely hard tasks that requires a working linguistics experience.