Hacker News new | ask | show | jobs
by chiefofgxbxl 1717 days ago
Thanks for linking this issue, and I'll certainly look into it. I agree with your analysis on Twitter: manually annotating the article should yield the following issues:

1. Focus: she was struck by a Toyota Corolla (emphasis is placed on the VRU)

2. Object: she was struck by a Toyota Corolla (agent is referred to as an object, rather than a person, e.g. driver of the Corolla)

3. Object: a 2010 Toyota Corolla being driven by a 29-year-old Brookline man, was driving east (again, the wording personifies the vehicle instead of assigning agency to the driver heading eastward)

I will need to debug this particular example further, but it appears the "2010 Toyota Corolla" and "Toyota Corolla" are not being classified as CARLIKE, a label I trained an NER model on to help deal with all the ways you can name a vehicle: year-make-model, make-model, year-model, model, model-ish, generic terms like truck/pickup/pick-up etc., short-hands like Chevy instead of Chevrolet, etc.

Furthermore, the tool is not identifying "woman" as a VRU. It's a little bit ambiguous because while "[she] entered the roadway", it's not 100% clear that means she is a VRU: a driver/vehicle can enter a roadway too.

1 comments

Interesting context, thanks! Cool tool, will be glad to see it get tuned further :)