I just don’t have time for long posts these days
T_T so here goes: a bulleted list of sound-bites.
Jun’ichi Tsujii
- As engineers, you’d want to create applications that just work (so stats-based systems fit the bill nicely).
- As scientists though, you’d want to know when and why things don’t work sometimes, and is it worth your while to try handle these special (if they are) cases at all.
- And for this purpose, stats-based models can be like black-boxes that you can’t really inspect… and this is where computational linguistics can provide insights.
Hans Uszkoreit
- Don’t believe in the F1 score blindly. For some applications you really just want the precision; for others recall is much more important.
- Example: for a system detecting possible reports of terrorist activities, a 1.0 recall means much more than 0.01 precision, because you can certainly afford a human perusing 100 reports and confirming that 1 genuine terrorist report.
- Unfortunately many inexperienced journal reviewers (Hans actually said that) would still insist that the system’s F1 is too low and can’t be convinced of the more important role of recall in such cases.
Diana McCarthy
- Is word sense disambiguation still relevent especially when we talk about tagging with reference to a sense inventory? Because we increasingly feel this phenomena about words having a continuum of meanings rather than discrete ones (except in the case of homographs).
- I don’t have an answer for this though. Thoughts welcome!
(Thanks Prof McCarthy. I hope my idea works! Even if only somewhat.)
Chris Manning
- No, we don’t have special funding for the Stanford NLP tools development. T_T
- Yes, it’s really hard to maintain the code and projects as people leave. T_T
- Yes, it’d be nice if we had a long-term development team to take care of the software engineering and maintenance stuff so that the research students can concentrate on doing research, but this is really hard (impossible?) in reality. T_T
(Prof Manning… what can I
say except that I/we feel the pain too?
T_T)
