Hi!
I think this is a good time to introduce the project I have been
working on - YodaQA:
http://pasky.or.cz/yodaqa-poster2015.pdf
https://github.com/brmson/yodaqa
I just wrote the first paper on it recently; it's an UIMA-based
question answering pipeline inspired by the DeepQA IBM papers.
It has reasonable performance of 32.6% accuracy, 79.3% recall
(30s/answer) on a "curated" TREC dataset. Also, all the databases
are available over the network (though overseas the latency can make
things a little slow) and I try to meticulously maintain detailed
installation/setup documentation.
Compared to OAQA, I make much heavier use of UIMA in my architecture
and provide an advanced reference pipeline. Compared to WatsonSim,
I work on TREC-like task, not Jeopardy-like task. Compared to OpenQA,
my primary focus is on unstructured data sources and YodaQA is not meant
as a portfolio of pipelines, but with various components sharing the
rest of the answer pipeline.
I have two aims with my system. The main aim is to create a common
research test-bed where other people working in QA can test how their
algorithms perform in a full end-to-end pipeline.
I also hope that now the "baseline" version is finished, I can go do
some interesting new science using YodaQA too. :-)
(The second aim is some hope to be able to find interesting commercial
applications for my system and competing with IBM ought to be fun! But
well, I just have a lot of very vague ideas right now, so research is my
primary focus now.)
Kind regards,
Petr Baudis