Developing a Question Answering Application Quickly Using NVIDIA Jarvis

Originally published at: https://developer.nvidia.com/blog/developing-a-question-answering-application-quickly-using-jarvis/

There is a high chance that you have asked your smart speaker a question like, “How tall is Mount Everest?” If you did, it probably said, “Mount Everest is 29,032 feet above sea level.” Have you ever wondered how it found an answer for you? Question answering (QA) is loosely defined as a system consisting…

Have you tried Jarvis QA yet? If not, you should.
Did you notice Jarvis QA fail to answer your question? If so, please mind that Jarvis QA has token size limit of 384 (w/ BERT Large). If you tried with the passage longer than that, the answer may be incorrect or missing. Nevertheless, you should not be discouraged with such limitation. See if you can come up with a way to cover the context in its entirety to find the answer.

has anyone ran into an issue where running jarvis_nlp.NaturalQuery crashes the tritonserver? Looking at the start-jarvis script inside of the container, you cant tell which process actually failed.

while sleep 5; do
ps aux | grep tritonserver | grep -q -v grep
PROCESS_1_STATUS=$?
ps aux | grep jarvis_server | grep -q -v grep
PROCESS_2_STATUS=$?
if [ $PROCESS_1_STATUS -ne 0 -o $PROCESS_2_STATUS -ne 0 ]; then
echo “One of the processes has exited unexpectedly. Stopping container.”
exit 1
fi
done

Just tried using a different GPU and now things are working.

@mika.ayenson – Good to hear! Let us know if you have any other problems going through this post’s solution.

Anyone who would have answered IR is QA in my AI class would have gotten a 0 on the essay question. QA requires NLU. For example, You are sailing on a small boat with your two friends. You hop off the boat at the dock and look back at the boat. How many people do you see?

You can’t answer such questions with IR, you have to understand implicit context, situational understanding, etc.

Using Wikipedia as a source of knowledge is risky as while it may be good for IR demo’s it is will also contain incorrect knowledge.