8/31/2020
This experiment uses the Wikipedia API and the Tensorflow.js QnA model to provide answers to natural language questions.
To use, simply enter a topic, a question about it, and press Get answer! to get an answer from Wikipedia! The results might not be always correct, but they do work most of the time.
Topic: wikipedia. Question: who runs it? Answer: wikimedia foundation
Topic: hacker news. Question: what is it? Answer: a social news website
Topic: hacker news. Question: who runs it? Answer: paul graham's investment fund
Topic: beethoven. Question: when did he die? Answer: 1827
The next feature I'd like to implement is topic extraction, so that instead of entering tensorflow as topic and asking the question what is it?, you can ask what is tensorflow?.
If you want to copy this experiment, you'll need to install @tensorflow/tfjs and @tensorflow-models/qna. Make sure you use qna version 1.0.0-alpha1, as version 1.0.0 returns far worse results. You can read more about it in this issue.