Question answering for many tasks
The very nature of many languages allows us to bring many different NLP tasks into a simple paradigm, QA. For example, given a sentiment classification task, one can argue that instead of classifying input into classes (positive, negative, and neutral), we can use QA-based approaches to solve it. We can redefine the input in this way:
Context: "I loved this movie!"
Question: "What best describes the sentiment of this text (Positive, Negative, Neutral)?"
Answer: "Positive"
In this way, not only a single NLP task but also many other NLP tasks can be combined with only a single token classifier. For example, different questions can be used to handle different NLP tasks. As described in an earlier section, a set of questions and respective answers are required. But in this very specific demonstration, not all of the answers come from the given context, but the answer can be from the question itself!
Another...
