Commonsense Knowledge Bases contain facts known to everyone, like say – Lemons are sour. But is this knowledge flexibly robust in the machine?
Here is a simple scheme (with its philosophy) to make them so :
What are the Wh-words in language?
Who, When, Where, What, Which, How, Why, Whom, Whose.
- Take a piece of commonsense entry in the KB. Say, ‘lemons are sour’. Now take various perspectives on it by asking all sorts of Wh-questions to it. (This is machine-generatable). This will generate 2 kinds of questions, one kind being very special (we shall soon see).
So, firstly, this whole set of questions will be –
- Who – INVALID.
- When do lemons taste sour?
- Where do lemons taste sour?
- What lemons taste sour?
- Which lemons taste sour?
- How do lemons taste sour?
- Why do lemons taste sour?
- Whom do lemons taste sour to?
- Whose lemons taste sour?
So we have 2 kinds of questions – the ones in blue and the ones in red – each primarily serving a purpose :
- The blue questions – They help the machine understand a fact from different points of view – from the points of view of some basic allied aspects to the fact.
- The red ones – What is the nature of these questions? They are all grammatically perfect and hence technically sound. But commonsensical-ly they are weird and a joke (e.g. – whose lemons taste sour?). It seems like someone is unnecessarily trying to be logical and scientific about something very simple and obvious (commonsense). One would say – what do you mean by whose lemons taste sour? Irrespective of whose lemons they are they are produced in farms and they bear a sour taste. What do you mean by whom do lemons taste sour to? To everyone (every human) who tastes them. These seemingly tall questions can be countered and cracked by sheer commonsense if the commonsense is well in place.
Now consider this – Commonsense facts are mostly unchallengeable-y well-known and well-understood. Something like lemons are sour. But think for a while – when can the validity of this sure and solid piece be challenged? When there are logically and scientifically “weird” questions attacking them. You doubt the basics; you become unsure of the very obvious via say some Cognitive Psychological state or phenomenon. It is these very weird questions which the red ones amongst those generated in the above exercise are! So, if the system can stand competent against these attacks, we can truly say that it is flexible and robust – that too, in a Cognitive sense.
- Answer these questions either via a web search engine (blue) or manually (red). Note that being a commonsense knowledge base, the blue ones would be answered at a very basic level. For example, a blue question like ‘why is the sky blue?’ in the set of questions to the fact – the sky is blue – would be answered as something like – something to do with light’s properties (and not the actual answer).
Maybe the manual ones (red) will noticeably outnumber the blue ones.
2 small ancillary purposes of the question sets:
- The red questions test the flexibility and faith of commonsense thinking, considering the (unduly) twisted nature of theirs.
- A machine which is completely logical will face (and self-generate) such logical (grammatically, combinatorially possibilities/cases-like) questions like the ones in red. So it better have answers to such questions also!
- The exercise can be iterated over the answers of the questions also, for breadth and depth.