What lies at the heart of Commonsense NLP is that the meaning of the sentence contains MORE than what is actually stated in the sentence in terms of words. What is this more? Can we capture it? Here is an idea.
I) Lets look at this ‘MORE’, elementally. Make all the possible pairs of words (since that is the least possibility i.e. 2) occurring together in sentences, attached to each other. Then see what is the MORE in meaning of (AB) as against [meaning of (A) + meaning of (B)]
Consider this example first. It is of verb + noun type.
Now, meaning(cutting vegetables) > meaning(cutting) + meaning(vegetables)
You may know what cutting in general is (making parts by using mostly a tool) and what vegetables are. But that is not sufficient to express/understand the pair-phrase “cutting vegetables”. The missing part is ‘how vegetables are cut’ i.e. placing them horizontally and applying repeated vertical strokes from the top. This is the commonsense part. This will never be present in data (unless in the rarest of cases where this is mentioned explicitly, relevant in some context). This is left to human interpretation which a machine would miss, and hence be less equipped to answer questions about the data, involving an understanding of it.
This is true for all ‘verb+ noun’ type pairs. That is, the ‘HOW’ is missing. (e.g. kick ball. No one mentions how the ball is kicked i.e. with the front of the toes of the foot).So we will have to make a database of hand-typed commonsense of the HOWs for the ‘verb+noun’ pairs.
II) Now wait. First we have to make a list of all the possible pairs from amongst N, Adj, V, Adv.
They are –
1) Verb + Adverb / Adverb + Verb
e.g. – cut fast
MORE = what fashion the stroke is in
2) Adj. + Noun
E.g. – fresh vegetables
This capture everything (as far as the COMING TOGETHER of the words is concerned Linguistically) – vegetables which are fresh
Hence nothing more.
3) Noun + Verb (in terms of doer doing an action)
E.g. – John kicked
MORE = How
4) Verb + Adj.
E.g. – turned blue
5) Noun + Verb (in the sense of an entity being on the receiving end of an action)
E.g. – mouse moved
More = How.
III) Now, all the possible pairs in English language for the above types won’t have to be examined. There is a trick for that. Here it is –
We need only those pairs which have an implicit additional meaning in them. Not those whose additional meaning will be provided in the specific data in which the pair occurs.E.g. Consider the type ‘Noun + Verb’ (type 3 above) Take these 2 pairs – ‘Man destroyed’ & ‘Pendulum swung’
In the case of the first, how man destroyed (whatever man “destroyed”, in any sense) will be given in the sentence as information, whereas how the pendulum swung won’t be stated explicitly. So we need to work on pairs like the latter only, and skip the former types. How to identify which is which?Do Google Image Search on each. If Google comes up with appropriate images then it means that there is implicit understanding in that pair (that’s why Google could come up with the image of the pair because it didn’t need any more information present in the complete sentence/data). This is true for ‘pendulum swung’. (It gives appropriate images of a swinging pendulum). Whereas, ‘man destroyed’ (which is effectively – man who destroyed) doesn’t give the appropriate images. (It gives images of a man ‘who is destroyed’, rather).
IV) Finally, an example of such pairs in a sentence –
The solution immediately turned blue.
PAIRS : solution turned – Noun + Verb (type 5)
turned blue – Verb + Adjective (type 4)
immediately turned – Adverb + Verb (type 1).